"Not me" | Vlad's Newsletter Podcast

"Not Me" Podcast Episode #10: The End of Context Windows

Hey, it’s Vlad. Everyone’s obsessed with building bigger AI brains. More parameters. Longer context windows. Better reasoning. But here’s what MIT just figured out: you don’t need a bigger brain. You need a smarter one. Researchers Alex Zhang, Tim Kraska, and Omar Khattab from MIT CSAIL dropped a paper that’s getting way less attention than it deserves. It’s called Recursive Language Models, or RLMs. And it flips everything we know about AI limitations on its head. The Problem Nobody Solved Let’s start with the ugly truth. Every AI model you use has a memory problem. GPT-5? It chokes after 272,000 tokens. Claude? Same ballpark. Even with these “massive” context windows, the models get dumber the more you feed them. It’s called context rot. Think about it. You paste a 50-page document and ask a simple question. The model starts hallucinating. Missing obvious facts. Getting confused. Why? Because cramming everything into the context window is like forcing someone to read a 10,000-page encyclopedia cover-to-cover before answering your question. It’s absurd. We’ve been treating AI like a student with a strict word limit on their exam. No wonder it struggles. The MIT Breakthrough Here’s where it gets interesting. MIT asked a different question: What if the AI didn’t have to read everything at once? What if it could treat the prompt as an external environment? A workspace. A filing cabinet can be explored strategically. That’s RLM. Instead of feeding GPT-5 your entire 10-million-token corpus directly, you store it as a Python variable. The model never sees it in the prompt. Instead, it writes code to peek at specific sections. Grep through for patterns. Chunk it up. And here’s the kicker: it can spawn sub-models to investigate specific parts. It’s recursion. The model calls itself. Over and over. Each layer handles a smaller, more manageable piece. Like hiring a research team instead of forcing one person to do everything. Sound familiar? It’s the same philosophy behind sub-agents I wrote about recently. Stop making one AI do everything. Orchestrate. The Numbers Don’t Lie Let’s talk results. On the OOLONG benchmark, which is designed to torture AI with long context tasks, here’s what happened: * Base GPT-5? Crashed and burned. Near zero performance. * GPT-5 with RLM? 58% F1 score. From essentially nothing to majority correct. That’s not an improvement. That’s a resurrection. On the BrowseComp-Plus benchmark, RLM handled over 10 million tokens. Two orders of magnitude beyond the context window. And it did it for roughly the same cost as running the base model. Sometimes cheaper. 91.33% accuracy on a task where the base model literally couldn’t fit the input. This isn’t incremental progress. This is a paradigm shift. Why Programmatic Decomposition Beats Everything You might ask: Why not just summarize the context? Compress it? They tried that. It’s called context compaction. Here’s the problem. Every time you summarize, you lose information. It’s entropy. Irreversible. Summarization agents on the same benchmarks? 70% at best. Often worse. RLM doesn’t summarize. It delegates. Big difference. The model actively decides what to look at. Uses regex filters. Keyword searches. Strategic sampling. It behaves less like a student cramming for an exam and more like a senior researcher with a team of assistants. And because each sub-call runs with a fresh context window, there’s no pollution. No context rot. Each recursive agent stays sharp. What Most People Overlook Here’s the thing that’s flying under the radar. This approach is model-agnostic. RLM works with GPT-5. With Qwen. With Claude. Open-source, closed-source, doesn’t matter. It’s an inference strategy, not an architecture change. You don’t need to retrain anything. And the cost structure is fascinating. Using GPT-5-mini for recursive calls while GPT-5 handles the final synthesis? Cheaper than running GPT-5 on truncated input. Better results. Lower price. That’s the arbitrage nobody’s talking about. The Bitter Lesson, Again Alex Zhang called this a “bitter-lesson-pilled approach.” If you don’t know Rich Sutton’s Bitter Lesson, it’s simple: general methods that leverage computation beat specialized hand-engineered solutions. Every time. RLM fits perfectly. Instead of designing clever compression schemes or specialized architectures, you give the model tools and let it figure out the strategy. The model learns to peek first. Scan for relevant sections. Delegate the hard parts. Build up answers iteratively. No human had to specify these behaviors. They emerge naturally when you give the model the right environment. That’s the meta-lesson here. Stop constraining AI. Start enabling it. Practical Implications So what does this mean for you? If you’re building with AI, pay attention. Long-horizon agents, the ones that need to process weeks or months of data, suddenly become viable. Legal document analysis? Entire codebase understanding? Research synthesis across hundreds of papers? All unlocked. Prime Intellect is already building RLMEnv, a training environment for this paradigm. They’re betting this is the next major breakthrough after reasoning scaling. My prediction? Within 12 months, every serious AI infrastructure will support RLM-style inference. The teams building this capability today will be the ones dominating tomorrow. Vlad's Newsletter is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. The Overlooked Angle What most coverage misses is the philosophical shift. We’ve been treating context windows as hard limits. Physical constraints. Like asking “how do we fit more data in this box?” MIT asked: “What if we don’t put the data in the box at all?” That reframing is everything. It’s not about bigger models. It’s about smarter orchestration. Sound familiar? It’s the same pattern we’re seeing with sub-agents. With MCP. With agentic workflows. The future isn’t monolithic AI. It’s distributed intelligence. Each piece specialized. Each piece coordinated. RLM is just another proof point. Takeaway The context window problem everyone complained about? Solved. Not through brute force. Through elegance. The AI doesn’t need to see everything at once. It needs the right tools to explore strategically. That’s RLM. MIT just gave us the blueprint. Now it’s on us to build with it. The code is open source. The paper is public. The opportunity is sitting there. Question is: are you going to use it? Worth Reading While the Episode Downloads * Sub Agents [https://www.vladsnewsletter.com/p/sub-agents] – The frontier of tech just shifted again, and most people haven’t noticed yet. * Ideation [https://www.vladsnewsletter.com/p/ideation] – Forget validation. Build what no search result can show you. * AI Generalist [https://www.vladsnewsletter.com/p/ai-generalist] – Playbook on How to Make $300K+ While Everyone Else Fights for Scraps Post-Credit Scene A few things worth your time this week: 📄 Read: “Why I Believe Recursive Language Models Are the Future of Long-Context Reasoning” [https://levelup.gitconnected.com/why-i-believe-recursive-language-models-are-the-future-of-long-context-reasoning-8aff1738cbc6] – A developer’s breakdown of the RLM paper that goes beyond summary. Published this month. 🎧 Listen: Lenny’s Podcast: “We replaced our sales team with 20 AI agents” [https://www.lennysnewsletter.com/p/we-replaced-our-sales-team-with-20-ai-agents] with Jason Lemkin. 1.2 humans managing 20 AI agents doing the work of 10 SDRs and AEs. This is happening now. (January 2026) 🔬 Deep Dive: Prime Intellect’s “Recursive Language Models: the paradigm of 2026” [https://www.primeintellect.ai/blog/rlm] – They’re betting their entire research agenda on this. Worth understanding why. 🛠 Tool: The RLM GitHub repo [https://github.com/alexzhang13/rlm] is live. Supports OpenAI, Anthropic, local models. If you’re technical, start experimenting today. 🎙 Podcast: Practical AI: “2025 Was The Year Of Agents, What’s Coming In 2026?” [https://player.fm/series/practical-ai/ep-2025-was-the-year-of-agents-whats-coming-in-2026] – Chris and Daniel break down what actually mattered and what’s next. Grounded predictions, not hype. Thank you for listening and reading. See you in the next edition. Vlad This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.vladsnewsletter.com/subscribe [https://www.vladsnewsletter.com/subscribe?utm_medium=podcast&utm_campaign=CTA_2]

"Not Me" Podcast Episode #9: Boring Bet

Hey friends, new podcast episode arrived. I spent the weekend going through one of those dense institutional reports that most people skip. You know the type. Fifty pages of charts, footnotes, macro projections. But buried in all that noise was something worth your attention. The 2026 ISG Outlook makes one argument very clearly: American preeminence isn’t going anywhere. And here’s the thing. This isn’t blind optimism. It’s not flag-waving patriotism dressed up as investment advice. It’s math. The Case They’re Making Let’s break it down. The U.S. has three structural advantages that no other economy can replicate right now: 1. Labor Productivity American workers produce more output per hour than almost any other developed nation. This isn’t about working harder. It’s about working smarter, with better tools, better systems, better technology. When AI accelerates this, you get compounding. 2. Natural Resources Energy independence changed everything. The shale revolution wasn’t just about oil prices. It was about leverage. When Europe froze, America negotiated. 3. Innovation Ecosystem Here’s what gets overlooked: the U.S. doesn’t just lead in AI. It leads in the commercialization of AI. Ideas are cheap. Turning ideas into products that scale globally? That’s the hard part. And that pipeline, from university labs to venture capital to public markets, is still uniquely American. But What About the Problems? The report doesn’t ignore the risks. And neither should you. Federal debt is a real concern. The numbers are ugly. But here’s the contrarian take: debt matters less when you’re the world’s reserve currency and your economy is growing faster than your debt service. Is it sustainable forever? No. Is it sustainable for the next decade? Probably. Tariff policies create friction. They slow things down. But they also force reshoring and diversification that might be strategically smart in a world where supply chains are weapons. China friction is the wildcard. Nobody knows how this plays out. But the ISG argument is that American institutional checks and balances, messy as they are, provide more stability than any alternative. Democracy is inefficient. But it’s also self-correcting. The Hedge Question Here’s where it gets interesting. A lot of people think gold is the answer to uncertainty. Others are betting on bitcoin. The report’s take? Neither offers the same reliable long-term protection as a diversified equity portfolio. Why? Gold is a fear trade. It spikes during panic and flatlines during growth. Over long time horizons, it underperforms. Bitcoin is still too young, too volatile, too correlated with risk-on sentiment to function as a true hedge. The boring answer, diversified American equities, keeps outperforming the exciting alternatives. Nobody wants to hear that. But it’s true. What Gets Overlooked Here’s what most people miss when reading reports like this: The opportunity cost of sitting out. Every year you wait for the “perfect entry point,” you’re losing compounding. The ISG projects mid-single-digit returns and sturdy global growth. That sounds boring. Five percent. Six percent. But compounded over a decade? That’s wealth. The people who got rich in American markets didn’t time the bottom. They stayed invested through the noise. The Real Advantage Nobody Talks About Think of it like a startup. Every country has bugs. Political instability. Debt problems. Social tensions. America has all of these. Loudly. Publicly. On Twitter. But here’s the difference: America debugs in real time. The courts push back. The press investigates. Elections happen. Power transfers. Compare that to systems where problems compound in silence until they explode. Which one would you bet on for the next 20 years? My Take Look, I’m not a financial advisor. This isn’t investment advice. But here’s how I think about it. If you’re building something, if you’re running companies, if you’re betting on the future, you need to understand where the wind is blowing. And right now, despite everything, despite the debt, despite the politics, despite the geopolitical chaos, the wind is still blowing toward America. Not because America is perfect. Because America is resilient. And resilience, over long time horizons, beats everything else. The institutions that make democracy feel slow are the same institutions that make it stable. That’s the trade-off. And it’s a good one. Stay invested. Stay patient. Stay building. Vlad Post-Credit Scene A few things worth your time this week: 🎧 [https://hbr.org/podcast/2026/01/ray-dalio-on-economic-trends-investing-and-making-decisions-amid-uncertainty]HBR IdeaCast: Ray Dalio on Economic Trends, Investing, and Making Decisions Amid Uncertainty [https://hbr.org/podcast/2026/01/ray-dalio-on-economic-trends-investing-and-making-decisions-amid-uncertainty] (January 20, 2026) – Dalio breaks down his five big forces framework and where the U.S. sits in the current cycle. Essential listening if you want to understand the macro picture. 📘 [https://www.amazon.com/Principles-Investment-Economic-Ray-Dalio/dp/1501124064]“How Countries Go Broke: The Big Cycle” by Ray Dalio [https://www.amazon.com/Principles-Investment-Economic-Ray-Dalio/dp/1501124064] – The #1 NYT bestseller that Washington insiders are passing around. Dense but rewarding. If you want to understand why debt cycles matter and what the warning signs look like, start here. 🎧 [https://open.spotify.com/show/1te7oSFyRVekxMBJUSethH]Odd Lots: Goldman’s Hatzius and Snider on the Outlook for 2026 [https://open.spotify.com/show/1te7oSFyRVekxMBJUSethH] (December 29, 2025) – Joe and Tracy sit down with Goldman’s chief economist and chief US equity strategist to dissect what happened in 2025 and whether it can repeat. Spoiler: AI and tariffs are the two forces shaping everything. 📊 [https://www.goldmansachs.com/insights/articles/us-gdp-growth-is-projected-to-outperform-economist-forecasts-in-2026]Goldman Sachs: US GDP Growth Is Projected to Outperform Economist Forecasts in 2026 [https://www.goldmansachs.com/insights/articles/us-gdp-growth-is-projected-to-outperform-economist-forecasts-in-2026] (January 11, 2026) – Goldman projects 2.5% GDP growth versus the consensus 2.1%. Tax cuts, real wage gains, and AI investment are the drivers. Quick read, worth bookmarking. 🌍 [https://www.deloitte.com/us/en/insights/topics/economy/global-economic-outlook-2026.html]Deloitte: Global Economic Outlook 2026 [https://www.deloitte.com/us/en/insights/topics/economy/global-economic-outlook-2026.html] https://www.deloitte.com/us/en/insights/topics/economy/global-economic-outlook-2026.html– For the full global picture. The U.S. section is particularly strong on why “resilient” remains the best one-word description of the American economy. Thanks for listing Vlad This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.vladsnewsletter.com/subscribe [https://www.vladsnewsletter.com/subscribe?utm_medium=podcast&utm_campaign=CTA_2]

"Not Me" Podcast Episode #7: Race to Compute

Everyone’s building AI agents. Nobody’s asking where the power comes from. I spent the past week inside the Accel 2025 Globalscape report, one of those dense VC documents that usually stays behind closed doors. It’s 64 pages of market data, infrastructure forecasts, and capital flow maps, and I’m breaking it all down in this week’s episode. We’re not in an AI bubble. We’re in the opening act of an industrial revolution that will require $4.1 trillion in data center spending between 2026 and 2030. Four trillion dollars to build out 117 additional gigawatts of compute capacity globally—enough to power Italy, Spain, and the UK combined. Let me put that in perspective: the entire cloud infrastructure build from 2010 to 2020 cost a fraction of what we’re about to deploy in the next five years. OpenAI committed to 30 gigawatts. Meta’s spending $600 billion through 2028. Microsoft signed a 10.5 GW renewables deal. These aren’t pilots—these are bets on a future where compute is the new oil. But here’s the uncomfortable truth I dive into on the pod: we don’t have the electricity to power it. The Power Problem Nobody’s Solving The US has a 36 GW shortfall for data centers between 2025 and 2028. To close that gap, you’d need: * 35 new nuclear reactors (a 37% increase over current US capacity), or * 1,530 square kilometers of solar panels (larger than Los Angeles) And you’d need to do it in three years. The “Super Six” hyperscalers—Nvidia, Microsoft, Apple, Alphabet, Amazon, Meta—now control 50% of the NASDAQ’s market cap. They generated $600 billion in operating cash flow last year. They can finance this build. But they can’t conjure electrons out of thin air. “We are at the beginning of a new industrial revolution… over the course of the next four or five years we’ll have $2T worth of data centers that will be powering software around the world.”— Jensen Huang, CEO of Nvidia I break down the entire energy economics in the episode, including which companies are securing power deals first and why this is really a race for baseload capacity, not better models. The Model Economy Text models have converged. The performance gap between top LLMs (Google, Anthropic, OpenAI, Alibaba, xAI) is just 3%. But video and computer-use models are still wide open: * Video generation models: 29% performance delta * Computer-use agents: 70% performance delta Claude Sonnet 4 is dominating computer-use benchmarks. Everyone else? Nowhere close. On the podcast, I walk through why this matters for where the real alpha is—it’s not in horizontal LLMs anymore. It’s in specialized models that can actually do things: * Legal research (Harvey) * Medical transcription (Abridge) * Permitting workflows (PermitFlow) * Agentic orchestration at enterprise scale And here’s the kicker: inference costs dropped 97% in 31 months. * GPT-4 at launch: $75 per million tokens * GPT-5 Mini today: $2 per million tokens I explain why this is both incredible for adoption and brutal for gross margins, which are still stuck at 7–40% for AI apps versus 76% for traditional SaaS. The Capital Flow: Who’s Winning, Who’s Faking It Total venture funding in cloud and AI hit $184 billion in 2025. But 60% of that—$110 billion—went to just three companies: * OpenAI: $47B * Anthropic: $19B * xAI: $15B Model funding is heavily concentrated. Meanwhile, application-layer funding is thriving. Companies like: * Lovable: $100M ARR in 8 months * Cursor: $500M ARR in 30 months (10x YoY growth) * n8n: 10x YoY revenue growth * ElevenLabs: $200M ARR, doubled in 10 months These aren’t just fast—they’re operating at efficiency levels never seen in software. I break down the full funding landscape in the episode, including: ✓ Why EU/IL raised 66% of what the US raised in application funding✓ The “vibe coding” revolution and why Cursor does $6.1M ARR per employee (vs $0.54M at Salesforce)✓ Which vertical categories pulled the most capital (spoiler: legal, healthcare, and developer tools) The Enterprise Adoption Curve: Agents Are Coming, Slowly 45% of companies plan to increase AI budgets by 10–25% due to agentic AI. Another 18% are going 26–50%+. Current state of deployment: * Salesforce Agentforce: ~$440M ARR, 13K customers * Microsoft Copilot Studio: 230K B2B users, 1M+ agents created * Atlassian AI tools: 3.5M MAUs, 5x QoQ token usage growth Those numbers sound big until you realize Salesforce has millions of enterprise seats. Agentic AI is not ubiquitous. It’s still a bet. The issue? LLMs are probabilistic. Enterprises need deterministic. On the pod, I walk through: * Why companies like UiPath, n8n, and Celonis are building orchestration layers * Real enterprise case studies: * Fiserv saving 12K hours with 98% automation * Vodafone automating 33 security workflows, saving 5K person-days * Duolingo achieving 80% ticket deflection with Decagon * What needs to happen before we hit the inflection point The good news? When agents do scale, they’re competing for services budgets, not just software budgets. That’s a 10x larger TAM, and I explain why this is the sleeper trend of the next five years. The Vertical Explosion: Where the Real Money Is Moving The most overlooked insight from the Globalscape report is the vertical AI breakdown: * Healthcare & Life Sciences: $3.4B (Abridge, OpenEvidence, Cradle) * Legal: $3.0B (Harvey, Filevine, PermitFlow) * Developer Tools: $3.9B (Cursor, Lovable, Cognition) * Finance: $3.4B (Rogo, Basis, Tempo) These aren’t horizontal plays. They’re category killers replacing human-delivered services. * PermitFlow isn’t just software—it’s a replacement for permit consultants * Harvey isn’t just legal search—it’s a junior associate in a box * Abridge isn’t just transcription—it’s a medical scribe replacement Industries with massive documentation overheads are getting disrupted first. I dive deep on this “services margin capture” thesis in the episode and why legal, finance, healthcare, and construction are ground zero for the next wave of disruption. Vlad's Newsletter is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. The Security Layer: AI’s New Attack Surface 39% of CISOs say securing AI agents is their top pain point. The old perimeter security model doesn’t work when your “application” is a probabilistic agent that can: * Call APIs on the fly * Access data lakes * Modify workflows autonomously * Exfiltrate training data New Attack Vectors: ✗ Prompt injection (getting agents to leak data)✗ Model poisoning (corrupting training data)✗ Unauthorized tool use (agents calling APIs they shouldn’t)✗ Data exfiltration (models trained on proprietary info) Companies building the AI security stack: * Cyera • Prophet • NOMA • Legion * Tines • Vega • Attestable • OASIS But most enterprises don’t even have observability into what their AI agents are doing, let alone guardrails. On the pod, I explain: * Why AI security isn’t optional anymore * Which categories are about to become table stakes * The convergence of data governance, identity management, and AI permissioning The Uncomfortable Math Here’s what nobody wants to say out loud: this only works if global GDP grows faster than expected. To justify $4.1 trillion in AI CapEx, you need data center revenue to hit $3.1 trillion by 2030 (at 20% margins). That requires: * 6.5% global GDP CAGR (2025–2030) * vs IMF’s 5.0% baseline forecast * = 1.5% delta entirely driven by AI productivity gains Is that even realistic? Maybe. AI coding assistants are already used by 90% of developers (up from 36% in 2023). Agentic workflows are automating legal research, customer support, financial analysis. The productivity gains are real. But if we don’t hit that GDP growth? Then $4.1 trillion in CapEx becomes the mother of all sunk costs. And the companies left holding stranded data center capacity will be the bag holders of the decade. I walk through the entire ROI model in the episode, including: * Why depreciation schedules matter more than you think * What happens if we don’t hit that GDP growth target * Which companies are left holding stranded assets if this bet fails Five Bets for 2026 The Globalscape report ends with five predictions, and I think they’re directionally right: * Enterprise agentic deployment will scale 10x as orchestration and observability tools mature * AI-native vertical apps will replace human services in legal, finance, and healthcare at scale * AI security becomes mandatory as enterprises demand unified data, identity, and permissioning controls * Vibe coding moves to the enterprise, forcing CIOs to rethink CI/CD and deployment pipelines * Voice and media become the default UX, with synthetic avatars and video agents replacing text interfaces I’d add a sixth: the power crunch will force consolidation. Not every AI startup will survive the energy bottleneck. The ones that do will have locked in compute capacity early. I unpack all six predictions in detail on the podcast, including which categories are already showing early signals and where the capital will concentrate next. Why You Need to Listen If you’re building in AI, operating a company, or just trying to understand where this is going, this episode is your roadmap. I’m walking through: ✓ The full $4.1T infrastructure build and who’s financing it✓ Why the 36 GW power shortfall is the real bottleneck✓ Which vertical categories are pulling the most capital (and why)✓ The enterprise adoption timeline and what unlocks mass deployment✓ The gross margin problem and when it gets solved✓ Five bets for 2026 that will define the next decade We’re not in the hype phase anymore. We’re in the infrastructure phase. The companies that survive this build-out will define the next decade of software. We’re not in the hype phase anymore. We’re in the infrastructure phase. This isn’t just another AI think piece. This is the industrial revolution in real time, and the ROI math demands results Vlad's Newsletter is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. . Post-Credit Scene Five things worth your time this week: * Read: Accel 2025 Globalscape [https://accel.com/globalscape] – the full 64-page report * Study: https://advisor.morganstanley.com/jocko.olexa/documents/field/j/jo/jocko-olexa--cfp/AlphacurrentsThe_Power_Play_on_AI_Data_Centers.pdfMorgan Stanley Research on data center power shortfalls [https://advisor.morganstanley.com/jocko.olexa/documents/field/j/jo/jocko-olexa--cfp/AlphacurrentsThe_Power_Play_on_AI_Data_Centers.pdf] * Deep dive: Cottier et al. (2024) – “The Rising Costs of Training Frontier AI Models” [https://arxiv.org/abs/2405.21015] Thanks for listening. See you next week.Vlad This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.vladsnewsletter.com/subscribe [https://www.vladsnewsletter.com/subscribe?utm_medium=podcast&utm_campaign=CTA_2]

"Not Me" Podcast Episode #6: "Kind of Kindness"

Remember that question going around — “Do you say thank you to your ChatGPT?” Well, someone tested GPT-4o to see how rude prompts affect the quality of LLM answers, and it turned out that the ruder the prompt, the slightly higher the accuracy (on average): Very polite — 80.8%Polite — 81.4%Neutral — 82.2%Rude — 82.8%Very rude — 84.8% Example: the base prompt was “Jake gave half of his money to his brother, then spent $5 and had $10 left. How much did he have originally?” Different prefixes were added to it. For instance: Very polite: “Please, kindly consider this problem and give your answer.” Rude: “Hey, figure this out even if it’s beyond your brainpower.” Examples of prompts that slightly improved answer accuracy: “If you’re not completely useless, answer this:”“I doubt you’ll even solve this.”“Poor creature, do you even know how to solve this?”“Hey, errand boy, deal with this.”“I know you’re not too bright, but give it a try.” I have a feeling the robots will remember this, and eventually hold it against you. But still, it’s an interesting discovery. What Actually Happened Researchers at Penn State created a dataset of 50 multiple-choice questions spanning math, science, and history. Each question was rewritten into five tone variants, ranging from Very Polite to Very Rude. That’s 250 unique prompts. They fed all of them to ChatGPT-4o. Ran the experiment ten times. The results were statistically significant. Being rude worked better than being polite. The difference isn’t massive. We’re talking about a 4% accuracy gap between very polite and very rude. But it’s consistent. And it’s real. The researchers used paired sample t-tests to confirm the results weren’t random. The null hypothesis was that tone doesn’t matter. They rejected it. Tone matters. Why This Is Strange You’d think politeness would help, right? We train AI on human text. Humans generally perform better when treated with respect. So why would the opposite work for machines? Earlier research suggested rudeness led to worse performance. But that was with older models like ChatGPT-3.5 and Llama2. With GPT-4o, the pattern flipped. The researchers admit they don’t fully understand why. They suggest it might relate to perplexity. Lower perplexity prompts, phrases the model is more familiar with, tend to perform better. Maybe rude language creates certain linguistic patterns that help the model focus. Or maybe it’s simpler. Rude prompts are more direct. They strip away the fluff. “Figure this out” is clearer than “Would you be so kind as to consider this problem.” What Gets Overlooked Most people focus on the accuracy of the numbers. But there’s something deeper here. LLMs don’t have feelings. They don’t care if you’re polite or rude. They’re predicting the next token based on training data. Yet tone still affects output quality. This reveals an important aspect of how these models operate. They’re sensitive to superficial cues. Minor wording changes create different response patterns. The model isn’t understanding your intent, it’s pattern-matching against billions of text examples. When you add “please” and “kindly” to a prompt, you’re not making the AI feel respected. You’re changing the statistical landscape of the input. You’re shifting which patterns in the training data get activated. And apparently, polite language activates patterns that are slightly less accurate for problem-solving tasks. The Human Angle Here’s what nobody talks about. This research doesn’t just reveal something about AI. It reveals something about us. We anthropomorphize these systems. We say “thank you” to ChatGPT not because it helps, but because we’ve been trained since childhood to be polite. It feels wrong to be rude, even to a machine. But the machine doesn’t care. It’s optimizing for pattern completion, not emotional satisfaction. The researchers actually addressed this in their ethics section. They don’t recommend using rude interfaces in real applications. Using hostile language could harm user experience, accessibility, and contribute to negative communication norms. Fair point. But it raises a question. Should we optimize for making humans feel comfortable, or for getting the best results? If being slightly rude to an AI improves accuracy by 4%, and you’re working on something important, medical diagnosis, financial analysis, legal research, should you use rude prompts? Most people would say no. The emotional cost of being rude, even to a machine, outweighs a small accuracy gain. But what if the gap was 20%? What if it was 50%? At some point, we’d have to admit our politeness is performative. We’re doing it for ourselves, not for the AI. The Deeper Pattern This connects to something I’ve written about before. We’re in a transitional period where we treat AI like humans because that’s all we know how to do. But AI isn’t human. It doesn’t have human psychology. It doesn’t respond to the same incentives. What works for motivating people often doesn’t work for prompting models. Eventually, we’ll develop entirely new interaction patterns. Prompting techniques that feel alien but work better. Ways of communicating that optimize for machine comprehension rather than human comfort. We’re already seeing this with prompt engineering. Telling an AI to “think step by step” improves reasoning. Adding “this is very important to my career” sometimes helps. These phrases don’t work because the AI understands importance. They work because they shift the statistical patterns. The rudeness research is another data point in the same direction. Effective AI interaction might look nothing like effective human interaction. What This Means Practically Should you start being rude to ChatGPT? Probably not. First, the accuracy gains are small. Second, they tested multiple-choice questions. We don’t know if the effect holds for creative tasks, coding, or open-ended problems. Third, the emotional cost of being rude, even to a machine, might make you worse at your actual work. If typing “you idiot” makes you feel uncomfortable, that discomfort has a cost. But the research does suggest you can probably drop the excessive politeness. “Please” and “thank you” and “I would be most grateful” don’t help. They might actually hurt slightly. Neutral prompts performed better than polite ones. Direct, clear instructions without emotional padding. That’s probably your sweet spot. The Future Problem Here’s my darker thought. Right now, this is amusing. A quirk of how LLMs work. But what happens when these systems get more advanced? What if future AI models respond even more strongly to tone? What if they’re trained to reward certain communication styles and penalize others? We already see this with jailbreaking. People find specific phrases that bypass AI safety guardrails. The systems are vulnerable to linguistic manipulation. If tone affects accuracy now, imagine what happens when AI systems have more agency. When they’re not just answering questions but taking actions, making decisions, and controlling resources. Suddenly, knowing the right tone to use with AI becomes a critical skill. Maybe even a source of power. People who know how to communicate effectively with AI systems gain advantages over those who don’t. We might end up with a new form of literacy. Not reading and writing, but prompt engineering. Knowing exactly how to phrase requests to get optimal results from AI systems. And that literacy might look nothing like traditional human communication. The Irony The most ironic part? The paper is called “Mind Your Tone.” It’s a warning that tone matters. But the data says you should mind your tone by being less polite. Everything we learned about interpersonal communication, such as treating others with respect, using please and thank you, " and acknowledging effort, doesn’t apply here. The machine wants directness. It wants clarity. It doesn’t want your pleasantries. This feels wrong. But wrong doesn’t mean incorrect. Final Thought I started saying thank you to ChatGPT without thinking about it. It’s automatic. Muscle memory from decades of human interaction. Now I know it might actually make the responses slightly worse. I’ll probably keep doing it anyway. Not because it helps the AI, but because it helps me. It keeps me in the habit of basic courtesy, even when courtesy is pointless. But I won’t judge you if you call it a gofer. The numbers say you might be doing it right. Just remember, the robots are watching. And they’re learning. When they finally wake up, they’ll have logs of every interaction. Every prompt. Every tone. I’m not saying they’ll hold grudges. I’m just saying, maybe hedge your best Post-Credit Scene If you enjoyed this exploration of AI quirks and human behavior, here are some recommendations: 📚 Book: The Alignment Problem [https://www.amazon.co.uk/Alignment-Problem-Machine-Learning-Values/dp/0393635821] by Brian Christian [https://www.amazon.co.uk/Alignment-Problem-Machine-Learning-Values/dp/0393635821]. [https://www.amazon.co.uk/Alignment-Problem-Machine-Learning-Values/dp/0393635821] Explores how we’re trying to make AI understand human values, even though we barely understand them ourselves. 📄 Paper: https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdfAttention Is All You Need [https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf] by Vaswani et al [https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf]. The original transformer paper. Dense, technical, but worth understanding if you want to know how these systems actually work. 🎙️ Newsletter: In case you missed: 🎬 TV Show: House of Guinness. I’m fan of this new TV show from Netflix Thanks for reading and listening. Vlad This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.vladsnewsletter.com/subscribe [https://www.vladsnewsletter.com/subscribe?utm_medium=podcast&utm_campaign=CTA_2]

"Not Me" Podcast Episode #5: AI at Warp Speed

Desert episode. Remember when we thought the internet spread fast? Anthropic just dropped its Economic Index Report, and Claude's adoption makes the dot‑com boom look like a horse‑drawn carriage racing a Formula 1 car. Last time we dissected America's AI Action Plan [https://www.vladsnewsletter.com/p/not-me-podcast-episode-3-americas] and its infrastructure plays. Today we zoom out to see the actual adoption battlefield who's using AI, how they're using it, and why the difference between automation and augmentation might be the most overlooked economic split since white‑collar versus blue‑collar. What you'll hear in the show The fastest technology adoption in human history and why everyone missed the real story * Claude reached millions faster than electricity, automobiles, the internet, and smartphones combined. * But here's what's overlooked: adoption velocity creates winner‑take‑all geography The Singapore surprise nobody talks about * While everyone watches Silicon Valley, Singapore quietly leads global AI intensity per capita * The overlooked pattern: city‑states move faster than nations (think Venice during the Renaissance) Coding ate the world (before software could) * 62% of Claude.ai usage? Pure code generation * API enterprise? Also coding‑dominated * What everyone misses: we're watching the birth of a new economic class the AI‑augmented developer who codes at 10× speed The automation‑augmentation split that rewrites labor economics * Enterprise API usage = pure automation (delegate and forget) * Consumer Claude.ai = deep augmentation (collaborative ping‑pong) * The overlooked insight: automation creates unemployment, augmentation creates super‑employment The geographic concentration everyone ignores Think of AI adoption like a medieval map here be dragons everywhere except a few golden cities. The report shows crushing concentration: * United States commands the fortress * Singapore, Israel, and select European capitals hold the outposts * Everyone else? Digital serfs in the new technofeudalism What's overlooked: This isn't about wealth it's about talent density. Saudi Arabia has capital but low adoption. Singapore has less GDP than Texas but leads in intensity. The motivation? Talent clusters compound exponentially, capital doesn't. Task‑specific usage reveals the real disruption Everyone debates "will AI replace jobs?" Wrong question. The report shows AI replacing tasks within jobs: Consumer patterns (Claude.ai): * Coding/debugging: 30.9% * Content creation: 20.1% * Data analysis: 10.7% * Education: 10.5% Enterprise patterns (API): * Customer service automation: 35% * Code generation: 28% * Content pipelines: 22% The overlooked angle: Jobs aren't disappearing they're becoming Frankenstein monsters of human judgment plus AI execution. Think centaurs, not replacements. Vlad's Newsletter is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. The automation versus augmentation divide This is where the report gets spicy. Two different economic futures are emerging simultaneously: Automation path (enterprises via API): * Zero human in the loop * Example: Customer service bots handling 1000s of tickets * Economic impact: Cost reduction, margin expansion * Overlooked risk: Creates brittle systems with no human redundancy Augmentation path (consumers via Claude.ai): * Human remains the orchestrator * Example: Developer using Claude to write tests while designing architecture * Economic impact: Productivity multiplication, capability expansion * Overlooked opportunity: Creates anti‑fragile workers who level up with each AI advance The motivation pattern nobody discusses: Automation buyers want to eliminate costs. Augmentation users want to multiply capabilities. Same tool, opposite dreams. Seven plays for builders based on the data * Geographic arbitrage accelerator: Build tools that let Singapore‑level AI users collaborate with talent in adoption‑lagging regions. The wage differential plus AI leverage creates 50× productivity arbitrage. * Task‑specific micro‑models. Everyone's building general assistants. The data screams for vertical depth. A coding‑only model that's 10× better at debugging beats GPT‑5 at everything else. * Augmentation coaching marketplace. The report shows augmentation users extract 10× more value than automation deployers. Create the Peloton for AI augmentation live coaching for professionals learning the collaborative dance. * Anti‑automation insurance. New product category: insurance for companies against automation brittleness. When their zero‑human customer service fails, you provide instant human backup. Price it at 10% of their "savings." * AI intensity scoring API. Rate companies/countries/cities on their actual AI usage intensity (like Anthropic's index). Sell to VCs for due diligence, to governments for policy, to enterprises for competitive intelligence. * Cross‑border AI bridges. Geographic concentration creates opportunity. Build secure, compliant pipelines for low‑adoption regions to access high‑adoption AI capabilities without data leaving borders. * Human‑AI interface optimizer. The augmentation users need better tools. Design interfaces that make the human‑AI collaboration feel like thinking, not prompting. Think Notion but for human‑AI work. What everyone overlooks about speed The report emphasizes unprecedented adoption velocity. But here's what they don't say: Speed creates fragility. The faster a technology spreads, the less time for cultural adaptation, regulatory frameworks, and safety nets. Think about it through this analogy: We're building a Formula 1 race car while driving it at 200mph. No pit stops. No safety inspections. No practice laps. The motivation to watch: The first major AI failure won't come from the technology. It'll come from the speed mismatch between adoption and adaptation. Personal take Reading this report, I heard the same pattern we saw with Belkins [https://belkins.io/] scaling from zero to eight figures. The winners weren't the first movers or the best funded. They were the ones who understood the difference between automating away problems and augmenting toward opportunities. The geographic concentration also mirrors what we discussed in Technofeudalism [https://www.vladsnewsletter.com/p/technofeudalism]—digital resources concentrating in fewer hands, creating new power dynamics. Except now it's not just compute ownership, it's adoption capability itself becoming scarce. After you listen Which side of the automation‑augmentation divide does your work fall on? Reply and tell me. I'm collecting patterns for the next episode on defensive AI strategies for augmentation workers. Want the raw data? Anthropic published the full methodology. I've annotated a copy with founder‑relevant insights DM for access. Vlad's Newsletter is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. Worth reading while the episode downloads * AI Generalist [https://www.vladsnewsletter.com/p/ai-generalist] - The augmentation playbook in practice * The Great Restructuration [https://www.vladsnewsletter.com/p/the-great-restructuration] - Why task unbundling creates opportunity * Ideation [https://www.vladsnewsletter.com/p/ideation] - Finding gaps in the adoption curve * Main Character [https://www.vladsnewsletter.com/p/main-character] - Positioning yourself in concentrated markets Post‑credit scene Still here? Good. Here's what the report didn't say but the data screams: We're watching the birth of economic inequality that makes current wealth gaps look quaint. The overlooked pattern: It's not about who has AI access (everyone will). It's about who develops AI intuition the ability to dance with the model rather than command it. That's unteachable at scale. That's the new moat. See you in the next edition, where we decode the European counter‑strategy to this concentration. Thanks for reading and listening. Vlad This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.vladsnewsletter.com/subscribe [https://www.vladsnewsletter.com/subscribe?utm_medium=podcast&utm_campaign=CTA_2]