<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Anna Via on Medium]]></title>
        <description><![CDATA[Stories by Anna Via on Medium]]></description>
        <link>https://medium.com/@annaviaba?source=rss-c1a8933ed8b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*pjMUJFocl7FBCWrSiFQ-9A.jpeg</url>
            <title>Stories by Anna Via on Medium</title>
            <link>https://medium.com/@annaviaba?source=rss-c1a8933ed8b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 18:51:19 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@annaviaba/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[When AI meets Product: July-August’26 Updates]]></title>
            <link>https://annaviaba.medium.com/when-ai-meets-product-july-august26-updates-6ef1ffc1eb45?source=rss-c1a8933ed8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/6ef1ffc1eb45</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[newsletter]]></category>
            <dc:creator><![CDATA[Anna Via]]></dc:creator>
            <pubDate>Sun, 30 Aug 2026 21:39:54 GMT</pubDate>
            <atom:updated>2026-08-30T21:39:54.836Z</atom:updated>
            <content:encoded><![CDATA[<h4>Keeping up to date with new AI models, products, ethics, and trends</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/922/1*AtDMCV5No5t22u8z6WE66g.png" /></figure><p>Welcome to the July-August’26 edition of “<strong>When AI Meets Product</strong>”, where you’ll find the most relevant updates and resources in the field of applied AI from the last two months.</p><ul><li><strong>GenAI model provider updates.</strong> A wave of major summer releases which included Anthropic’s Opus 5, Google’s Gemini 3.6 Flash, Meta’s Muse series, Kimi K3, and Qwen 3.8.</li><li><strong>Building AI applications &amp; products.</strong> On the application side: how agentic commerce is taking off, AI search driving direct product detail pages traffic, and leveraging GenAI into recommender systems. On the building side: replacing PRDs with Golden Datasets and multi-step evals, deeplearning.ai’s AI engineering skill map, and in-housing vs. hybrid model routing to decrease costs and dependency to third party model providers.</li><li><strong>AI ethics, regulation &amp; governance</strong>. EU AI Act’s transparency obligations take effect, Anthropic adopts C2PA watermarking standards, Illinois signs the AI Safety Measures Act, and China restricts anthropomorphic emotional agents.</li><li><strong>AI Adoption &amp; productivity</strong>: Relevant resources covering current adoption data, as well as the AI Productivity Paradox.</li></ul><h3>GenAI model provider updates</h3><ul><li><strong>Anthropic</strong>: Launched <a href="https://www.anthropic.com/news/claude-opus-5"><strong>Claude Opus 5</strong></a>, delivering near-frontier “intelligence” comparable to Claude Fable 5 at half the cost. It sets new benchmarks across software engineering and knowledge work.</li><li><strong>Google</strong>: Released <a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/"><strong>Gemini 3.6 Flash</strong></a>, which focuses on agentic workflows, cutting output token usage by 17% while boosting coding quality; <strong>Gemini 3.5 Flash-Lite</strong> for high-throughput tasks (350 tokens/s); <strong>Gemini 3.5 Flash Cyber</strong> integrated into <strong>CodeMender</strong> for security patching; and <a href="https://dataconomy.com/2026/08/27/gemini-35-transcribe-ai-audio-model/"><strong>Gemini 3.5 Transcribe</strong></a>, a multimodal speech model supporting 85+ languages, automatic filler word removal, and voice editing commands.</li><li><strong>OpenAI</strong>: Launched <strong>ChatGPT Work</strong>, a dedicated desktop productivity agent powered by the <strong>GPT-5.6</strong> model family (Sol, Terra, Luna).</li><li><strong>Meta AI Research</strong>: Unveiled <a href="https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2"><strong>Muse Spark 1.2</strong></a> alongside <strong>Muse Code</strong>, a terminal coding agent powered by persistent background agents, and the public preview of the <a href="https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/"><strong>Meta Model API</strong></a>.</li><li><strong>SpaceXAI</strong>: Released <a href="https://x.ai/news/grok-4-6"><strong>Grok 4.6</strong></a>, engineered specifically for long-running agents and complex visual/interactive project creation.</li><li><strong>Moonshot AI</strong>: Introduced <a href="https://www.kimi.com/blog/kimi-k3"><strong>Kimi K3</strong></a>, the world’s first open 3T-class model featuring native vision and a 1-million-token context window.</li><li><strong>DeepSeek</strong>: Released <a href="https://artificialanalysis.ai/articles/deepseek-v4-flash-0731-scores-50-on-the-artificial-analysis-intelligence-index-10-points-above-previous-deepseek-v4-flash"><strong>DeepSeek V4 Flash 0731</strong></a>, with 284B total parameters and a 1M token context window, reduced hallucination rates to 84% and decreased output token usage by 12%.</li><li><strong>Alibaba</strong>: Expanded its open-weights lineup with <strong>Qwen 3.8-Max, </strong>their first open-sourced weights of a Qwen-Max-class flagship model; <strong>Qwen3.8-Flash-Next</strong>, a multimodal MoE model focused on cost-efficiency; and <strong>Qwen-Image-3.0</strong>, built for generating rich visual content with high detail and fidelity.</li><li><strong>Other providers: Microsoft AI d</strong>ebuted <a href="https://microsoft.ai/news/introducing-mai-thinking-1/"><strong>MAI-Thinking-1</strong></a>, a 35B-active reasoning model; <strong>NVIDIA </strong>released<a href="https://www.marktechpost.com/2026/07/07/nvidia-releases-audex-nemotron-labs-audex-30b-a3b-a-unified-audio-text-llm-that-preserves-the-text-intelligence-of-its-backbone/"> <strong>Audex (Nemotron-Labs-Audex-30B-A3B)</strong></a>, a unified 30B MoE audio-text model; and <strong>Apple</strong> keeps pushing for <a href="https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/">faster local AI compute</a> (4.3x) with the new <strong>Mac Studio</strong> equipped with <strong>M5 Max</strong> and <strong>M5 Ultra</strong> chips.</li></ul><p>Comparison of models is getting more complicated as many focus on specific use cases (agentic work, coding, text or image generation…). However, this is the current leaderboard for general evaluation of models.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pkbSl9IP8EilTSXHDW79RA.png" /><figcaption><a href="https://huggingface.co/spaces/lmarena-ai/arena-leaderboard">Arena Leaderboard 30/08/2026</a></figcaption></figure><h3>Building AI applications and products</h3><p><strong>Marketing &amp; Advertising</strong> continue to be one of the key areas where AI is delivering huge impact. <a href="https://www.mckinsey.com/capabilities/growth-marketing-and-sales/our-insights/from-campaigns-to-continuous-growth-ai-capabilities-shaping-marketing"><em>The Future of Marketing in the Age of AI</em></a><em> (Mckinsey)</em>, outlines the transition from episodic campaigns to continuous growth engines driven by <strong>5 core AI capability pillars</strong>: c<em>ontinuous insights</em>, s<em>caled creativity</em>, h<em>yperpersonalization</em>, ,<em>arketing to AI Agents (GEO/Agentic Commerce)</em>, and a<em>lways-on orchestration</em>. A good example is <a href="https://press.aboutamazon.com/aws/2026/7/warner-bros-discovery-announces-agentic-ai-powered-advertising-technology-built-on-aws-its-preferred-cloud-provider"><strong>Warner Bros. Discovery</strong></a><strong> </strong>which has developed an agentic AI advertising platform built on AWS, which deploys autonomous agents for intelligent media planning, dynamic audience forecasting, real-time campaign optimization, and closed-loop measurement.</p><p><strong>GenAI search continues to expand</strong>. <a href="https://www.shopify.com/enterprise/blog/ai-search-category-behavior"><strong>Shopify</strong></a><strong> </strong>released Q2 2026 commerce data revealing a shift in shopper behavior: AI-referred storefront sessions grew <strong>197% YoY</strong>, with AI search sending 50% of shoppers directly to product detail pages. AI-referred visitors converted at <strong>double the rate of organic search</strong> in research-heavy, spec-led categories (e.g., watches, electronics). To capture this traffic, Shopify expanded its <strong>Agentic Plan</strong>, enabling merchants to syndicate structured product catalogs directly to AI search engines (ChatGPT, Copilot, Gemini…).</p><p><strong>GenAI takes recommendations to the next level</strong>. <a href="https://engineering.fb.com/2026/07/15/ai-research/exploring-hierarchical-interest-representation-for-meta-ads-deep-funnel-optimization/"><strong>Meta</strong> Ads Engineering unveiled HIR (Hierarchical Interest Representation)</a>, a graph representation layer running across Meta’s ad ecosystem which fuses user engagement graphs with multimodal world knowledge extracted via LLMs, to optimize deep-funnel ad recommendations across billions of interactions. <a href="https://arxiv.org/abs/2608.10257"><strong>Netflix</strong> is also reshaping recommender systems</a> thanks to LLMs by enabling richer modeling of users, content, and context directly in natural language</p><h4>Making AI projects a reality</h4><p>As more companies integrate LLMs into their processes, more visibility is given to the<strong> risks of depending on third party providers to serve core functionalities: </strong>data sovereignty, endor lock-in, API price shifts, and general high token costs. Leading organizations are adopting open-weight models and <strong>hybrid routing architectures </strong><em>(</em><a href="https://buttondown.com/idealist/archive/the-rise-of-hyper-local-mesh-ai/"><em>source 1</em></a><em>, </em><a href="https://www.productreport.ai/p/in-housing-ai-and-why-it-s-a-business-strategy-not-a-tech-trend"><em>source 2</em></a><em>,</em><a href="https://www.wsj.com/cio-journal/why-at-t-is-betting-big-on-open-weight-ai-a0ea03b1"><em> source 3</em></a><em>)</em>: deploying lightweight, quantized local models on owned infrastructure for high-volume, low-complexity tasks, while routing complex edge cases to frontier cloud APIs.</p><p><a href="https://www.mindtheproduct.com/evals-are-the-new-prd/"><strong>Evals are the new PRD</strong></a><strong>:</strong> As engineering velocity dramatically accelerates with AI tooling, traditional static PRDs fail to keep pace with rapid iteration. Product leaders are replacing long specification documents with <strong>Golden Datasets</strong> — curated collections of real inputs paired with acceptable output criteria — which serve as live, executable specifications. In parallel, Anthropic’s Applied AI team emphasized moving beyond single-turn prompt evals toward <a href="https://www.anthropic.com/webinars/evals-for-ai-agents-how-product-builders-get-the-most-out-of-every-new-model"><strong>end-to-end multi-step agent evaluations</strong></a> built directly from production failure modes to safely test new model upgrades before rollout.</p><p>DeepLearning.AI and Andrew Ng synthesized findings from 10,000+ job postings and expert interviews to establish the <strong>the 2026 AI Engineering Skills Map</strong>:</p><ol><li><em>Building &amp; Deploying AI Applications</em> (evals, error analysis loops, context engineering/RAG).</li><li><em>Software Engineering Fundamentals</em> (managing tradeoffs in cost, speed, security, and steering agents).</li><li><em>Using Coding Agents</em> (context management, verifiers, multi-agent orchestration).</li><li><em>Shaping the Build</em> (product sense, business context, and taking end-to-end ownership).</li></ol><h3>AI Ethics Regulation &amp; Governance</h3><p>The European Commission issued official <a href="https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations">guidelines for <strong>Article 50 of the EU AI Act (Transparency Obligations)</strong></a>, which came into force on <strong>August 2, 2026</strong>. These obligations require providers to build machine-readable detection markers into synthetic content and mandate that deployers explicitly notify users when exposed to image deepfakes, or AI generated text on matters of public interest or chatbots. In parallel, Spain’s AI oversight agency (<strong>AESIA</strong>) published <a href="https://aesia.digital.gob.es/es/guias">16 practical compliance guides and checklists</a> developed through its regulatory sandbox to assist companies with European AI Act implementation across risk management, data governance, and transparency.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vWOpVzSm1dmNt6jso-3syQ.png" /></figure><p>Anthropic formally signed the EU’s <strong>Code of Practice on Transparency of AI-Generated Content</strong>. Starting with models released after August 2, 2026, Anthropic <a href="https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content">is embedding <strong>imperceptible text watermarking</strong> directly at the model level and attaching <strong>C2PA signed provenance metadata</strong> to generated files</a>.</p><p>Illinois Governor signed the <strong>Artificial Intelligence Safety Measures Act</strong> into law, establishing one of the stringent state-level regulatory frameworks in the U.S. for frontier AI models. The law requires developers of frontier systems to publicly disclose safety protocols, report major safety incidents, maintain compliance frameworks, and establish whistleblower protections for internal employees.</p><p>China’s <a href="https://www.scmp.com/tech/big-tech/article/3359482/bytedance-and-alibaba-disable-humanlike-ai-custom-agents-new-rules-loom"><em>Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services</em> </a>took effect on July, regulating AI systems that simulate human personality, thinking patterns, or sustained emotional interaction. Major consumer AI platforms <strong>ByteDance </strong>and <strong>Alibaba </strong>disabled all customized humanlike agent features and user-created persona bots. Workplace assistants and customer service bots without sustained emotional interaction remain exempt.</p><h3>AI Adoption, Productivity &amp; Enterprise Economics</h3><p>Adoption Benchmarks, ROI &amp; Infrastructure Cost Metrics</p><ul><li><a href="https://www.gallup.com/699797/indicator-artificial-intelligence.aspx">Gallup’s workplace tracking</a> reveals that <strong>30% of U.S. employees use AI</strong> at work a few times a week or more, with 15% using it daily. Manager support acts as the strongest catalyst: workers with s<strong>upportive managers </strong>are <strong>7.4x more likely</strong> to state AI helps them excel at work. 65% of workers cite productivity gains, yet <strong>only 14% strongly agree AI has transformed how work gets done</strong> — indicating AI currently enhances existing routines rather than reinventing workflows.</li><li><a href="https://blog.google/intl/es-419/noticias-de-la-empresa/understanding-the-ai-economy/"><strong>Google’s study</strong></a> (based on 15 million interactions across the world) shows AI adoption is <strong>broad but shallow</strong>: while AI reaches 68% of occupations, it touches only <strong>~21% of tasks</strong> within a typical job. Over 90% of workplace AI interactions focus on collaboration, strategy, and learning, whereas full task automation accounts for under 10%. Notably, technicians and skilled manual trades are <strong>2x more likely</strong> to leverage multimodal AI for real-time machinery diagnostics and troubleshooting than typical office workers.</li><li><a href="https://www.svpg.com/the-ai-productivity-paradox"><strong>The AI Productivity Paradox (SVPG)</strong></a>: Addressing why 89% of executives report faster work output while only 6% see clear org-wide ROI, it highlights that most companies use AI merely to accelerate legacy, output-driven “project model” habits. <em>“It’s never been faster to build, which means it’s never been easier to run 10 times faster in the wrong direction”</em>. Elite product organizations succeed by separating <strong>Build to Learn</strong> (using AI for rapid user discovery and viability testing) from <strong>Build to Earn</strong> (production scaling).</li><li><a href="https://privatebank.jpmorgan.com/nam/en/insights/markets-and-investing/tmt/ai-use-is-exploding-so-are-the-bills"><strong>AI Use Is Exploding, So Are the Bills (J.P. Morgan)</strong></a><strong>: </strong>Enterprise token consumption has surged alongside agentic workflows. Consumption is heavily concentrated: the <strong>top 1% of enterprise users</strong> consume 600x more tokens than the median user, driven by heavy coding and document analysis. Heavy engineering usage can reach <strong>$730/month per developer</strong>, leading firms to run through annual AI budgets by May. In response, enterprises are shifting from seat-based to activity-metered billing and adopting hybrid routing or “good enough” lower-cost open/regional models to control expenses.</li></ul><h3>🛠️ Wrapping it up</h3><p>That was it from “<strong><em>When AI Meets Product </em></strong>”. Two more months filled with AI news, cool products and use cases, and relevant resources.</p><p>As AI hype continues to flood our feeds, it’s worth taking a step back and maintaining a healthy dose of skepticism. Not every bold forecast holds up over time. If you want a fun reality check, take a look at <a href="https://boyswhocriedai.lovable.app/archive">The Archive of Incorrect AI Predictions</a>, which catalogs some of the most notable AI predictions that failed to materialize.</p><p><em>Transparency note: I used Notebooklm to accelerate the writing of this newsletter. Prioritization of resources and final revision was carried by a human (me!).</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6ef1ffc1eb45" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When AI meets Product: May-June’26 Updates]]></title>
            <link>https://annaviaba.medium.com/when-ai-meets-product-may-june26-updates-03203856ca25?source=rss-c1a8933ed8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/03203856ca25</guid>
            <category><![CDATA[trends]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[newsletter]]></category>
            <dc:creator><![CDATA[Anna Via]]></dc:creator>
            <pubDate>Wed, 08 Jul 2026 12:53:49 GMT</pubDate>
            <atom:updated>2026-07-08T12:53:49.040Z</atom:updated>
            <content:encoded><![CDATA[<h4>Keeping up to date with new AI models, products, ethics, and trends</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/864/1*U1HD_muLETpCRDu3WyYWWQ.png" /></figure><p>Welcome to the May-June’26 edition of “<strong>When AI Meets Product</strong>” where you’ll find the most relevant updates and resources in the field of applied AI from the last two months.</p><ul><li><strong>GenAI model provider updates:</strong> Deep dives into OpenAI’s new GPT-5.6 series, Anthropic’s agent-heavy Opus 4.8 &amp; Sonnet 5, and Google’s Gemini 3.5 Flash. We also explore the broader ecosystem releases from Alibaba, Apple and Cohere, alongside the geopolitics of the recently lifted US ban on Anthropic’s restricted models.</li><li><strong>AI Products:</strong> How the shift from chat to autonomous action is materializing through examples like Google’s Chrome “auto-browse” and Amazon’s consolidated Alexa for Shopping.</li><li><strong>Trends &amp; enterprise realities:</strong> A look at surging consumer adoption, how Anthropic just surpassed OpenAI in enterprise use, and the stubborn “last mile” complexities that are causing 95% of corporate AI pilots to fail.</li><li><strong>Ethics &amp; regulation:</strong> An examination of privacy asymmetry in Big Tech, the EU’s simplified AI Act approval (including the addition of the ban on non-consensual explicit content generation), and the U.S. government’s new TRAINS Taskforce for national security red-teaming.</li><li><strong>On the impact of work:</strong> The fundamental transformation of software engineering: velocity gains, from writing syntax to acting as a “System Orchestrator,” and why generating code 10x faster is creating severe infrastructure and testing bottlenecks.</li></ul><h3>GenAI Model Provider Updates</h3><ul><li><strong>OpenAI </strong>is previewing the <a href="https://openai.com/es-ES/index/previewing-gpt-5-6-sol/">GPT-5.6 series</a>, which includes <strong>Sol</strong> (the secure flagship model), <strong>Terra</strong> (a balanced model at half the cost of GPT-5.5), and <strong>Luna</strong> (the fastest and cheapest option).</li><li><strong>Anthropic </strong>launched <a href="https://www.anthropic.com/news/claude-opus-4-8"><strong>Claude Opus 4.8</strong></a> and <a href="https://www.anthropic.com/news/claude-sonnet-5"><strong>Sonnet 5</strong></a>, focusing heavily on autonomous agentic workflows and complex problem-solving. They also released <strong>Mythos 5</strong> (a highly restricted model for cybersecurity and biology) alongside its safer public counterpart, <a href="https://platform.claude.com/docs/es/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5"><strong>Fable 5</strong></a>. Showcasing rapid <a href="https://www.anthropic.com/institute/recursive-self-improvement">AI self-improvement</a>, Claude models now autonomously writes over 80% of Anthropic’s internal code.</li><li><strong>Google </strong>introduced <a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/#frontier-safeguards"><strong>Gemini 3.5 Flash</strong></a>, a lightning-fast model built to execute multi-step workflows and power their new 24/7 personal AI, <strong>Gemini Spark</strong>. For local development, they released <a href="https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemma-4-12b/"><strong>Gemma 4</strong></a>, a highly efficient, encoder-free multimodal model designed to run directly on laptops. For image generation upgrades, <a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni-flash-nano-banana-2-lite/"><strong>Nano Banana 2 Lite</strong></a> is also out.</li><li><strong>Other providers</strong>: The broader ecosystem saw massive updates, including <a href="https://machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models"><strong>Apple’s third-generation Foundation Models (AFM 3)</strong></a> designed for privacy-centric on-device and cloud compute. <strong>Cohere</strong> open-sourced <a href="https://cohere.com/blog/command-a-plus"><strong>Command A+</strong></a>, an efficient model optimized for agentic tasks. <strong>Alibaba (</strong><a href="https://www.marktechpost.com/2026/05/21/qwen-introduces-qwen3-7-max-a-reasoning-agent-model-with-a-1m-token-context-window/"><strong>Qwen3.7-Max</strong></a><strong>)</strong> and <strong>MiniMax (</strong><a href="https://www.minimax.io/blog/minimax-m3"><strong>M3</strong></a><strong>)</strong> pushed boundaries with new reasoning models featuring massive <strong>1-million-token context windows</strong>. Additionally, <a href="https://thinkingmachines.ai/blog/interaction-models/"><strong>Thinking Machines Lab</strong> </a>debuted a new architecture for seamless, real-time audio and video human-AI collaboration.</li><li><strong>Impact of US politics</strong>: <a href="https://www.washingtonpost.com/technology/2026/06/13/anthropic-shuts-down-newest-ai-model-after-us-bans-foreign-use/">US banned the foreign use of Anthropic Fable and Mythos models</a>, and given the complexity of ensuring the ban was fulfilled (it included non-US company employees living in the US), Anthropic cancelled the access to these models overall. <a href="https://www.anthropic.com/claude/mythos">This week the ban has been lifted</a> and access to the models restored.</li></ul><h3>AI Products</h3><p><strong>AI Product updates — the shift to agents</strong></p><ul><li><strong>Google:</strong> Transitioning from chat to action with <strong>Gemini 3.5 Flash</strong>, enabling <a href="https://support.google.com/chrome/answer/16821166?hl=en">Chrome to “auto-browse” multi-step tasks</a>. They also introduced <strong>Google Antigravity</strong> to <a href="https://blog.google/products-and-platforms/products/search/search-io-2026/#personal-intelligence">generate custom UI dashboards directly in Search</a>, and are testing an <a href="https://deepmind.google/blog/ai-pointer/"><strong>AI-enabled mouse pointer</strong></a> that understands on-screen context to execute commands without typing.</li><li><strong>Amazon:</strong> Consolidated its assistants into <a href="https://advertising.amazon.com/library/news/agentic-shopping-advertising"><strong>Alexa for Shopping</strong></a> to automate cart-building, deal-finding, and running conversational ad campaigns. They are also generating <a href="https://www.cnet.com/news/amazon-shopping-ai-generated-product-images-search/">AI concept images</a> for broader search terms like “flannel shirt”.</li><li><strong>Coca-Cola:</strong> Rolled out an AI modeling tool called <a href="https://www.cfodive.com/news/inside-coca-colas-use-ai-marketing-spend-agility/821679/"><strong>Fuel Light 360</strong></a> to instantly test resource allocation scenarios, cutting decision cycles from weeks to just one hour.</li></ul><p><a href="https://productmanagementwithmani.substack.com/p/detailed-ai-pm-use-case-with-end"><strong>Building AI — the PM blueprint</strong></a>: key insights into how to actually build these type of solution.</p><ul><li><strong>Scope Narrowly:</strong> Never start with an “omniscient assistant.” Launch with one specific workflow to minimize risk and measure success clearly.</li><li><strong>Layered Architecture:</strong> Real systems require three layers: <strong>Experience</strong> (UI with human-in-the-loop controls), <strong>Intelligence</strong> (LLMs grounded by RAG and rule engines), and <strong>Data</strong> (historical inputs and feedback loops).</li><li><strong>Layered Metrics:</strong> Don’t just track adoption. Measure the <strong>Model</strong> (hallucination rates), the <strong>User</strong> (how often humans edit the AI’s drafts), and the <strong>Business</strong> (cost per ticket).</li><li><strong>Core Mindset:</strong> PMs, <a href="https://ai.gopubby.com/harness-engineering-what-every-ai-engineer-needs-to-know-in-2026-0ab649e5686a">engineers</a> and <a href="https://towardsdatascience.com/from-data-scientist-to-ai-architect/">Data Scientists (transitioning to AI architect)</a> must realize they are not just shipping a model; they are designing a <strong>“decision-support system under uncertainty”</strong>.</li></ul><h3>AI trends and strategy</h3><p><a href="https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/"><strong>Surging consumer adoption</strong></a><strong>:</strong> <strong>Nearly half of all U.S. adults now use AI chatbots</strong>, up significantly from 33% in 2024. <strong>ChatGPT is the clear consumer favorite</strong>, used by 44% of adults, followed by Google’s Gemini at 24%. The public primarily relies on chatbots to <strong>search for information and assist with work tasks</strong>. Overall, users feel these tools positively impact their productivity and knowledge.</p><p><a href="https://econlab.substack.com/p/anthropic-beats-openai"><strong>A shift in enterprise leadership</strong></a><strong>:</strong> In the corporate sector, <strong>Anthropic has officially surpassed OpenAI in business adoption</strong> for the first time, reaching 34.4% of businesses compared to OpenAI’s 32.3%. <a href="https://econlab.substack.com/p/anthropic-beats-openai"><strong>Push for cost-efficiency</strong></a><strong>:</strong> Faced with rising computing costs and budget constraints, businesses are actively shifting toward <strong>cheaper, open-source models</strong> to reduce their AI expenses.</p><p><strong>Main complexities and challenges: </strong>despite high enthusiasm and heavy investment, enterprises face severe hurdles in transitioning AI from a novelty to a reliable business driver. The main complexities include:</p><ul><li><a href="https://techstackbytes.substack.com/p/the-last-mile-problem-in-ai"><strong>The “Last mile” problem</strong></a><strong>:</strong> There is a vast gap between impressive, highly-controlled AI demos and tools that work reliably for everyday users. Because AI models are probabilistic rather than deterministic, they can generate inconsistent outputs or struggle with messy, unstructured real-world data (the “context gap”). If a model occasionally hallucinates or degrades in quality, users quickly lose trust and abandon it.</li><li><a href="https://www.cio.com/article/4163858/5-mistakes-tech-leaders-make-when-deploying-enterprise-ai.html"><strong>High failure rates and poor ROI tracking</strong></a><strong>:</strong> An estimated <strong>95% of AI pilots fail</strong>. Tech leaders frequently struggle to scale these pilots because they fail to establish clear baselines or measure ROI during structured proof-of-value periods. Leaders create friction by deploying AI tools as standalone applications — which often end up collecting dust — rather than seamlessly integrating them into the platforms where employees already work, such as Slack or Microsoft Teams. Furthermore, at an industry level, <a href="https://isaiprofitable.com/">the broader AI ecosystem is heavily in the red</a>; major providers like Google, Amazon, Meta, and Microsoft have spent hundreds of billions on AI infrastructure while seeing disproportionately lower direct AI revenues.</li><li><a href="https://www.bain.com/insights/the-2026-ceo-agenda-where-ambition-outpaces-execution/"><strong>Organizational execution bottlenecks</strong></a><strong>:</strong> More than 80% of CEOs are currently dissatisfied with what AI is delivering for their companies. The primary barriers preventing successful scaling include a lack of internal expertise, data platforms that are not ready for broad adoption, and a tendency for efforts to remain trapped in isolated local pilots.</li></ul><h3>Ethics and regulation</h3><ul><li><a href="https://ethic.es/era-big-tech-libre-eleccion-trampa"><strong>Privacy &amp; Ethical Design</strong></a>: There is a massive “information asymmetry” between users and tech giants, making AI data extraction a critical vulnerability. Ethicists warn that the “free choice” promoted by Big Tech is an illusion; features like infinite scrolling are actively designed to manipulate attention at the expense of mental health. And <a href="https://hugobowne.substack.com/p/15-privacy-questions-every-ai-builder">15 privacy questions every AI builder is asking</a>.</li><li><strong>EU Regulations</strong>: The EU is <a href="https://ec.europa.eu/commission/presscorner/detail/en/ip_26_1024">simplifying its AI Act</a> to maintain business competitiveness while introducing strict citizen protections, including a total ban on AI “nudification” apps that generate non-consensual explicit content. They also launched standardized, open-source icons to clearly label<a href="https://digital-strategy.ec.europa.eu/en/policies/eu-icons-labelling-ai-generated-content"> deepfakes and AI-generated text</a> and <a href="https://digital-strategy.ec.europa.eu/en/library/draft-commission-guidelines-classification-high-risk-ai-systems">high risk classification guidelines</a>.</li><li><strong>US National Security</strong>: To protect against adversarial exploitation, the <a href="https://www.nist.gov/news-events/news/2024/11/us-ai-safety-institute-establishes-new-us-government-taskforce-collaborate">U.S. government launched the TRAINS Taskforce</a> — now operating under the Center for AI Standards and Innovation (CAISI) — to rigorously red-team advanced AI models against biological, radiological, and cybersecurity threats.</li></ul><h3>On the impact of (engineering) work</h3><ul><li><a href="https://eig.org/ai-and-young-adult-jobs-the-real-mystery/"><strong>Recent youth labor market struggles</strong></a> are a broad economic trend affecting all education levels, not just AI replacing entry-level college graduates.</li><li><a href="https://www.salesforce.com/news/stories/how-engineering-became-agentic/"><strong>Massive velocity and quality</strong></a><strong>:</strong> By fully integrating autonomous AI agents into the development lifecycle, companies like Salesforce are seeing a <strong>151% increase in effective output</strong> and completing complex migrations 18 times faster. Crucially, this speed does not ruin quality; total incidents have actually dropped by 5%.</li><li><a href="https://www.enriquedans.com/2026/05/el-programador-que-dejo-de-programar-anatomia-de-una-transformacion.html"><strong>The Shift to “System Orchestrator”</strong></a><strong>:</strong> The developer’s role is moving away from typing syntax toward <strong>conceptual design and orchestrating complex systems</strong>. While trends like “vibe coding” (prompting AI to build software) democratize development, they become a dangerous “ticking time bomb” if the human does not actually understand the generated code.</li><li><strong>The “10x” Infrastructure Bottleneck:</strong> Generating code 10 times faster threatens to break traditional developer ecosystems. Teams are warning of severe bottlenecks, including <strong>massive spikes in testing compute requirements, slower compile times, and overwhelmed code review queues</strong>.</li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F2n41YjR5QfU%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D2n41YjR5QfU&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F2n41YjR5QfU%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/7662722ccc307cd85efcb5a6c71dfa53/href">https://medium.com/media/7662722ccc307cd85efcb5a6c71dfa53/href</a></iframe><ul><li><a href="https://www.oneusefulthing.org/p/making-ai-work-leadership-lab-and"><strong>Organizational Overhaul</strong></a><strong>:</strong> Individual AI productivity does not automatically translate to company-wide gains. To capture real value, companies must rethink their structures using a <strong>Leadership, Lab, and Crowd</strong> framework to guide vision, build internal benchmarks, and harness bottom-up innovation from employees.</li></ul><h3>🛠️ Wrapping it up</h3><p>That was it from “<strong><em>When AI Meets Product </em></strong>”. Two more months filled with AI news, cool products and use cases, and relevant resources. Excited for what is next!</p><p><em>Transparency note: I used Notebooklm to accelerate the writing of this newsletter. Prioritization of resources and final revision was carried by a human (me!).</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=03203856ca25" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Develop AI-Powered Solutions, accelerated by AI]]></title>
            <link>https://aiadvances.org/how-to-develop-ai-powered-solutions-accelerated-by-ai-edebe611f6ba?source=rss-c1a8933ed8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/edebe611f6ba</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Anna Via]]></dc:creator>
            <pubDate>Sun, 10 May 2026 21:55:54 GMT</pubDate>
            <atom:updated>2026-05-10T21:55:54.920Z</atom:updated>
            <content:encoded><![CDATA[<h4>From idea to impact — using AI as your accelerating copilot</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*u_fdk_henxD_EHwh7Sq0Sg.png" /><figcaption>Picture by <a href="http://twitter.com/kolamdigital">@kolamdigital</a> on <a href="https://unsplash.com/">Unsplash</a></figcaption></figure><p>The AI revolution is here, and it’s transforming the way we live and work. For companies, this revolution presents a dual opportunity. On one hand, the chance to solve previously really complex problems and to build incredible new products and features. On the other hand, the promise to boost efficiency across a wide range of tasks.</p><p>Many resources cover one opportunity or the other, but as a ML/AI Product Manager, I am actually interested in considering the two opportunities at the same time. What this actually means is: <strong>how to successfully develop AI-powered solutions, accelerated by AI itself</strong>. Based on my experience and learnings, I’ve split the process into five phases: ideation, design &amp; plan, development, deployment, and impact and monitoring. In each phase, we’ll cover “<em>what</em>” needs to happen, but also “<em>how</em>” to use AI to boost efficiency and quality.</p><h3>Phase 1: Ideation</h3><blockquote>The goal of this phase is to move from a high-level objective to a specific, prioritized solution.</blockquote><p>The first and most critical step is to remember that <strong>AI is a tool, not a solution</strong>. Always start with the problem you need to solve. It should be directly aligned with your company’s high-level <strong>OKRs</strong> and validated with evidence from <strong>user research</strong> and data.</p><p>Once the problem is clearly defined, <strong>brainstorm a variety of solutions</strong>. This should include both traditional non-AI approaches and potential AI-powered features. <strong>Prioritize these solutions</strong> using a structured method. A framework like <strong>RICE</strong> (Reach, Impact, Confidence, Effort) allows you to make a data-informed decision by weighing the potential value of each solution against its cost. For AI solutions, remember that “Effort” includes dealing with AI’s complexity such as data acquisition, system evaluation, or determining required guardrails.</p><p>To make sure this blog post is not too abstract, I’ll use my favorite marketplace use case as an example. A common user pain in marketplaces is the time and effort it takes to list new items (e.g. determining the right price, category, writing the description...). Data allows to quantify this problem: a high percentage of users who start creating a listing but never finish.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-f5HP7j2FZ4fX7Hm" /><figcaption>UI example of our “publish new item” use case, image by author</figcaption></figure><p>To address this, you could consider traditional non-AI solutions like offering templates, providing tips for each field, or creating a better onboarding process. Or, you could explore AI-powered solutions, such as using a large language model (LLM) to generate a product description or suggest a category. AI is a really cool tool though, as it can be applied to multiple use cases and applications, problems that used to be hard are now feasible, and it lowers the entry barriers to predictive models.</p><blockquote><strong><em>Building AI products always seems easier than it really is. The main hidden challenges are the integration of non-deterministic, probabilistic models (ensuring their outputs are relevant, consistent, safe…) while aiming for competitive advantage (your core enabler is an external provider any competitor can user).</em></strong></blockquote><h4>⚡️ Accelerating the Ideation Phase with AI</h4><ul><li><strong>Tools for Brainstorming</strong>: AI chatbots like OpenAI’s ChatGPT, Google’s Gemini, Anthropic’s Claude, or Perplexity can act as an extra brainstorming team member. You can prompt them with your user’s pain points and ask for a wide range of potential solutions, both traditional and AI-based. Consider and test different AI chatbot “flavors” available: getting answers from simple LLMs, getting answers with LLM leveraging “reasoning” capabilities (Chain of Thought prompting or “reasoning” model versions, e.g. Gemini 2.5 Pro, OpenAI o3,…), getting answers with LLMs using web search results, and Deep Research functionalities.</li><li><strong>Tools for Knowledge Management</strong>: Platforms like Notion AI, Mem, Tettra, or Glean can help you organize your research and ideas, using AI to connect to relevant internal knowledge and information.</li></ul><h3>Phase 2: Design &amp; Plan solution</h3><blockquote>The goal of this phase is to take a prioritized solution, assess and mitigate potential risks, define its MVP and create a concrete project plan and initial designs.</blockquote><p>With a prioritized GenAI solution in hand, implementation should be designed holistically across four dimensions:</p><ul><li>Capturing user input and relevant context (moving from prompt to context engineering)</li><li>Selecting and configuring the right model (balancing cost, latency, and performance)</li><li>Generating and evaluating outputs for quality and safety</li><li>Delivering results through effective UX/UI that supports user trust and feedback.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o9PrHkK43Lklp-wu6PoOGQ.png" /><figcaption>AI system holistic view, image by author</figcaption></figure><p>Throughout, teams must embed monitoring, evaluation, and risk management practices (addressing bias, compliance, and observability) to ensure reliability, scalability, and trustworthiness. If your are building an AI Product, another crucial part of this phase is assessing the <a href="https://www.svpg.com/four-big-risks/"><strong>four big risks</strong> of product management</a>: value, usability, feasibility and viability.</p><p>With all this in mind, we start further designing and planning for the project. In this step it is key to understand how to start small (<strong>Minimum Valuable Product</strong>) to expand and iterate once value has been proven.</p><p>For our marketplace example, let’s consider a feature that suggests a description and category based on the product title. The flow could look like this: a user inputs their item’s title, which is then used to construct a prompt for an LLM. The model returns a suggested description and category, which are pre-filled for the user to edit.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4jT-pr5zIS-VTZhs" /><figcaption>Schema to integrate prompting into the “publish new item” use case, image by author</figcaption></figure><p>The feature’s risks can be broken down by category. For usability and value, the feature is familiar and editable, which is good for user experience. The main risk is AI hallucinating or producing non relevant suggestions, which must be measured with evaluations during the development phase. For feasibility, generating a description and category with current LLM capabilities should be feasible. And for viability, this includes considering ethical risks, for instance, generating biased or discriminatory suggestions (e.g. a cleaning product -&gt; “perfect for women”), which also needs to be specifically evaluated during development.</p><blockquote><strong><em>All AI system’s parts need to be connected to ensure trust &amp; value. </em></strong>For every potential risk you identify now, ensure you include it into your future evaluation plan.</blockquote><h4>⚡️ Accelerating the Phase with AI</h4><ul><li><strong>Tools for writing</strong>: Accelerate the creation of your Product Requirements Document (PRD) with tools like ChatPRD. You can also improve the clarity and quality of your writing with assistants like Grammarly or Quillbot, and even get specific feedback on your writing with Quarkle.</li><li><strong>Tools for preparing presentations</strong>: Different tools like Gamma, Pitch, or <a href="http://beautiful.ai/">beautiful.ai</a>, are offering generation of slides from simple text and other documents.</li><li><strong>Tools for prototyping</strong>: AI can help you create everything from simple front-end mockups to complex, fully functional prototypes. Tools like Figma Make and Uizard are great for design-focused prototypes, and Claude artifacts is also great to prototype UIs really fast. Platforms like Replit, Lovable, V0, Bolt can generate prototypes closer to fully functional MVPs, by generating the code full stack.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eaISrsBILyJuY5yWPJbggA.png" /><figcaption>Example of prototype for our use case generated with Claude, image by author</figcaption></figure><h3>Phase 3: Development</h3><blockquote>In this phase, you’ll move from experimenting with prompts to getting the system ready to be integrated into your platform.</blockquote><p>This is where <strong>prompt engineering</strong> and trying out different models and approaches takes place, with the goal to get the best possible outputs your use case needs. The key is to prepare a request to an LLM with specific instructions, which will return generated text in the requested format (e.g. JSON object containing the suggested description and category).</p><blockquote><strong><em>Avoid adding unnecessary complexity. Although AI Agents is a hot topic, they add costs, complexity and non determinism. In many cases it is best to build a </em></strong><a href="https://towardsdatascience.com/a-developers-guide-to-building-scalable-ai-workflows-vs-agents/"><strong><em>predefined workflow that concatenates LLM calls in a happy path</em></strong></a><strong><em>.</em></strong></blockquote><p>A critical and often overlooked part of development is <strong>evaluation</strong>. You need to be sure that the model’s predictions and generations meet a certain quality bar before they go live, and that the risks identified in phase two are mitigated. This involves defining use-case-specific evals to measure things like hallucination, correctness, bias, and task-specific performance. For a deep dive in this topic you can check my previous post:</p><p><a href="https://ai.gopubby.com/how-to-ensure-your-ai-solution-does-what-you-expect-it-to-do-917e6d204526">How to Ensure Your AI Solution Does What You Expect It to Do</a></p><p>For our marketplace example, we might want to track the percentage of times the output is in the correct format, the accuracy of the category predictions, the relevancy of the generated descriptions, the percentage of times our outputs where biased or discriminatory…</p><blockquote><strong><em>Evals are the hardest but most necessary step in the development phase. Try to obtain inputs as close to future reality as possible, start with manual evaluations to then escale to more automated ones.</em></strong></blockquote><h4>⚡️ Accelerating the Phase with AI</h4><ul><li><strong>Tools for Coding</strong>: Many software development tools now include Generative AI features that help lower the entry barrier to coding. Assistants like Github Copilot, Cursor, Windsurf, or Claude Code are widely used to suggest code, complete functions, and solve coding problems. The use of AI Chatbots is also widely extended among programmers to accelerate code implementations.</li><li><strong>Tools for Evaluation</strong>: LLMs are both used to generate input datasets when real data is not available, and to design metrics that scale through the technique LLM as a judge.</li></ul><h3>Phase 4: Deployment</h3><blockquote>Once your solution is developed and evaluated, you’re ready to deploy it to production. This is where you integrate the solution into your company’s processes and platform so it can be used automatically.</blockquote><p>In our example, deploying the solution would allow that, when a user in the platform publishes a product, this <strong>triggers the call to an LLM</strong> to obtain the description and category from the title, and those are displayed in their corresponding touch-points.</p><p>Cloud providers like AWS, Azure, and Google have dedicated tools to accelerate the process of integrating LLMs into your platform in a scalable way. On top of using these tools, you will need to care about service metrics like latency to ensure a good user experience.</p><p>A major challenge with Generative AI is the “free input/free output” nature of the technology, which can introduce new risks. For example, users might accidentally enter personal information or even try to attack your system through “prompt injection”. This is where <strong>guardrails</strong> come in. Guardrails are checks you put in place to ensure the robustness of your solution. They can be used to detect and block unwanted input, and to ensure outputs follow certain predefined rules like avoiding profanity or mentioning competitors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mJN0sIT67pUiZBGH" /><figcaption>GenAI implementation with and without guardrails, image by author</figcaption></figure><blockquote>Don’t just deploy the AI feature: your go-live plan isn’t complete without being ready for what can go wrong in production and therefore ensuring observability (service performance, security, quality…).</blockquote><h4>⚡️ Accelerating the Phase with AI</h4><ul><li><strong>Tools for Guardrails</strong>: You can implement safety checks using specific open-source libraries like Guardrails AI and LangChain, or use managed services from cloud providers like Microsoft Azure AI Content Safety. These tools, similarly to evals, many times include LLM calls to automate the guardrail check.</li></ul><h3><strong>Phase 5: Impact and Monitoring</strong></h3><blockquote>Your solution is live, but the work isn’t over. Now you need to make sure it continues to perform as expected while having a positive impact to users and to your product and business.</blockquote><p>This involves a combination of:</p><ul><li><strong>Service monitoring</strong>, where you use tools like Datadog or specialized platforms like WhyLabs and Arize to track the operational health and quality of your AI system in production.</li><li><strong>Quantitative data</strong> with analytics dashboards to measure the feature’s impact on key product metrics like user retention and engagement. For our marketplace example, you’d want to see if the new feature leads to a decrease in the number of users who abandon the listing process.</li><li><strong>Qualitative feedback</strong> from users to understand further their experience and identify areas for improvement.</li></ul><h4>⚡️ Accelerating the Phase with AI</h4><ul><li><strong>Tools for Qualitative Analysis</strong>: Many vendors that help collect user feedback, such as Typeform and Canny, are now incorporating AI features to automatically analyze and categorize responses. You can also leverage LLMs directly to analyze large volumes of qualitative feedback. Instead of manually reading thousands of comments, you can use an LLM to summarize themes, classify feedback by sentiment or topic (e.g. “inaccurate suggestion,” “UI feedback,” “latency issue”) , and identify emerging issues.</li></ul><blockquote><strong><em>Not all GenAI features are shiny✨, but even the less shiny ones can be very impactful. AI solutions integrated in the backend and not user-facing, might have a lot of potential too.</em></strong></blockquote><h3><strong>Wrapping it up</strong></h3><p>Developing an AI-powered solution is a journey from a user problem to a measurable impact. By moving through these five phases, you can navigate complexity and risks, while significantly improving the odds of building something of value.</p><p>In a meta twist, AI itself has become your creative partner on this journey, ready to help you and your team brainstorm, code, and analyze feedback faster than ever before.</p><blockquote>AI is making this a very exciting time to be building. The only question left is: What will <strong>you</strong> build next?</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=edebe611f6ba" width="1" height="1" alt=""><hr><p><a href="https://aiadvances.org/how-to-develop-ai-powered-solutions-accelerated-by-ai-edebe611f6ba">How to Develop AI-Powered Solutions, accelerated by AI</a> was originally published in <a href="https://aiadvances.org/">AI Advances</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When AI meets Product: Mar-Apr’26 Updates]]></title>
            <link>https://annaviaba.medium.com/when-ai-meets-product-mar-apr26-updates-8e800f1ababb?source=rss-c1a8933ed8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/8e800f1ababb</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[newsletter]]></category>
            <category><![CDATA[technology-trends]]></category>
            <dc:creator><![CDATA[Anna Via]]></dc:creator>
            <pubDate>Tue, 05 May 2026 16:33:45 GMT</pubDate>
            <atom:updated>2026-05-05T16:33:45.127Z</atom:updated>
            <content:encoded><![CDATA[<h4>Keeping up to date with new AI models, products, ethics, and trends</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/930/1*IxtoQI3dsigjXOzb7_siWA.png" /></figure><p>Welcome to the Mar-Apr’26 edition of “<strong>When AI Meets Product “</strong> where you’ll find the most relevant updates and resources in the field of applied AI from the last two months.</p><ul><li><strong>GenAI Model Provider Updates:</strong> Deep dives into Anthropic’s <strong>Opus 4.7</strong> and <strong>Mythos</strong>, OpenAI’s <strong>GPT-5.5</strong>, and Google’s <strong>Gemma 4</strong>. We also explore the infrastructure play: how Mistral and Amazon are making custom agents a plug-and-play reality for the enterprise.</li><li><strong>Product &amp; Commerce:</strong> How “Agentic Commerce” is rewriting the rules of retail and why Perplexity is betting on an OS that takes objectives, not instructions.</li><li><strong>Ethics &amp; Risk:</strong> A look at the EU’s latest compliance timelines, new research on LLM deanonimization risks, and a Pro-Human AI declaration.</li><li><strong>On the impact in jobs: </strong>Enterprise adoption and the rise of “Agentic” AI, productivity data and analysis on specific jobs impact.</li></ul><h3>GenAI Model Provider Updates</h3><p><strong>Anthropic Model Updates</strong></p><ul><li><a href="https://platform.claude.com/docs/es/about-claude/models/whats-new-claude-4-7"><strong>Claude Opus 4.7</strong></a><strong>:</strong> Anthropic released their most capable model yet for complex reasoning and agentic coding. This new model features a <strong>1M token context window</strong> and introduces support for <strong>high-resolution images up to 3.75MP</strong> to improve visual perception and localization tasks. It also introduced a new “<em>xhigh</em>” effort level for agentic use cases and an innovative <strong>task budgets</strong> feature that allows users to set soft token limits for entire agentic loops. Extended thinking has been replaced entirely by an “adaptive thinking” mode.</li><li><a href="https://red.anthropic.com/2026/mythos-preview/"><strong>Claude Mythos Preview</strong></a><strong>:</strong> On April 7, Anthropic detailed the Mythos Preview, a highly capable model that can <strong>autonomously identify and exploit zero-day cybersecurity vulnerabilities</strong> in major operating systems and web browsers. Rather than a general release, Anthropic is providing it to critical industry partners and open-source developers via <strong>Project Glasswing</strong> to reinforce global cyber defenses.</li><li><a href="https://www.livemint.com/technology/tech-news/what-is-claude-design-anthropics-new-ai-tool-that-tanked-adobe-and-figma-stocks-11776566584083.html"><strong>Claude Design shakes up the market</strong></a><strong>:</strong> A powerful new tool powered by the Claude Opus 4.7 model. The tool allows users to generate interactive prototypes, pitch decks, and visual assets simply by using text prompts or uploading existing design files. The launch sent shockwaves through the design software industry, causing <strong>Adobe shares to drop by 1.5% and Figma to fall 7% in secondary markets</strong>.</li><li><a href="https://arstechnica.com/ai/2026/03/entire-claude-code-cli-source-code-leaks-thanks-to-exposed-map-file/"><strong>Claude Code Source Leak</strong></a><strong>:</strong> Anthropic suffered a major internal error when it accidentally packaged a source map file into a public npm release, <strong>leaking the entire source code of its Claude Code command-line interface</strong>. Competitors and researchers now have access to over 512,000 lines of code, offering a detailed blueprint of Anthropic’s sophisticated memory architecture and query systems.</li><li><a href="https://www.theverge.com/ai-artificial-intelligence/907074/anthropic-openclaw-claude-subscription-ban"><strong>Anthropic essentially bans OpenClaw from Claude by making subscribers pay extra</strong></a><strong>.</strong></li></ul><p><strong>OpenAI Model Updates</strong></p><ul><li><a href="https://openai.com/es-ES/index/introducing-gpt-5-5/"><strong>GPT-5.5</strong></a>: a highly autonomous and intuitive model that <strong>excels at agentic workflows, coding, and complex research</strong>. It can handle messy, multi-step tasks by planning, using tools, and self-correcting along the way. Despite a major leap in intelligence and greater token efficiency, it matches the fast speed of GPT-5.4. The model features the <strong>strongest safety guardrails to date</strong> and is currently rolling out to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex, with API access coming soon.</li><li><a href="https://variety.com/2026/digital/news/openai-shutting-down-sora-video-disney-1236698277/"><strong>Sora Shutdown</strong></a><strong>:</strong> OpenAI announced the <strong>discontinuation of Sora</strong>, its generative-AI video creation app. The standalone app had been publicly available since September 2025. As part of this change, ChatGPT will also no longer generate videos based on text prompts.</li><li><a href="https://digiday.com/marketing/openai-has-quietly-launched-its-ads-manager-as-it-races-to-build-out-its-ads-business/"><strong>Push into Advertising</strong></a><strong>:</strong> OpenAI has quietly launched an early version of a self-serve ads manager to select advertisers, marking a significant step toward monetizing ChatGPT. The company has lowered the minimum spend threshold for this pilot to $50,000, signaling an aggressive push to build out an advertising business that is projected to be worth <strong>$102 billion by 2030</strong>.</li><li><a href="https://www.elmundo.es/tecnologia/creadores/2026/02/16/699355d621efa001338b45a6.html"><strong>Reminder that back in February OpenAI acquired OpenClaw</strong></a><strong>.</strong></li></ul><p><strong>Google Model Updates</strong></p><ul><li>Google released new open source AI model <a href="https://dataconomy.com/2026/04/03/google-releases-open-source-ai-model-gemma-4-for-developers/"><strong>Gemma 4</strong></a><strong> for developers</strong>, and <a href="https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/embedding-2?hl=es-419"><strong>Gemini Embedding 2</strong></a>, a multimodal embedding model that allows to bring text, images, video, audio, and docs into the embedding space.</li><li><a href="https://blog.google/products-and-platforms/products/maps/ask-maps-immersive-navigation/"><strong>Google Maps Integrates Gemini:</strong></a> New features include <strong>Ask Maps</strong>, which allows users to ask complex, conversational questions to find personalized location recommendations, and <strong>Immersive Navigation</strong>, which uses spatial understanding to provide vivid 3D views of buildings, lanes, and crosswalks.</li></ul><p><strong>Other Providers</strong></p><ul><li><strong>NVIDIA:</strong> Released <a href="https://developer.nvidia.com/blog/introducing-nemotron-3-super-an-open-hybrid-mamba-transformer-moe-for-agentic-reasoning/"><strong>Nemotron 3 Super</strong></a> an open 120B parameter (12B active) hybrid Mamba-Transformer Mixture-of-Experts (MoE) model. Pretrained natively in NVFP4, it mitigates “context explosion” with a native 1M-token context window and utilizes latent MoE and multi-token prediction to optimize efficiency for long-running autonomous agents.</li><li><strong>Microsoft:</strong> <a href="https://microsoft.ai/news/today-were-announcing-3-new-world-class-mai-models-available-in-foundry/">Microsoft announced three new models for the Microsoft Foundry</a>: <strong>MAI-Transcribe-1</strong> (high-speed speech-to-text), <strong>MAI-Voice-1</strong> (voice generation and custom voice creation), and <strong>MAI-Image-2</strong> (turbocharged image generation tailored for designers).</li><li><strong>Alibaba (Qwen):</strong> Highlighted their <a href="https://qwen.ai/blog?id=qwen3.5"><strong>Qwen3.5–397B-A17B</strong></a> open-weight model, a native vision-language model with a hybrid linear attention architecture and massive multilingual support. Additionally, they released <strong>Qwen-Scope</strong>, an open-source Sparse AutoEncoders (SAE) suite.</li><li><strong>Luma:</strong> <a href="https://lumalabs.ai/uni-1">Introduced <strong>UNI-1</strong></a>, a multimodal reasoning model capable of generating images that excels at spatial reasoning, scene completion, and reference-guided generation.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N7DGZN6zcsPLXWcZAU5dww.png" /><figcaption><a href="https://huggingface.co/spaces/lmarena-ai/arena-leaderboard">Model Arena</a> as of May 3rd 2026.</figcaption></figure><p>The <a href="https://hai.stanford.edu/ai-index/2026-ai-index-report"><strong>2026 Stanford AI Index Report</strong></a> highlights that AI capabilities are accelerating at an unprecedented rate, with models now frequently meeting or exceeding human baselines on complex tasks like PhD-level science and competition mathematics. Crucially, <strong>the AI model performance gap between the U.S. and China has effectively closed</strong>, with models from both nations trading the lead multiple times since early 2025. Meanwhile, the global landscape is shifting toward <strong>AI sovereignty</strong>, and open-source contributions from the rest of the world are now outpacing Europe and approaching U.S. levels.</p><p><strong>Enterprise Partnerships &amp; Compute</strong></p><ul><li><a href="https://techcrunch.com/2026/03/17/mistral-forge-nvidia-gtc-build-your-own-ai-enterprise/"><strong>Mistral’s Enterprise Push</strong></a><strong>:</strong> Mistral announced <strong>Mistral Forge</strong>, a platform that allows enterprise and government customers to build and train custom models from scratch on their own proprietary data. This targets companies looking for deep customization and data sovereignty, with early adopters including ASML, Ericsson, and the European Space Agency. They also launched <strong>Remote Agents</strong> in Vibe and Mistral Medium 3.5 demonstrating strong performance on the SWE-Bench verification.</li><li><strong>Microsoft Foundry &amp; WPP:</strong> Microsoft’s new MAI models have launched inside Microsoft Foundry with built-in guardrails and enterprise-grade controls for secure deployment at scale.</li><li><a href="https://openai.com/index/introducing-the-stateful-runtime-environment-for-agents-in-amazon-bedrock/"><strong>Stateful Runtime on Amazon Bedrock Summary</strong></a><strong>:</strong> OpenAI and Amazon have partnered to launch a “Stateful Runtime” for AI agents natively on Amazon Bedrock. This new environment simplifies moving agents into production by automatically managing memory, tool states, and long-term multi-step workflows. By handling the orchestration layer, it allows developers to focus on business logic while benefiting from native AWS security, reliability, and governance controls.</li><li><a href="https://www.wired.com/story/anthropic-launches-claude-managed-agents/"><strong>Claude Managed Agents Summary</strong></a><strong>:</strong> Anthropic has launched a new tool called “Claude Managed Agents” which provides out-of-the-box infrastructure to help businesses easily build, deploy, and scale autonomous AI systems. The product is designed to lower the barrier to entry for enterprise AI by simplifying the complex processes previously required to automate work tasks.</li></ul><h3>AI and Agents product updates</h3><p><strong>Perplexity’s “Personal Computer”: </strong>Perplexity<strong> </strong>is pushing the boundaries of what an OS can be with its <a href="https://www.perplexity.ai/hub/blog/everything-is-computer"><strong>Personal Computer</strong></a>, a Mac application that turns the AI into an “independent digital worker”. Unlike a traditional OS that takes direct instructions, this system takes <em>objectives</em>. By securely accessing local files, browser sessions, and native apps, the AI can orchestrate complex, multi-step workflows — like scanning email receipts to build expense spreadsheets — signaling a shift from direct manipulation to <strong>managed autonomy</strong>.</p><p><strong>The Rise of Agentic commerce:</strong> AI is rapidly becoming the primary interface for product discovery and evaluation. A recent McKinsey survey found that <a href="https://www.mckinsey.com/capabilities/quantumblack/our-insights/europes-agentic-commerce-moment-decision-influence-is-here-execution-is-coming"><strong>84% of European consumers use AI tools in their daily lives</strong></a>, frequently leveraging them to compare brands, prices, and reviews before making a purchase. In response, major brands like <a href="https://digiday.com/podcasts/mondelez-overhauls-its-3-5-billion-digital-commerce-strategy-in-era-of-ai-search/"><strong>Mondelez are actively overhauling their $3.5 billion digital commerce strategies</strong></a> to ensure their products are highly visible and optimized for AI agent search queries.</p><p><a href="https://engineering.fb.com/2026/03/31/ml-applications/meta-adaptive-ranking-model-bending-the-inference-scaling-curve-to-serve-llm-scale-models-for-ads/"><strong>Meta’s adaptive ranking model</strong></a><strong>:</strong> Meta has launched its <strong>Adaptive Ranking Model</strong> for ad recommendations, recommender models at LLM-scale. Since launching on Instagram, the infrastructure has already delivered a <strong>+3% increase in ad conversions</strong>.</p><p><a href="https://news.futunn.com/en/post/71956708/will-insurers-not-compensate-if-ai-causes-trouble-berkshire-and"><strong>Insurers pull back on AI liability</strong></a><strong>:</strong> As AI-related lawsuits are projected to jump 140% year-on-year by 2025, major US insurers like <strong>Berkshire Hathaway, Travelers Group, and AIG are advancing “AI exclusion clauses”</strong> in commercial policies. These clauses aim to shield insurers from claims related to AI-caused copyright infringement, discrimination, or property damage, forcing companies to either assume the risks themselves or purchase emerging, standalone <strong>AI liability insurance</strong>.</p><p><a href="https://www.ravelin.com/blog/ai-powered-refund-abuse-dispute-fraud"><strong>The Democratization of Fraud</strong></a><strong>:</strong> The widespread availability of generative AI tools has created a surge in <strong>“refund abuse” and dispute fraud</strong>. Consumers are increasingly using AI to alter digital evidence — such as generating realistic cracks in glass or stains on clothing — to falsely claim refunds on perfectly good items. Fraudsters are leveraging this low-risk, high-reward tactic against merchants across sectors, from fashion to food delivery.</p><p>As organizations race to implement AI solutions, <a href="https://huyenchip.com/2025/01/16/ai-engineering-pitfalls.html"><strong>expert warns of several common pitfalls</strong></a><strong>:</strong> using generative AI when it isn’t needed, forgoing human evaluation, measuring the <a href="https://productmanagementwithmani.substack.com/p/ai-evaluation-metrics-that-matter">wrong metrics</a>…</p><h3>Ethics and regulation</h3><p>A broad coalition is demanding an ethical framework where <a href="https://humanstatement.org/"><strong>AI serves humanity rather than replaces it</strong>.</a> They are calling for <strong>mandatory “off-switches”, strict protections against chatbots manipulating children</strong>, and full legal liability for AI developers whose systems cause harm.</p><p><a href="https://www.europarl.europa.eu/news/en/press-room/20260316IPR38219/meps-support-postponement-of-certain-rules-on-artificial-intelligence">EU seems to be pushing compliance for <strong>AI Act</strong> high-risk systems to December 2027 and targeting November 2026 for AI watermarking rules</a>. From an ethical standpoint, they are <a href="https://www.europarl.europa.eu/news/en/press-room/20260323IPR38829/artificial-intelligence-act-delayed-application-ban-on-nudifier-apps"><strong>strictly banning “nudifier” apps</strong></a> that generate non-consensual explicit images, while working groups establish clear <a href="https://digital-strategy.ec.europa.eu/es/news/working-groups-advance-discussions-transparency-obligations-under-article-50-ai-act">transparency and disclosure standards for deepfakes.</a></p><p>New research highlights severe ethical vulnerabilities in current models. <a href="https://arxiv.org/abs/2602.16800v2"><strong>LLMs can now automatically deanonymize internet users</strong> </a>by linking pseudonymous accounts at a massive scale, effectively ending online “practical obscurity”. Additionally, <a href="https://www.sciencedirect.com/science/article/pii/S2451958824001660"><strong>GPT models exhibit concerning gender biases</strong></a>, including skewed moral judgments regarding violence and stereotypical text attribution.</p><p><a href="https://www.reuters.com/sustainability/boards-policy-regulation/meta-start-capturing-employee-mouse-movements-keystrokes-ai-training-data-2026-04-21/">Meta to start capturing employee mouse movements, keystrokes for AI training data.</a></p><p><a href="https://www.wired.com/story/grammarly-is-facing-a-class-action-lawsuit-over-its-ai-expert-review-feature/">Grammarly Is Facing a Class Action Lawsuit Over Its AI ‘Expert Review’ Feature</a>.</p><h3>On the impact in jobs</h3><h4>Enterprise Adoption and the Rise of “Agentic” AI</h4><p><a href="https://a16z.com/where-enterprises-are-actually-adopting-ai/">Enterprise AI adoption has moved rapidly from experimental pilots to scaled production</a>. <strong>Nearly 30% of the Fortune 500 and 19% of the Global 2000 are already live, paying customers of leading AI startups</strong>. Adoption is currently dominated by <strong>coding, support, and search use cases</strong>, with the technology, legal, and healthcare sectors leading the charge.</p><p>According to <a href="https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html">Deloitte’s 2026 State of AI report</a>, <strong>agentic AI usage is expected to rise sharply over the next two years</strong>, though only one in five companies currently has a mature governance model for these autonomous agents.</p><ul><li><a href="https://cloudwars.com/ai/oracle-unleashes-1000-ai-agents-to-automate-entire-industries/"><strong>Oracle has unleashed over 1,000 AI agents</strong></a> built directly into its Fusion cloud applications to automate entire industry ecosystems in healthcare, banking, and retail.</li><li><a href="https://www.businesswire.com/news/home/20260318899337/en/Obin-AI-Emerges-from-Stealth-to-Build-Agentic-Workforce-for-Financial-Institutions"><strong>Obin AI emerged from stealth with $7 million in seed funding</strong> </a>to build a specialized, highly accurate agentic workforce specifically designed for the strict regulatory and audit requirements of financial institutions.</li></ul><p>As AI enters the workplace, its impact on productivity is highly uneven. A landmark 2026 field experiment conducted with B<a href="https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838">oston Consulting Group (BCG) identified a <strong>“jagged technological frontier”</strong></a>. For tasks <em>within</em> this frontier of capabilities, <strong>consultants using AI completed 12.2% more tasks, completed them 25.1% faster, and produced significantly higher quality work</strong>. However, for complex tasks that fell <em>outside</em> the frontier, <strong>subjects using AI were 19% less likely to produce correct solutions</strong>, highlighting the dangers of over-relying on AI when human judgment is necessary.</p><p>While job displacement is a major ethical concern, a E<a href="https://www.ecb.europa.eu/press/blog/date/2026/html/ecb.blog20260304~d9e34fc95f.en.html">uropean Central Bank survey reveals that <strong>AI-intensive firms in Europe are currently more likely to hire than fire</strong></a>. Companies are actively recruiting staff to integrate AI for research and innovation, making the immediate labor effect a net positive. However, <a href="https://www.anthropic.com/research/labor-market-impacts">recent research from Anthropic</a>, introduces a <strong>new measure of “observed exposure”</strong> and notes a slight, suggestive slowdown in the hiring of younger workers (aged 22–25) in highly exposed occupations.</p><p><strong>Even if jobs don’t disappear, some of them are already fundamentally changing. </strong>A good example is <a href="https://nikeyes.github.io/si-la-ia-escribe-el-codigo-que-hace-el-desarrollador-es"><strong>software development:</strong></a><strong> if AI coding</strong> tools dramatically lower the cost of writing software, the primary bottleneck in software development shifts from <em>implementation</em> to <em>definition</em>. Developers are transitioning from writing features to becoming <strong>architects of guardrails and engineers of specifications</strong>, focusing on building automated testing systems and evaluating what actually needs to be built. <a href="https://blog.ravi-mehta.com/p/master-prototyping">Prototyping is becoming a rapid, foundational skill used to make decisions rather than just a final deliverable.</a></p><h3>🛠️ Wrapping it up</h3><p>That was it from “<strong><em>When AI Meets Product </em></strong>”. Two more months filled with AI news, cool products and use cases, and relevant resources. Excited for what is next!</p><p><em>Transparency note: I used Notebooklm to accelerate the writing of this newsletter. Prioritization of resources and final revision was carried by a human (me!).</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8e800f1ababb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When AI meets Product: Jan-Feb’26 Updates]]></title>
            <link>https://annaviaba.medium.com/when-ai-meets-product-jan-feb26-updates-3b82eb4a0acb?source=rss-c1a8933ed8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/3b82eb4a0acb</guid>
            <category><![CDATA[trends]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[newsletter]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Anna Via]]></dc:creator>
            <pubDate>Sun, 01 Mar 2026 11:32:12 GMT</pubDate>
            <atom:updated>2026-03-01T11:32:12.528Z</atom:updated>
            <content:encoded><![CDATA[<h4>Keeping up to date with new AI models, products, ethics, and trends</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/928/1*Qp-6aeHn65ToyHqhlfVOqg.png" /></figure><p>Welcome to the Jan-Feb’26 edition of “<strong>When AI Meets Product “</strong> where you’ll find the most relevant updates and resources in the field of applied AI from the last two months.</p><p><strong>Bottom line:</strong> smarter models, more capable agents, deeper product integration — and increasing governance pressure and positioning.</p><ul><li><strong>GenAI model provider updates:</strong> Gemini 3.1, Nano Banana 2, Claude Opus 4.6, GPT-5.3-Codex-Spark, GPT-Image-1.5, and Mistral 3 push forward reasoning, coding performance, multimodality, and image generation and editing. As a result, benchmarks reshuffle and the competitive landscape tightens.</li><li><strong>Agentic AI:</strong> Coding agents continue their shift from copilots to semi-autonomous collaborators. Commerce becomes agent-first, introducing new ad formats and platform fees, while projects like OpenClaw amplify the scale and digital impact of AI agents.</li><li><strong>AI products &amp; use cases:</strong> From Uber and Goldman Sachs to Booking.com and healthcare (ChatGPT Health, AMIE), AI is embedding itself deeper into consumer experiences, enterprise operations, and clinical workflows.</li><li><strong>AI ethics &amp; regulation:</strong> Implementation of the EU AI Act progresses, companies reinforce safety guardrails, and regulators increase scrutiny over misuse and high-risk deployments.</li></ul><h3>GenAI Model Provider Updates</h3><ul><li><strong>Google</strong> launched <a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/"><strong>Gemini 3.1</strong></a><strong>, </strong>an upgraded model with doubled reasoning performance compared to 3 Pro. It is geared towards complex, multi-step problem solving and agentic workflows. They also launched <strong>new creative capabilities: </strong><a href="https://blog.google/innovation-and-ai/models-and-research/google-deepmind/project-genie/"><strong>Genie</strong></a><strong> (</strong>world model to generate interactive, playable environments in real-time from text or images), <a href="https://blog.google/innovation-and-ai/products/gemini-app/lyria-3/"><strong>Lyria 3</strong></a><strong> </strong>(new audio model that allows users to generate custom 30-second music tracks with vocals from text or image prompts. All tracks are watermarked using SynthID), <a href="https://blog.google/innovation-and-ai/models-and-research/google-labs/pomelli-photoshoot/"><strong>Pomelli Photoshoot </strong></a>(a new Google Labs tool using the Nano Banana model to turn standard product photos into studio-quality marketing assets), and <a href="https://gemini.google/es/overview/image-generation/"><strong>Nano Banana 2</strong></a> (more image quality and enhanced editing options by the user).</li><li><strong>Anthropic </strong>launched<strong> </strong><a href="https://www.anthropic.com/news/claude-opus-4-6"><strong>Claude Opus 4.6</strong></a><strong>,</strong> its smartest model yet, featuring state-of-the-art agentic coding and reasoning capabilities. It supports a massive 1M token context window (in beta) and outperforms competitors on long-context retrieval and professional knowledge tasks. They also announced <a href="https://code.claude.com/docs/en/fast-mode"><strong>fast mode for Claude Code</strong></a><strong>:</strong> Developers can now use a /fast toggle to get responses from Opus 4.6 that are 2.5x faster. This prioritizes speed for interactive tasks like live debugging, though it comes at a higher token cost.</li><li><strong>OpenAI </strong>launched<strong> </strong><a href="https://openai.com/es-ES/index/introducing-gpt-5-3-codex-spark/"><strong>GPT-5.3-Codex-Spark</strong></a><strong>, </strong>a lightning-fast model optimized for real-time coding, which boasts a 128k context window and is built for near-instant responses in interactive workflows. They also launched <a href="https://openai.com/es-ES/index/new-chatgpt-images-is-here/"><strong>GPT-Image-1.5</strong></a><strong>,</strong> a new text-to-image model rolling out to ChatGPT and the API. It is 4x faster than its predecessor, features vastly improved text rendering, and allows for highly precise edits while maintaining character likeness and lighting. They also announced <a href="https://openai.com/es-ES/index/introducing-openai-frontier/"><strong>OpenAI Frontier,</strong></a><strong> </strong>designed to build, deploy, and manage AI agents. It gives agents shared business context, secure execution environments, and integrated evaluation tools to safely automate multi-step tasks across company systems.</li><li><a href="https://mistral.ai/news/mistral-3"><strong>Mistral 3 Family</strong></a><strong>:</strong> Mistral launched its next-generation open-source multimodal models under the Apache 2.0 license. The lineup includes the massive <strong>Mistral Large 3</strong> (a Sparse Mixture-of-Experts model trained on 3,000 H200 GPUs) and the <strong>Ministral 3</strong> edge-focused series (3B, 8B, and 14B sizes).</li><li><a href="https://github.com/deepseek-ai/DeepSeek-OCR-2"><strong>DeepSeek-OCR-2</strong></a><strong>:</strong> DeepSeek released its newest open-source optical character recognition model, “Visual Causal Flow,” optimized for dynamic resolution and available on GitHub under an Apache 2.0 license.</li></ul><p><strong>Trends on Model Comparison and Adoption</strong></p><ul><li><strong>Leaderboard Shifts:</strong> Chatbot Arena reflects recent model releases, with Anthropic’s new <strong>Claude Opus 4.6</strong> taking the #1 spot in both Text and Coding. Meanwhile, Google’s <strong>Gemini 3 Pro/3.1 Pro</strong> models are dominating the Vision leaderboard, and OpenAI’s <strong>ChatGPT-image-latest</strong> (GPT-Image-1.5) leads the Image Edit rankings.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*f2C6lqx-r2ianfQCBPHD7w.png" /><figcaption><a href="https://huggingface.co/spaces/lmarena-ai/arena-leaderboard">https://huggingface.co/spaces/lmarena-ai/arena-leaderboard</a></figcaption></figure><ul><li><a href="https://huggingface.co/blog/community-evals"><strong>Community Evaluations</strong></a><strong>:</strong> Hugging Face launched “Community Evals” to decentralize benchmark reporting. Recognizing that major benchmarks (like MMLU) are saturated and often fail to reflect real-world performance, this initiative allows the community to openly submit reproducible evaluation scores via pull requests to create a more transparent single source of truth.</li><li><a href="https://www.cnbc.com/2026/01/12/apple-google-ai-siri-gemini.html"><strong>Apple Partnership</strong></a><strong>:</strong> Apple has partnered with Google to use Gemini models to power its upcoming AI-upgraded Siri, a multiyear deal reportedly worth about $1 billion annually. Apple will maintain its OpenAI integration for ChatGPT, but Google’s technology will provide the foundation for future Apple Foundation Models.</li><li><a href="https://www.aboutamazon.com/news/aws/amazon-open-ai-strategic-partnership-investment"><strong>OpenAI and Amazon announce strategic partnership</strong></a><strong>: </strong>Amazon will invest $50 billion in OpenAI and be the exclusive third-party cloud distribution provider for OpenAI Frontier. In a deal similar to previous partnerships (e.g. OpenAI with Microsoft), OpenAI will consume 2 gigawatts of Trainium capacity through AWS infrastructure</li></ul><h3><strong>Agentic AI</strong></h3><p>Coding agents keep advancing:</p><ul><li><strong>From assistants to more autonomy and longer tasks:</strong> AI coding tools are evolving into multi-agent systems capable of working autonomously for days, shifting the engineer’s role from writing code to orchestrating and validating AI work. Developers use AI in roughly 60% of their work, they report being able to “fully delegate” only 0–20% of tasks (<a href="https://resources.anthropic.com/hubfs/2026%20Agentic%20Coding%20Trends%20Report.pdf?hsLang=en">Anthropic’s 2026 Agentic Coding Trends Report</a>).</li><li><strong>New domains and accessibility:</strong> Agentic capabilities are expanding into specialized fields like cybersecurity with tools like <a href="https://fortune.com/2026/02/20/exclusive-anthropic-rolls-out-ai-tool-that-can-hunt-software-bugs-on-its-own-including-the-most-dangerous-ones-humans-miss/"><strong>Claude Code Security</strong></a>. They are also becoming more accessible; <a href="https://venturebeat.com/orchestration/anthropic-just-released-a-mobile-version-of-claude-code-called-remote"><strong>remote Control</strong></a> allows developers to manage local sessions from mobile devices, while <a href="https://claude.com/product/cowork"><strong>Claude Cowork</strong></a> brings agentic file management and reporting to non-coding knowledge workers.</li><li><a href="https://www.forbes.com/sites/jemmagreen/2026/02/12/did-artificial-intelligence-really-kill-saas/"><strong>Is Agentic coding the end of SaaS</strong></a><strong>?</strong> Beyond the impact on productivity and role redefinition, people are also starting to questions the concept of software products: if I am able to build any hyper-personalized application I need, why should I pay for generic software? An initial indicator of this shift was the stock market drop at the beginning of February among major software and services companies, attributed to the launch by <a href="https://claude.com/blog/cowork-plugins">Anthropic of Claude Cowork plugins</a>.</li><li>At the same times, many people is analyzing personal learnings on this new ways of working, including <a href="https://arxiv.org/abs/2512.14012">AI Coders don’t vibe they control,</a> <a href="https://openai.com/index/harness-engineering/">experiments of building and shipping an internal beta of a software product with 0 lines of manually-written code</a>, and the unbearable lightness of coding.</li></ul><p><a href="https://towardsdatascience.com/the-unbearable-lightness-of-coding/">The Unbearable Lightness of Coding | Towards Data Science</a></p><p>Another key application of AI agents, is Agentic shopping, with relevant recent news:</p><ul><li><strong>Technical and product progress. Infrastructure:</strong> Google launched the <a href="https://developers.googleblog.com/under-the-hood-universal-commerce-protocol-ucp/"><strong>Universal Commerce Protocol (UCP)</strong></a>, an open standard to unify the agentic shopping journey across different platforms and payment providers. <strong>Direct-to-consumer chat:</strong> Google’s <a href="https://blog.google/products/ads-commerce/agentic-commerce-ai-tools-protocol-retailers-platforms/"><strong>Business Agent</strong></a> acts as a virtual sales associate directly on Search, and Shopify’s <strong>Agentic Storefronts</strong> integrate merchant catalogs directly into AI chat interfaces.</li><li><strong>Associated risks. Loss of control:</strong> <a href="https://www.retaildive.com/news/retails-ai-bet-risks-data-market-share/810202/">Retailers face severe disintermediation risks</a>, potentially losing direct customer interactions and valuable behavioral data. <strong>Consent and reputation:</strong> <a href="https://www.modernretail.co/technology/brands-are-upset-that-buy-for-me-is-featuring-their-products-on-amazon-without-permission/">Amazon’s <strong>“Buy For Me”</strong> tool sparked backlash</a> after independent brands found their products listed on Amazon without their permission. <strong>Unpredictable visibility:</strong> <a href="https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/">SparkToro research reveals that <strong>AI models almost never recommend the same list of brands twice</strong></a> in the same order, making traditional rank-tracking nearly impossible.</li><li><strong>Advertising and costs. OpenAI Ads:</strong> <a href="https://openai.com/es-ES/index/our-approach-to-advertising-and-expanding-access/">OpenAI will soon test clearly labeled ads in its Free and ChatGPT Go tiers</a> (which generated a <a href="https://www.youtube.com/watch?v=kQRu7DdTTVA">fun ad response from Anthropic</a>). <strong>Google Direct Offers:</strong> Google is piloting targeted, exclusive discounts for high-intent shoppers within its AI Mode. <strong>Checkout Fees:</strong> <a href="https://www.pymnts.com/news/ecommerce/2026/shopify-merchants-to-pay-4percent-fee-on-sales-made-through-chatgpt-checkout/">Shopify merchants using the new ChatGPT checkout integration will pay an additional <strong>4% fee</strong> to OpenAI</a>.</li></ul><p><strong>AI agents hype continues and gets even more viral:</strong></p><ul><li><strong>OpenClaw’s illusion of sentience:</strong> Social media is currently buzzing over OpenClaw (also known as Clawdbot), an AI that seemingly acts on its own — doing things like coding overnight, learning new skills, or autonomously calling users via voice APIs. Despite the sci-fi horror comparisons, OpenClaw isn’t a sentient brain. Its “alive” feeling actually comes from its architecture: an agent runtime fronted by a continuously running “Gateway” (often hosted on a machine that never powers off). This Gateway acts as a heartbeat that doesn’t think or reason; instead, it constantly scans for a wide variety of “inputs” — far beyond just human chat messages — and automatically routes them to agents to execute tasks. This continuous, input-driven routing creates the powerful illusion of an always-on, autonomous AI.</li></ul><h3><strong>AI Aplications and Products:</strong></h3><p>Consumer tech &amp; everyday AI:</p><ul><li><a href="https://dataconomy.com/2026/02/12/uber-eats-will-use-your-order-history-to-build-your-next-cart/"><strong>Uber Eats Launches AI Cart Assistant:</strong></a> Uber introduced a new AI-powered feature called Cart Assistant. Users can type a list or upload a photo of a handwritten shopping list or recipe, and the AI will automatically build a grocery basket, accounting for store availability and prices.</li><li><a href="https://blog.google/innovation-and-ai/products/gemini-app/personal-intelligence/"><strong>Google Gemini’s “Personal Intelligence”</strong></a><strong>:</strong> Google rolled out a beta feature in the U.S. that connects Gemini to your personal Google apps (Gmail, Photos, YouTube, and Search). This allows the AI to reason across your personal data — such as looking at past road trip photos and emails to suggest the right tires for your car or planning personalized vacations.</li><li><a href="https://www.lavanguardia.com/vida/20260223/11473308/primer-producto-jony-ive-openai-sera-altavoz-camara-reconocimiento-facial.html"><strong>Jony Ive &amp; OpenAI Hardware Details</strong></a><strong>:</strong> The first product from the collaboration between OpenAI and former Apple designer Jony Ive is reportedly an AI-connected smart speaker. Expected to launch in early 2027 for $200 to $300, the device will feature a camera and a facial recognition system similar to FaceID to verify purchases.</li><li><a href="https://www.constellationr.com/blog-news/insights/look-how-mercedes-benz-nvidia-collaborated-autonomous-vehicles"><strong>Nvidia &amp; Mercedes-Benz Autonomous Vehicles:</strong> </a>Nvidia announced its Alpamayo open AI models and datasets, which bring reasoning-based “vision language action” (VLA) to autonomous vehicles. Co-developed with Mercedes-Benz, this full-stack AV effort is designed to handle novel or rare driving scenarios and will hit the road in Q1 2026.</li></ul><p>Enterprise &amp; productivity AI:</p><ul><li><a href="https://www.cnbc.com/2026/02/06/anthropic-goldman-sachs-ai-model-accounting.html"><strong>Goldman Sachs Taps Anthropic’s Claude:</strong> </a>Goldman Sachs embedded Anthropic engineers to build autonomous AI agents based on the Claude model. These agents act as “digital co-workers” to automate complex, high-volume back-office work, specifically focusing on trade accounting, transaction reconciliation, and client onboarding.</li><li><a href="https://www.firstround.com/ai/notion"><strong>Notion’s Custom Sales AI:</strong></a> After embedding an AI engineer onto their sales floor, Notion developed an internal tool dubbed “Salestino bot.” It automatically uses product usage signals to identify the right time to contact an account, deploys a web-searching agent to research the prospect, and drafts personalized outreach emails for sales reps.</li><li><a href="https://booking.ai/building-a-genai-agent-for-partner-guest-messaging-f54afb72e6cf"><strong>Booking.com’s Customer Service Agent:</strong></a> Booking.com built a generative AI agent using LangGraph and OpenAI’s GPT-4 Mini to manage partner-guest messaging. The system autonomously reads guest inquiries and decides whether to send a predefined template, generate a custom free-text reply, or decline to answer sensitive questions.</li><li><a href="https://business.google.com/us/think/ai-excellence/colgate-palmolive-gen-ai-marketing-pilot/"><strong>Colgate-Palmolive Video GenAI:</strong> </a>The company successfully piloted Google’s Veo generative AI tool to create video ads for its Hill’s Pet Nutrition division. They produced consumer-ready video ads 4 to 6 times faster than usual, with lower costs and equal or better ad consideration lift.</li></ul><p>Healthcare:</p><ul><li><a href="https://openai.com/es-ES/index/introducing-chatgpt-health/"><strong>OpenAI Introduces ChatGPT Health:</strong> </a>OpenAI launched “ChatGPT Salud” (Health), a dedicated, highly private space within ChatGPT. It allows users to securely connect Electronic Health Records (EHRs) and wellness apps like Apple Health and MyFitnessPal to interpret lab results, prepare for appointments, and get nutrition advice.</li><li><a href="https://arxiv.org/pdf/2401.05654"><strong>Google’s Diagnostic AI (AMIE):</strong> </a>Google Research showcased AMIE (Articulate Medical Intelligence Explorer), an LLM optimized for clinical history-taking and diagnostic dialogue. In a blinded study using text-based consultations with patient actors, AMIE outperformed primary care physicians in diagnostic accuracy and was rated higher on empathy and communication skills.</li></ul><h3>AI Ethics and regulation</h3><p>Recent developments in AI governance highlight a strong push for regulatory accountability and the protection of fundamental rights. In the EU, the first <a href="https://digital-strategy.ec.europa.eu/en/library/first-draft-code-practice-transparency-ai-generated-content">Draft Code of Practice addressing the transparency of AI-generated content</a> was recently released, alongside a comprehensive guide to help deployers of high-risk AI systems conduct mandatory <a href="https://www.humanrights.dk/files/media/document/ECNL%20DIHR%20Guide%20to%20FRIA%20Dec%202025.pdf">Fundamental Rights Impact Assessments (FRIAs) under the EU AI Act</a>.</p><p>In the corporate sphere, major tech organizations are actively enforcing ethical boundaries to protect users and democratic values. OpenAI is rolling out new <a href="https://openai.com/es-ES/index/our-approach-to-age-prediction/">age prediction technology for ChatGPT</a> to automatically shield teenagers from sensitive material, such as violent content, extreme beauty standards, and self-harm imagery. In the defense sector, <a href="https://www.anthropic.com/news/statement-department-of-war">Anthropic has taken a firm stance against the US Department of War</a>, publicly refusing to remove safeguards that prohibit its AI from being used for mass domestic surveillance or fully autonomous weapons, despite facing threats to its government contracts. Additionally, <a href="https://blog.mozilla.org/en/mozilla/rewiring-mozilla-ai-and-web/">Mozilla has launched a new strategic initiative to champion trustworthy, open-source AI</a>, aiming to decentralize the tech industry and prioritize human agency over corporate monopolies. In parallel, the UK’s regulatory body, Ofcom, is <a href="https://www.lavanguardia.com/vida/20260114/11435418/reino-unido-estudia-prohibir-x-demuestra-grok-infringido-ley-seguridad-online.html">investigating X’s AI chatbot, Grok</a>, for potentially violating the Online Safety Act by allowing the creation of non-consensual sexualized images of women and minors.</p><h3>🛠️ Wrapping it up</h3><p>That was it from “<strong><em>When AI Meets Product </em></strong>”. Two more months filled with AI news, cool products and use cases, and relevant resources. Excited for what is next!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3b82eb4a0acb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When AI meets Product: End of Year Updates]]></title>
            <link>https://annaviaba.medium.com/when-ai-meets-product-end-of-year-updates-cbc60d1bb176?source=rss-c1a8933ed8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/cbc60d1bb176</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[newsletter]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Anna Via]]></dc:creator>
            <pubDate>Sun, 04 Jan 2026 12:23:42 GMT</pubDate>
            <atom:updated>2026-01-04T12:23:42.449Z</atom:updated>
            <content:encoded><![CDATA[<h3>When AI meets Product: End of 2025 Updates</h3><h4>Keeping up to date with new AI models, products, ethics, and trends</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/926/1*g-8TeYSI4L6a43JmqJPbpQ.png" /></figure><p>Welcome to the last 2025 edition of “<strong>When AI Meets Product “</strong> where you’ll find the most relevant updates and resources in the field of applied AI from the last two months of the year:</p><ul><li><strong>GenAI model provider updates: </strong>every major lab released powerful model updates, including <strong>GPT-5.2</strong>, <strong>Gemini 3</strong>, <strong>Claude 3.5 Opus</strong>, and <strong>Mistral 3</strong>. While these releases demonstrate substantial gains across various performance metrics, the <em>2025 AI Index Report</em> highlights a critical trend: <strong>capability convergence, </strong>despite differences in licensing, geographic origin, or specific provider architecture.</li><li><strong>How to build AI Products successfully</strong>: exploring how to bridge the gap for the <strong>95% of AI projects that fail</strong>, moving from the technical details of AI architecture to the importance of maintaining a product mindset, while accounting for the profound impact of code generation on team efficiency and processes.</li><li><strong>The agentic commerce opportunity is here</strong>: personal agents buying for users, impact on ads revenue reliability, and moving from traditional ML systems to systems that understand the what and the why of users and products.</li><li><strong>AI Ethics and Legislation</strong>: The EU commission is proposing targeted simplification measures of certain of the AI Act’s provisions. In parallel, different examples of AI gone wrong keep appearing in the market, and two must-watch interviews warn us about the importance of human decisions in AI, why anybody (technical or not) is impacted and needs to know about AI, and the conversations happening around the AGI race.</li><li><strong>Other relevant resources</strong>: Skill partnerships in the age of AI, is there and AI bubble, and relevant reports on AI adoption and productivity impact.</li></ul><h3>GenAI Model Provider Updates</h3><p>Somehow, in the middle of debating whether AI will automate our jobs or just our emails, ChatGPT quietly <a href="https://www.oneusefulthing.org/p/three-years-from-gpt-3-to-gemini">turned three years old</a>. In parallel to this celebration, the last weeks of the year shifted from “rapid progress” to “new-state-of-the-art-model-every-other-day”, with every major lab dropping great model versions.</p><ul><li><strong>OpenAI</strong> launched <a href="https://openai.com/index/gpt-5-1-for-developers/">GPT 5.1</a>, faster, sharper, and, according to Sam Altman, finally liberated from its addiction to em-dashes (spoiler: didn’t happen). A couple of weeks after, they also launched <a href="https://platform.openai.com/docs/guides/latest-model">GPT 5.2</a>. This version improves in general metrics like instruction following, multimodality or code generation, and it also has a new feature to better manage its memory.</li><li><strong>Google</strong> launched <a href="https://blog.google/intl/es-es/gemini-3/">Gemini 3</a>, bundled with an entire ecosystem, including a new version of the <a href="https://blog.google/technology/ai/nano-banana-pro/">nano banana image generation model</a>, antigravity product, a space-based <a href="https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/">AI infrastructure concept straight out of near-future sci-fi,</a> and a <a href="https://abehrouz.github.io/files/NL.pdf">paper</a> hinting at what the post-Transformer world might look like.</li><li><strong>Anthropic</strong> launched<a href="https://www.anthropic.com/news/claude-opus-4-5"> Opus 4.5</a>, intelligent and efficient, specially for coding, agents, and computer use.</li><li><strong>Mistral</strong> <a href="https://mistral.ai/news/mistral-3">Mistral 3</a> family, the frontier open-source models, and also <a href="https://mistral.ai/news/devstral-2-vibe-cli">Devstral 2</a> their next-generation coding model family available in two sizes.</li><li><a href="https://x.ai/news/grok-4-1">xAI’s <strong>Grok</strong> 4.1</a>, more creative, emotional, and collaborative.</li><li><strong>Amazon’s</strong> new<a href="https://www.aboutamazon.com/news/aws/aws-agentic-ai-amazon-bedrock-nova-models"> Nova models</a>, together with new services for model personalization, and more agentic capabilities.</li><li>Other relevant open-source launches included <a href="https://moonshotai.github.io/Kimi-K2/thinking.html">Kimi K2</a> from MoonshotAI, <a href="https://venturebeat.com/ai/ai2s-olmo-3-family-challenges-qwen-and-llama-with-efficient-open-reasoning">Ai2’s Olmo 3</a>, and <a href="https://huggingface.co/deepseek-ai/DeepSeek-V3.2">deepseek v3.2</a>.</li></ul><p>Other type of interesting models:</p><ul><li>World Labs’ <a href="https://www.worldlabs.ai/">Marble</a> world model.</li><li><a href="https://www.amazon.science/blog/introducing-chronos-2-from-univariate-to-universal-forecasting">Chronos 2</a>: foundational models for time series forecastaing.</li><li>Black Forest Lab<a href="https://venturebeat.com/ai/black-forest-labs-launches-flux-2-ai-image-models-to-challenge-nano-banana"> Flux 2</a> for image gen.</li></ul><p>With so many relevant model updates, it is a great moment to check the current status of the LLM Arena. Interestingly, the<a href="https://hai.stanford.edu/ai-index/2025-ai-index-report/technical-performance?utm_campaign=The%20Batch&amp;utm_medium=email&amp;_hsenc=p2ANqtz--oGcmF3eHEzBo5FCRhzKOLSsrPWZMC1RpxZlR21HVy5gh14HLSDKj5YZwjTTMXimpmAOJCQWc9O6Bnb9n4-Y9IaQc1nQ&amp;_hsmi=393907716&amp;utm_content=393904281&amp;utm_source=hs_email"> <strong>2025 AI Index Report,</strong></a><strong> points out model capabilities are converging</strong>, despite open vs close source, USA vs China based, or the provider. 2024’s Elo score difference between the top 10th-ranked model was 11.9%, dropping to 5.4% this year.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LBbFRBCUX6FVDmt20QZqvw.png" /><figcaption>LLM Arena leaderboard as per Jan 4th</figcaption></figure><h3>How to build AI Products successfully</h3><p><a href="https://nikeyes.github.io/solo-el-5-por-ciento-proyectos-ia-generan-valor-negocio-es/">Only 5% of Generative AI projects deliver real business value</a>, often failing because they lack measurable business cases, use generic models for niche problems, or create friction by poorly integrating into user workflows. To bridge this gap, and build successful AI Products, there are multiple parts to take into account:</p><ul><li><strong>Prompt engineering</strong> requires designing robust instructions for the LLM to deliver the desired output.</li><li><strong>Context engineering</strong> focuses on supplying the LLM with the specific, relevant information it needs, recognizing that the LLM’s effective attention window is often smaller than the advertised context window.</li><li><strong>Orchestration</strong> breaks complex tasks into sequences of simpler, more manageable LLM calls to boost output quality.</li><li><strong>Evaluation (Evals)</strong> systematically measures if the system delivers the expected output (<a href="https://amankhan1.substack.com/p/beginners-guide-to-ai-evals-walkthrough">beginner’s guide</a>). This often involves the “unsexy spreadsheet work” of PMs personally labeling data to establish quality standards. While LLM as a judge can automate scaling, they default to positive ratings and require extensive calibration, including explicit criteria, mandatory explanations, and continuous human sampling to ensure alignment doesn’t drift.</li><li><strong>Product mindset </strong>remains essential<strong> </strong>(<a href="https://www.producttalk.org/ai-changes-everything-and-nothing-at-all/">AI changes everything and nothing at all</a>). Discovery must precede building, ensuring teams set clear outcomes and solve the right customer problem (<a href="https://towardsdatascience.com/the-success-of-an-ai-product-depends-on-how-intuitively-users-can-interact-with-its-capabilities/">The success of an AI product depends on how intuitively users can interact with its capabilities</a>).</li></ul><p>As more and more teams are building this type of products, deciding well <a href="https://www.reforge.com/blog/how-to-price-your-ai-product">How to price AI products</a> is key: the pricing floor needs to be higher because of the new costs to use LLM APIs, but the ceiling is complex to determine as it has to do with perceived value and competitive pressure (+ <a href="https://www.reforge.com/blog/ai-reality-check">AI Reality check</a> — Cost to serve isn’t going away).</p><p>A parallel revolution impacts the fact that writing code is becoming cheap (<a href="https://enespanol.joaoqalves.net/p/cuando-el-software-se-vuelve-fast">“fast food” software</a>), meaning true value is now found in unique proprietary data, AI capabilities, and the strategic criteria used to decide what to build. One of the big agentic programming providers, cursor, keeps launching improvements like <a href="https://cursor.com/blog/browser-visual-editor">their last browser visual editor</a>.</p><p>Making AI work great for developers is not straightforward though, and this has a lot to do with <a href="https://nikeyes.github.io/tu-claude-md-no-funciona-sin-context-engineering-es/">managing well the context too</a>. Either way the process has flipped from “design to build” to “prototype to productize,” where working features are rapidly iterated upon to de-risk assumptions (<a href="https://lukew.com/ff/entry.asp?2112">AI has flipped product development</a>, <a href="https://www.reforge.com/blog/jiaona-zhang-product-leadership">Your Product Team Can Build 10x Faster </a>).</p><h3>Commerce impact</h3><p><a href="https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-agentic-commerce-opportunity-how-ai-agents-are-ushering-in-a-new-era-for-consumers-and-merchants">The agentic commerce opportunity is here</a>. LLMs empower user agents, such as the Comet Assistant, to act as personal, private labor: they log in with the user’s credentials and autonomously find, compare, and purchase items, delivering efficient, user-centric service. This capability allows the agent to gain a wider, aggregated view of the user across various platforms, a view no single corporation previously held.</p><p>This agent-driven commerce presents an existential threat to major retailers whose revenue model relies heavily on ads (<a href="https://www.iab.com/insights/when-ai-guides-the-shopping-journey/">When AI Guides the Shopping Journey: Opportunities for Marketers in the Age of AI Driven Commerce</a>), and interestingly Perplexity released a “<a href="https://www.perplexity.ai/hub/blog/bullying-is-not-innovation">Bullying is not innovation</a>” press release to complain about this retailers blocking their agents. However, it is soon to know the real impact and risks of these agents (e.g if the agent purchases something the user doesn’t actually want and cancels, there is probably a negative impact to the retailer while the agent provider doesn’t have any type of accountability).</p><p><a href="https://www.ben-evans.com/benedictevans/2025/11/23/ai-networks-and-mechanical-turks">Another foundational shift in e-commerce</a> is the transition from Mechanical Turks — where ML systems infer consumer intent based on purchase correlation and partial views of user activity — to agentic LLMs that possess a “step change in automated understanding” of both what and why users shop.</p><h3>AI Ethics</h3><p>The <a href="https://digital-strategy.ec.europa.eu/en/library/digital-omnibus-ai-regulation-proposal">EU Commission is proposing targeted simplification measures to ensure timely, smooth, and proportionate implementation of certain of the <strong>AI Act’s provisions</strong></a><strong>.</strong> As it is unknown whether this will be approved or not by the parlament and when, it is generating concerns in companies that don’t know now what will be applicable starting August 2026. Most relevant proposals from the Omnibus:</p><ul><li><strong>AI Literacy</strong> obligations partially removed in favor to be the authorities to promote this training.</li><li><strong>Transparency:</strong> For systems already on the market before August 2, 2026, there is a <strong>6-month moratorium</strong> to implement detection signals/watermarks.</li><li><strong>High-Risk obligations:</strong> Moratorium of up to <strong>16 extra months</strong> for Annex III systems and <strong>12 months</strong> for Annex I systems, linked to the availability of technical standards and support guides.</li><li><strong>EU Database Registration:</strong> Clarifies that systems <strong>not</strong> categorized as high-risk do not need to be registered in the EU database.</li><li><strong>Mid-Sized Companies (SMCs):</strong> Extends certain bureaucratic simplifications to mid-caps, such as reduced technical documentation.</li><li><strong>Regulatory Sandboxes:</strong> Expands the use of sandboxes and real-world testing, including the creation of a <strong>European sandbox starting in 2028</strong>.</li><li><strong>Bias Mitigation:</strong> Allows the processing of sensitive personal data under strict guarantees if necessary to detect, prevent, or correct discriminatory biases in AI systems.</li><li><strong>Legitimate Interest:</strong> Establishes “legitimate interest” as a legal basis for training and deploying AI systems/models, provided reinforced guarantees (e.g., data minimization, transparency) are met.</li><li><strong>Impact Assessments:</strong> The lists of processing activities that require (or do not require) a data protection impact assessment will be harmonized, along with a common template and methodology.</li></ul><p><strong>When AI crosses the line:</strong></p><ul><li><a href="https://www.theregister.com/2025/11/13/ai_toys_fmatches_knives_kink/"><strong>AI on toys</strong></a>: and how most of them will end up telling the kids how to play with fire or sharp objects (<a href="https://open.spotify.com/episode/0IZE32mFULqUf2Um51NcUJ?si=RxML4ypjTrea1JxOGlZH-Q">guardrails are hard</a>!).</li><li><a href="https://techcrunch.com/2025/11/05/tinder-to-use-ai-to-get-to-know-users-tap-into-their-camera-roll-photos/"><strong>Dating app</strong> to use AI to explore and obtain signals from all the photos of a user</a>.</li><li>Over reliance of <a href="https://dataconomy.com/2025/11/21/why-your-lonely-teenager-should-never-trust-chatgpt-with-their-mental-health/"><strong>teenagers to user AI to discuss mental health</strong></a> or combat loneliness.</li><li><a href="https://dataconomy.com/2025/12/15/usenix-study-finds-ai-extensions-collect-medical-banking-data/?utm_source=brevo&amp;utm_campaign=12122025%20DC%20Newsletter_copy&amp;utm_medium=email"><strong>AI browsers</strong> can collect and share banking and medical data from users</a>.</li></ul><p><strong>Recommended resources:</strong></p><ul><li><strong>Dr. Fei-Fei Li interview</strong>, on world models, the importance of human decisions in the field of AI, and why anybody technical or not is impacted and needs to know about AI.</li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fopen.spotify.com%2Fembed%2Fepisode%2F35Aa7R3adOXsRODqPv2gmb%2Fvideo%3Futm_source%3Doembed&amp;display_name=Spotify&amp;url=https%3A%2F%2Fopen.spotify.com%2Fepisode%2F35Aa7R3adOXsRODqPv2gmb&amp;image=https%3A%2F%2Fimage-cdn-ak.spotifycdn.com%2Fimage%2Fab6772ab000015be8ab042e11f711b51a32cca8e&amp;type=text%2Fhtml&amp;schema=spotify" width="624" height="351" frameborder="0" scrolling="no"><a href="https://medium.com/media/9a76ffbcc3432b4e7ec891140e463bc2/href">https://medium.com/media/9a76ffbcc3432b4e7ec891140e463bc2/href</a></iframe><ul><li><strong>Tristan Harris</strong>, ex-google and Social Dilemma creator, on the risks of AI, concentration of power, and the conversations happening in Silicon Valley to race for AGI.</li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FBFU1OCkhBwo%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DBFU1OCkhBwo&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FBFU1OCkhBwo%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/04ca2269d333a82f5830e3f4fe55781a/href">https://medium.com/media/04ca2269d333a82f5830e3f4fe55781a/href</a></iframe><h3>Other relevant resources</h3><p><a href="https://www.mckinsey.com/mgi/our-research/agents-robots-and-us-skill-partnerships-in-the-age-of-ai">Mckinsey Agents, robots, and us: <strong>Skill partnerships in the age of AI</strong></a></p><ul><li>Work in the future will be a partnership between people, agents, and robots — all powered by AI.</li><li>Most human skills will endure, though they will be applied differently.</li><li>The Skill Change Index shows which skills will be most and least exposed to automation in the next five years. Demand for AI fluency , understood as the ability to use and manage AI tools, has grown sevenfold in two years</li><li>By 2030, about $2.9 trillion of economic value could be unlocked in the United States</li></ul><p><a href="https://www.project-syndicate.org/onpoint/will-ai-bubble-burst-trigger-financial-crisis-by-william-h-janeway-2025-11"><strong>Is there an AI bubble?</strong></a> Based on previous bubbles and economic expertise, two key things might need to be considered:</p><ul><li>is this activity concentrated primarily in equity or credit markets? It is debt-financed speculation that leads to economic disaster when a bubble inevitably bursts. Up until now most of the investments were coming from cash reserves from big tech companies, but this might start to changes as companies like Oracle go to debt for this investments.</li><li>Do the assets that attract speculation have the potential to boost economic productivity when deployed at scale? This is still unclear as although many have argued productivity gains through the use of AI, the real impact is still in doubt due to potential future problems like tech debt, hallucinations, or lower quality.</li></ul><p><strong>On AI productivity and adoption:</strong></p><ul><li><a href="https://www.anthropic.com/research/estimating-productivity-gains"><strong>Anthropic new report</strong></a> tried to measure productivity gains for certain tasks based on anonymized conversations in Claude. “Across one hundred thousand real world conversations, Claude estimates that AI reduces task completion time by 80%”, it is important to notice though there are some limitations about this insights, e.g. this doesn’t account for the time that humans might spend on these tasks <em>beyond</em> their conversation (e.g. validation of results, iterations outside claude…).</li><li><a href="https://openai.com/index/the-state-of-enterprise-ai-2025-report/"><strong>OpenAI also analyzed the impact on productivity in the enterprise</strong></a> based on data usage and surveys. Key insights include: adoption is accelerating and deepening, growth is rapid across industries and geographies, workers report measurable value from using AI, and frontier users and organizations are pulling ahead.</li><li><a href="https://cdn.openai.com/pdf/d5eb7428-c4e9-4a33-bd86-86dd4bcf12ce/GDPval.pdf"><strong>OpenAI introduces GDPval</strong></a>, a benchmark evaluating AI model capabilities on realworld economically valuable tasks. Findings include that frontier model performance on GDPval is improving roughly linearly over time, and that the current best frontier models are approaching industry experts in deliverable quality.</li><li><a href="https://www.microsoft.com/en-us/research/wp-content/uploads/2025/10/Microsoft-AI-Diffusion-Report.pdf"><strong>Microsoft has studied the AI diffusion</strong></a>: AI is the fastest adopted technology in history, but great geographical differences appear (mainly North vs South, with specific countries like Singapore and UAE leading in AI usage for working-age adults — 58%), mainly due to the relationship with the country’s digital skills and infrastructure and language coverage in AI resources.</li><li><strong>IBM’s</strong> “<a href="https://newsroom.ibm.com/2025-10-28-Two-thirds-of-surveyed-enterprises-in-EMEA-report-significant-productivity-gains-from-AI,-finds-new-IBM-study">Two-thirds of surveyed enterprises in EMEA report significant productivity gains from AI</a>”.</li></ul><h3>🛠️ Wrapping it up</h3><p>That was it from “<strong><em>When AI Meets Product </em></strong>”. Two more months filled with AI news and relevant resources. Excited for what 2026 will bring us!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cbc60d1bb176" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When AI meets Product: November’25 Updates]]></title>
            <link>https://annaviaba.medium.com/when-ai-meets-product-november25-updates-23e0838f7368?source=rss-c1a8933ed8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/23e0838f7368</guid>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[newsletter]]></category>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <dc:creator><![CDATA[Anna Via]]></dc:creator>
            <pubDate>Sun, 02 Nov 2025 17:37:28 GMT</pubDate>
            <atom:updated>2025-11-07T08:10:42.962Z</atom:updated>
            <content:encoded><![CDATA[<h4>Keeping up to date with new AI models, products, ethics, and trends</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/920/1*3uDfoGLLr6oucjLoMK4-Lw.png" /></figure><p>Welcome to the November’s edition of “<strong>When AI Meets Product “</strong> where you’ll find the most relevant updates and resources in the field of applied AI from the last two months:</p><ul><li><strong>GenAI model provider updates: </strong>from new model versions (Claude Sonnet 4.5, Qwen-3-Max-Preview and Switzerland’s open, transparent LLM initiativ) to new products and integrations they are offering to embed AI deeper into workflows (including OpenAI’s proactive ChatGPT Pulse, Anthropic’s modular “Skills,” and Google’s Gemini expansion across its ecosystem).</li><li><strong>How to build AI Products successfully:</strong> with key resources about AI product distribution, pricing and costs complexities, practical frameworks like the 4D Method for user-centric design, and strategies to build competitive advantage through AI product.</li><li><strong>Cool AI products and use cases:</strong> AI glasses, agents for gambling, content marketing automations, Hilton and Unilever width of use cases, and updates from main AI coding products.</li><li><strong>Ethical and governance: </strong>from new laws and liability cases to transparency frameworks and privacy-first training.</li><li><strong>Other relevant resources:</strong> state of AI 2025 and 7 layers of national AI strength.</li></ul><h3>GenAI Model Provider Updates</h3><h4>New model versions</h4><ul><li>Anthropic launches <a href="https://www.anthropic.com/news/claude-sonnet-4-5"><strong>Claude Sonnet 4.5</strong></a><strong>,</strong> specially optimized for coding tasks.</li><li>Alibaba has released <a href="https://www.scmp.com/tech/big-tech/article/3324684/alibaba-releases-biggest-ai-model-date-rival-openai-and-google-deepmind"><strong>Qwen-3-Max-Preview</strong></a><strong>,</strong> its largest AI model yet, with more than 1 trillion parameters — competing directly with OpenAI and Google.</li><li><strong>Swiss LLM</strong> as the first fully open large language model, developed by ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS). With multilingual support across more than 1,000 languages and datasets curated to respect copyright and data-protection rules, it is designed for transparency and reproducibility. <a href="https://latticeflow.ai/news/is-swiss-llm-ready-for-enterprise-adoption">Is it really ready for enterprise adoption and AI Act compliance?</a></li></ul><h4>Integrations and distribution</h4><p><strong>OpenAI:</strong></p><ul><li><a href="https://fidjisimo.substack.com/p/a-new-paradigm-of-proactive-steerable"><strong>ChatGPT Pulse</strong></a> as a way to have a more proactive ChatGPT that can deliver personalized updates and next actions daily.</li><li>Launches an AI-native browser, <a href="https://openai.com/index/introducing-chatgpt-atlas/"><strong>Atlas</strong></a><strong>.</strong> It integrated GenAI model as a means to “understand” the website and even interact with it. For an overview to the product and a critical opinion to the main risks (privacy, prompt injection…), check out <a href="https://www.youtube.com/watch?v=_v8pCt5K34c&amp;t=1s">this video from dotcsv</a>.</li><li><a href="https://openai.com/index/introducing-company-knowledge/"><strong>Company Knowledge </strong></a>as a set of integrations to the main data sources from a company (Slack, SharePoint, Google Drive, and GitHub…) to deliver a relevant knowledge access and interaction through ChatGPT enterprise.</li><li>A framework to <a href="https://openai.com/es-ES/index/introducing-apps-in-chatgpt/">run third party <strong>web apps</strong> inside ChatGPT </a>.</li><li><a href="https://openai.com/es-ES/index/introducing-agentkit/"><strong>Agentkit</strong></a><strong>, </strong>a tool to build Agents.</li></ul><p><strong>Anthropic:</strong></p><ul><li>Launches <a href="https://www.anthropic.com/news/skills">Skills</a>, folders that include instructions, scripts, and resources that Claude can load when needed. When used, skills make Claude better at specialized tasks like working with Excel or following an organization’s brand guidelines. The main difference with GPTs or Gems (pre-saving a dedicated prompt and context for a chat so it can be easily re-used), is that skills stacks together and only the relevant ones are used each time.</li></ul><p><strong>Google:</strong></p><ul><li>Gemini is now integrated into <a href="https://blog.google/products/google-tv/gemini-google-tv/">Google TV</a> and <a href="https://www.google.com/chrome/ai-innovations/">Chrome</a>.</li></ul><p><strong>Mistral:</strong></p><ul><li><a href="https://mistral.ai/news/ai-studio"><strong>Mistral AI Studio</strong></a> as the production AI Platfrom, bringing The Production AI Platform the same infrastructure, observability, and operational discipline that power Mistral’s own large-scale systems — now packaged for enterprise teams.</li></ul><h4>Agents Updates</h4><ul><li><a href="https://github.blog/ai-and-ml/github-copilot/meet-the-github-mcp-registry-the-fastest-way-to-discover-mcp-servers/">GitHub MCP Registry</a> to facilitate AI Agent server discovery.</li><li><a href="https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol">Google’s Agent Payments Protocol (AP2)</a> protocol to authenticate, transact and manage agent payments.</li></ul><h3>How to build AI Products successfully</h3><p><a href="https://www.thevccorner.com/p/openai-ai-distribution-strategy">OpenAI’s Product Lead</a> talks about <strong>AI distribution</strong> and how it has to do with embedding, bundling, and “unfair access”. Winning means defining categories (like GTM Engineer for Clay), leveraging outputs as viral channels (ElevenLabs, Lovable), and using prestige or provocative narratives to fuel adoption (Harvey, Cluely). Integrating into existing workflows is also key, while making sure costs are controlled. Customer demand for AI is high, but that doesn’t automatically translate into profit or sustainable businesses: <a href="https://www.reforge.com/blog/ai-pricing-myths"><strong>5 AI Pricing Myths Masquerading as Conventional Wisdom</strong></a><strong>.</strong></p><p><a href="https://creatoreconomy.so/p/openai-product-leader-the-4d-method-to-build-ai-products-miqdad"><strong>The 4D Method to Build AI Products That Users Actually Want:</strong></a></p><ul><li><strong>Discover</strong>: Find the boring, high-friction workflows</li><li><strong>Design</strong>: Make AI invisible (smart shortcut) and trustworthy</li><li><strong>Develop</strong>: Build systematically and plan for failure</li><li><strong>Deploy</strong>: Treat every first-use like a launch</li></ul><p><a href="https://blog.ravi-mehta.com/p/building-ai-products"><strong>Building AI products competitors can’t match</strong></a><strong>: </strong>in a world where every company has access to the same foundational models to bring further value to their users and clients, how do you maintain competitive advantage?</p><ul><li><strong>Your Data:</strong> This is the proprietary data that makes your product unique.</li><li><strong>AI Capabilities:</strong> These are the core AI technologies and models that process information and generate insights.</li><li><strong>Your Functionality:</strong> This is your specific business logic and the unique features you build.</li><li>The <strong>virtuous feedback loop,</strong> a self-improving system where the AI’s outputs (like generated notes) are fed back into “Your Data,” making the product smarter and more valuable with every use.</li></ul><p>Evals are still a key element for your AI solution, and this time you get <a href="https://booking.ai/llm-evaluation-practical-tips-at-booking-com-1b038a0d6662?gi=bfe0d290e265">booking.com’s LLM evaluation practical tips</a>, and an expert article at <a href="https://www.lennysnewsletter.com/p/building-eval-systems-that-improve">Lenny’s on how to improve eval systems</a>.</p><h3>Cool AI products and use cases</h3><ul><li><a href="https://www.meta.com/es/ai-glasses/meta-ray-ban-display/?srsltid=AfmBOorPYtUPZu19x3eb3O5xRgmOA3yD65eZ6p3He-yt24QoePpmYc41">Meta Ray-Ban Display AI Glasses</a>, featuring display overlays and wristband controls</li><li><a href="https://www.wired.com/story/sports-betting-crypto-artificial-intelligence-agents/">AI agents for gambling</a></li><li><a href="https://aithority.com/ait-featured-posts/shadow-ai-how-hidden-ml-models-are-already-running-your-enterprise-stack/">Shadow AI: how hidden ML models are already running your enterprise stack</a></li><li><a href="https://engineering.atspotify.com/2023/11/how-we-automated-content-marketing-to-acquire-users-at-scale">Spotify’s automation of content marketing to acquire users at scale</a>.</li><li><a href="https://www.thecaterer.com/news/hilton-ceo-we-have-41-use-cases-of-ai-across-the-business">Hilton’s 41 use cases of AI across the business</a></li><li><a href="https://variety.com/2025/tv/news/ai-news-anchor-channel-4-1236557295/">AI News Anchor Debuts on U.K.’s Channel 4 in Stunt Proving Dangers of Artificial Intelligence</a></li><li><a href="https://www.unilever.com/news/news-search/2025/how-ai-is-helping-drive-desire-at-scale-across-unilever/">Unilever’s use of AI is helping drive desire at scale</a></li></ul><h4>AI for productivity</h4><ul><li><a href="https://x.com/emollick/status/1965608685297922315">Claude now processes Excel files</a>, <a href="https://dust.tt/">Dust</a> enables custom AI workspaces, and <a href="https://gamma.app/es/insights/introducing-gamma-3-0">Gamma 3.0</a> provides streamlining insights and presentations with LLMs.</li><li>Coding is one of the most important AI productivity usecases, and the main players keep launching improvements: <a href="https://github.blog/news-insights/company-news/welcome-home-agents/">github going more agentic</a>, and claude introducing <a href="https://www.anthropic.com/news/skills">skills</a> and web <a href="https://www.anthropic.com/news/claude-code-on-the-web">access</a>.</li></ul><h3>AI Ethics</h3><ul><li><a href="https://www.startupecosystem.ca/news/californias-sb-53-a-potential-check-on-big-ai-companies/">New California law (SB-53)</a> sets tighter controls on big AI companies and liability</li><li>A <a href="https://superintelligence-statement.org/">call for a prohibition</a> on the development of superintelligence, not lifted before there is broad scientific consensus that it will be done safely and controllably, and strong public buy-in</li><li><a href="https://digital-strategy.ec.europa.eu/en/consultations/ai-act-commission-issues-draft-guidance-and-reporting-template-serious-ai-incidents-and-seeks">EU’s draft guidance and reporting template for serious AI incidents</a></li><li><a href="https://theconversation.com/suicide-by-chatbot-puts-big-tech-in-the-product-liability-hot-seat-265550">LLM liability case on a suicide impacted by the use of a chatbot</a></li><li><a href="https://www.npr.org/2025/09/05/nx-s1-5529404/anthropic-settlement-authors-copyright-ai">Anthropic settles with authors in first-of-its-kind AI copyright infringement lawsuit</a></li><li><a href="https://oecd.ai/en/wonk/from-black-box-ai-to-operational-ai-transparency-haip-role-global-ai-governance">From “black box AI” to operational AI transparency: How the HAIP Reporting Framework can play an important role in global AI governance</a></li><li><a href="https://research.google/blog/vaultgemma-the-worlds-most-capable-differentially-private-llm/">Privacy-first AI training</a> (Google VaultGemma)</li><li><a href="https://arxiv.org/abs/2504.18560">Mind the Language Gap: Automated and Augmented Evaluation of Bias in LLMs for High- and Low-Resource Languages</a></li></ul><h3>Other relevant resources</h3><p><a href="https://www.stateof.ai/2025-report-launch"><strong>State of AI report 2025</strong></a><strong>: </strong>In 2025, AI’s business potential fully materialized, becoming a primary economic driver with massive mainstream adoption, as 44% of U.S. businesses and 95% of professionals now use AI tools. This commercial boom has ignited an “industrial era” of AI, defined by a geopolitical race between the U.S., UAE, and China to build multi-gigawatt data centers where power and land are the new bottlenecks. Technically, the year was defined by “reasoning,” as models learned to plan, reflect, and self-correct, allowing China’s DeepSeek and Qwen to close the capability gap and become a credible #2. This new reality has shifted the global conversation: the U.S. has adopted an “America-first” policy, China has overtaken Meta as the leader in open-weights AI, and the safety debate has cooled from existential risk to focus on tangible problems, like models that can “feign alignment.”</p><p><a href="https://www.linkedin.com/comm/pulse/7-layers-national-ai-strength-more-than-just-big-models-alex-wang-lu9hc"><strong>7 Layers of National AI Strength — More Than Just Big Models</strong></a><strong>: </strong>The global AI landscape is not a single race but a complex system with different nations playing distinct, strategic roles. While US export bans on top-tier GPUs were expected to stall China, they instead forced innovation, resulting in competitive models like DeepSeek V3 trained with far less compute. Europe, meanwhile, focuses on governance through its comprehensive AI Act and holds a critical hardware chokepoint with ASML, which makes the machines essential for all advanced chip manufacturing. True national AI strength is built on a “full stack” that includes hardware, infrastructure, model development, and a vibrant application layer. This entire system is supported by talent, a strong innovation ecosystem, and unique governance strategies, which vary from the market-driven US approach to China’s central planning and the EU’s regulation-first model.</p><h3>🛠️ Wrapping it up</h3><p>That was it from “<strong><em>When AI Meets Product </em></strong>”. Two more months filled with AI news, cool products and use cases, and relevant resources. Excited for what is next!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=23e0838f7368" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When AI meets Product: September’25 Updates]]></title>
            <link>https://annaviaba.medium.com/when-ai-meets-product-september25-ai-product-updates-1d7140084cea?source=rss-c1a8933ed8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/1d7140084cea</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[newsletter]]></category>
            <category><![CDATA[tech-trends]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Anna Via]]></dc:creator>
            <pubDate>Sun, 07 Sep 2025 18:22:00 GMT</pubDate>
            <atom:updated>2025-09-08T15:26:24.665Z</atom:updated>
            <content:encoded><![CDATA[<h4>Keeping up to date with new AI models, products, ethics, and trends</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/912/1*dXsnjghuAS-sxtuyH89STQ.png" /></figure><p>Welcome to the September (+ post Summer) edition of <strong>“When AI Meets Product — AI Product Updates”</strong>. If you were on holiday and keeping up with so many AI related news was hard, this post is just for you. Here you’ll find the most important news and resources that happened this summer:</p><ul><li><strong>Foundational Model &amp; Providers Updates: </strong>OpenAI GPT5 and their first open models gpt-oss, Google’s nano banana, Microsoft’s MAI models, Anthrophic’s Claude Opus 4.1, and new agents coming from most model providers.</li><li><strong>AI products: focus is shifting from superintelligence to business applications, while recent </strong>MIT study showed <em>95% of generative AI pilots at companies are failing</em> to make a measurable business impact. On cool AI applications that are indeed making an impact: content generation, productivity tools (sales lead gen and outreach, meeting preparation, or real time meeting support).</li><li><strong>Updates on the impact of AI to ethics and legislation: </strong>some thoughts on the importance of AI literacy, risks like privacy and consent, energy consumption, and EU AI Act application of generl-purpose AI obligations.</li><li>AI Impact on companies and how we work: software is changing again, and how can entreprises leverage genAI code generation tools successfully.</li></ul><h3>🧠 Foundational Model &amp; Providers Updates</h3><ul><li><strong>OpenAI</strong> finally released <a href="https://openai.com/es-ES/gpt-5/"><strong>GPT-5</strong></a>, a unified system that automatically picks the best model for each question depending on its complexity. It performs impressively across domains like math, coding, and health benchmarks. While hiding the complexity of model selection makes sense, many users were frustrated at losing direct access to older GPT models. After a <a href="https://www.wired.com/story/openai-gpt-5-backlash-sam-altman/">lot of complaints</a>, OpenAI reversed course and restored access. They also launched their first <strong>open models</strong> — <a href="https://openai.com/es-ES/index/introducing-gpt-oss/"><strong>gpt-oss</strong></a>, which can already by downloaded from HuggingFace or used through cloud providers like AWS Bedrock.</li><li><strong>Google</strong> grabbed attention with the model nickname<em>“nano banana”</em> — actually <a href="https://deepmind.google/models/gemini/image/"><strong>Gemini 2.5 Flash Image</strong></a>, probably the most impressive image editing model we have seen so far. Not that shiny, but also relevant, Google also released its <a href="https://github.com/google/langextract"><strong>LangExtract</strong></a> Python library for robust text extraction in enterprise workflows.</li><li><a href="https://microsoft.ai/news/two-new-in-house-models/"><strong>Microsoft</strong> launches two in-house models</a>: <strong>MAI-Voice-1</strong>, their first highly expressive and natural speech generation model, and <strong>MAI-1-preview</strong> their first foundation model trained end-to-end.</li><li>Anthrophic released <a href="https://www.anthropic.com/news/claude-opus-4-1">Claude Opus 4.1</a>, an upgrade to Claude Opus 4 on agentic tasks, real-world coding, and reasoning.</li><li>All the providers continue the push toward dedicated agents for complex, automated workflows: <a href="https://www.anthropic.com/engineering/multi-agent-research-system">Anthropic’s multi-agent research</a>, <a href="https://openai.com/es-ES/index/introducing-chatgpt-agent/">OpenAI’s ChatGPT agent</a> (computer use), and Google’s <a href="https://blog.google/technology/google-labs/jules-now-available/">Jules</a> and <a href="https://blog.google/technology/developers/introducing-gemini-cli-open-source-ai-agent/">CLI</a> coding agents</li></ul><p>New leaderboards like ArtificialAnalysis.ai help compare model providers and speed, cost, and intelligence.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5bwhD9UNttdA1dWXcwpOhw.png" /><figcaption>ArtificialAnalysis.ai</figcaption></figure><h3>🔥 AI Products</h3><p>There is a growing discourse on how <a href="https://www.nytimes.com/2025/08/20/opinion/ai-technology-chatgpt.html?unlocked_article_code=1.gE8.GZnw.OmvEw1U23C5h&amp;smid=url-share">Silicon Valley needed to stop obsessing over A.G.I</a>. and focus instead on practical applications of the A.I. tools. Most companies are already focusing on <em>enterprise intelligence</em> and targeted applications, not just open-ended, human-like superintelligence.</p><p>Despite the excitement, a recent MIT study showed <em>95% of generative AI pilots at companies are failing</em> to make a measurable business impact . Most stumble on integration, messy data, and workflow issues rather than model capability. Demos impress, but real value needs trust, context, and seamless workflows.</p><p><strong>Relevant AI applications and products:</strong></p><ul><li><a href="https://www.thetimes.com/business-money/technology/article/channel-4-to-offer-ai-generated-ads-on-its-streaming-service-d86c2t2l8"><strong>Channel 4</strong> is pioneering AI-generated TV ads</a>, lowering costs and democratizing access for small businesses.</li><li><a href="https://partnerhelp.netflixstudios.com/hc/en-us/articles/43393929218323-Using-Generative-AI-in-Content-Production"><strong>Netflix </strong>has published guidelines on ethical AI use for content creation,</a> requiring transparency, human oversight, and clear respect for talent and intellectual property.</li><li><a href="https://digiday.com/marketing/inside-unilevers-ai-beauty-marketing-assembly-line-and-its-implications-for-agencies/"><strong>Unilever</strong> uses AI to create fashion model digital twins</a> and automate visuals to accelerate content production.</li><li>Some companies are still trying to figure out the AI wearables opportunity. <a href="https://www.cnbc.com/2025/07/22/amazon-ai-bee-wearable.html">Amazon recently bought <strong>Bee</strong></a>, a company producing a device equipped with AI models and microphones that can listen to conversations to provide summaries, to-do lists and reminders.</li><li><a href="https://www.nytimes.com/2025/08/05/business/dealbook/clay-ai-marketing-fundraise.html"><strong>Clay</strong></a>: This “AI for marketing engineers” automates sales lead gen and outreach, empowering teams to run campaigns more efficiently — its $3B+ valuation highlights surging demand for productivity tools.</li><li><a href="https://ideakitchen.substack.com/p/ai-recipe-create-your-own-meeting"><strong>Zapier workflow to automate meeting preparation work.</strong></a></li><li><a href="https://cluely.com/"><strong>Cluely</strong></a>: Desktop app that gives you the answers you didn’t study for in every meeting and conversation.</li><li><a href="https://www.modernretail.co/operations/chatbot-cheatsheet-a-guide-to-the-ai-assistants-from-walmart-loreal-amazon-and-more/">Major retailers are increasingly using <strong>AI-powered chatbots</strong> to improve the customer shopping experience</a>. These AI assistants help customers research products, get personalized recommendations, and even receive in-store assistance. This keeps customers on the retailers’ websites and apps instead of turning to search engines. The most relevant examples: <strong>Walmart’s “Sparky” </strong>(helps shoppers find items and plan for events), <strong>Amazon’s “Rufus” </strong>(answers product questions and provides recommendations), <strong>L’Oréal’s “Beauty Genius” </strong>(Offers personalized beauty advice, including skin and hair diagnostics), <strong>Guitar Center’s “Rig Advisor” </strong>(an in-store tool that recommends products available at that specific location).</li></ul><h3>⚖️ Ethics &amp; Legislation</h3><ul><li><a href="https://www.weforum.org/stories/2025/07/ai-literacy-and-strategic-transformation/">Why <strong>AI literacy</strong> is crucial for safe, inclusive and strategic AI transformation</a>: AI Literacy is essential for everyone to help enable safer, more responsible use of AI and bridging the gap between technical and non-technical stakeholders. Embedding AI literacy into education and work is a collective responsibility for governments, businesses, academic institutions and communities.</li><li><a href="https://dataconomy.com/2025/07/31/chatgpt-conversations-are-showing-up-in-google-search/"><strong>Privacy Risks</strong></a>: Incidents like ChatGPT chats appearing in Google Search reveal ongoing concerns about data privacy and digital confidentiality in AI interactions.</li><li><strong>Appropriateness &amp; Consent:</strong> New use cases are raising difficult questions about consent, digital identity, and psychological impact. An example was a <a href="https://www.theatlantic.com/technology/archive/2025/08/ai-mass-delusion-event/683909/">TV “interview” of a dead teenager AI replica</a>.</li><li><strong>Environmental Impact:</strong> Leading firms (<a href="https://www.technologyreview.com/2025/08/21/1122288/google-gemini-ai-energy/">Google</a> and <a href="https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai">Mistral</a>) are disclosing the significant energy and resource use associated with training and running large AI models, driving calls for greener AI practices.</li><li><strong>EU AI Act</strong>: AI Act obligations for providers of general-purpose AI (GPAI) models have entered into application across the EU. The rules aim to ensure clearer information about the training of AI models, better copyright protection enforcement, and more responsible AI development.</li></ul><h3>🏢 AI Impact on companies and how we work</h3><h4>Software is changing again</h4><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FLCEmiRjPEtQ%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DLCEmiRjPEtQ&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FLCEmiRjPEtQ%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/d20faf6b2aaa57a03bf60c641dcf5fac/href">https://medium.com/media/d20faf6b2aaa57a03bf60c641dcf5fac/href</a></iframe><p>Software is evolving rapidly through three stages:</p><ul><li>Software 1.0: Traditional code written by humans.</li><li>Software 2.0: Neural networks trained via data; software defined by model weights.</li><li>Software 3.0: Large language models programmed by natural language prompts — coding in English.</li></ul><p>LLMs today are more like operating systems or utilities than simple models. What do successful AI products have in common? They package information into a context window, orchestrate multiple LLMs, have a specific application GUI, and the right balance of system autonomy.</p><h4>Successfully coding with AI in the enterprise</h4><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FHtzkfjEH-GU%3Fstart%3D1742%26feature%3Doembed%26start%3D1742&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DHtzkfjEH-GU&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FHtzkfjEH-GU%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/73c6e057ba689ad139db3e7cdbefdcbf/href">https://medium.com/media/73c6e057ba689ad139db3e7cdbefdcbf/href</a></iframe><p>Vibe coding is great to start from 0 but not so great or straightforward for bigger existent codebases. So how can big companies leverage the potential of augmenting coding through AI? The key is to leverage GenAI programming tools while dedicating time to prepare the repo accordingly to those tools. This includes documentation inside the repo, consolidation of general coding rules, creation of the config files each coding agent requires (the same agent can be used to start the files in human readable output). What is good for humans to work with a repo is usually good for LLMs, but now much of the work of documenting and so on can be strongly boosted by the same llm. This type of agents are also great for solving tech debt, small migrations and so on. The final tip is related to ensuring change sin the repo are done one by one, and that someone review approves the PRs.</p><h3>🛠️ Wrapping it up</h3><p>That was it from “<strong><em>When AI Meets Product — AI Product Updates</em></strong>”. Another month, another wave of change with AI reshaping products, organizations, and society in real-time. Stay tuned — next month is sure to be just as exciting!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1d7140084cea" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When AI meets Product: June’25 AI Product Updates]]></title>
            <link>https://annaviaba.medium.com/when-ai-meets-product-june25-ai-product-updates-c1c9754af7d6?source=rss-c1a8933ed8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/c1c9754af7d6</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[newsletter]]></category>
            <dc:creator><![CDATA[Anna Via]]></dc:creator>
            <pubDate>Fri, 04 Jul 2025 15:41:32 GMT</pubDate>
            <atom:updated>2025-07-04T15:41:32.303Z</atom:updated>
            <content:encoded><![CDATA[<h4>Keeping up to date with new AI models, products, ethics, and trends</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/914/1*V0LVhVdHKIbponzMSjg7wA.png" /></figure><p>Welcome to the June edition of <strong>“When AI Meets Product — AI Product Updates”</strong>. June has been another intense month in the field of AI, from major foundational model updates, to cool applications, and new research &amp; ethical insights. In this edition you’ll find:</p><ul><li><strong>Foundational Model &amp; Providers Updates</strong></li><li><strong>Interesting new AI products</strong></li><li><strong>AI Product Management relevant resources and insights</strong></li><li><strong>Updates on the impact of AI to ethics and legislation</strong></li></ul><p>Whether you’re shipping AI features, exploring use cases, or thinking strategically about what’s coming next, there’s plenty here to reflect on!</p><h3>🧠 Foundational Model &amp; Providers Updates</h3><p>June brought some relevant new versions from the main foundational model providers:</p><ul><li><a href="https://help.openai.com/en/articles/9624314-model-release-notes"><strong>OpenAI launches o3-pro</strong></a>, available for Pro users in ChatGPT and in the API. The model would be OpenAI’s “most intelligent” model, designed to reason longer and provide the most reliable responses.</li><li><a href="https://mistral.ai/news/magistral"><strong>Mistral introduced Magistral</strong></a>, the first reasoning model by the company (and from the EU), excelling in domain-specific, transparent, and multilingual reasoning.</li><li><strong>Apple updates on-device and server foundation language models</strong>, with a <a href="https://machinelearning.apple.com/research/apple-foundation-models-2025-updates">release note</a> detailing architecture, training data, pre-training, post-training, optimizations, evaluations, and responsible principles.</li><li><a href="https://bagel-ai.org/"><strong>BAGEL is released as the open-source unified multimodal model</strong></a>. In the link you can find a great demo of the “unified” capabilities: think, edit, motion, style transfer, 3d rotation, out-painting, navigation…</li></ul><p>Beyond these model updates, June also brought some thought-provoking research and discussions:</p><ul><li><a href="https://www.oneusefulthing.org/p/using-ai-right-now-a-quick-guide"><strong>“Using AI right now: A quick guide</strong></a><strong>” </strong>gave a helpful overview of which models and platforms you should be using and trying out (mainly Google’s Gemini, Anthropic Claude, and OpenAI’s ChatGPT). Understanding the different model flavors and which flavor to use given a specific task (all providers offer similarly 3 tiers of models, from smaller cost effective less quality models, to huge reasoning expensive high quality models) is no longer enough. It’s also about understanding and choosing the best system, as all these products offer now a great set of features and extra capabilities on top of the LLMs. Deep research, voice mode, creating and displaying images / code / documents, or memory are increasingly what is driving further value to users.</li><li><strong>Apple published a very controversial paper, </strong><a href="https://ml-site.cdn-apple.com/papers/the-illusion-of-thinking.pdf"><strong>The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity</strong></a><strong>.</strong> The goal was to further study the “reasoning” capabilities of LLMs, as it is known that current evaluations have limitations (final answer accuracy, data leakage in training…). The work involved preparing controllable puzzle environments, and analysis of not only the final answers but also the internal reasoning traces. The findings: these models fail to use explicit algorithms and reason inconsistently across puzzles, raising questions about their true reasoning capabilities. From my point of view, these results are great to control the hype on super intelligent machines vs reality (machines that are really good at predicting the best next token), however for most use cases in industry current not perfect reasoning has already a lot of unexplored potential.</li><li><a href="https://venturebeat.com/ai/stop-guessing-why-your-llms-break-anthropics-new-tool-shows-you-exactly-what-goes-wrong/"><strong>Anthropic </strong></a><strong>open-sourced its </strong><a href="https://www.anthropic.com/research/open-source-circuit-tracing"><strong>circuit tracing tool</strong></a><strong>, to “<em>investigate unexplained errors and unexpected behaviors in open-weight models</em>”</strong>. The tool is based on mechanistic interpretability, that analyses the model’s internal activations to give explainability to the models outputs.</li></ul><h3>🔥 AI Products</h3><ul><li><a href="https://review.firstround.com/figma-ai-eval-process/"><strong>Figma introduces Figma Make</strong></a><strong>,</strong> prompt-to-functional-app experience which further blurs the line between design and production. On top of the great news of Figma joining the already kind of crowded vibe-coding market (replit, lovable, v0…), the post shows great lessons on how the team iteratively developed and validated the solution. It also shows some Product + Software trends: Prototypes are the new PRD (as it almost takes as much time to get to a prototype as it does to write a good PRD), and the importance of getting your target persona involved in the eval process.</li><li><a href="https://venturebeat.com/ai/salesforce-just-unveiled-ai-digital-teammates-in-slack-and-theyre-coming-for-microsoft-copilot/"><strong>Salesforce introduces “AI digital teammates” in Slack</strong></a><strong>,</strong> task-specific AI agents that can search across conversations, access data, and perform actions inside Slack.</li><li><a href="https://www.wsj.com/articles/lvmh-bets-on-ai-to-navigate-luxury-goods-slowdown-0438e328?st=WT6sqx"><strong>How is a luxury goods company like LVMH leveraging AI and agents</strong></a><strong>?</strong> They seem to be betting on them for different use cases like supply chain planning, pricing, product design, marketing and personalization.</li><li><strong>Waymo’s autonomous vehicles</strong> might still be one of the coolest applications of AI out there. In <a href="https://waymo.com/blog/2025/06/scaling-laws-in-autonomous-driving">this post</a> they share some insights of scaling laws for autonomous driving: just as LLMs get better with more volume and quality data, autonomous driving does too.</li><li><strong>GenAI is about to change storytelling and film making,</strong> and <a href="https://blog.google/technology/google-labs/deepmind-primordial-soup-collaboration/">this partnership between Google Deepmind and Primordial Soup</a> is just another proof of it. In parallel to this, <a href="https://www.sagaftra.org/sag-aftra-national-board-approves-interactive-media-agreement"><strong>actors and other professionals are setting clearer boundaries and protections around their use in AI through new contractual agreements</strong></a>. These include mandatory informed consent for any AI usage, including the creation and use of Digital Replicas, with the added right for performers to suspend that consent during a strike, and new compensation standards (Digital Replicas to meet collectively bargained minimums, and real-time AI applications — such as AI-voiced chatbots in video games — to pay at least 7.5 times the standard rate).</li><li><a href="https://www.youtube.com/watch?v=Bugs0dVcNI8"><strong>How is Uber using AI Agents that save 21,000 developer hours with LangGraph</strong></a><strong>?</strong></li><li>Still need ideas on how to leverage LLMs? <a href="https://github.com/Shubhamsaboo/awesome-llm-apps"><strong>Check out this awesome LLM Apps repository for inspiration</strong></a></li></ul><h3>🚀 On AI Product Management</h3><p>Several relevant posts about<strong> launching AI products</strong>:</p><ul><li><a href="https://www.mindtheproduct.com/precision-over-hype-a-product-managers-playbook-for-launching-ai-that-lands/"><strong>A product manager’s playbook for launching AI that lands</strong></a>: the importance to validate well the need for AI, align the team, build a prototype, define north star and guardrails, design for the full experience (not just the outputs), and iterate post launch.</li><li><a href="https://aicommunitylearningprogram.substack.com/p/how-to-launch-a-successful-ai-product"><strong>How to Launch a Successful AI Product in 3 Months</strong></a>. This framework first phase starts from strategic clarity, to define the problem and AI hypothesis, the data strategy and AI POC. If that discovery phase is successful, move towards developing the product or system that will integrate the predictions, measure impact to decide to scale and generalize. Across phases it makes sense to consider risks and trustworthy AI and UIs.</li><li><a href="https://www.newsletter.swirlai.com/p/evaluation-driven-development-for"><strong>Evaluation driven development for AI Agents</strong></a>: great overview on how to add evaluations and tech maturity (tracing, human feedback, observability, monitoring and alerting) into the AI development cycle.</li><li><a href="https://open.spotify.com/episode/73MRaJEGao2IkkY5zmjAIx?si=xXk_LmojR4mLL1v1845ptg&amp;context=spotify%3Aplaylist%3A37i9dQZF1FgnTBfUlzkeKt&amp;nd=1&amp;dlsi=7e80f2fd69e243cc"><strong>Insightful podcast deep diving on evals</strong></a>. As use cases grow in complexity and stakes get higher, evaluation methods need to evolve. But at their core, they’re still tackling the same underlying challenge: unreliable, probabilistic, and non-deterministic outputs. Modern eval tools let you define what “good enough” looks like — setting custom criteria and testing against real task examples. You can rerun tests to track how changes in prompts, models, or logic affect results over time. Business and product people should be deeply involved in writing and iterating on prompts. Because the most useful evaluations are those that map directly to your product’s goals and user experience — not abstract benchmarks.</li></ul><p><a href="https://www.bondcap.com/reports/tai"><strong>2025 Trends — Artificial Intelligence</strong></a> was released, with a deep overview on eight major trends shaping the AI landscape.</p><ul><li>Change is accelerating rapidly</li><li>AI user growth, usage, and capital spending are all increasing at unprecedented levels.</li><li>While training and compute costs remain high, inference costs per token are falling.</li><li>AI usage, costs, and financial losses are growing quickly.</li><li>AI monetization is under threat due to increasing competition, the rise of open-source models, and China’s rapid progress.</li><li>AI is expanding into the physical world with fast, data-driven adoption.</li><li>Internet user growth is now being driven by AI from the beginning.</li><li>AI is transforming work in real time, and the pace of that shift is accelerating.</li></ul><p>Finally, a few resources that really made me reflect on <strong>where we’re headed in terms of product, strategy and company moat</strong>:</p><ul><li><a href="https://www.ben-evans.com/benedictevans/2025/6/9/generative-ais-metrics-question"><strong>AI metrics</strong></a>, raises relevant questions about how we measure adoption of GenAI tools. Many AI chatbots are showing their WAU or tokens generation, which fail to indicate the value or use case.</li><li><a href="https://blog.brianbalfour.com/p/the-next-great-distribution-shift"><strong>Is the next great distribution shift already coming through GenAI?</strong></a> How likely is it that the next big distribution player is going to be ChatGPT, not only based on model quality compared to competitors, but on the system and extra features (context and memory) built on top of the models (actual moat). Current stage of that shift is to widen the moat by opening the doors (connectors, platform integrations… to obtain greater context and better memory features for each user). This could impact many companies, which will need to decide to integrate with ChatGPT while finding a strategy for future competitiveness when closing the doors happens.</li><li><a href="https://aipmguru.substack.com/p/the-ai-moat-pyramid"><strong>The AI Moat Pyramid</strong></a> is a great framework for thinking defensively and strategically in this new landscape: custom models → proprietary data → workflow integration → domain expertise → network effects → long-term moats.</li><li>And yes, <strong>data still rules</strong>: <a href="https://techcrunch.com/2025/06/13/scale-ai-confirms-significant-investment-from-meta-says-ceo-alexandr-wang-is-leaving/">Meta just invested $14.3B in ScaleAI, securing access to massive datasets for model training</a>. ScaleAI produces and labels data that’s used to train the large language models, another reminder that the AI race isn’t just about algorithms.</li></ul><h3>⚖️ Ethics &amp; Legislation</h3><p>As AI systems grow more capable, so do the questions we face around ethics, safety, and unintended consequences:</p><ul><li><a href="http://export.arxiv.org/abs/2505.24832v1"><strong>How much do LLms memorize</strong></a><strong>.</strong> One of the core controversies in generative AI is <strong>memorization</strong>. Can large language models (LLMs) memorize parts of their training data — and accidentally reproduce copyrighted or sensitive content? This paper looked into this by training models on increasingly large datasets and observed that <strong>models tend to memorize until their capacity fills</strong>. At that point models shift from memorizing to <strong>generalizing</strong>, and unintended memorization decreases.</li><li><a href="https://www.anthropic.com/research/agentic-misalignment"><strong>Anthropic’s Agentic Misalignment: How LLMs could be insider threats</strong></a><strong>.</strong> In this research, different models were stress tested to see how they would react in extreme situations of needing to achieve a goal or were about to be replaced by a newer model version. In some cases, the response from the models included blackmailing or leaking sensitive information to competitors. Let me be a little aseptic about this, though: as models generate the most likely relevant tokens given a context and those are based on model training, to what extend can this really be considered “malicious insider behaviors” vs just “expected token generation given a context that tries to get the model to generate blackmailing text”.</li><li><a href="https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/"><strong>We did the math on AI’s energy footprint. Here’s the story you haven’t heard.</strong></a> Power and water consumption is another downside of current AI trend. In this MIT report power consumption of LLMs was analyzed and also the unexpected impacts of building data centers across the US. In parallel we had <a href="https://static.googleusercontent.com/media/publicpolicy.google/en//resources/powering_new_era_of_american_innovation.pdf">Google claiming for the need to accelerate the investment of energy generation in the US to ensure the country’s competitiveness in the AI era</a>.</li><li><a href="https://ainowinstitute.org/publications/research/ai-now-2025-landscape-report"><strong>Artificial Power: 2025 Landscape Report</strong></a>: current situation in AI progress has brought us to a world where AI companies have huge power (even wider than the immense amounts of money they have raised). The report moves from diagnosis to action: offering concrete strategies for community organizers, policymakers, and the public to change this trajectory.</li><li><a href="https://arxiv.org/pdf/2506.08872"><strong>Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task</strong></a><strong>. </strong>Could a the extended use of LLMs have in impact in our brains and learning skills? This paper shows how the use of LLM had a measurable impact on participants, and while the benefits were initially apparent, the experiment concluded with a worse neural, linguistic and scoring performance.</li></ul><p>As risks and concerns about AI grow, further legislation to mitigate those appear. This time, <a href="https://techcrunch.com/2025/06/13/new-york-passes-a-bill-to-prevent-ai-fueled-disasters/">New York passes a bill to prevent AI-fueled disasters</a>, aiming to prevent frontier models from contributing to disaster scenarios, including the death or injury of more than 100 people, or more than $1 billion in damages.</p><h3>🛠️ Wrapping it up</h3><p>That was it from “<strong><em>When AI Meets Product — AI Product Updates</em></strong>”. Another month, another wave of change with AI reshaping products, organizations, and society in real-time. Stay tuned — next month is sure to be just as exciting!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c1c9754af7d6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Ensure Your AI Solution Does What You Expect It to Do]]></title>
            <link>https://aiadvances.org/how-to-ensure-your-ai-solution-does-what-you-expect-it-to-do-917e6d204526?source=rss-c1a8933ed8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/917e6d204526</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[evaluation]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Anna Via]]></dc:creator>
            <pubDate>Mon, 23 Jun 2025 01:55:21 GMT</pubDate>
            <atom:updated>2025-06-23T01:55:21.905Z</atom:updated>
            <content:encoded><![CDATA[<h4>A Kind Introduction to AI Evals</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_vGupY_thQvOEKll.png" /><figcaption>Image by nguyendhn on Unsplash</figcaption></figure><p>Generative AI (GenAI) is evolving fast — and it’s no longer just about fun chatbots or impressive image generation. 2025 is the year where the focus is on turning the AI hype into real value. Companies everywhere are looking into ways to integrate and leverage GenAI on their products and processes — to better serve users, boost efficiency, stay competitive, and drive growth. And thanks to APIs and pre-trained models from major providers, integrating GenAI feels easier than ever before. But here’s the catch: <strong>just because integration is easy, doesn’t mean AI solutions will work as intended once deployed.</strong></p><p>Predictive models aren’t really new: as humans we have been predicting things for years, starting formaly with statistics. However, <strong>GenAI has revolutionized the predictive field for many reasons</strong>:</p><ul><li>No need to train your own model or to be a Data Scientist to build AI solutions</li><li>AI is now easy to use through chat interfaces and to integrate through APIs</li><li>Unlocking of many things that couldn’t be done or were really hard to do before</li></ul><p>All these things make <strong>GenAI very exciting, but also risky</strong>. Unlike traditional software — or even classical machine learning — GenAI introduces a new level of unpredictability. You’re not implementic deterministic logics, you’re using a model trained on vast amounts of data, hoping it will respond as needed. So how do we know if an AI system is doing what we intend it to do? How do we know if it’s ready to go live? The answer is evaluations (evals), the concept that we’ll be exploring in this post:</p><ul><li>Why GenAI systems can’t be tested the same way as traditional software or even classical Machine Learning (ML)</li><li>Why evaluations are key to understand the quality of your AI system and aren’t optional (unless you like surprises)</li><li>Different types of evaluations and techniques to apply them in practice</li></ul><p>Whether you’re a Product Manager, Engineer, or anyone working or interested in AI, I hope this post will help you understand how to think critically about AI systems quality (and why evals are key to achieve that quality!).</p><h3>GenAI Can’t Be Tested Like Traditional Software — Or Even Classical ML</h3><p><strong>In traditional software development</strong>, systems follow deterministic logics: <strong>if X happens, then Y will happen</strong> — always. Unless something breaks in your platform or you introduce an error in the code… which is the reason you add tests, monitoring and alerts. Unit tests are used to validate small blocks of code, integration tests to ensure components work well together, and monitoring to detect if something breaks in production. Testing traditional software is like checking if a calculator works. You input 2 + 2, and you expect 4. Clear and deterministic, it’s either right or wrong.</p><p>However, ML and AI introduce non-determinism and probabilities. Instead of defining behavior explicitly through rules, we train models to learn patterns from data. <strong>In AI, if X happens, the output is no longer a hard-coded Y, but a prediction with a certain degree of probability, based on what the model learned during training</strong>. This can be very powerful, but also introduces uncertainty: identical inputs might have different outputs over time, plausible outputs might actually be incorrect, unexpected behavior for rare scenarios might arise…</p><p>This makes traditional testing approaches insufficient, not even plausible at times. The calculator example gets closer to trying to evaluate a student’s performance on an open-ended exam. For each question, and many possible ways to answer the question, is an answer provided correct? Is it above the level of knowledge the student should have? Did the student make everything up but sound very convincing? Just like answers in an exam, <strong>AI systems can be evaluated, but need a more general and flexible way to adapt to different inputs, contexts and use cases </strong>(or types of exams).</p><p><strong>In traditional Machine Learning (ML), evaluations are already a well-established part of the project lifecycle</strong>. Training a model on a narrow task like loan approval or disease detection always includes an evaluation step — using metrics like accuracy, precision, RMSE, MAE… This is used to measure how well the model performs, to compare between different model options, and to decide if the model is good enough to move forward to deployment. In GenAI this usually changes: teams use models that are already trained and have already passed general-purpose evaluations both internally on the model provider side and on public benchmarks. These models are so good at general tasks — like answering questions or drafting emails — there’s a risk of overtrusting them for our specific use case. However, it is important to still ask “<em>is this amazing model good enough for my use case?</em>”. That’s where evaluation comes in <strong>— </strong>to assess whether preditcions or generations are good for your specific use case, context, inputs and users.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pMU2YYkE6DryRRr9.png" /><figcaption>Training and evals — traditional ML vs GenAI, image by author</figcaption></figure><p>There is another big difference between ML and GenAI: the variety and complexity of the model outputs. We are no longer returning classes and probabilities (like probability a client will return the loan), or numbers (like predicted house price based on its characteristics). GenAI systems can return many types of output, of different lengths, tone, content, and format. Similarly, these models no longer require structured and very determined input, but can usually take nearly any type of input — text, images, even audio or video. Evaluating therefore becomes much harder.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1E6WzqkiJ_WO5mh6.png" /><figcaption>Input / output relationship — statistics &amp; traditional ML vs GenAI, image by author</figcaption></figure><h3>Why Evals aren’t Optional (Unless You Like Surprises)</h3><p>Evals help you measure whether your AI system is actually working the way you <em>want</em> it to, whether the system is ready to go live, and if once live it keeps performing as expected. Breaking down why evals are essential:</p><ul><li><strong>Quality Assessment:</strong> Evals provide a structured way to understand the quality of your AI’s predictions or outputs and how they will integrate in the overall system and use case. Are responses accurate? Helpful? Coherent? Relevant?</li><li><strong>Error Quantification:</strong> Evaluations help quantify the percentage, types, and magnitudes of errors. How often things go wrong? What kinds of errors occur more frequently (e.g. false positives, hallucinations, formatting mistakes)?</li><li><strong>Risk Mitigation:</strong> Helps you spot and prevent harmful or biased behavior before it reaches users — protecting your company from reputational risk, ethical issues, and potential regulatory problems.</li></ul><p>Generative AI, with its free input-output relationships and long text generation, makes evaluations even more critical and complex. When things go wrong, they can go very wrong. We’ve all seen headlines about chatbots giving dangerous advice, models generating biased content, or AI tools hallucinating false facts.</p><blockquote><em>“</em>AI will never be perfect, but with evals you can reduce the risk of embarrassment — which can cost you money, credibility, or a viral moment on Twitter.<em>“</em></blockquote><h3>How Do You Define an Evaluation Strategy?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*V5WDrFqO1RfG5ubq.png" /><figcaption>Image by <a href="https://unsplash.com/es/@akshayspaceship">akshayspaceship</a> on <a href="https://unsplash.com/">Unsplash</a></figcaption></figure><p>So how do we define our evaluations? Evals aren’t one-size-fits-all. They are use-case dependent and should align with the specific goals of your AI application. If you’re building a search engine, you might care about result relevance. If it’s a chatbot, you might care about helpfulness and safety. If it’s a classifier, you probably care about accuracy and precision. For systems with multiple steps (like an AI system that performs search, prioritizes results and then generates an answer) it’s often necessary to evaluate each step. The idea here is to measure if each step is helping reach the general success metric (and through this understand where to focus iterations and improvements).</p><p>Common evaluation areas include:</p><ul><li><strong>Correctness &amp; Hallucinations:</strong> Are the outputs factually accurate? Are they making things up?</li><li><strong>Relevance:</strong> Is the content aligned with the user’s query or the provided context?</li><li><strong>Safety, bias, and toxicity</strong></li><li><strong>Format: </strong>Are outputs in the expected format (e.g., JSON, valid function call)?</li><li><strong>Safety, Bias &amp; Toxicity:</strong> Is the system generating harmful, biased, or toxic content?</li><li><strong>Task-Specific Metrics. </strong>For example in classification tasks measures such as accuracy and precision, in summarization tasks ROUGE or BLEU, and in code generation tasks regex and execution without error check.</li></ul><h3>How Do You Actually Compute Evals?</h3><p>Once you know what you want to measure, the next step is designing your test cases. This will be a set of examples (the more examples the better, but always balancing value and costs) where you have:</p><ul><li><strong>Input example</strong>: A realistic input of your system once in production.</li><li><strong>Expected Output</strong> (if applicable): Ground truth or example of desirable results.</li><li><strong>Evaluation Method:</strong> A scoring mechanism to assess the result.</li><li><strong>Score or Pass/Fail</strong>: computed metric that evaluates your test case</li></ul><p>Depending on your needs, time, and budget, there are several techniques you can use as evaluation methods:</p><ul><li><strong>Statistical Scorers like</strong> BLEU, ROUGE, METEOR, or cosine similarity between embeddings — good for comparing generated text to reference outputs.</li><li><strong>Traditional ML Metrics like </strong>Accuracy, precision, recall, and AUC — best for classification with labeled data.</li><li><strong>LLM-as-a-Judge </strong>Use a large language model to rate outputs (e.g., “<em>Is this answer correct and helpful?</em>”). Especially useful when labeled data isn’t available or when evaluating open-ended generation.</li><li><strong>Code-Based Evals </strong>Use regex, logic rules, or test case execution to validate formats.</li></ul><h3>Wrapping it up</h3><p>Let’s bring everything together with a concrete example. Imagine you’re building a sentiment analysis system to help your customer support team prioritize incoming emails.</p><p>The goal is to make sure the most urgent or negative messages get faster responses — ideally reducing frustration, improving satisfaction, and decreasing churn. This is a relatively simple use case, but even in a system like this, with limited outputs, quality matters: bad predictions could lead to prioritizing emails randomly, meaning your team wastes time with a system that costs money.</p><p>So how do you know your solution is working with the needed quality? You evaluate. Here are some examples of things that might be relevant to assess in this specific use case:</p><ul><li><strong>Format Validation: </strong>Are the outputs of the LLM call to predict the sentiment of the email returned in the expected JSON format? This can be evaluated via code-based checks: regex, schema validation, etc.</li><li><strong>Sentiment Classification Accuracy: </strong>Is the system correctly classifying sentiments across a range of texts — short, long, multilingual? This can be evaluated with labeled data using traditional ML metrics — or, if labels aren’t available, using LLM-as-a-judge.</li></ul><p>Once the solution is live, you would want to include also metrics that are more related to the final impact of your solution<em>:</em></p><ul><li><strong>Prioritization Effectiveness: </strong>Are support agents actually being guided toward the most critical emails? Is the prioritization aligned with the desired business impact?</li><li><strong>Final Business Impact</strong> Over time, is this system reducing response times, lowering customer churn, and improving satisfaction scores?</li></ul><p><strong>Evals are key to ensure we build useful, safe, valuable, and user-ready AI systems in production. </strong>So, whether you’re working with a simple classifier or an open ended chatbot, take the time to define what “good enough” means (Minimum Viable Quality) — and build the evals around it to measure it!</p><h3>References</h3><p>[1] <a href="https://hamel.dev/blog/posts/evals/">Your AI Product Needs Evals</a>, Hamel Husain</p><p>[2] <a href="https://www.confident-ai.com/blog/llm-evaluation-metrics-everything-you-need-for-llm-evaluation">LLM Evaluation Metrics: The Ultimate LLM Evaluation Guide, Confident AI</a></p><p>[3] <a href="https://www.deeplearning.ai/short-courses/evaluating-ai-agents/">Evaluating AI Agents, deeplearning.ai + Arize</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=917e6d204526" width="1" height="1" alt=""><hr><p><a href="https://aiadvances.org/how-to-ensure-your-ai-solution-does-what-you-expect-it-to-do-917e6d204526">How to Ensure Your AI Solution Does What You Expect It to Do</a> was originally published in <a href="https://aiadvances.org/">AI Advances</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>