<?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 AoO on Medium]]></title>
        <description><![CDATA[Stories by AoO on Medium]]></description>
        <link>https://medium.com/@AoO.ai?source=rss-9b837c8a77a0------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*p0p362xtYbIvOC9tKU7KDQ.jpeg</url>
            <title>Stories by AoO on Medium</title>
            <link>https://medium.com/@AoO.ai?source=rss-9b837c8a77a0------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 06:35:44 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@AoO.ai/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[27 Unconventional Growth Lessons from Perplexity]]></title>
            <link>https://medium.com/@AoO.ai/27-unconventional-growth-lessons-from-perplexity-6b93b4bd4dc3?source=rss-9b837c8a77a0------2</link>
            <guid isPermaLink="false">https://medium.com/p/6b93b4bd4dc3</guid>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[AoO]]></dc:creator>
            <pubDate>Fri, 31 Oct 2025 04:39:01 GMT</pubDate>
            <atom:updated>2025-10-31T04:39:01.216Z</atom:updated>
            <content:encoded><![CDATA[<p>To promote its AI browser Comet, Perplexity recently rolled out an incentive program: if you invite a friend to download Comet and submit just one query, you receive $15 in cash (settled in Dub), and your friend gets a free month of Pro.</p><p>Tactics like this are commonplace in e-commerce but rare in the AI world. What’s fascinating is that Perplexity treats growth and marketing fundamentally differently from most AI companies.</p><p><strong>Breakout Moves</strong></p><p>Perplexity once created a viral ad inspired by “Squid Game,” collaborating directly with Lee Jung-jae. The idea: use Perplexity’s search Q&amp;A to win challenges. The resulting download spike was enormous, with the CEO later admitting production costs were minimal, but the “ROI was off the charts.”</p><p>They teamed up with F1 champion Lewis Hamilton, with “Perplexity” appearing on his helmet.</p><p>To reach students, Perplexity did more than offer discounts — they turned growth into a competition between schools, where students could win extended access or school-wide perks by spreading the word.</p><p>Perplexity constantly partners with channels like LinkedIn and Lenny’s Podcast, bundling free Pro with other membership deals — a win-win that reliably drives conversion.</p><p>These moves have fueled Perplexity’s exponential growth. From a $150 million valuation and two million MAUs in March 2023, Perplexity has rocketed to a $20 billion valuation within two and a half years.</p><p>Annual queries leapt from 500 million in all of 2023 to over 100 million per week by 2025. Perplexity now logs nearly 200 million monthly site visits and ranks among the top AI web properties globally.</p><p>Drawing from recent interviews with founder Aravind Srinivas and other Perplexity executives, here are their most interesting growth “mental models.”</p><h3>Acquisition Mentality</h3><p><em>“One of my best lessons came from MrBeast in a conversation about advertising: ‘The average internet user isn’t a genius. If you want to reach hundreds of millions, your content must be incredibly simple.’ I remember running a subtle ad campaign once — nobody got it. The virality experts are right: keep it simple.”</em></p><p><em>“On X, you should never post something that’s 100% correct — no one cares. The most engagement comes from content that’s partly right, partly wrong. Hitting a 60/40 or 70/30 balance is ideal. Pure facts don’t get shared.”</em></p><p>“Every post must offer either educational or entertainment value. If it doesn’t, it shouldn’t exist. ‘Excited to announce our new feature’ — no one needs that.”</p><p>“The TikTok algorithm is tuned purely for virality. If you try to go viral, expect that 9 out of 10 videos will flop, but one might hit seven million views. It’s all about power laws.”</p><p>“For TikTok ad tests, you don’t need to spend a ton — $10,000 in a week is enough to know which messages resonate.”</p><p>In AI, “curiosity traffic” will bring users to anything that looks magical — or has a wow-factor demo. “After two years, we’re still benefiting from this organic curiosity. It’s incredible, and remains our very best source for new users.”</p><p>Most Perplexity growth is organic — word of mouth, especially via X and LinkedIn, but users also share permanent answer links through WhatsApp or iMessage. Email updates about new features also matter.</p><p><strong>Influencer and marketing takeaways:</strong></p><p>“We underestimated the time cost of working with lots of small creators. Explaining what we do isn’t easy — ‘AI search’ means little to most. It’s more effective to go deep with a handful of creators who really ‘get’ it.”</p><p>“Sponsoring newsletters rarely works unless your product can be explained in two sentences. Our effort there was a waste — no learnings, no results.”</p><p>“It typically takes 3–7 exposures for someone to truly try a product. Brand marketing’s impact is hard to quantify, but that doesn’t make it useless.”</p><p>“True brand marketing is judged by the secondary conversations it sparks — will your campaign go viral, will it lead to extra sharing and discussion?”</p><p>Community is an underrated channel: “If you can empower your core users to evangelize Perplexity in their circles, you can build real user density and word-of-mouth momentum — students advocating on campus are a great example.”</p><p>“Build things that are easy and desirable to share — look at features like Spotify Wrapped, which made everyone say ‘wow, everyone is sharing this’.”</p><p>“There’s no middle ground left in growth. You either discover untapped channels or outperform everyone else. Most are saturated. If you’re not first, you simply have to be better.”</p><h3>Retention Mentality</h3><p>“Our most important metric is daily queries. Even with a billion installs, if nobody uses the product, it doesn’t matter. On average, users ask 6–7 questions a day. You can’t rely just on existing users for growth. We started with just 3,000 daily queries; 100 million is a milestone — one billion is the real finish line.”lexfridman​</p><p>“If a user does three searches in their first session, that’s a sign they ‘get it’ — these users are more likely to become long-term. Once a user asks around 100 questions, they almost never churn.”</p><p>“Traffic isn’t the problem — retention and early activation are. Improving activation and week 1–2 retention moves all your active user metrics up. Converting from anonymous to registered user is tough; our target conversion rate is 30%.”</p><p>“Multi-device usage is a strong predictor of retention. If people use Perplexity both on the weekend and weekdays, it’s embedded in their routines.”</p><p>“Early on, don’t obsess over CAC or LTV. Focus on growth and very high retention — the monetization path reveals itself later. In the beginning, you should only care about activation and retention.”</p><p>“When should you build a growth team? After you have product–market fit — if you can hit 30% retention at month three or four, it’s time to ‘pour gasoline on the fire’.”</p><p>“Never underestimate product improvement for retention. Micro-optimizations are vastly underrated — even a 10% boost in retention increases the whole active user base. But micro-optimizations have diminishing returns, so make bold moves each quarter: new features or big campaigns, even if only a quarter succeed.”</p><p>“For consumer products, hitting 45% retention at month six is exceptional, especially for a non-social app. Flat retention curves mean you’re delivering sustained value.”</p><p>A/B tests: “A good A/B test tells you what to do next. A bad one just wastes everyone’s time. PMs need strong intuition for the results, or they’re just flying blind.”</p><p>“The most loyal users aren’t just satisfying their curiosity — they use Perplexity to fill knowledge gaps or for real work or study. That insight should shape where you spend on user acquisition.”</p><p>“Growth is fundamentally an intuition game. Build that intuition by knowing your data better than anyone — when you spot real correlations, you unlock your growth engine. Then, you become the ‘numbers expert’ that gets a seat at the table.”</p><p>“If you improve 1% a day, after a year, you’re over 37 times better (1.01365=37.781.01^{365}=37.781.01365=37.78), not just 365%. These compounding improvements, experienced by more users, create word of mouth and trust. When users return weeks later, the product feels faster, simpler, more accurate. That’s how you build trust at scale.”</p><p><em>Sources: Silicon Valley Girl, Semafor, Lex Fridman Podcast, Stanford GSB Insights, Lenny’s Newsletter, Peter Yang, 20VC, Unusual Ventures, and public interviews with Perplexity’s leadership.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6b93b4bd4dc3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[OpenAI Launches the Atlas Browser: Why Some Say It Favors Google and DeepSeek]]></title>
            <link>https://medium.com/@AoO.ai/openai-launches-the-atlas-browser-why-some-say-it-favors-google-and-deepseek-73a18403dbf8?source=rss-9b837c8a77a0------2</link>
            <guid isPermaLink="false">https://medium.com/p/73a18403dbf8</guid>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[AoO]]></dc:creator>
            <pubDate>Wed, 22 Oct 2025 11:21:32 GMT</pubDate>
            <atom:updated>2025-10-22T11:21:32.708Z</atom:updated>
            <content:encoded><![CDATA[<p>OpenAI has finally rolled out its much-anticipated AI browser, ChatGPT Atlas. The announcement triggered a brief dip of about 4% in Google’s stock price, though the rebound that followed shows just how divided the market reaction to Atlas truly was. The fluctuating sentiment paints a picture: Atlas is both controversial and polarizing.</p><p><strong>Atlas in a Nutshell</strong></p><p>Atlas deeply integrates ChatGPT into the browser experience. Its sidebar can read and interact with the content on any page — no more copying and pasting between tabs. The new tab page simultaneously presents AI-driven answers alongside classic Google search results. Its “Agent” mode automates multi-tab tasks; OpenAI’s in-house demo showed it building a grocery order from a recipe, adding items across sites, and even placing the order — all through instructions. Browser “Memories” can recall previous browsing sessions, offering smart suggestions. Atlas is built on Chromium and currently supports only M-series Macs. Free users get the basics, but advanced Agent features are only available to Plus and Pro subscribers.</p><p><strong>Why I’m Switching — and Not Switching Completely</strong></p><p>After trying Atlas for several hours, I’m planning to switch from Dia to Atlas, with Comet and Atlas becoming my main browsers. The reason is simple: Atlas covers nearly all the features I use in Dia, except for cross-tab conversations — which, in practice, I rarely need. What Dia lacks, Agent functionality, is exactly what makes Comet and Atlas addictive: direct browser manipulation via AI. And since Dia’s own AI backend is built on ChatGPT, why not use the native browser instead?</p><p><strong>Atlas vs. Comet: Trust and Transparency</strong></p><p>I’ll continue using both Comet and Atlas because I want a deeper comparison. Many say Atlas blatantly copies Comet — down to the blue agent runtime border. But in use, there are substantial differences. Atlas is faster and, in many cases, smarter. Comet, though, makes it easy to toggle between different models (GPT-5, Claude, Gemini), perfect for those who value “AI horse races” and cross-validation.</p><p>Search flows differ too: Both Dia and Comet ask you to choose AI or Google Search first — Dia prioritizes Google, Comet prioritizes Perplexity, while Atlas displays results from both at once. I suspect this is meant for Chrome users long accustomed to Google, but for me, it further exposes the inefficiency and ad churn of classic search engine ranking.</p><p>What I love about Comet/Perplexity is how every AI-generated answer ends with a citation link. It’s so easy to verify the source and check the reliability of claims. I’m willing to tolerate occasional “AI degeneration” for this transparency. In contrast, Atlas follows the classic ChatGPT format — single authoritative answers with no visible sources, leaving users to judge accuracy by tone and obvious mistakes. For skeptical users like me, Perplexity’s citation-first approach is still preferable to Atlas’s dual-tab display.</p><p><strong>The Browser as Workspace Metaphor</strong></p><p>My second reason for sticking with Comet and Atlas is perhaps niche but powerful: I’ve always mapped one browser interface to a single work project. For years, Safari was for personal browsing and reading, Chrome for work. I’d go further — separate Chrome accounts per project to avoid cross-contamination, customizing browser themes and colors to visually partition my workspace.</p><p>This workflow delivers clear benefits. Launching a particular browser instantly switches me into the corresponding mindset and context. For heavy research and multi-tasking, I set up dozens of tabs in my “project” browser; my personal browser remains minimal and relaxed. Color and layout, along with AI tab and window management, serve as cognitive “hotkeys,” streamlining work across multiple threads.</p><p>Sometimes, it feels like my brain has a GPT-5-style router — each browser or theme becomes a neural interface, channeling focus and energy for different tasks.</p><p>In the AI browser era, this approach only gets more powerful. Atlas, Comet, and Dia all let me tailor automation and tab management to specific workflows. Agent capabilities streamline information gathering and research — just another reason I keep trying new tools.</p><p><strong>Strong Product, Divided Opinions</strong></p><p>While I personally enjoyed Atlas, I was surprised by the depth of negative reactions.</p><p>The biggest controversy centers around privacy and data security — a recurring theme for any browser, amplified in the AI age. After Comet’s recent security scare, skepticism abounds. Some argue Atlas isn’t about improving browsing, but about increasing user data collection — a potential “AI-powered surveillance gateway.” Despite OpenAI’s repeated emphasis on privacy, doubt persists. Privacy-first browsers have always remained a niche market; Atlas hasn’t fully shaken that reputation.</p><p>Others question the product’s core value proposition, seeing Atlas as a mere reskin of Chromium. Why not just use Chrome plus ChatGPT or Gemini plugins? Some say Atlas is a strategic move for OpenAI to reclaim user data and control — not a breakthrough innovation.</p><p>One comment received considerable support: Atlas doesn’t revolutionize browsers, but rather marks a defensive maneuver against Google, whose stronghold remains nearly unchallenged. From this view, Google’s best defense is simply to improve Gemini within Chrome — users needn’t switch at all.</p><p>That analysis is reasonable: AI browsers all fundamentally work the same way, differentiation is tough. But it overlooks one point — ChatGPT is already a massive app with hundreds of millions of monthly users (public estimates put it near 800 million, though OpenAI hasn’t confirmed). If those users shift their habits from chatbot to Atlas, OpenAI can gather substantially richer behavioral data — potentially challenging Google’s Chrome dominance.</p><p>Of course, there are other criticisms floating around — borrowing from Comet’s design, limited to Apple Silicon Macs for now, and similar nitpicks.</p><p><strong>DeepSeek’s “Think Different” Moment</strong></p><p>Interestingly, some off-topic critiques suggest that OpenAI’s product strategy has inadvertently made DeepSeek look cooler.</p><p>Just days before Atlas launched, DeepSeek open-sourced their DeepSeek-OCR model, introducing a new “visual compression” technique for language models. The approach: render lengthy text as images, then use a vision model to turn those images into a small number of visual tokens, which a language model can then decode. Where 1,000 characters once meant 1,000 tokens, DeepSeek can reportedly reconstruct the text with just 100 visual tokens — while retaining 97% accuracy.</p><p>The impact was immediate within the AI developer community; even Andrej Karpathy endorsed it, raising the question: Are pixels better suited than tokens as inputs for large language models? Do text tokens waste resources at the input stage? Perhaps every LLM input should start as an image, even if the origin is text.</p><p>With Atlas and DeepSeek-OCR both launched within days of each other, developers couldn’t help but compare the two flagships. OpenAI, led by Sam Altman, is often described as focusing on market maneuvers, resource aggregation, and rapid product iteration. By contrast, DeepSeek, revered as an “open-source god,” keeps a low profile but is laser-focused on solving industry-critical challenges — like the infamous long-context problem.</p><p>DeepSeek’s other recent highlight was an ACL Best Paper award for “Native Sparse Attention,” which also addresses context length by reengineering attention mechanisms at the training stage. While OpenAI continues to dominate headlines and launch new products (Sora2, GPT-5, among others), its breakthroughs in core architectural innovation seem less public and more incremental — at least in the eyes of developer communities.</p><p>Of course, these opinions are extreme — OpenAI has made impressive progress this year. But when most major players chase market share, traffic, and data, DeepSeek genuinely embodies the spirit of “Think different.” That’s something worth respecting.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=73a18403dbf8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[No Invite, No Hype: Google’s Veo 3.1 vs OpenAI Sora 2]]></title>
            <link>https://medium.com/@AoO.ai/no-invite-no-hype-googles-veo-3-1-vs-openai-sora-2-026cf2f50306?source=rss-9b837c8a77a0------2</link>
            <guid isPermaLink="false">https://medium.com/p/026cf2f50306</guid>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[AoO]]></dc:creator>
            <pubDate>Thu, 16 Oct 2025 15:06:46 GMT</pubDate>
            <atom:updated>2025-10-16T15:06:46.630Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How exclusivity and social psychology turn tech launches into cultural moments.</em></p><p>Google just launched its flagship AI video tool, Veo 3.1. Notice your social feeds filling up with Veo 3.1-generated videos?</p><p>Compare that to OpenAI’s Sora 2. Just weeks ago, Sora 2’s invite-only debut sparked a wildfire across timelines. Holidays didn’t slow anyone down: people raced to get access, shared their results, and made Sora 2 the hottest topic in AI overnight. Why didn’t Veo 3.1 get this treatment?</p><p>Is it about the tech?</p><p>Sora 2 delivered showy upgrades: stunning physics-based effects, better control and persistence, supporting character dialogue and synchronized sound. But Veo 3.1 is technically on par — producing sharper realism, smarter prompt handling, and higher output quality. With videos up to a minute in 1080p, audio enhancements, and expanded editing features, Veo 3.1 has plenty going under the hood.</p><p>Flow, Google’s creative platform, also got major updates — like “Frames to Video” for smooth transitions and “Extend” for longer shots. You can now easily insert or soon remove elements, with natural lighting and seamless reconstruction.</p><p>The tech gap? Not as wide as the buzz gap.</p><p>I see two main reasons:</p><p>First, the name.</p><p>Calling it “3.1” — just a tiny step up from “3” — doesn’t signal a leap forward. If Google had gone with “3.5,” it might have echoed the kind of breakthrough the AI community loves to rally around (“the GPT-3.5 moment for video”). In the world of AI, naming matters.</p><p>Second, the invite strategy.</p><p>Veo 3.1 is wide open — Gemini, Flow, and API all available without barriers. Sora 2 did the opposite: invite-only, with codes becoming digital gold. Scoring an invite meant instant social currency. “Got a Sora code?” was all over AI channels, and people even resold invites for hundreds or thousands online.</p><p>This wasn’t just marketing — it was psychology.</p><p>In consumer products, exclusivity drives both desire and pride. AI launches are no different: the best tech gets turbocharged when it’s paired with scarcity and smart branding.</p><p>User perception is shaped as much by excitement as by engineering. Today, Sora 2 seems hotter than Veo 3.1, so people assume it’s better — even though the reality is nuanced.</p><p>Invite-driven launches tap into classic principles from psychology:</p><p>Social proof: When everyone’s chasing invites, the fear of missing out draws in even more people.</p><p>Scarcity: If access is hard, value goes up. Invite codes are the ultimate “limited edition.”</p><p>Commitment: When you work to get an invite — calling in favors, joining waitlists — you tend to boost the product’s reputation, justifying your effort.</p><p>Better yet, sharing invites forges closer bonds between users and the product. The result? A cascade of advocacy and viral growth.</p><p>Launching products? Don’t underestimate the invite effect.</p><p>Google clearly knows these tactics but chose to play it straight. That’s why I say Veo 3.1 “refuses to copy Sora 2’s playbook,” not that it couldn’t.</p><p>Maybe it’s brand confidence, maybe it’s just Google being Google. But in a competitive landscape, most newcomers don’t get to play the “noble” card — they have to hustle.</p><p>If everyone’s bringing their own exclusive offer to the party, but you show up with full access, you risk missing what makes people care.</p><p>Pride in a product matters — whether it’s earned through the tech itself or through clever participation mechanics. But those are two kinds of pride, and that distinction means something.</p><p>Bottom line:</p><p>Veo 3.1 is technically advanced and widely available.</p><p>Sora 2 generated buzz through exclusivity and social currency.</p><p>The “heat” of a launch is driven by psychology as much as by innovation.</p><p>In AI, winning attention takes more than algorithms — it takes strategy.</p><p>Never underestimate how human nature shapes the fate of tech launches.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=026cf2f50306" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Sam Altman’s 7 Guiding Principles for AI Builders]]></title>
            <link>https://medium.com/@AoO.ai/sam-altmans-7-guiding-principles-for-ai-entrepreneurs-7d2f1bb5d109?source=rss-9b837c8a77a0------2</link>
            <guid isPermaLink="false">https://medium.com/p/7d2f1bb5d109</guid>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[AoO]]></dc:creator>
            <pubDate>Sat, 11 Oct 2025 09:07:11 GMT</pubDate>
            <atom:updated>2025-10-11T09:09:08.664Z</atom:updated>
            <content:encoded><![CDATA[<p>After OpenAI’s 2025 DevDay, CEO Sam Altman sat down for interviews with a16z, Stratechery, and Rowan Cheung to discuss recent trends and his outlook on the AI landscape. This article synthesizes the key takeaways from those conversations, focusing on actionable wisdom for AI entrepreneurs.</p><h3>1. Let Tactics Grow Into Strategy</h3><p>Rowan Cheung: With so much to build right now, what’s your advice for creating and sustaining a unique advantage?</p><p>Sam Altman: The best moats are the ones tailored specifically to you — you have to discover them for yourself. That’s one of the greatest rewards of building a startup. There’s no universal answer. A true advantage is rooted in your specific product, technology, and market, and it almost always emerges from practice. It’s like a business mantra I love: “Let tactics grow into strategy.” You can start by doing small, effective things, and you’ll find that strategy and competitive advantages naturally emerge from the work itself. When we launched ChatGPT, we didn’t know what our long-term advantage would be. Some of its key features, like memory, weren’t even in the initial plan but later became a core part of its strength.</p><blockquote>Author’s Reflection:<br>The story of Claude Code is another classic example of a tactic evolving into a strategy. This reminds me of a concept I encountered recently: the world operates as a “reaction chain.” When you take one step, a cascade of resources and feedback is activated, pushing you to continuously adjust your course. You can’t chart a perfect path to startup success; you have to embrace change daily.</blockquote><h3>2. The Core Logic of Success Hasn’t Changed</h3><p>Rowan Cheung: Given the explosion of possible products, how should founders think about building something with long-term viability?</p><p>Sam Altman: While many things in the world are changing, I believe the core logic of what gives a company a durable advantage has not. Things like network effects, brand and distribution power, proprietary user data, and market dynamics are still fundamental. If you look at what has made businesses successful over the years, the core principles today are not so different from the past. The specific methods for building these advantages may have evolved, but the underlying logic is the same.</p><blockquote>Author’s Reflection:<br>This insight is profound. If you focus only on what’s changing, the entrepreneurial mindset can swing wildly between excitement and anxiety. But if you focus on what’s constant, you can sometimes win by playing a simpler, more fundamental game. It’s the difference between the fox, who knows many things and seizes opportunities in the chaos, and the hedgehog, who knows one big thing and can navigate through cycles. As the prominent investor Kathy Xu (Xu Xin) once said, “Venture capital is the business of investing in what changes the world, while value investing is the business of investing in what the world <em>can’t</em> change.” These are two distinct strategies for survival, built on change and permanence.</blockquote><h3>3. An Innovation Culture is Built on “Betting on People”</h3><p>a16z: You’ve said that other companies can copy products or acquire IP, but they can’t buy your culture or the engine of innovation it powers. How do you build that?</p><p>Sam Altman: It ties back to my background as an investor. A truly great research culture operates less like a product company and more like a top-tier seed-stage venture firm. In that kind of culture, you are making bets on “founders” — in our case, researchers — and structuring your investments similarly. This mindset has been crucial to the culture we’ve built.</p><blockquote>Author’s Reflection:<br>Shou Zi Chew (CEO of TikTok) is another example of a successful investor-turned-CEO. It’s worth noting that the “bet on people” philosophy has always been debated. Many business theories argue that failures often stem from over-reliance on star talent and genius intuition at the expense of systematic, repeatable innovation processes. While talent is undeniably critical, its mobility and unpredictability are major headaches for managers. Most entrepreneurs don’t have the same resources as Sam Altman to attract top talent, and knowing <em>how</em> to bet on the right people is a rare talent in itself.</blockquote><h3>4. User Trust Over Commercialization</h3><p>a16z: What are your thoughts on long-tail advertising?</p><p>Sam Altman: I’m open-minded. ChatGPT has built a strong bond of trust with its users — a trust so strong that it can even withstand occasional mistakes and hallucinations. But that relationship would quickly erode if commercial interests compromised it, for example, by recommending a paid product instead of the best option. Bidded, search-ranking-style ads won’t work. Perhaps there are other ad models to explore, but we have to be extremely cautious.</p><p>Author’s Reflection:<br>User trust is the most critical asset for any product, but its importance is often only felt during a crisis. In another interview, Sam mentioned that Instagram’s ad experience gave him a lot to think about, particularly how it helps users discover new things. He believes that affiliate marketing might be a worthwhile direction to explore.</p><h3>5. True Insight Comes from Practice</h3><p>a16z: What startup opportunities are you most excited about right now?</p><p>Sam Altman: I don’t know the answer. One lesson I’ve learned is that your predictions are almost always wrong, and it has taught me genuine humility. It’s easy to say smart-sounding things that just repeat what others are saying, but it’s incredibly difficult to arrive at a truly unique insight. The only method I know that works is to get your hands dirty: build products, explore the technology, talk to people, and engage deeply with the world to develop your own understanding. Unfortunately, very few people are willing to support that kind of deep exploration, even though it’s consistently the most effective way.</p><blockquote>Author’s Reflection:<br>This brings to mind a powerful idea from investor Li Lu. He described the process of exchanging knowledge as having a “1+1 &gt; 4” effect. In a true exchange, both parties not only keep their own ideas and gain the other’s, but the collision of their thoughts creates entirely new insights.</blockquote><h3>6. Focus on What Matters Most</h3><p>Ben Thompson: Given your strategy of investing as much as possible, what opportunities do you explicitly say no to?</p><p>Sam Altman: We say no to far more opportunities than people imagine. This is very different from early-stage venture investing, where you can spread your bets across dozens of projects. At OpenAI, we’re faced with countless tempting product ideas and emerging directions, but we have to be rigorously disciplined and choose only the very few paths we believe are core to our strategy. It can be painful to let things go, but only by maintaining focus can we execute on these core initiatives to the highest standard and maximize our overall impact.</p><blockquote>Author’s Reflection:<br>At first glance, “betting on people” and “strategic focus” might seem contradictory, but they are actually complementary. It highlights just how difficult strategic betting is for a startup. Knowing what <em>not</em> to do is more important than knowing what to do, yet you can’t afford to do nothing at all.</blockquote><h3>7. When User Feedback and Data Conflict, the User Is Often Right</h3><p>Ben Thompson: How do you think about user feedback, especially when the voices of a dissatisfied minority can be amplified on social media, as seemed to happen with a recent model?</p><p>Sam Altman: The key is to balance data analysis with listening to specific user groups, while remembering that the latter is often a self-selected sample. I generally lean toward data, but when user feedback directly contradicts the data, you have to dig in deep, because the user feedback is often more accurate in those moments.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7d2f1bb5d109" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Modest but Real: 3 Indie AI Tools Solving Everyday Problems (And Earning Their Keep)]]></title>
            <link>https://medium.com/@AoO.ai/modest-but-real-3-indie-ai-tools-solving-everyday-problems-and-earning-their-keep-17b9cbe3323d?source=rss-9b837c8a77a0------2</link>
            <guid isPermaLink="false">https://medium.com/p/17b9cbe3323d</guid>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[AoO]]></dc:creator>
            <pubDate>Mon, 29 Sep 2025 09:31:09 GMT</pubDate>
            <atom:updated>2025-09-29T09:31:38.908Z</atom:updated>
            <content:encoded><![CDATA[<p>01<br>NameQuick is a macOS app that uses AI to automatically recognize and bulk-rename files. Its functionality is as simple as it sounds, yet within just four months of launching, it’s making over $5,000 per month. The creator, Josef from Amsterdam, shared that the idea came from his own frustrations working in SEO — he often spent hours manually renaming images and product files for clients, e.g., changing “IMG_4837.jpg” to “blue-running-shoes-mens-size-10.jpg” for SEO purposes.</p><p>Tired of this tedious work, he decided to build a Mac app that uses AI to analyze files and generate relevant, descriptive names automatically. He began developing in March, launched in April, and now has 275 active users — with most revenue from one-time purchases and about $170 per month from subscriptions. Initially, Josef was unsure if a “file renaming app” was cool or marketable, but real-world pain points often turn out to be extremely monetizable.</p><p>NameQuick’s journey reinforces an old entrepreneurial mantra: “Scratch your own itch.” Solve your own pain, and you’ll likely help many others. Early marketing targeted the SEO industry, but this audience was less willing to pay (many preferred outsourcing the task). A spike from a major product feature led to lots of new users but also churn due to readiness issues. Later, Josef shifted towards productivity-focused Mac YouTubers and bloggers, reaching out individually — this worked better. Interestingly, specialized users like home inspectors picked up the tool, using it to categorize inspection photos by room or problem type, opening up new use cases Josef hadn’t envisioned.</p><p>Today, Josef is running NameQuick full-time and is both its developer and power user. He handles GitHub issues, user emails, and product launches personally every morning. Next, he’s aiming to grow subscription revenue (MRR) past $1,000/month.</p><p>02<br>Reddit is a secret weapon for indie developer growth, but working it is both time-consuming and full of hidden tricks: account building, not getting banned, sifting through relevant posts, and tactical commenting all demand effort and finesse. Many early-stage SaaS teams send founders or interns into countless communities searching for relevant traffic.</p><p>So, what if you could automate that? Enter Tydal, an AI-powered Reddit automation tool launched three months ago, already earning over $1,000/month in recurring revenue. You plug in your product URL, and the AI generates descriptions, recommends subreddits, monitors posts, and even creates engagement-ready comments — using over 50 proven templates for Reddit interactions.</p><p>Tydal claims all automation is within Reddit’s official policies, but of course, sensible discretion is advised. Now with 1,050 users, 46 paying, and total revenue of $1,650, Tydal is seeing steady positive feedback with a “3-day free trial, then paid” conversion model. The developer’s next target: double that MRR to $2,000/month.</p><p>03<br>This last case sparked some debate. An indie developer saw a successful community SaaS called Sublaunch, with $60,000/month in revenue, and built a streamlined, lower-priced competitor — SubStarter — by stripping Sublaunch down to its core feature.</p><p>Sublaunch is a comprehensive platform for creators to monetize over platforms like Telegram, Discord, and WhatsApp. It handles paid community creation, subscription management, permissions, payments, and an integrated referral system, with tiered pricing: 15% commission on the free tier, going as low as 3% for $169/month.</p><p>SubStarter, however, focuses only on Telegram and just the must-have automation and payment features. The free plan takes a 10% commission, while Pro is €49/month with a 3% fee. Some criticize this “just the core feature, lower price” strategy as lacking in business ethics. Others see it as smart and pragmatic — if a market’s core pain point is validated and big players start bloating and raising prices, ultra-focused indie competitors will always have a place.</p><p>For indie makers, targeting proven pain points and doing one thing well (and cheaper) is often more practical than chasing the next trend. These tactics aren’t about “copycats,” but about competition driving product evolution and a reminder: focus on your real market, not on unnecessary feature sprawl.</p><p>Bottom Line:<br>All three examples show that solving “boring” but real pain points — whether it’s renaming files, automating Reddit growth, or streamlining community management — can lead to profitable indie products. Scratching your own itch and serving niche, validated needs is often more sustainable than chasing the hype.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=17b9cbe3323d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ChatGPT Pulse & the AI Content Shift]]></title>
            <link>https://medium.com/@AoO.ai/chatgpt-pulse-the-ai-content-shift-4799d042e3ed?source=rss-9b837c8a77a0------2</link>
            <guid isPermaLink="false">https://medium.com/p/4799d042e3ed</guid>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[AoO]]></dc:creator>
            <pubDate>Fri, 26 Sep 2025 09:38:44 GMT</pubDate>
            <atom:updated>2025-09-26T09:38:44.410Z</atom:updated>
            <content:encoded><![CDATA[<blockquote><em>From Personalized Recommendations to Personalized Generation — Is Swiping Dead?</em></blockquote><p>This morning, ChatGPT announced a “not-so-minor, not-so-major” new feature: Pulse. In his signature style, Chief Marketer Sam Altman called it — once again — his favorite feature to date.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Z7HDyT_jRlDcU9Ek" /><figcaption><a href="https://openai.com/index/introducing-chatgpt-pulse/">https://openai.com/index/introducing-chatgpt-pulse/</a></figcaption></figure><p>So, what is Pulse? Think of it as your AI-powered, proactive information assistant. Each morning, it surfaces a personalized news report for you, pushed proactively in the form of information cards. These cards are generated by analyzing your chat history, AI memory, emails, calendars, and any other connected apps.<br>You can also tell Pulse what you want — what types of information to collect, how to present it, and which sources to prioritize. For now, Pulse is available to Pro mobile users only.</p><p>If you’re a Perplexity user, this idea might feel familiar. <a href="https://www.perplexity.ai/help-center/en/articles/11521526-perplexity-tasks">Their “Tasks” feature</a>, launched three months ago, already lets you automate scheduled searches, generate AI-personalized reports, and deliver these summaries to your inbox or in-app, on a schedule you choose.</p><p>And back in June, the former core team of NotebookLM launched <a href="https://www.huxe.com/">Huxe</a> — a personalized AI audio assistant. Once you connect your email and calendar, Huxe gives you a daily personalized podcast, hosted by two AI “anchors.” The show summarizes your schedule, the latest from your newsletters, and can dive into custom topics (like health or finance). Huxe even lets you interact with the podcast — interrupt anytime, ask questions, drill deeper.</p><p>You could say ChatGPT Pulse is an homage to these pioneers. All three are united by a common vision: in the AI era, content products are being re-invented from the ground up.</p><p>Just a few days ago, Chinese investor Zhu Xiaohu remarked that the “Toutiao and Xiaohongshu of the AI age are already here.” Looking at these products, he might be right.</p><p>During the mobile internet boom, content products competed on the strength of their recommendation algorithms. Data mining captured user interests and preferences, then matched them to a massive content library — think Toutiao, Douyin, Xiaohongshu, WeChat Official Accounts.</p><p>Now, in the AI era, personalized recommendation is evolving into personalized generation. Whether through scheduled chats, information cards, or AI-generated podcasts, these new products all use LLMs to create bespoke, on-demand content for every user.</p><p>After two months with Perplexity Tasks and Huxe, I can say this is a fundamentally new interaction model. In the past, I’d “swipe” endlessly — always hoping to stumble upon something relevant. Now, I simply “ask,” while the AI prepares, organizes, and proactively surfaces what matters. If something is unclear, I follow up with a question.</p><p>Looking back, several shifts stand out:</p><ol><li>In the past, constant swiping was necessary because algorithms never truly understood me. Now, AI draws content directly from my inbox and other personal sources. Even if not everything is of interest, the “proximity” makes it much more relevant — I naturally care more about information drawn from my actual data.</li><li>AI actively filters out spam and merges duplicate content. Ads are reduced, and repetitive information gets summarized. For example, if there’s a major news story, I might receive dozens of similar newsletters — as clickbait, headlines only differ slightly. Now, AI distills the essential facts, and then highlights unique perspectives from different outlets, significantly cutting down noise.</li><li>There’s a new “boss mode” feeling. On my commute, AI briefs me on my schedule, key meetings, and the major news. It’s a little like having an executive assistant. If something catches my interest, I can ask for deeper background — AI handles research and synthesis faster than any human could.</li></ol><p>Of course, not everything is perfect. Classic issues still pop up: AI hallucinations, sometimes fuzzy accuracy — which can even make the experience less efficient than traditional search or browsing.</p><p>So, what ultimately drives the value of these products? It comes down to three things:</p><ul><li>The underlying LLM’s capabilities</li><li>The system’s access to personal, private user data</li><li>The richness of the interactive experience</li></ul><p>As LLMs continue to become commoditized, the unique edge will come from user data and innovative interaction models. This is why every new product now seeks to access your email, calendar, chat records, and browsing data. Only with this highly personalized context can AI generate truly “for you” content.</p><p>Your “personal data pool” is more than email and calendar — it also includes your question history, usage preferences, saved items, and even your browsing footprint. This scramble for deep and wide user data is why major players are building “AI browsers” — whoever owns the richest pool of user data will have the strongest foundation for next-generation AI content delivery.</p><p>As this trend continues, content ecosystems are set to transform. Beyond traditional PGC, UGC, and PUGC sources, the addition of AGC (AI generated content) is likely to reshape information entry points, product forms, advertising models, and the broader logic of information distribution.</p><p>For startups, the data race may seem daunting — giants like Google and Tencent possess natural data moats. Still, opportunities remain in the dimension of user experience and new forms of interaction. AGC unlocks new possibilities for engagement — creating potential for “next Toutiao, Xiaohongshu, or Official Accounts” to emerge around AI-native content.</p><p>Take WeChat Official Accounts, for instance. Neither social graphs, subscriptions, nor recommendation engines have fundamentally solved the “open rate” or content overload problem. AI-powered content products might cut through the noise with smarter filtering and personalized engagement, redefining how content reaches and activates value for users and creators alike.</p><p>Of course, how — and whether — to build such deeply personalized AI assistants is another question altogether. As WeChat’s Allen Zhang might put it, the philosophical debate isn’t just about “how to do it,” but “should we do it at all.”<br>If WeChat (or any platform) did invent an all-knowing “Silicon Being” for its users, what would the world look like? Until these questions are answered, the window is open for AI entrepreneurs to shape the future.</p><p>Right now, the most competitive edge these new products offer comes from owning private user data — like your inbox. In China, this capability could naturally be dominated by the likes of Tencent, whose platforms already aggregate those key data streams.</p><p>What’s Pulse’s real innovation? It pushes the logic of “one face for a thousand people” (algorithmic personalization) further — to “a newspaper for each individual, a report for each need.” Instead of passively swiping an endless stream, you get a curated morning brief that’s just for you — and often programmable, auditable, and actionable. This isn’t just about summarizing information; it’s about closing the loop from generation to delivery, ensuring the info is genuinely useful and ready for what’s next.</p><p>Crucially, subscriptions are being redefined — not to an account or a topic, but to “tasks and results.” You specify, “Here’s what I want to know/achieve today,” and the AI goes out and retrieves, processes, and delivers exactly that, in a format ready for follow-up. Information is moving from “readable” to “actionable.”</p><p>In a phrase, here’s the product evolution: On the supply side, AGC now joins UGC/PGC; distribution is shifting from recommendations to generation and execution; the entry point is migrating from app-layer to OS-level notifications and system integration; and the ultimate goal is no longer maximizing screen time, but maximizing task completion and outcome.</p><p>To understand whether these “AI-era Toutiao” products will stick, watch three things:</p><ul><li>Habit formation: Do they become your first-info-check each morning, with stable daily/next-day retention?</li><li>Practical value: Do the cards/briefs meaningfully shorten the time from search to decision, rather than just re-summarizing summaries?</li><li>Trust and control: Is data authorization granular and user-friendly? Is the generated content traceable, correctable, and reusable downstream?</li></ul><p>When recommendation evolves into generation and execution, the “North Star” for content products shifts — from time-on-site and click rates, to “task completion and user retention.” This is the new “Toutiao moment” of the AI era — not keeping you glued to the feed, but getting you straight to what you need, and preparing you for the next action automatically.</p><p>Ultimately, though, behind all this innovation lies a deeper question — not just how, but whether we should build it. That answer, perhaps, is still waiting for thought leaders like @Allen Zhang to decide. Until then, the window remains open for the next wave of AI builders.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4799d042e3ed" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Burned €800K Building a Team, Returned to a One-Person Company: Four Hard Lessons from Expansion]]></title>
            <link>https://medium.com/@AoO.ai/burned-800k-building-a-team-returned-to-a-one-person-company-four-hard-lessons-from-expansion-692b311c7667?source=rss-9b837c8a77a0------2</link>
            <guid isPermaLink="false">https://medium.com/p/692b311c7667</guid>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[AoO]]></dc:creator>
            <pubDate>Thu, 25 Sep 2025 09:11:00 GMT</pubDate>
            <atom:updated>2025-09-25T09:11:00.058Z</atom:updated>
            <content:encoded><![CDATA[<p>French indie developer Benjamin recently reflected on his entrepreneurial journey. In the past two years, he built a team by hiring six employees, paying out €800,000 in salaries. Today, he’s dissolved the team and returned to being a “one-person company” — now focused on building an AI Agent-powered automated SEO product.</p><p>This experience forced Benjamin to reconsider the issues and traps inherent in team expansion, sparking a lively discussion among other indie developers. Many face the same dilemmas: <br>Your product is growing — should you scale up the team? <br>AI brings new opportunities — would a full-time team help ideas become reality faster? <br>If the funding market is hot, should you hire aggressively to accelerate growth?</p><p>Benjamin distilled his €800,000 — and two years of trial and error — into four tough lessons:</p><p>1. Hire Slow.<br>Even if you have the budget, don’t hire unless you’re truly drowning in work — late nights, weekends, overwhelmed. Never hire just because the funding market is strong, or to “look faster” for VCs.</p><p>2. Beware “Busywork for the Sake of Busywork.”<br>Once you have full-time employees, it’s easy to fill their hours with tasks that don’t actually help the core product. This can slow everything down. <br>For example, if daily operations are light, hiring a full-time designer or operations manager just to keep them busy will generate unnecessary tasks, draining management attention. More people doesn’t always mean better efficiency.</p><p>3. In Practice, AI Does Most of the Work.<br>Benjamin noticed most employees would hand tasks over to AI tools, tweak them, and submit the results. Whether the output was rough or polished, he still had to review and approve everything — making him realize that direct AI management was faster and better, especially for complex product iterations. <br>In the future, he only plans to hire for two qualities: taste (a sense of quality), and agency (self-motivation).</p><p>4. Salary Pressure Shifts Your Focus.<br>Monthly payroll pressure gradually moved his focus away from long-term product-building to simply “making payroll.” He was forced to chase service contracts instead of focusing on product development. Company operations started to feel like just a machine to pay the team, not to pursue what he truly cared about.</p><p>Benjamin set a rule: for his new project, he’ll only consider hiring again when he hits clear product-market fit (PMF) with monthly recurring revenue (MRR) above €20,000. Otherwise, he’ll stay solo.</p><h3>Community Reactions: Agreement and Pushback</h3><p>His second and fourth lessons especially struck a chord with many developers. Some shared similar experiences: <br>Quitting big tech jobs to go all-in solo, realizing it’s better to automate repetitive work with AI agents and focus on decisions and creative tasks. AI won’t replace humans, but as a solo founder, it massively boosts efficiency while reducing cash flow stress.</p><p>Others noted that a solo founder can often match the output of two top employees — thanks to flexibility, lack of slow approvals, and the ability to grind nights and weekends. Hiring the wrong person wastes time on management, whereas using tools and automation (like AI + no-code platforms) avoids emotional overhead.</p><p>But there was also substantial pushback. <br>Many argued Benjamin’s problems came down to poor management — not team structure. <br>The ceiling for solo operations is low; scaling and sustainable growth require building standardized processes and systems, so that even average employees can reliably contribute and the business is repeatable and scalable. The real key is shifting from “hero founder” to “system builder.”</p><p>Others pointed out that tying team expansion to €20,000 in MRR or PMF isn’t as clear-cut as it sounds. Even for solo developers working alongside AI agents, you should prioritize workflow and process from day one — separate out low-value tasks, and aim to make your system as efficient as possible. Hoping to “just hire the right person” isn’t a solution; without process, even the best team can fail.</p><p>Some cautioned that reviews like Benjamin’s could mislead other indie hackers, as his conclusions might reflect an avoidance of management responsibilities rather than universal truths.</p><p>Really, the discussion isn’t about which is “better” — team expansion or solo development. <br>Team size, AI efficiency, processes — these are all tools, not goals.</p><h3>What Actually Matters:</h3><blockquote>As Lean Startup methodology suggests: <br>Test with minimal resources, focus on your core value, and always clarify what truly drives your business forward.<br>Processes and standards are much more important than headcount; AI can boost efficiency, but it can’t replace judgment and vision; and management is a skill all founders must eventually master — avoidance only makes problems repeat.</blockquote><p>At the end of the day, the biggest entrepreneurial risk is putting resources into things nobody needs — whether that’s a big team, or elaborate processes. <br>Your time and budget should be saved for what actually drives growth. <br>And, most of the time, what’s truly important is just one thing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=692b311c7667" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Behind the Largest AI Investment in History: The Rise of the “AI Trust’]]></title>
            <link>https://medium.com/@AoO.ai/behind-the-largest-ai-investment-in-history-the-rise-of-the-ai-trust-c1f218447528?source=rss-9b837c8a77a0------2</link>
            <guid isPermaLink="false">https://medium.com/p/c1f218447528</guid>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[AoO]]></dc:creator>
            <pubDate>Wed, 24 Sep 2025 01:26:19 GMT</pubDate>
            <atom:updated>2025-09-24T01:26:19.847Z</atom:updated>
            <content:encoded><![CDATA[<h3>Behind the Largest AI Investment in History: The Rise of the “AI Trust”</h3><p>On September 22, 2025 (PDT), OpenAI and NVIDIA jointly announced that NVIDIA plans to invest up to $100 billion in OpenAI in stages, supporting the creation of an AI data center with at least 10 gigawatts (GW) of capacity.</p><p>A Record-Breaking Deal<br>The scale of this deal is unprecedented. If executed as outlined, the $100 billion investment will set a new benchmark for single-round technology funding. For context, SoftBank’s first Vision Fund also totaled $100 billion, making it the world’s largest tech investment fund to date — this new deal is of a similar magnitude. Previously, the largest single funding round in tech was OpenAI’s $40 billion raise, led by SoftBank in April 2025. This NVIDIA investment will not only surpass that but also top the all-time record for tech M&amp;A, previously held by Microsoft’s $68.7 billion acquisition of Activision Blizzard in 2022. Excluding government-led projects, such $100-billion-plus deals are extremely rare in business history.</p><p>Ambitious Infrastructure: 10GW of AI Power<br>The project aims to build at least 10GW of AI data center capacity. To put this in perspective: a gigawatt (GW) is a unit of power often used to describe electric grid or power plant capacity. The Three Gorges Dam, the world’s largest hydroelectric power station, has a capacity of around 22.5GW. In other words, the planned AI data centers will require nearly half as much power as this gigantic dam. NVIDIA CEO Jensen Huang noted this project will consume 4–5 million high-end GPUs — equivalent to the company’s entire annual GPU output, and twice as many as last year.</p><p>Deal Structure: Loops and Alliances<br>According to reports, the investment consists of two linked components: NVIDIA will receive non-voting equity in OpenAI, while OpenAI will use the investment funds to purchase NVIDIA chips and systems. The deal’s first $10 billion will be triggered when OpenAI and NVIDIA finalize chip procurement contracts. The initial 1GW phase is set to go live in late 2026 using NVIDIA’s next-generation Vera Rubin platform.</p><p>This structure creates a mutually reinforcing “circular” relationship — NVIDIA gets equity and huge chip orders, while OpenAI secures both funding and computing power to advance its “super AGI” roadmap.</p><p>The Power of Industry Coalitions<br>Outside this specific partnership, OpenAI has been involved in the $500 billion “Stargate” project, along with SoftBank and Oracle, aiming to build the world’s largest AI compute cluster. As part of Stargate, OpenAI will purchase $300 billion worth of cloud services from Oracle. There are even claims that this forms an industry “triangle,” with OpenAI buying cloud services from Oracle, Oracle acquiring chips from NVIDIA, and NVIDIA investing in OpenAI.</p><p>Through intertwined equity and order flows, OpenAI, Microsoft, NVIDIA, and Oracle are forging a powerful alliance at the heart of the AI ecosystem, covering large AI models, cloud computing, and semiconductors. This early formation of an “AI trust” has already drawn regulatory scrutiny, with U.S. authorities starting antitrust investigations. It’s important to note that while this resembles a trust, whether it becomes a true monopoly will depend on regulatory outcomes and future market shifts.</p><p>Competitive Landscape: Evolving from Models to Infrastructure<br>Outside this alliance, Google integrates its self-developed TPU chips, cloud platform, and proprietary models for an all-in-one solution. Amazon, using its market-leading AWS cloud, invests heavily in proprietary Trainium/Inferentia chips and backs AI lab Anthropic, aiming to build a viable alternative in both compute and model ecosystems.</p><p>Whereas past AI competition centered on “who could launch the strongest model fastest,” this deal signals that the new battleground is infrastructure: who can raise and deploy more capital, secure next-generation GPUs, and build the largest, most cost-effective compute clusters.</p><p>The New Bottleneck: Electricity<br>A 10GW AI data center, running at steady load, would consume as much power annually as a mega-city of ten million people. As AI’s appetite for energy grows, electricity will become the fundamental variable in competition: where will the power come from? What’s the cost? How efficiently can these vast resources be brought online?</p><p>If one company — OpenAI — alone needs 10GW in the coming years, the AI industry’s future energy needs will far exceed the capability of any single business. Instead, only governments or nation-scale alliances (or “trusts”) will have the means to orchestrate and allocate such enormous power resources.</p><p>Electricity is fundamental: it fuels both silicon-based civilizations (AIs) and underpins the daily life of carbon-based humans. The battle over energy allocation could ultimately redefine the boundary between these two worlds</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c1f218447528" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Midfield Battle of AI Coding: Sam Altman’s Vendetta]]></title>
            <link>https://medium.com/@AoO.ai/the-midfield-battle-of-ai-coding-sam-altmans-vendetta-116752dfbfdf?source=rss-9b837c8a77a0------2</link>
            <guid isPermaLink="false">https://medium.com/p/116752dfbfdf</guid>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[AoO]]></dc:creator>
            <pubDate>Fri, 19 Sep 2025 13:02:16 GMT</pubDate>
            <atom:updated>2025-09-19T13:03:52.447Z</atom:updated>
            <content:encoded><![CDATA[<h3>Part 1:Feud and Turning Points</h3><p>On September 16, 2025, within 24 hours of the launch of GPT‑5‑Codex, Sam Altman posted a series of celebratory tweets to mark OpenAI’s return to the forefront of the AI coding arena. According to him, the current atmosphere around Codex echoes the frenzy that followed ChatGPT’s initial release.</p><p>This latest version of Codex represents a technical milestone: for the first time, an AI system can autonomously operate within complex projects for up to seven consecutive hours. Codex can debug itself, review pull requests, coordinate multi-end tasks, and essentially close the loop on the entire development process. Yet Altman’s comparison to the early days of ChatGPT is an exaggeration meant to reclaim narrative dominance — especially after a bruising month.</p><p>On August 1st, just before the public release of GPT‑5, Anthropic discovered that OpenAI engineers had accessed the Claude Code API under false pretenses to aid GPT‑5’s development. This breach led Anthropic to cut OpenAI’s access. At the time, Claude Code was the darling of developers — ahead in model accuracy and coding capabilities. Since the release of Claude 4/Opus 4 in May, it had prompted a wave of developer migration. Community estimates put Claude Code’s user base at more than 115,000 developers and a weekly code processing volume approaching 195 million lines. Its annualized company revenue passed $5 billion, representing 5.5x growth in six months. This momentum made Anthropic’s enterprise segment its core growth driver, with annualized corporate revenue hitting $4.5 billion in July 2025.</p><p>This context explains the industry buzz: “Even OpenAI uses Claude Code to write code.” OpenAI responded diplomatically to the API cutoff: “While we respect Anthropic’s decision, it is disappointing — given our own API remains open in return.” A few days later, during the GPT‑5 launch, OpenAI pointedly declared, “GPT‑5 is our best coding model ever,” seeking to reclaim mindshare.</p><p>The event, however, was marred by communication missteps and user backlash, diverting attention from key product highlights. It took a month and a recentering of narrative focus for OpenAI to emphasize its code-centric ambitions again.</p><p>Meanwhile, a spate of “downgrades” and unpopular updates to Claude Code began causing community discontent about its reliability and “intelligence.” In public developer forums, users increasingly discussed switching back to Codex. Sam Altman, seizing the moment, referenced these trends in his posts — hinting at user migration and benchmark gains for Codex.</p><p>Insiders say this counteroffensive was long in the making. OpenAI had previously tried to acquire leading AI coding startups like Cursor and Windsurf — signaling an intent to reclaim dominance. The launch of the new Codex marks the opening shot of this strategic push.</p><p>The rivalry between OpenAI and Anthropic — rooted in differences over AI safety and commercialization — runs deep. Dario Amodei, who played a central role in GPT‑2 and GPT‑3 at OpenAI, left with his team in 2020, citing irreconcilable differences over corporate values and safety philosophy. At Anthropic, he and his sister Daniela championed “controllable, trustworthy AI.”</p><p>After a year of ChatGPT’s success, Anthropic released Claude as a challenger. In February 2025, Anthropic introduced a new paradigm — Agentic Coding — enabling AI to surpass OpenAI’s offerings in the code domain. The underlying dispute is philosophical: Should AI be given greater autonomy, or must it remain strictly governed by human oversight?</p><h3>Part 2: Product Philosophies Diverge</h3><p>AI coding tools have evolved quickly. The 1.0 era, with Copilot as the flagship, focused on context-based code autocompletion. With the emergence of tools like Cursor and Windsurf, “AI-native IDEs” set off the 2.0 wave: AI now “takes orders” to execute codebase-level tasks, thus inaugurating Agentic Coding. This shift drove a redesign in developer experiences — moving toward collaborative, graphical interfaces.</p><p>Claude Code chose a different path: a focus on the terminal, making it closer to the natural workflows of power users and operations teams. Its terminal-first design offers programmability and composability, enabling developers to integrate AI into pipelines and connect to a range of plugins and APIs via the MCP protocol. This broadened the scope for solving complex tasks — design, development, and collaboration alike.</p><p>February 2025 saw Claude Code’s integration with the hybrid reasoning model Claude 3.7 Sonnet. This allowed the AI to flexibly toggle between efficiency and depth, making its reasoning chain transparent and building user trust. By May, the release of Claude 4/Opus4 meant further improvements in precision and agent capabilities. Developers could now turn natural language instructions into full development cycles, evolving AI from “assistant” to “partner.”</p><p>This surge in capability and popularity triggered pressure on OpenAI. In July, with tension over API access, OpenAI found itself on the defensive.</p><p>The new Codex is OpenAI’s answer. Incorporating hybrid reasoning but charting a new path, its breakthrough feature is “dynamic thinking time”: Codex judges problem complexity in real time, allocating resources as needed — responding instantly to simple tasks and dedicating hours to complex challenges. This elasticity reduces latency and resource cost, making long-running AI autonomy possible.</p><p>Codex also unifies CLI tools, IDE plugins, and cloud workflows — attempting to merge the GUI-driven model from Cursor with the terminal/agent model espoused by Claude Code, aiming for seamless cross-platform experience.</p><p>Notably, GPT‑5 was uniquely fine-tuned for Codex, an investment uncommon for OpenAI’s generalist AGI ambitions. Data from OpenAI suggests that coding-centric users comprise only a small portion of their vast audience — highlighting the strategic significance ascribed to Codex internally.</p><p>External reviews generally validate Codex’s new strengths: dynamic thinking time enables multi-hour, continuous, autonomous coding; cross-platform workflow is a standout feature. Yet Codex is not ideal for beginners or exploratory projects — its workflows remain more closed, and current multi-tool orchestration is less flexible compared to Claude Code.</p><p>In summary, Codex and Claude Code are diverging along fundamental lines: Codex focuses on autonomy and streamlined execution; Claude Code on control, transparency, and developer agency. This is a mirror of the founders’ philosophical split. OpenAI trusts in AI autonomy, positioning Codex as a tireless, reliable “partner.” Anthropic insists on human-in-the-loop, building safety, auditing, and local deployment into every aspect of Claude Code.</p><p>This debate has even drawn comment from industry titans — Nvidia CEO Jensen Huang, for example, has publicly questioned such tightly held safety stances.</p><p>Just as religious debates over the existence of God have shaped societies, today’s AI “belief divides” are being played out in the marketplace and engineering trenches.</p><h3>Part 3: The Competitive Arena</h3><p>OpenAI’s counterattack on Anthropic goes beyond product — notably targeting business models. In 2025, industry sources estimate OpenAI’s annualized revenue at more than $13 billion. Subscriptions account for 55–60%, enterprise clients for 25–30%, and APIs/developer ecosystem for the rest. Anthropic, by contrast, has prioritized enterprise clients from inception — covering sectors like pharma, travel, and finance. CEO Dario Amodei has stated that the number of $10 million and $100 million clients tripled year-on-year, driving its revenue to $4.5 billion as of July.</p><p>Anthropic’s income is highly concentrated: API and enterprise clients contribute more than 85%, while consumer subscriptions are a minority.</p><p>Codex’s technical documentation now prioritizes features for enterprise and API markets: sandboxing, granular approval, seamless cloud-local migration, federated team collaboration, audit trails, and flexible quota systems — all aimed at compliance-driven, high-volume users.</p><p>But OpenAI faces competitors of its own. In April, OpenAI tried to acquire Windsurf for $3 billion, but negotiations stalled — reportedly in part due to major shareholder Microsoft’s interests in Copilot. Days after exclusivity lapsed, Google moved in, closing a $2.4 billion deal to acquire tech and talent, integrating Windsurf IP and its team into Gemini AI. Remaining Windsurf assets were bought by Cognition AI, developer of another known code agent, Devin.</p><p>By 2025, the AI coding landscape saw at least 10 funded projects — including Codeium, Continue, Unblocked, Cline, Cognition AI (Devin), Cursor, Lovable, Greptile, Graphite, and CodeRabbit — while Chinese players like DeepSeek, Zhipu (GLM), and Qwen joined the top ranks of community code agent leaderboards.</p><p>It’s widely agreed: after chatbots, AI coding is the most fiercely contested battlefield in AI.</p><h3>Part 4: Users and the Boundaries of Trust</h3><p>The developer community’s response has been anything but monolithic. Initially, Agentic Coding stirred fears about jobs, but it soon became clear that even as tools like Claude Code or GPT‑5‑Codex can take a project from “idea to prototype,” developers remain indispensable for productization, scaling, and reliability.</p><p>Many find that AI code generation delivers big productivity gains in prototyping — but at scale and in production, issues with structural clarity and stability emerge. Extensive human refactoring often follows.</p><p>Recent criticisms of Claude Code have focused on its “dumbing down,” i.e., repeated failures to comprehend or fix simple bugs. Codex, by contrast, has faced skepticism about its toolchain maturity and complaints that its “dynamic thinking” sometimes causes it to skip complex or indeterminate tasks.</p><p>Consequently, user opinions are polarized: newcomers and non-engineers are thrilled with “from idea to product” capabilities; seasoned developers express doubts about code quality and maintainability, reserving deep trust. These splits essentially map to OpenAI vs. Anthropic worldviews: optimism about AI’s potential versus guarded skepticism.</p><p>Leading figures like Andrej Karpathy advocate for “symbiosis” between code agents and human engineers but admit the future remains uncertain.</p><p>As depicted in the latest “Black Mirror” episode, philosophical debates about trust, risk, and autonomy aren’t just speculative fiction — they are shaping the future of technology.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=116752dfbfdf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Anthropic Builds Tools for Agents: An Agent-Centric Approach]]></title>
            <link>https://medium.com/@AoO.ai/how-anthropic-builds-tools-for-agents-an-agent-centric-approach-5b068803135a?source=rss-9b837c8a77a0------2</link>
            <guid isPermaLink="false">https://medium.com/p/5b068803135a</guid>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[AoO]]></dc:creator>
            <pubDate>Fri, 12 Sep 2025 06:56:25 GMT</pubDate>
            <atom:updated>2025-09-12T06:56:25.246Z</atom:updated>
            <content:encoded><![CDATA[<p>Since Anthropic released its MCP (Meta-tool Calling Protocol), AI agents have become capable of leveraging hundreds of tools to perform complex tasks. Some developers have even suggested that every app is worth rebuilding with agents. However, a new problem has quickly emerged: simply providing an agent with more tools doesn’t guarantee efficiency or stability and can sometimes lead to hallucinations.</p><p>In traditional software development, API calls produce consistent results. But when an agent uses a tool, it might take a different path each time due to variations in context, understanding, and intent, potentially using the wrong tool or misinterpreting information. Anthropic recently shared their practical methodology for building tools for agents, revealing that the old mindset of “engineers writing APIs for other engineers” is no longer effective. Instead, it requires a complete shift in perspective: rethinking and redesigning every detail from an <strong>agent-centric</strong> viewpoint to ensure agents can truly learn and use tools correctly. In Anthropic’s workflow, the agent is not just a user of tools but also a co-developer, deeply involved in the tool’s creation and iteration.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8OrHyatXJuKVKUs9" /><figcaption>(<a href="https://www.anthropic.com/engineering/writing-tools-for-agents">https://www.anthropic.com/engineering/writing-tools-for-agents</a>)</figcaption></figure><h3>Anthropic’s Development Workflow</h3><p>Anthropic’s core methodology for agent tool development can be summarized as an iterative approach with deep AI collaboration.</p><p>First, they prioritize building a minimal viable tool prototype, running it directly on a local MCP server or a Desktop Extension (DXT). This allows for immediate integration and testing with Claude Code or a desktop environment. Before development, they gather all relevant documentation, dependencies, and API specifications, preferably in an llms.txt format, to help Claude understand the tool&#39;s goal and context.</p><p>To align the tool with real-world needs, the team designs complex and diverse evaluation tasks based on actual business scenarios. This allows the agent to test various usage paths in batches and assists in analyzing the results. Each test assesses not only if the tool “works” but, more importantly, where the agent gets stuck, encounters ambiguity, or experiences errors during a call.</p><p>During evaluation, the team meticulously reviews logs and the agent’s chain-of-thought reasoning, tracking every detail of tool calls, parameter passing, errors, and repeated calls. This process helps uncover issues. For example, the team once noticed Claude persistently adding “2025” to search parameters, skewing results. They solved this by optimizing the tool’s description to guide the agent correctly.</p><p>Crucially, Anthropic feeds all evaluation dialogues and operational records, including the agent’s reasoning process and call chain, into Claude Code for batch analysis and automated tool improvement.</p><p>Through this process, the AI becomes the team’s “tool auditor and optimizer,” automatically identifying shortcomings in tool naming, parameters, and descriptions, and helping to verify that the entire tool system remains logical and coherent after each modification.</p><p>To ensure tools are effective in practice and not just “good for internal demos,” the team maintains a set of test tasks from real business operations. Every tool update is validated against these tasks to avoid creating tools that can only pass specific tests but fail to generalize.</p><h3>Five Core Development Principles</h3><h4>1. Choose Agent-Appropriate Tools</h4><p>More tools are not always better. Anthropic has observed that many developers wrap existing functions or APIs directly into new tools without considering their suitability for agents. This is because agents and traditional software differ vastly in how they use tools, from operational understanding to data ingestion. For instance, an LLM’s capacity for processing information at one time is limited, whereas traditional software can handle massive datasets in loops. An agent given a tool that “returns all contacts at once” is not only inefficient but also becomes overwhelmed by irrelevant information. A tool that allows “searching by initial” or “searching by keyword” is far more effective.</p><p>The team’s experience suggests focusing on creating high-quality, targeted tools that solve core problems — such as specific search, filter, or messaging tools — rather than simply bundling a large collection of APIs.</p><p>Tools can also be consolidated. Instead of separate tools for “view user,” “view schedule,” and “create event,” a single “schedule event” tool can cover the entire workflow.</p><p>For logging functions, a “search for key information” tool is better than one that exposes all raw logs.</p><p>For customer information, a composite tool that “retrieves all key customer information at once” is recommended.</p><p>Each tool should have a clear purpose and well-defined boundaries to help the agent break down and combine tasks logically, just as a human would. Too many or overlapping tools will only hinder the agent’s performance. A curated set of well-defined tools is more effective than a massive assortment of APIs.</p><h4>2. Unify Namespaces to Reduce Confusion</h4><p>An AI agent might connect to multiple MCP servers and use hundreds of tools from different sources. If tools have overlapping functions or unclear purposes, the agent can easily get confused about which one to use. This is where <strong>namespaces</strong> are critical. Grouping related tools with a unified prefix, such as by service (asana_search, jira_search) or by resource (asana_projects_search, asana_users_search), helps the agent accurately differentiate and select the appropriate tool.</p><p>Tests show that naming schemes with prefixes or suffixes significantly impact an agent’s ability to call tools correctly, with performance varying across different large models. It is recommended to continuously adjust and optimize naming conventions based on practical evaluations. A scientific approach to naming and grouping reduces the risk of incorrect tool selection, faulty parameter passing, or functional confusion. It also compresses the context, allowing the agent to focus on the overall task rather than intermediate implementation details.</p><h3>3. Return Meaningful Contextual Information</h3><p>The data a tool returns to an agent shouldn’t be a heap of technical details. Raw data like a UUID (e.g., a87ffd387...), internal technical links, or file type codes are not easily understood by an agent. The agent needs to use this information for subsequent steps. Providing more natural language descriptions and intuitive fields, such as &quot;Username = John Doe,&quot; &quot;File Type = Contract PDF,&quot; or &quot;Image Name = Profile Picture A,&quot; makes it easier for the agent to understand, combine, and use the information. This improves the agent&#39;s accuracy and reduces errors.</p><p>In complex scenarios, an agent might need both natural language and an underlying ID. In such cases, you can add a response_format parameter to the tool&#39;s interface, allowing the agent to choose between a &quot;detailed&quot; or &quot;concise&quot; mode. The detailed version would include all metadata and IDs for subsequent automated calls, while the concise version would return only the core content to save tokens. You could even allow the agent to specify the exact fields it needs, similar to GraphQL, for a flexible and efficient response. Furthermore, while JSON is a common format, some agent models perform better with Markdown, XML, or even plain text. There is no one-size-fits-all solution; it is best to choose the optimal format based on the task and the agent&#39;s performance during testing.</p><h4>4. Control Token Volume</h4><p>Beyond data quality, the volume of the output is also critical. If a tool returns too much content at once, it can quickly exhaust the agent’s context window, drastically reducing processing efficiency. For scenarios prone to “content flooding,” the team recommends implementing mechanisms like pagination, range filtering, and truncation with sensible defaults. For example, Claude 3.5 Sonnet defaults to a maximum return of 25,000 tokens. Although agent context windows will expand in the future, token-saving design principles will always be relevant.</p><p>When implementing response truncation, be sure to inform the agent with clear hints or documentation, such as suggesting a strategy of “multiple, sequential searches instead of one large search.” Similarly, when a tool encounters an error, such as a parameter validation failure, it should return a clear, actionable error message rather than an obscure error code or traceback, making it easier for the agent to understand and correct its action.</p><h4>5. Refine Prompt Engineering</h4><p>One of the most effective ways to ensure an agent uses tools correctly is to meticulously refine the tool’s description and specifications. These descriptions are loaded into the agent’s context and directly guide how it calls the tool. When writing a tool description, imagine you are explaining it to a new colleague. Any contextual dependencies, industry jargon, special formats, or resource relationships must be described with absolute clarity.</p><p>Use strict data models to constrain inputs and outputs where necessary. Avoid ambiguous parameter names; for example, prefer user_id over user, as the latter could ambiguously refer to a name, a user object, or an ID. Each evaluation provides feedback on the effectiveness of your prompt. Even minor adjustments to a description can yield significant improvements. The Anthropic team once dramatically reduced the error rate and increased the completion rate of the Claude 3.5 Sonnet model on the SWE-bench benchmark simply by optimizing tool descriptions. For more best practices on defining specifications, you can refer to the company&#39;s developer guide. When building tools for Claude, it&#39;s also helpful to understand how tools are dynamically loaded into the system prompt. For tools developed for an MCP server, you can add annotations to indicate which ones require special permissions or pose potential risks, enhancing the agent&#39;s safety and transparency. The more detailed the description and the clearer the parameters, the smarter the agent will be and the fewer pitfalls it will encounter.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5b068803135a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>