<?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 Vikas Kumar on Medium]]></title>
        <description><![CDATA[Stories by Vikas Kumar on Medium]]></description>
        <link>https://medium.com/@imvk45?source=rss-220a05870872------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*_zV4EqZTMPpegZhL26wflg.jpeg</url>
            <title>Stories by Vikas Kumar on Medium</title>
            <link>https://medium.com/@imvk45?source=rss-220a05870872------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 16:08:30 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@imvk45/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[Everyone Is Racing to Build Smarter Models.]]></title>
            <link>https://medium.com/@imvk45/everyone-is-racing-to-build-smarter-models-4da1c731abe7?source=rss-220a05870872------2</link>
            <guid isPermaLink="false">https://medium.com/p/4da1c731abe7</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[optimization]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[claude]]></category>
            <dc:creator><![CDATA[Vikas Kumar]]></dc:creator>
            <pubDate>Thu, 23 Jul 2026 14:43:44 GMT</pubDate>
            <atom:updated>2026-07-23T14:43:44.712Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Everyone Is Racing to Build Smarter Models. But the Real Advantage May Be Building Models That Can Act.</strong></h3><p>The AI industry is obsessed with reasoning.</p><p>Every major model release is judged through mathematics benchmarks, coding evaluations, test time compute, and increasingly long reasoning traces. The assumption is simple: the model that can think longer will eventually become the model that can do more useful work.</p><p>But after reading Moonshot AI founder Yang Zhilin’s recent interview, I think that assumption is incomplete. Reasoning ability and agentic ability are not the same capability, and confusing the two may be why many impressive AI demos still struggle in production.</p><p>A reasoning model receives a problem, performs computation inside its context window, and returns an answer. An agentic model operates through a continuous loop of reasoning, action, observation, and correction.</p><p>That difference sounds small. Technically, it changes everything.</p><p>An agent can search for information, call an API, execute code, inspect the result, identify a failure, modify its plan, and attempt the task again. The environment becomes part of the inference process. Instead of depending only on what the model knows internally, the system continuously gathers evidence from the outside world.</p><p>Consider a model asked to modify a production codebase.</p><p>A strong reasoning model may generate an impressive patch based on the files included in its context. An agentic system can inspect the repository, trace dependencies, modify the relevant implementation, run the test suite, analyze the stack trace, correct the error, and verify that the change did not introduce regressions elsewhere.</p><p>The first system predicts what should work.</p><p>The second system creates evidence that it actually works.</p><p>This is also why coding agents have improved faster than many other agent categories. Software provides unusually clear feedback signals. Code compiles or it does not. Tests pass or fail. An API returns the expected response or generates an error. These signals make coding a powerful environment for reinforcement learning because every action can produce measurable evidence about whether the system is moving closer to the objective.</p><p>Now compare that with product strategy, legal analysis, management, or design. The reward signals are delayed, subjective, and often influenced by factors outside the model’s control. A coding agent can learn from a failed test within seconds. A strategy agent may not know whether its recommendation was correct for several months.</p><p>The optimization problem is completely different.</p><p>Moonshot’s Kimi K2 research reflects this shift toward agentic intelligence. Kimi K2 is a Mixture of Experts model with one trillion total parameters while activating approximately 32 billion parameters for each token. However, the more interesting detail is not the model size. Moonshot designed it around tool use, coding, and agentic tasks rather than treating those capabilities as features added after pretraining.</p><p>The technical work behind the model is equally interesting.</p><p>Moonshot introduced MuonClip, an optimizer that combines the token efficiency of Muon with a mechanism called QK Clip. The purpose of QK Clip is to control instability in attention logits during large scale training. According to the Kimi K2 technical report, this helped the team pretrain the model on 15.5 trillion tokens without experiencing a single loss spike.</p><p>That may sound like a narrow training detail, but it points toward a much larger shift in AI research.</p><p>High quality data is not growing as quickly as the industry’s demand for it. If the amount of useful training data becomes a bottleneck, progress cannot depend only on feeding models more tokens. Researchers will have to extract more intelligence from every token through better optimizers, improved architectures, synthetic data, and reinforcement learning.</p><p>The next scaling advantage may come from learning efficiency, not only model size.</p><p>This changes how I evaluate AI products. Asking which foundation model a company uses is becoming less useful because every team can access increasingly capable models through an API. The more important question is what the system does after the model produces the wrong answer.</p><p>Can it recognize uncertainty? Can it gather more evidence? Can it inspect its own output? Can it recover from a failed tool call? Can it preserve state across multiple actions? Can it verify that the final result actually satisfies the original objective?</p><p>These questions reveal whether a company has built an agent or simply placed a chat interface around a model.</p><p>The architecture of a reliable agent extends far beyond the language model. It requires tool interfaces, state management, permission boundaries, observability, retry policies, evaluation systems, recovery mechanisms, and a strong verification layer. It also requires protection against reward hacking, where the system successfully optimizes a metric while violating the real intent of the task.</p><p>In other words, agent engineering is not just an LLM problem.</p><p>It is also a distributed systems problem, a control systems problem, a security problem, and an evaluation problem.</p><p>Yang described pure reasoning systems using the idea of a brain inside a tank. The brain may be extremely intelligent, but it cannot touch the world, observe the consequences of its decisions, or correct itself using external evidence. Many AI products today operate exactly like that. They generate an answer and stop precisely when the difficult engineering work should begin.</p><p>The next stage of AI may not be won by the model that produces the most impressive first answer.</p><p>It may be won by the system that can recognize when the first answer is wrong, understand why it failed, and continue working until the result has been verified.</p><p>Reasoning helps a model decide what it should do. Agency determines whether it can actually get the work done.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4da1c731abe7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When AI Meets Sustainability: The uncomfortable math behind the technology we are betting the…]]></title>
            <link>https://medium.com/@imvk45/when-ai-meets-sustainability-the-uncomfortable-math-behind-the-technology-we-are-betting-the-455e2b699461?source=rss-220a05870872------2</link>
            <guid isPermaLink="false">https://medium.com/p/455e2b699461</guid>
            <category><![CDATA[sustainability]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[energy]]></category>
            <category><![CDATA[math]]></category>
            <dc:creator><![CDATA[Vikas Kumar]]></dc:creator>
            <pubDate>Wed, 24 Jun 2026 19:07:29 GMT</pubDate>
            <atom:updated>2026-06-24T19:07:29.708Z</atom:updated>
            <content:encoded><![CDATA[<h3>When AI Meets Sustainability: The uncomfortable math behind the technology we are betting the future on</h3><p>We talk about artificial intelligence like it is weightless. A prompt goes in, an answer comes out, and the magic seems to happen somewhere else, in the cloud, in a data center, in a place we never have to look at.</p><p>But AI is not weightless. It runs on electricity, water, chips, cooling systems, and human oversight, and each of those has real environmental, social, and governance costs. If you want to understand AI honestly, ESG is a useful lens because it forces three questions: what it consumes, who it affects, and who is accountable when it fails.[<a href="https://iea.blob.core.windows.net/assets/de9dea13-b07d-42c5-a398-d1b3ae17d866/EnergyandAI.pdf">iea.blob.core.windows</a>]</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZvJhXy5ByDOSYskb" /></figure><p><strong>The energy story</strong></p><p>The cleanest story we tell ourselves about AI is the efficiency story. AI optimizes supply chains. It routes trucks better. It tunes power grids. All true.</p><p>And all beside the point.</p><p>Because while AI may make individual processes more efficient, the absolute energy demand of AI itself is climbing far faster than those savings. A single text AI prompt can use about 0.34 watt hours, while image generation and agentic reasoning can require much more compute. That sounds small until you scale it across billions of daily queries, such as ChatGPT handling around one billion requests per day.[<a href="https://www.youtube.com/watch?v=gq91MDbOeLs">youtube</a>][<a href="https://iea.blob.core.windows.net/assets/de9dea13-b07d-42c5-a398-d1b3ae17d866/EnergyandAI.pdf">iea.blob.core.windows</a>]</p><p>That is where the problem becomes structural. The IEA projects that data centers, heavily driven by AI workloads, could consume about 945 TWh of electricity per year by 2030 in its base case scenario. The same report notes that data centers already accounted for around 1.5 percent of global electricity consumption in 2024.[<a href="https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape/">brookings</a>]</p><p>Efficiency gains are real, but they do not cancel out demand growth. This is a live example of the Jevons paradox: cheaper inference tends to increase total use, so overall consumption can still rise even as cost per query falls.[<a href="https://tracebloc.io/blog/llm-benchmarking">tracebloc</a>]</p><p><strong>Water and infrastructure</strong></p><p>Energy is only half the story. Data centers also need water for cooling, which makes AI’s footprint physical in a very immediate way. Your source notes that Google’s data centers used about 8.1 billion gallons of water in 2024, a 32 percent increase over 2023, largely driven by AI cooling demand.[<a href="https://www.techtarget.com/searchdatacenter/tip/How-to-manage-data-center-water-usage-sustainably">techtarget</a>]</p><p>External analysis from EESI estimates that a medium sized data center can use roughly 110 million gallons of water per year, and typical facilities can consume on the order of hundreds of thousands of gallons per day. Much of this water evaporates in cooling towers instead of returning to local supplies, which can strain already stressed watersheds.[<a href="https://www.eesi.org/articles/view/data-centers-and-water-consumption">eesi</a>]</p><p>That means AI infrastructure can pressure both power grids and water systems at the same time. The sustainability discussion is no longer just about emissions; it is also about land use, grid capacity, and resource competition in the specific regions where data centers are being built.[<a href="https://www.akcp.com/2025/09/02/truth-about-data-water-footprint-of-data-centers/">akcp</a>]</p><p>There is a technical layer here that often gets ignored. Lower inference cost through quantization, mixture of experts, batching, and specialized accelerators reduces energy per token but can increase total load as usage expands. Optimization makes each call cheaper, but if organizations then run many more calls, system level impact can still grow.[<a href="https://iea.blob.core.windows.net/assets/de9dea13-b07d-42c5-a398-d1b3ae17d866/EnergyandAI.pdf">iea.blob.core.windows</a>]</p><p><strong>The social cost</strong></p><p>The social side is where AI stops being abstract. Earlier predictive systems such as COMPAS, which was used to assess criminal recidivism risk, and Amazon’s experimental resume screener showed that algorithmic tools can scale historical bias instead of eliminating it. In COMPAS, Black defendants were falsely flagged high risk at nearly twice the rate of white defendants with similar reoffending rates, while Amazon’s system learned to downrank résumés that contained the word “women’s” or came from women’s colleges.[<a href="https://www.artificialintelligencemadesimple.com/p/why-your-ai-fails-in-production-guest">artificialintelligencemadesimple</a>]</p><p>This is why Cathy O’Neil’s line in Weapons of Math Destruction has stuck: “Models are opinions embedded in mathematics”. These systems do not begin neutral and then become biased later. They learn from historical data, human labels, and institutional incentives, then reproduce those patterns at scale.[<a href="https://amzur.com/blog/reasons-for-ai-model-failure">amzur</a>]</p><p>The labor story is more complicated than “AI will replace everyone.” Your source notes that estimates from groups like McKinsey suggest that many jobs have at least 30 percent of tasks that could be automated, but that the impact varies by role. For some workers, such as financial analysts, AI may automate data gathering and drafting while humans focus on judgment and client relationships; for others, like customer service representatives, AI can handle 70 to 80 percent of queries, shrinking the human part of the job substantially.[<a href="https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6513481">papers.ssrn</a>]</p><p>That is why “AI will create new jobs” is not a complete answer. Aggregate projections from organizations like the World Economic Forum show net job growth but also tens of millions of roles displaced by 2030, especially in routine office and service work. Those new jobs often require advanced technical skills or different locations, which many displaced workers do not have.[<a href="https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6513481">papers.ssrn</a>]</p><p><strong>The governance gap</strong></p><p>Governance is the hardest part because AI systems are faster and less predictable than traditional software. A conventional program usually fails in consistent ways. AI does not. It can behave differently across prompts, across contexts, and across successive updates, and in its agentic form it can initiate actions without explicit human approval at every step.[<a href="https://labelstud.io/learningcenter/llm-evaluation-vs-llm-benchmarking/">labelstud</a>]</p><p>The Dutch childcare benefits scandal is one of the clearest examples of what happens when accountability is missing. Starting in 2013, a self learning fraud detection system wrongly accused more than 26,000 families of benefits fraud, later estimated closer to 35,000, often demanding large repayments and cutting access to other support. There was no meaningful appeal path, officials deferred to the algorithm, and the system ran for years before investigative journalism and a parliamentary inquiry forced it into the open.[<a href="https://www.artificialintelligencemadesimple.com/p/why-your-ai-fails-in-production-guest">artificialintelligencemadesimple</a>]</p><p>A smaller but revealing case comes from Air Canada’s customer service chatbot. When the bot gave a passenger incorrect information about a bereavement fare, the company argued at a tribunal that the chatbot was a separate legal entity responsible for its own actions. The tribunal called this “a remarkable submission” and held that Air Canada was responsible for all the information on its website, including the chatbot’s responses.[<a href="https://www.artificialintelligencemadesimple.com/p/why-your-ai-fails-in-production-guest">artificialintelligencemadesimple</a>]</p><p>Geoffrey Hinton summarized the governance risk clearly when he said he does not believe the “invisible hand” of the market will be enough to keep us safe, and that regulation is needed to ensure large companies invest seriously in safety research. His concern is echoed by the fact that many ESG ratings still reward policy documentation more than operational behavior, giving points for a Responsible AI PDF even when real oversight is thin.[<a href="https://www.youtube.com/watch?v=y2N9dIhcNrY&amp;vl=en">youtube</a>][<a href="https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape/">brookings</a>]</p><p>What responsible use looks like</p><p>None of this is an argument against AI. It is an argument against pretending it is free.</p><p>If you build, buy, or invest in these systems, the right questions are specific and technical. How much energy does each million tokens of inference consume in practice, and on what grid mix. How much water does your cooling strategy use per kilowatt hour, and where does that water come from. What data was used to train the model, how was it cleaned, and which populations are most affected by errors. What guardrails and kill switches exist for agentic systems, and who has the authority and the context to use them in time.[<a href="https://theconversation.com/data-centers-consume-massive-amounts-of-water-companies-rarely-tell-the-public-exactly-how-much-262901">theconversation</a>]</p><p>Those are the questions that turn AI from a hype cycle into a real accountability problem. A company can publish a Responsible AI policy and still deploy systems that cause harm if those questions are never answered in detail. It can report efficiency improvements and still increase absolute consumption if it does not track total load. It can automate decisions and still leave people without a meaningful right of appeal.[<a href="https://www.energy.gov/oe/clean-energy-resources-meet-data-center-electricity-demand">energy</a>]</p><p><strong>Closing</strong></p><p>AI’s biggest risk is not that it is evil. It is that it is fast, opaque, and expensive in ways we keep undercounting.[<a href="https://www.nature.com/articles/d41586-025-01113-z">nature</a>]</p><p>The fix starts with measurement. Not just the model, but the whole system. Not just the output, but the cost. Not just the promise, but the accountability.[<a href="https://www.eesi.org/articles/view/data-centers-and-water-consumption">eesi</a>]</p><p>If we want AI to be part of a sustainable future, we have to stop treating it like it exists above the physical world. It does not. It runs on it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=455e2b699461" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Switching to a Better AI Model Breaks Everything You Built]]></title>
            <link>https://medium.com/@imvk45/why-switching-to-a-better-ai-model-breaks-everything-you-built-7e14e1b34b22?source=rss-220a05870872------2</link>
            <guid isPermaLink="false">https://medium.com/p/7e14e1b34b22</guid>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm-benchmarks]]></category>
            <dc:creator><![CDATA[Vikas Kumar]]></dc:creator>
            <pubDate>Tue, 23 Jun 2026 15:24:21 GMT</pubDate>
            <atom:updated>2026-06-23T15:24:21.054Z</atom:updated>
            <content:encoded><![CDATA[<p>Machine learning benchmarks, fragile prompts, and why your “upgrade” feels like a rollback</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*L59pcLmWqtBKPc-A" /></figure><p>Modern AI stacks depend on one fragile assumption. The large language model that powers your workflow today will behave roughly the same way tomorrow.</p><p>That assumption breaks the moment you “upgrade” to a new model. Benchmarks say the model is strictly better, but your application behaves worse. Latency profiles shift, agents get stuck, structured outputs break, and your team starts asking why a higher scoring model on paper is failing more often in practice.[<a href="https://aifn.co/why-ai-benchmarks-fall-short-in-the-real-world">aifn</a>]</p><p>This article is about that gap between benchmark gains and production behavior. It explains why prompt tuned systems are so sensitive to model changes, how context and compaction policies introduce hidden variability, and why the only reliable fix is to treat model upgrades as controlled migrations rather than simple configuration changes.[<a href="https://arxiv.org/html/2502.06559v1">arxiv</a>]</p><h3>1. Benchmarks Say “Upgrade”, Production Says “Maybe”</h3><p>In the machine learning community, new models are announced with higher scores on standard benchmarks. Accuracy, F1, exact match, coding pass rates, and reasoning metrics all move in the right direction. Evaluation papers and model cards describe these improvements under controlled conditions with fixed datasets and fixed metrics.[<a href="https://arxiv.org/abs/2411.12990">arxiv</a>]</p><p>At the same time, developers integrating these models into real systems report a different pattern. Threads about recent LLM releases describe more literal instruction following, reduced persistence in long tasks, and degraded behavior in extended contexts, even though the benchmark numbers improved. This is a textbook example of the benchmark reality gap, where performance on curated tests diverges from performance under live, nonstationary workloads.[<a href="https://www.reddit.com/r/ClaudeAI/comments/1snhfzd/claude_opus_47_is_a_serious_regression_not_an/">reddit</a>]</p><h3>2. Prompts as Informal API Contracts</h3><p>A prompt is not just a string. It is an informal API contract between your application and a stochastic model. When you ship a prompt template, you have implicitly encoded assumptions about several dimensions of model behavior</p><p>Output structure, including how strictly the model adheres to JSON or other formal schemas used by downstream components[<a href="https://www.reddit.com/r/ArtificialInteligence/comments/1t2g1ok/why_do_ai_models_improve_rapidly_in_benchmarks/">reddit</a>]</p><p>Initiative and inference, including how much the model fills in missing details versus asking for clarification[<a href="https://www.reddit.com/r/claude/comments/1svoxgl/opus_47_is_a_regression_and_im_tired_of_the/">reddit</a>]</p><p>Task persistence, including whether the model continues through a long multi step instruction or pauses for repeated human confirmation[<a href="https://www.reddit.com/r/ClaudeAI/comments/1snhfzd/claude_opus_47_is_a_serious_regression_not_an/">reddit</a>]</p><p>Those assumptions are learned from a particular model version. When a provider replaces that model with a new variant, the conditional distribution over outputs changes. Even if the new distribution dominates the old one on generic benchmarks, it can be misaligned with your prompt contract. As a result, previously stable chains break, because your orchestration and parsing logic was tuned for the old behavior profile.[<a href="https://labelstud.io/blog/why-benchmarks-matter-for-evaluating-llms/">labelstud</a>]</p><h3>3. Context Windows, Compaction, and Effective Memory</h3><p>Large context windows are often marketed as raw token numbers, but real systems interact with an effective working memory governed by compaction policies, truncation strategies, and cost constraints. To control latency and spend, many LLM stacks summarize or drop earlier turns once a conversation reaches a fraction of the advertised limit.[<a href="https://www.anthropic.com/news/claude-opus-4-7">anthropic</a>]</p><p>Engineering reports from teams building long horizon agents show a consistent pattern. Once cumulative context crosses a certain threshold, the model starts forgetting specific details, misinterpreting earlier instructions, or repeating work, even though the nominal token limit has not been reached. When model upgrades modify compaction heuristics or context prioritization, agents that previously behaved coherently across long sessions can start exhibiting sudden loss of relevant information, despite no change in application side code.[<a href="https://www.reddit.com/r/ClaudeAI/comments/1nw9cuo/how_does_context_window_and_autocompacting_work/">reddit</a>]</p><h3>4. Benchmark Design and the Measurement Problem</h3><p>Recent interdisciplinary reviews of AI benchmarks have highlighted structural issues in how they are constructed and interpreted. Common problems include narrow coverage of task distributions, reliance on static datasets that quickly become saturated, and vulnerability to shortcut exploitation by large models.[<a href="https://neurips.cc/virtual/2024/poster/97566">neurips</a>]</p><p>This does not make benchmarks useless, but it does mean they are weak proxies for production performance. Benchmarks typically assume independent, identically distributed inputs, clear ground truth labels, and single step outputs. Real systems deal with feedback loops, multi step interactions, shifting user behavior, and compositional workflows where small deviations accumulate over time. An upgrade that improves benchmark scores may still degrade stability or controllability in these more complex regimes.[<a href="https://aifn.co/why-ai-benchmarks-fall-short-in-the-real-world">aifn</a>]</p><h3>5. Models as Moving Dependencies, Not Static Libraries</h3><p>From an engineering perspective, the safest mental model is to treat the LLM as a moving dependency rather than a static library. Its behavior can drift over time due to training changes, safety adjustments, sampling defaults, and context management updates, even if the API surface remains the same.[<a href="https://labelstud.io/blog/why-benchmarks-matter-for-evaluating-llms/">labelstud</a>]</p><p>Teams that have adapted to this reality tend to converge on three practices. They build small but representative evaluation sets seeded from production data, including both typical and high risk requests. They run canary deployments where a fraction of traffic is routed to the new model while outputs, error rates, and business metrics are monitored against the previous baseline. They insert abstraction and routing layers that decouple application logic from any single provider, enabling model A or B to be swapped per task without rewriting the entire stack.[<a href="https://dev.to/amitksingh1490/how-we-extended-llm-conversations-by-10x-with-intelligent-context-compaction-4h0a">dev</a>]</p><h3>6. Why “Better” Still Breaks Things</h3><p>From this angle, the apparent paradox disappears. A model can be strictly better on many benchmarks and still break your system because the training objective, evaluation datasets, and deployment environment are not aligned. The new model optimizes for improved performance on a set of controlled tasks. Your system optimizes for stable behavior in a noisy, evolving environment.[<a href="https://aifn.co/why-ai-benchmarks-fall-short-in-the-real-world">aifn</a>]</p><p>Every upgrade modifies the conditional behavior at the core of your application. If you treat that modification as a toggle, you inherit all the drift with none of the protection. If you treat it as a migration, with explicit tests, staged rollout, and the option to route around regressions, the upgrade becomes manageable. In that framing, the question stops being “Why did the better model break everything” and becomes “Why did we ever ship probabilistic components without the same discipline we apply to every other critical dependency”.[<a href="https://arxiv.org/html/2502.06559v1">arxiv</a>]</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7e14e1b34b22" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Changed as My AI Product Grew from 10 to 10,000 Users]]></title>
            <link>https://medium.com/@imvk45/what-changed-as-my-ai-product-grew-from-10-to-10-000-users-8975483e02af?source=rss-220a05870872------2</link>
            <guid isPermaLink="false">https://medium.com/p/8975483e02af</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[distributed-systems]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Vikas Kumar]]></dc:creator>
            <pubDate>Mon, 08 Jun 2026 18:34:51 GMT</pubDate>
            <atom:updated>2026-06-08T18:34:51.057Z</atom:updated>
            <content:encoded><![CDATA[<h3>I Thought Scaling Would Be an AI Problem</h3><p>When I started building AI products, I assumed most engineering discussions would revolve around prompts, models, and inference. For the first few weeks, that assumption felt correct. Every improvement came from experimenting with prompts, trying different models, and tweaking workflows.</p><p>As usage grew, however, I found myself spending less time thinking about the model and more time thinking about everything surrounding it. <strong>Authentication, databases, caching, observability, security,</strong> and cost gradually became larger concerns than the model itself. The AI remained the feature users talked about, but it stopped being the component that consumed most of the engineering effort.</p><h3>At 10 Users, Almost Everything Can Be Simple</h3><p>At the beginning, simplicity is an advantage.</p><p>The architecture for many early AI products can fit on a whiteboard. A frontend communicates with a backend, the backend calls an LLM provider, and responses are stored in a database. Most requests are synchronous, most debugging is manual, and most operational issues can be solved by directly inspecting logs.</p><p>At this stage, users are effectively helping discover whether the product should exist. Building extensive infrastructure too early usually creates more problems than it solves because the bottleneck is product validation rather than scale.</p><p>What mattered most for me was shortening the feedback loop. If a user encountered a problem, I wanted to understand it quickly, deploy a fix, and learn from the interaction.</p><h3>At 100 Users, The Product Starts Acting Like a Real System</h3><p>The architecture may look similar, but user behavior starts changing dramatically.</p><p>People forget passwords. Sessions expire. Requests arrive simultaneously. Different browsers behave differently. Unexpected usage patterns emerge. Some users become power users while others barely engage.</p><p>This is where authentication, authorization, and user management become meaningful parts of the system. I found that understanding who was doing what became almost as important as understanding what the AI was doing.</p><p>Basic monitoring also became necessary. Earlier, I could manually inspect problems. At 100 users, that approach stopped scaling. The challenge shifted from fixing issues to detecting them quickly enough.</p><h3>At 1,000 Users, System Design Starts Mattering</h3><p>The transition to 1,000 users felt less like growth and more like entering a new phase of engineering.</p><p>Database queries that were previously insignificant started showing up in performance metrics. API endpoints that seemed fast enough became noticeable bottlenecks. Features that required multiple model calls introduced latency that users could actually feel.</p><p>This was also the point where I started separating concerns more deliberately. Authentication, business logic, retrieval, model orchestration, and storage all benefited from clearer boundaries.</p><p>Observability became critical because debugging distributed systems is fundamentally different from debugging local applications. When a request passes through multiple services before reaching a model and then returns through additional processing layers, understanding failures requires more than reading server logs.</p><p>The challenge wasn’t building features anymore. The challenge was understanding how the entire system behaved under real-world usage.</p><h3>At 10,000 Users, Small Decisions Become Expensive</h3><p>The most surprising thing about reaching larger scale was how quickly small inefficiencies became visible.</p><p>An extra database query that nobody noticed at 100 users suddenly appeared millions of times a month.</p><p>A slightly longer prompt translated into meaningful token costs.</p><p>A slow endpoint affected enough users to become a product issue rather than an engineering issue.</p><p>At this stage, architectural discussions became heavily influenced by economics. Caching became valuable. Queue-based processing became useful. Model routing started making sense. Background jobs replaced synchronous workflows wherever possible.</p><p>The focus shifted from making features work to making them work efficiently.</p><h3>Security Evolves With Scale</h3><p>One thing I underestimated early on was how much security changes as a product grows.</p><p>At 10 users, security often means protecting API keys and ensuring users can log in.</p><p>At 10,000 users, security becomes part of almost every architectural decision.</p><p>User data needs clear boundaries. Access controls become more important. Audit trails become useful. Sensitive information can appear in prompts, logs, retrieval systems, and analytics pipelines.</p><p>For AI products specifically, prompt injection, data leakage, and retrieval security introduce concerns that don’t exist in many traditional applications.</p><p>The larger the system becomes, the more security moves from being a feature to being a design principle.</p><h3>The Biggest Lesson</h3><p>Looking back, every growth stage introduced a different bottleneck.</p><p>At 10 users, I was trying to determine whether the product was useful.</p><p>At 100 users, I was trying to make it reliable.</p><p>At 1,000 users, I was trying to understand it.</p><p>At 10,000 users, I was trying to operate it efficiently and securely.</p><p>The AI model remained important throughout the journey, but it gradually became one component within a much larger ecosystem.</p><p>That’s probably the biggest misconception about scaling AI products. People assume the challenge is making the model smarter. In my experience, the harder challenge is building an entire system around the model that users can depend on every day.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8975483e02af" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[75+ Coffee Chats, 200 Cold Emails, and Still No Job]]></title>
            <link>https://medium.com/@imvk45/75-coffee-chats-200-cold-emails-and-still-no-job-5c5c3e6b912c?source=rss-220a05870872------2</link>
            <guid isPermaLink="false">https://medium.com/p/5c5c3e6b912c</guid>
            <category><![CDATA[indian-students]]></category>
            <category><![CDATA[american-indians]]></category>
            <category><![CDATA[f1]]></category>
            <category><![CDATA[life]]></category>
            <category><![CDATA[jobs]]></category>
            <dc:creator><![CDATA[Vikas Kumar]]></dc:creator>
            <pubDate>Tue, 28 Apr 2026 23:43:28 GMT</pubDate>
            <atom:updated>2026-04-28T23:43:28.435Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The side of US student life nobody posts about.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6iqk9uWnUxojsYhpVv9ikQ.png" /><figcaption>Three minutes before everything shifted</figcaption></figure><p>It’s 2 AM. The email says: “we’ve decided to move forward with other candidates.”</p><p>Final round. Weeks of prep. A task I stayed up three nights to finish. And this is the reply.</p><p>I close the laptop. I don’t cry right away. That comes later, somewhere between brushing my teeth and lying down, when the silence in the room gets too loud.</p><p>This is the part of being an international student in the US that nobody posts about.</p><p>Not too long ago, I thought I had things figured out.</p><p>Almost a perfect GPA. Building AI tools. Shipping actual engineering work, not just class projects. Pull requests merged into open source. A two-way AI interview platform I built from scratch, end to end. Real GitHub contributions people could actually look at.</p><p>The kind of profile that looks clean on paper and convincing in a resume line. The kind of student people assume will “be fine.”</p><p>I had a role lined up with one of the FAANG companies. The kind of name that makes conversations easier, that makes your parents proud before you even explain what you actually do.</p><p>I hadn’t even fully processed it myself before things shifted. A restructure. Team matching phase. And just like that, before I could even properly tell my parents, the job was gone.</p><p>No dramatic ending. No clear explanation. Just silence replacing what felt like certainty.</p><p>There’s also a quieter irony I’ve been sitting with. I’m an AI engineer trying to break into a market where AI is starting to take the kinds of entry-level engineering roles people like me are applying for. The thing I build is, in some small way, the thing reshaping the ground I’m trying to stand on. I don’t say that with bitterness. It’s just a strange position to be in, and one I think a lot of people in this field are quietly navigating right now.</p><p>That was the beginning of a very different phase.</p><p>People don’t really talk about what happens in between these moments. Especially as an international student, where everything runs on a clock you don’t control.</p><p>You start doing what everyone tells you works.</p><p>You meet professors. Follow every lead. Around 75 coffee chats with people across companies, roles, and time zones. Then 200+ cold emails. Alumni, founders, engineers, recruiters.</p><p>Sometimes people respond. Sometimes they give you a take-home. You build it. You ship it. You go beyond what was asked, because that’s what you’ve always done.</p><p>And then, nothing. No reply. No feedback. Just silence again.</p><p>You keep going because stopping feels worse.</p><p>After a while, the rejections stop feeling like events. They become a pattern.</p><p>You wake up, reach for your phone, and the first thing you read before getting out of bed is some version of “we’ve decided to move forward with other candidates.” Sometimes there are two or three of them stacked in your inbox from overnight. You read them, you mark them as read, you get up. That’s the morning.</p><p>Starting your day that way, again and again, does something to you that’s hard to describe. It’s not dramatic. It’s just a slow, quiet weight that settles in before you’ve even had coffee.</p><p>The strange part is, on paper, things look like they should be working. I’ve had career coaches, mentors, and peers tell me the resume is strong. The experience is solid. I’ve rewritten the LinkedIn summary more times than I can count, tightened the bullet points, reframed every role to sound sharper, more outcome-driven. Over 102 recruiters have viewed the profile in recent weeks. None have reached out.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZGoUJbmPXMZ8i34HjW5d5A.png" /><figcaption>102 views. Zero messages</figcaption></figure><p>I’ve reached out to them instead. Polite, specific, tailored messages. Most don’t reply.</p><p>That’s the part that’s hard to sit with. When you’re doing the things you’re supposed to do, taking the advice, optimizing the profile, putting in the volume, and the outcome still doesn’t move. There’s no obvious thing to fix. No clear lever to pull. Just the quiet realization that effort and outcome aren’t always connected the way you were taught they would be.</p><p>And that’s the part that slowly gets to you.</p><p>There are days you don’t feel like talking to anyone. Even small conversations feel like effort. Some nights I just sat there, not doing anything, letting everything sink in. Some of those nights turned into tears I didn’t expect.</p><p>Some processes do move forward. You make it to later rounds. You start to believe again.</p><p>There was one where I reached the final round. Gave it everything I had. Prepared harder than I ever had. And still, no offer.</p><p>That one stays with you a little longer than the others.</p><p>There was McKinsey too. Case prep. Mock interviews. Late nights trying to think sharper, faster, better. Still didn’t convert.</p><p>It’s strange how quickly things swing. One day you feel hopeful. The next, you’re back in uncertainty.</p><p>And then there’s the quieter, more personal side of all this.</p><p>I come from a humble background. Everything I’ve achieved has been through scholarships. That’s something I’m deeply proud of. But it doesn’t take away the work behind it. The long hours. The consistency. The pressure to keep proving yourself again and again.</p><p>Being here is different.</p><p>I’ve spent most of my life in structured environments, surrounded by people, friendships, conversations that came naturally. Here, things feel more distant. You see people around you living similar lives, but the experience isn’t the same.</p><p>You start noticing small differences.</p><p>Someone eats pizza because they enjoy it. I eat pizza because it’s what’s available.</p><p>It sounds small. It isn’t.</p><p>Even without loans or rent weighing me down, there’s a different kind of pressure. The kind that comes from expectations. From distance. From wanting to make something meaningful out of the opportunity you’ve been given.</p><p>And in the middle of all this, you start questioning yourself more than you’d like to admit.</p><p><em>Am I doing enough?</em> <em>Am I missing something?</em> <em>Why does it feel like things are working for others, but not for me right now?</em></p><p>But here’s the part I want to remember.</p><p>I didn’t stop.</p><p>Even on the days I felt completely drained, I still showed up. Wrote another commit. Sent another message. Pushed another PR. Applied to another role. Followed up again.</p><p>Because this phase, as uncomfortable as it is, is real.</p><p>This is the side of the US that doesn’t get posted about. The uncertainty. The waiting. The near-misses. The effort that doesn’t immediately translate into outcomes. Especially for people who are used to things working out when they put in the work.</p><p>Right now, things are still unclear. A few possibilities in motion. Nothing fully certain yet.</p><p>But maybe that’s the point of this phase.</p><p>Not to have everything figured out. But to learn how to keep going when it isn’t.</p><p>If a future version of me ever reads this, I hope he doesn’t just look at where things ended up. I hope he remembers what it took to get there. The effort. The frustration. The quiet resilience that didn’t always feel like strength at the time.</p><p>Because this version of me, right now, is still in it.</p><p>I tried. I built. I shipped. And I’m still struggling.</p><p>Not finished. Not settled. Just still trying.</p><p>And maybe, for now, that’s enough.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5c5c3e6b912c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Was Wasting 6 Hours a Week Typing Prompts Until I Fixed This]]></title>
            <link>https://medium.com/@imvk45/i-was-wasting-6-hours-a-week-typing-prompts-until-i-fixed-this-5747449272f3?source=rss-220a05870872------2</link>
            <guid isPermaLink="false">https://medium.com/p/5747449272f3</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[wispersync-for-voice]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[optimization]]></category>
            <dc:creator><![CDATA[Vikas Kumar]]></dc:creator>
            <pubDate>Mon, 06 Apr 2026 00:35:40 GMT</pubDate>
            <atom:updated>2026-04-06T00:35:40.200Z</atom:updated>
            <content:encoded><![CDATA[<p>If you use Claude daily and still type every prompt, you are probably slowing yourself down without realizing it. The issue is not just typing speed. It is how typing changes the way you communicate. You shorten things, skip context, and assume the model will figure it out. Most of the time, it does not. That is why simple tasks turn into multiple follow ups.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/780/0*jEtXEUG1p42N9OGa" /></figure><p>I changed one simple thing. I stopped typing and started speaking.</p><h3>The Setup</h3><p>I use a tool called Wispr Flow. It converts your voice into clean text anywhere on your system. It does not connect to Claude directly and does not require any setup inside it. Wherever your cursor is, it simply types what you say.</p><p>Setting it up takes a few minutes. You install the app, allow microphone and accessibility permissions, and set a hotkey. Once that is done, it works across everything you use.</p><h3>How It Works</h3><p>The workflow stays exactly the same. You open Claude, click into the input box, press your hotkey, and speak your prompt. When you release the key, your words appear instantly. You press enter and get your response.</p><p>There is no new interface or process. You are just replacing typing with speaking.</p><h3>Why This Works Better</h3><p>The biggest difference is clarity.</p><p>When you type, you compress your thinking. You skip details and leave gaps. Then the model tries to interpret what you meant, which often leads to incomplete or slightly off responses.</p><p>When you speak, you explain things more naturally. You include context without forcing it. Your instructions are clearer, which leads to better outputs on the first try and fewer follow ups.</p><h3>Where It Helps Most</h3><p>This becomes especially useful when your prompts are more detailed. Instead of carefully typing and editing, you can just explain your idea in one flow.</p><p>There is also a hands free mode that lets you speak continuously for a few minutes. This is helpful when you are breaking down a problem or describing something step by step.</p><p>If you use Claude for coding, it works well with technical language too. You can describe what you want changed in plain terms and it keeps up without any extra effort.</p><h3>Small Things That Make a Difference</h3><p>Speaking normally works better than overthinking your words. Giving slightly more context improves results immediately. For longer prompts, hands free mode feels more natural than holding a key the whole time.</p><h3>What Changes After This</h3><p>After a while, it stops feeling like you are writing prompts. You are just explaining what you want, like you would to a teammate.</p><p>That small shift makes Claude much more effective because it finally has the context it needs.</p><p>Set it up once and try it for a day. You will notice the difference quickly.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5747449272f3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Give Me 20Minutes and I’ll Change How You Use AI Forever]]></title>
            <link>https://medium.com/@imvk45/give-me-18-minutes-and-ill-change-how-you-use-ai-forever-7551afc3f629?source=rss-220a05870872------2</link>
            <guid isPermaLink="false">https://medium.com/p/7551afc3f629</guid>
            <category><![CDATA[optimization]]></category>
            <category><![CDATA[ai-tools]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[productivity]]></category>
            <dc:creator><![CDATA[Vikas Kumar]]></dc:creator>
            <pubDate>Wed, 25 Mar 2026 16:56:43 GMT</pubDate>
            <atom:updated>2026-03-25T17:05:14.048Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The uncomfortable truth about convenience, and what the people actually pulling ahead are doing instead.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fObXFwMzwmK8IG8DbGjpzA.png" /></figure><p>You open your laptop. Something needs writing, so you ask AI. Something needs summarizing, so you ask AI. Things get done. The inbox clears.</p><p>And somewhere in the back of your head, a small voice whispers: did I actually do anything just now?</p><p>That voice is not imposter syndrome. It is your brain noticing something real.</p><p>AI does not make you smarter by default. It makes you faster. Speed without depth is just a more efficient way to stay shallow. The people pulling ahead are not using AI more. They are using it differently.</p><p><strong>Your brain is chasing the wrong dopamine</strong></p><p>Your brain has a glitch called completion bias. It releases roughly the same dopamine hit for finishing any task regardless of whether that task matters. Reformatting a slide. Cleaning up a doc no one will read. Each one feels like progress. None of it compounds.</p><p>This is why smart people end up exhausted and stuck. Not because they are not working hard. Because they are applying full effort to things with a ceiling.</p><p><strong><em>| Not all work is equal. AI should widen that gap, not flatten it.</em></strong></p><p>Some tasks have a quality threshold: clear the bar and extra effort yields nothing. Internal emails, routine decks, data reformatting. Nobody promotes you because your status update was beautifully written. Stop when it is good enough.</p><p>Other tasks have no ceiling at all. A key relationship, a product decision, an idea only you can bring. Being ten percent better here does not yield ten percent better results. It solves problems three levels downstream.</p><p>Use AI aggressively on the first category. Save yourself for the second.</p><p><strong>Four things worth outsourcing immediately</strong></p><p>Drafting. Use AI to get past zero. The output will be average. That is the point. You now have something to react to and rebuild in your own voice.</p><p>Research. Deep research tools synthesize across dozens of sources and flag their own gaps. What used to take a week takes twenty minutes.</p><p>Analysis. Let AI take the first pass at finding patterns in messy data. Then bring your judgment to what it surfaces.</p><p>Grunt work. Reformatting, translating, tabulating. Mechanical tasks that consume time without building anything. Give them away completely.</p><p>One rule: the moment a task requires your judgment or your name on it in a meaningful way, none of this applies. That is yours to carry.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*HifnZ9N9sUvX9aMl" /></figure><p><strong>Stop prompting like everyone else</strong></p><p>Most people treat AI like a search engine with better grammar. One question, one answer, move on. It works. But it leaves most of the value on the table.</p><p>Here is what you need to understand: AI is a probability engine, not a calculator. It generates the most statistically likely response to your input, which means vague inputs produce vague outputs. And it will sound confident either way.</p><p>You are talking to a drunk genius. Learn to drive the car.</p><p>Give it examples of what good looks like. Ground it in your own previous work so it stops averaging across its training and starts pattern matching to you. Ask it to show its reasoning before it produces anything. That one habit alone separates most people.</p><p>Then go further. Chain tasks together. Research this, cross reference it, find the three most important insights, write a memo. One prompt, multiple cognitive steps, one coherent output. Most people never get here. That is exactly why it is worth learning.</p><p><strong>The part nobody tells you about</strong></p><p>Once you have stripped friction from low ceiling work, the instinct is to keep stripping friction everywhere. Summarize this. Write my reflection. Draft my cover letter.</p><p>This is where the atrophy starts.</p><p>When astronauts spend months in zero gravity, they lose up to twenty percent of their muscle and bone density. The environment removed all resistance, so the body stopped maintaining itself. Passive AI use does the same thing to your thinking.</p><p>If AI never makes you uncomfortable, you are using it to stay exactly where you are.</p><p>For information tasks, use AI to remove friction. For transformation tasks, use AI to add it. Ask it to quiz you. Ask it to increase the difficulty until you are being grilled at the edge of what you know. The discomfort is not a sign something is going wrong. It is the signal that learning is actually happening.</p><p><strong>The most underrated skill you can build right now</strong></p><p>The final obstacle is not lack of information. It is the performance of certainty.</p><p>You nod along in a meeting because asking a basic question feels like admitting you do not belong. The performance of competence slowly replaces the pursuit of it. Over time, the mask becomes the ceiling.</p><p>When Satya Nadella took over Microsoft in 2014, the company had missed mobile and was losing the cloud race. He made one cultural shift: from know-it-alls to learn-it-alls. The market cap went from $300 billion to over $3 trillion in a decade. One permission to say “I don’t know yet.” Ten times the value.</p><p>AI gives you a private room to practice this. No judgment, no raised eyebrows. Ask the question you would never ask in a meeting. Ask it to explain something like you are ten years old. Ask three times. Notice what you now understand that you did not before.</p><p><strong>The divide is not the tool. It is the intention.</strong></p><p>At a glance, everyone is using the same tools. They are not getting the same results.</p><p>Use AI to clear the runway on work with a ceiling. Prompt it like someone who knows it can be wrong. Let it challenge you, not just serve you. Ask it the questions you are too proud to ask anyone else.</p><p>That small voice asking whether you actually did anything?</p><p>Listen to it. Then use it to do something about it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7551afc3f629" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The 5 Minute Setup That Turned My Terminal Into an AI Engineer]]></title>
            <link>https://medium.com/@imvk45/the-5-minute-setup-that-turned-my-terminal-into-an-ai-engineer-aedd334fffec?source=rss-220a05870872------2</link>
            <guid isPermaLink="false">https://medium.com/p/aedd334fffec</guid>
            <category><![CDATA[writing-prompts]]></category>
            <category><![CDATA[optimization]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[ai-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Vikas Kumar]]></dc:creator>
            <pubDate>Sun, 15 Mar 2026 22:18:00 GMT</pubDate>
            <atom:updated>2026-03-15T22:18:00.193Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*EI19oUfI8Dc2XZEM" /></figure><p>I opened my laptop late at night telling myself I would just check one thing before going to sleep. That is the classic lie developers tell themselves. Five minutes later you are supposed to close the laptop, but somehow two hours pass and you are still there, typing prompts into your terminal while Claude reads your repository and suggests fixes faster than you can scroll through the files.</p><p>If you have not tried Claude Code inside your terminal, it is one of the easiest upgrades you can make to your workflow. Instead of constantly switching between your editor, documentation, and search results, you can just ask questions directly in the terminal while Claude reads your project files.</p><p>This guide shows how to set it up in about five minutes so you can start using it inside your own projects.</p><h3>Step 1: Check Your Claude Plan</h3><p>Before installing anything, make sure your Claude account supports Claude Code. The feature is currently available only on <strong>paid plans</strong>, so if you are on the free plan the installation will work but authentication will fail later. This is the most common point where people get stuck.</p><p>Go to <strong>claude.ai</strong>, log in, and check your account plan. You should see something like <strong>Claude Pro, Max, Teams, or Enterprise</strong> listed in your account settings. If the plan shows as Free, upgrade first and then continue with the setup.</p><p>Once you confirm the account has a supported plan, you are ready to install Claude Code.</p><h3>Step 2: Install Claude Code</h3><p>Open your terminal and run the installation command. This downloads and installs Claude Code automatically.</p><pre>curl -fsSL https://claude.ai/install.sh | bash</pre><p>Your terminal will start printing installation logs as the script runs.</p><pre>Downloading Claude Code...<br>Installing dependencies...<br>Configuring environment...<br>Claude Code installed successfully</pre><p>If the installation finishes without errors, the tool is now installed on your system.</p><h3>Step 3: Verify the Installation</h3><p>Next, confirm that the command works properly. This step ensures your terminal recognizes the Claude executable.</p><pre>claude --version</pre><p>If everything is installed correctly, you should see a version number printed.</p><pre>claude 1.x.x</pre><p>If you see <strong>command not found</strong>, simply close the terminal and open a new one so the environment reloads. After restarting the terminal, run the command again and the version number should appear.</p><h3>Step 4: Log In</h3><p>Now you need to authenticate your Claude account so the terminal tool can connect to your subscription.</p><pre>claude</pre><p>When you run this command for the first time, Claude will ask you to log in through your browser.</p><pre>Opening browser for authentication...<br>Please log in to your Claude account<br>Authentication successful<br>Connected to Claude</pre><p>A browser window should open automatically. Log in using the same account that has your paid plan. Once the login is completed, the terminal session will confirm the connection.</p><h3>Step 5: Open Your Project Folder</h3><p>Claude works best when it understands your repository. That means you should run it inside the folder containing your project instead of your home directory.</p><p>Navigate to your project folder first.</p><pre>cd my-project</pre><p>Then start Claude again.</p><pre>claude</pre><p>Claude will scan the project and load the context.</p><pre>Scanning project files...<br>Indexing repository...<br>Project loaded successfully<br>You can now ask Claude about your code</pre><p>After this step, Claude understands your codebase and is ready to answer questions about it.</p><h3>Step 6: Add a CLAUDE.md File</h3><p>To make Claude even more useful, create a small file called <strong>CLAUDE.md</strong> in your project root. This file tells Claude how your project works and what commands or rules it should follow.</p><p>Example:</p><pre>## Commands<br>npm run dev<br>npm test</pre><pre>## Notes<br>TypeScript strict mode enabled<br>Do not modify the legacy auth folder</pre><p>Whenever Claude runs inside your repository, it reads this file automatically. That means you do not have to explain your project structure every time you start a session.</p><h3>Step 7: Ask Claude Something Real</h3><p>At this point everything is ready and you can start interacting with Claude directly in the terminal. Instead of searching documentation or digging through stack traces alone, you can ask questions about your actual code.</p><p>For example:</p><pre>Why is this function returning undefined?</pre><p>Claude reads the relevant files and responds with suggestions based on your project.</p><pre>Analyzing function...<br>Possible issue found in userService.ts line 42<br>Variable may be undefined before return<br>Suggested fix available</pre><p>This is usually the moment when it clicks. Instead of generic answers, Claude is responding with context from your repository.</p><h3>Final Thoughts</h3><p>The setup takes only a few minutes, but the workflow change is noticeable almost immediately. You stay inside your terminal, ask questions about your code, and get responses that understand your project instead of generic explanations.</p><p>Once you start using it regularly, it becomes hard to go back to the old workflow of constantly switching between tools. The terminal stops being just a place to run commands and starts feeling more like a conversation with your codebase.</p><p>If you followed the steps above, Claude Code should now be running in your terminal and ready to help you build faster.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aedd334fffec" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Hiring Pipeline Is Not Broken. It Is Being Quietly Restructured.]]></title>
            <link>https://medium.com/@imvk45/the-hiring-pipeline-is-not-broken-it-is-being-quietly-restructured-c088b92c2ea4?source=rss-220a05870872------2</link>
            <guid isPermaLink="false">https://medium.com/p/c088b92c2ea4</guid>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[hiring]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[freshers]]></category>
            <dc:creator><![CDATA[Vikas Kumar]]></dc:creator>
            <pubDate>Sat, 07 Mar 2026 03:30:53 GMT</pubDate>
            <atom:updated>2026-03-07T03:30:53.237Z</atom:updated>
            <content:encoded><![CDATA[<p>A few months ago I sent out applications for technical roles and heard almost nothing back. Not rejections. Just silence. I optimized my resume, leaned on referrals, and tailored cover letters. The silence continued. I assumed it was the market. High interest rates, cautious hiring, slower funding cycles. The usual explanations.</p><p>Then I read a research report from Anthropic on AI and labor markets and something clicked. The market did slow down. But there is also something more structural happening underneath it, and most people entering technical fields right now are not seeing it clearly because it does not show up in the obvious places.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/770/1*twu8DqW0WaCZqbXw40sqEQ.png" /></figure><p>It does not show up in unemployment numbers. It shows up in hiring pipelines. And that distinction matters enormously if you are early in your career.</p><p><strong>The framework that changed how I think about this</strong></p><p>Most research on AI and jobs asks the wrong question. It asks what AI could theoretically automate, decomposes occupations into tasks, estimates which tasks are feasible for large language models, and produces dramatic percentages that generate headlines but do not always match what we observe in the real world.</p><p>Anthropic takes a different approach with a metric they call observed exposure. Instead of measuring what AI could do, they measure what AI is actually being used for inside real professional workflows right now. The difference sounds subtle but the implications are significant.</p><p>Their methodology combines three sources. The O*NET occupational database, which breaks down hundreds of jobs into granular task descriptions. Prior research estimating which tasks are theoretically feasible for large language models. And real usage data showing how AI systems are being deployed in professional contexts.</p><p>Mathematically, the framework treats jobs as a weighted aggregation of tasks. If a job consists of tasks t1, t2, t3 …… tn, and each task carries a time weight wi representing the fraction of working hours spent on it, exposure can be approximated as:</p><p>Exposure = Σ (wi × ui)</p><p>Where ui captures whether that task is currently being performed with AI assistance. Tasks that are fully automated receive higher weight than tasks where AI only augments human effort. The result is a single number representing how much of a job’s workflow is currently touched by AI in practice.</p><p><strong>The gap nobody is talking about</strong></p><p>When you apply this framework across occupations, the most striking finding is not which jobs are exposed. It is how large the gap is between theoretical capability and actual adoption.</p><p>Computer and mathematics roles are a good example. Large language models could theoretically assist with the vast majority of tasks in those occupations. Observed usage today covers roughly one third of them. We are at 33% of potential deployment in one of the most AI-adjacent fields that exists.</p><p>That gap is not a reassuring statistic. It is a preview of what is still coming. And it is arriving into labor markets that are already adjusting their hiring behavior in response to the first third.</p><p><strong>What the data actually shows</strong></p><p>The report finds no systematic increase in unemployment among workers in highly exposed occupations since ChatGPT launched. If you stopped reading there you might conclude that AI disruption is overhyped.</p><p>But the more important signal is in hiring. Workers aged 22 to 25 are being hired into AI-exposed roles at roughly 14% slower rate compared to before 2023. The front door is not slamming shut. It is closing slowly and most people are not noticing because nobody is announcing it.</p><p>Economically this pattern makes complete sense. When AI increases worker productivity, firms do not need to fire existing employees to capture the gains. They simply need fewer new hires to sustain the same level of output. The workforce stays flat or grows slowly while productivity rises. No headlines. No mass layoff announcements. Just a gradual tightening of the entry point.</p><p>For early career professionals this is precisely the worst version of disruption because it is invisible until you are already inside it.</p><p><strong>Where I think the analysis undersells the risk</strong></p><p>The report is careful and honest about its limitations, but there is one aspect I think deserves more emphasis. Unemployment is a lagging indicator and it is the primary signal being tracked here. Labor markets adjust through many channels before unemployment moves. Job posting volumes, skill requirement shifts, hiring cycle lengths, and offer rates all change earlier. By the time unemployment data shows a clear signal, the structural shift will already be well underway.</p><p>The second thing worth sitting with is the adoption curve. Most technologies follow an S-shaped diffusion pattern. Early adoption is slow because tools are unfamiliar, workflows are unoptimized, and integration is clunky. But once AI becomes embedded into standard software platforms and daily professional tools, adoption can accelerate rapidly. We are almost certainly still in the flat early portion of that curve. The steep part has not arrived yet. When it does, the gap between 33% observed exposure and theoretical capability will close faster than the current data suggests.</p><p><strong>What this actually means for people entering technical fields</strong></p><p>Here is where I land after sitting with this research for a while. The disruption is not going to feel like a wave. It is going to feel like the water level slowly rising while everyone debates whether it is actually rising.</p><p>Individual tasks get automated or augmented. Productivity per worker increases. Companies revise their headcount projections downward slightly. Nobody calls it an AI layoff because it is not one. It is just math applied quietly to hiring plans over several consecutive quarters.</p><p>The professionals who navigate this well will not be the ones who compete with AI on task execution. They will be the ones who understand how to use AI as leverage across an entire workflow. Domain expertise combined with AI fluency combined with the judgment to know when the output needs a human in the loop is a combination that is genuinely difficult to replicate and genuinely valuable to organizations figuring out what their new productivity structure looks like.</p><p>The window to build that combination is open right now. The 67% of AI potential that has not yet been deployed is not a comfort. It is a clock.</p><p>I sent out applications into that silence and eventually found my footing. But I think about the people sending applications two or three years from now, into a market where that 33% has become 50% or 60%, and whether they will understand what changed and why. That is the conversation I think we should be having more loudly right now.</p><p>Full research available at anthropic.com/research/labor-market-impacts</p><p>#AI #LaborMarket #DataScience #EarlyCareer #MachineLearning #Analytics</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c088b92c2ea4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Cheapest Prompt You Can Add to Any LLM and Why It Might Save More Than You Think]]></title>
            <link>https://medium.com/@imvk45/the-cheapest-prompt-you-can-add-to-any-llm-and-why-it-might-save-more-than-you-think-e100089a0ada?source=rss-220a05870872------2</link>
            <guid isPermaLink="false">https://medium.com/p/e100089a0ada</guid>
            <category><![CDATA[ai-efficiency]]></category>
            <category><![CDATA[optimization]]></category>
            <category><![CDATA[genrative-ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <dc:creator><![CDATA[Vikas Kumar]]></dc:creator>
            <pubDate>Mon, 02 Mar 2026 16:33:20 GMT</pubDate>
            <atom:updated>2026-03-02T16:33:20.054Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*KeUYpne2S3VBGKXJ.png" /></figure><p>There is a five-word phrase you can append to almost any LLM prompt that costs roughly 10 tokens and could save you hundreds.</p><p>“Do you have any questions?”</p><p>No fine-tuning. No new framework. No API changes. Just a single sentence that shifts the dynamic from “model guesses and answers” to “model asks, then answers.”</p><p>It sounds too simple to matter. But when you follow the research trail, this small intervention sits at the center of one of the most under-studied problems in applied AI: how much do we waste by letting models answer before they actually understand what we want?</p><h3>The Problem Nobody Is Measuring</h3><p>Every time you send an ambiguous prompt to an LLM, one of two things happens. Either the model makes an assumption and answers, sometimes correctly and often not quite, or it hedges and returns a long response that half-answers three different versions of your question.</p><p>In both cases you end up typing a follow-up. Each correction is a new round of tokens, new inference, new compute, new energy consumed.</p><p>Nobody is systematically counting those repair loops. The AI efficiency conversation has been almost entirely focused on the model side: smaller weights, faster inference, better hardware. Very little attention has been paid to the conversational waste that accumulates at the prompt level, turn by turn, across millions of daily interactions.</p><h3>What the Research Shows</h3><p>A few researchers noticed this pattern and started building benchmarks around it. What they found is hard to ignore.</p><p>ClarifyMT-Bench, a 2025 benchmark built around 6,120 multi-turn dialogues and a five-dimensional ambiguity taxonomy, found that models have a systematic under-clarification bias. They answer prematurely even when a prompt is genuinely ambiguous. Accuracy degrades sharply as a conversation deepens, dropping from roughly 80 to 90 percent on the first turn to as low as 20 to 45 percent by the third turn depending on model and ambiguity type.</p><p>AskBench tested models on highly ambiguous tasks that required asking before answering. General-purpose LLMs that skipped clarification and answered immediately achieved roughly 1 percent accuracy on hard cases. Models trained to ask first performed dramatically better.</p><p>The quality benefit of clarification is not marginal. In genuinely ambiguous settings it is often the difference between a useful answer and a useless one. But quality is only half the story. The more pressing question for anyone building or using these systems at scale is what clarification actually costs in tokens, and whether that cost is worth it.</p><h3>The Token Math</h3><p>A clarification round costs tokens upfront. The model asks one or two questions, you answer them, and that exchange adds roughly 100 to 300 tokens to the conversation.</p><p>But when a model answers without clarifying and gets it wrong, a single correction loop on a medium-complexity task can easily consume 500 to 2,000 tokens. Two or three repair loops and you have spent an order of magnitude more than a clarification round would have cost.</p><p>The net expected token cost per successfully resolved task can be expressed as:</p><p><strong>E[tokens] = P(clarify) × T(clarification round) + P(direct and correct) × T(direct answer) + P(error) × T(repair loops)</strong></p><p>If you are building or operating any LLM-powered product at scale, this equation matters. The middle term is the one most teams optimize for. The third term is the one most teams ignore entirely. But the weight of each term is not fixed. It depends heavily on one variable that most prompt design frameworks still do not account for explicitly.</p><h3>Ambiguity Is the Key Variable</h3><p>Not every prompt carries the same risk of getting it wrong, and that changes the calculation entirely.</p><p>“What is the capital of France” needs no follow-up. But most real-world prompts are not like that. “Write me a report on AI” is ambiguous on scope, audience, depth, and format. “Help me with my project” is severely underspecified. The kind of prompts that make up the majority of actual usage in coding assistants, support tools, and research applications sit firmly in ambiguous territory.</p><p>The higher the ambiguity, the higher the probability that a direct answer misses the mark, and the higher the return on a single clarifying question upfront. ClarifyMT-Bench models this across five ambiguity types. AskBench operationalizes it as a threshold: below some ambiguity level answer directly, above it ask first.</p><p>What does not yet exist is that threshold expressed quantitatively in terms of expected tokens saved per clarification question asked. That is the number that would actually change how system prompts are designed at scale. And once you start thinking in tokens saved, you quickly realize that tokens are not just a cost metric.</p><h3>The Energy Dimension</h3><p>They are an energy metric too.</p><p>Recent inference measurements put energy use at roughly 0.39 to 4 joules per output token depending on hardware and optimization level. Because energy cost is approximately linear in total tokens at fixed hardware settings, any net reduction in tokens per completed task translates proportionally into energy savings per task.</p><p>If clarification-first prompting reduces average tokens per resolved conversation by 20 percent on ambiguous tasks, scaled across millions of daily interactions the aggregate reduction is not trivial. Nobody has published that number yet, which means nobody has made the business or environmental case for this intervention in concrete terms. But you do not need a published study to test whether the pattern holds on your own workload.</p><h3>Try It on Your Next Prompt</h3><p>Add “Do you have any questions before you start?” to your next complex or ambiguous request. Watch whether the questions that come back surface assumptions you had not made explicit. Watch whether the final answer is more targeted and requires fewer corrections than what you would normally get.</p><p>For high-ambiguity tasks the pattern holds consistently. The model surfaces a real ambiguity, you resolve it in one sentence, and what follows is more precise and shorter than what a direct-answer-then-repair loop would have produced. The total token spend across the full conversation is lower, the answer is better, and you got there in fewer turns.</p><p>Five words. Ten tokens. Try it today.</p><p><em>This article is part of ongoing research into clarification-first prompting, token efficiency, and the energy cost of LLM inference.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e100089a0ada" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>