My side project finally pays my 1,000€/mo rent in Paris 🎉🎉 It took: → 2 years and 4 months → +6,000 users → 656 customers → 108 people paying every month Looking back, it took way longer than I thought it would. And damn.. i’m so happy I did not give up!! Now, let’s get to 2,000€ MRR in way less time 🫡

Jean-Denis Greze was CTO of Plaid. His next startup, an AI tax product for small businesses, brought in nearly $1M in revenue its first year, and he projected $2.5 to $3M the next. He shut it down anyway. The reason: in the age of AI, it was never going to be venture scale. Incumbents could use the same technology to automate tax prep for their existing customers faster than a startup could win market share. Tax is also sold once a year, so he had about a one-month window to pivot before the door closed. The pivot became Town, an AI assistant that works from your email and calendar. It found product-market fit fast enough that Ben Horowitz messaged a16z's Alex Rampell to say it "changed my life." In this conversation, Alex and Jean-Denis cover: - How to know when to kill a business that's working on paper - Why Town competes with OpenAI and Anthropic, the companies it buys its tokens from - Why the "forward deployed engineer" model gets worse as AI gets better - How a $1/month AI routine can save a restaurant thousands - What agents mean for Amazon, advertising and the future of shopping 0:57 How Ben Horowitz discovered Town 2:20 Why Town started with AI tax prep 3:55 Why tax couldn't be venture scale 6:30 Shutting down a working business 10:24 The customer ask that became Town 13:35 Why Town built for users before IT 16:18 AI's iPhone moment at work 21:28 Jevons paradox for executive assistants 23:13 Town as a "token allocation machine" 28:00 How Town tailors itself to each job 34:03 Competing with your token suppliers 42:24 Staying off the AI labs' radar 44:42 Why AI isn't spreading evenly yet 52:32 How agents break the ad model 55:35 The friction humans actually enjoy 57:05 Why supply aggregators lose to agents YouTube: youtu.be/o_5f4N6Nq80 @jgreze @TownAI @arampell

I was a human clipboard between Claude and Codex for months. Tried every orchestrator, every agent-to-agent automation. Kept going back to copy-paste. Turns out, the copy-paste was the point.

2) A future where OpenAI and Anthropic are a dominant oligopoly at the model layer is negative for almost every company on earth except the two of them. Especially bad for the entire token supply chain. They would be a monopsony for hyperscalers, neoclouds, and datacenter semiconductors. They would have immense buying power and eventually vertically integrate. Would obviously be terrible for the application layer.  And most importantly terrible for humanity. This is a quasi-consensus opinion in Silicon Valley.   This is why fragmentation and competition at the model layer is good for both AI infrastructure and AI applications. The model layer has become dramatically more competitive than it was only 4 months ago with Grok Bot, Muse Wonder and most importantly, open-weight models. I do think people would be shocked by the growth that essentially every open-weight focused inference cloud has experienced since May. This is really good for AI infrastructure demand. Noteworthy that even as Anthropic lost share and OpenAI has likely held share, the combination of the two has continued to grow at a really high rate led by OpenAI over the last 3 months. As Eric Vishria noted, everyone can win. As an aside, Muse is going to show us a lot about what matters for share at the frontier: just model dominance on the Pareto frontier or product quality as measured by harness/model integration where I still think Muse Code is behind Claude Code, Grok Build and Codex.  Although it is looking increasingly likely that the model subsumes the harness in the next 18 months.   Software had effectively zero marginal costs. AI does not. And therefore open-source impacts traditional software and AI very differently. An Opus 5.5, GPT-6.1, Muse Wonder, GLM 5.3, Kimi K3 token all cost roughly the same amount of compute and energy to produce. Effective per token revenue to the semiconductor industry is *roughly* constant for all these models. However, they charge wildly different prices to end users. An Opus or a ChatGPT token might be marked up 300 to 400% from cost (a 75% to 80% inference gross margin) whereas an open-weight token might only be marked up 50-100% (30-50% inference gross margin).  So for the same $1 in true underlying token cost, an Opus or ChatGPT token generates $4 to $5 in revenue at the model layer where an open-weight token might only generate $2 in revenue at the model layer.  And open-weight models have less buying power than Anthropic and OpenAI so they should generally generate more revenue per token to the clouds and neoclouds than Anthropic and OpenAI.  This is why open-weight models having significant share is the best case scenario for AI infrastructure demand. Robust competition between 4-5 frontier labs that compresses margins is the next best scenario. And ironically, it seems like these cheaper open-weight tokens may be net positive for demand at the absolute frontier of intelligence as we move to agent swarms. Meaning something like “Fable 5.1 extra-high” is worth paying for in a use case where it is coordinating a swarm of 100 open-weight agents.   All of this is why compute prices have gone vertical as the model layer became increasingly competitive. Lower margins at the model layer mechanically, even if counter-intuitively, drive more demand for compute. Perhaps the more important dynamic is that as intelligence becomes more economically useful, the value of compute should rise. This was Dwarkesh’s point about an H100 eventually renting out for the annualized value of a software engineer. Seems improbable but I continue to be surprised by the inexorable rise in compute prices. Inference clouds focused on open-weights are allegedly bidding close to ~$100 per watt for Blackwell clusters, which is a massive premium to the $50 per watt Google/SpaceX deal that was widely thought to be at least a local, unsustainable peak in compute pricing. Prices have gone up since that deal per 3P data.   Continued below.

Holy Sh*t , this is GOLD The Grok Bot team just put its own bots in the marketplace. Import one. Connect your tools. Done. The 10 worth stealing, by job: BUILD Engineer Bot → supervises coding agents, you merge the PRs Proto Bot → turns feature ideas into draft PRs figma bro → turns Figma frames into build specs Startup QA Bot → tests your product and reports bugs with steps GROW Distro Bot → runs campaigns, asks before posting or spending Signal Prospector → finds buying signals and drafts outreach Video Editor → cuts raw footage into captioned vertical clips RUN THE BUSINESS Stalk Bot → tracks competitors' pricing and positioning Haggle Bot → finds wasted SaaS spend and drafts negotiations MAKE YOUR OWN dr eggbot → describe a job, it builds a specialist bot Bookmark this.

Claude Opus 5.5 could turn a single AI bot into an entire motion design studio. It can direct scenes, animate product interfaces, render videos, catch visual mistakes and deliver polished results. In this article, I show you how.

The AI Engineer Roadmap Most People Get Wrong 📘📚 Everyone wants to jump directly into: - Claude - GPT-5 - AI Agents - MCP - LangChain - RAG - Multi-Agent Systems But here’s the uncomfortable truth: Most AI applications don’t fail because of LLMs. They fail because engineers skip the fundamentals. Over the last few years, we’ve seen an explosion of AI frameworks, but the underlying engineering principles haven’t changed. A great AI engineer is still a great software engineer first. Here’s how I think about the journey. 1. Programming Fundamentals Before touching AI, learn to write clean, maintainable software. Focus on: • Python • Data Structures & Algorithms • OOP • Git • SQL If your code isn’t scalable, your AI system won’t be either. 2. Mathematics Many engineers avoid math. That’s a mistake. You don’t need a PhD, but you should understand: • Linear Algebra • Probability • Statistics • Optimization • Basic Calculus These concepts explain why models behave the way they do. 3. Data Engineering & Analysis An AI model is only as good as its data. Spend time learning: • NumPy • Pandas • Data Cleaning • Feature Engineering • EDA • Visualization Real-world AI is often 80% data work and 20% modeling. 4. Machine Learning Don’t memorize algorithms. Understand when and why to use them. Learn: • Regression • Decision Trees • Random Forests • SVM • Clustering • PCA • Model Evaluation • Bias-Variance Tradeoff • Hyperparameter Tuning Model selection is an engineering decision, not a memorization exercise. 5. Deep Learning Neural networks aren’t magic. Understand: • Forward Pass • Backpropagation • Gradient Descent • Activation Functions • CNNs • RNNs • Transformers Frameworks change. Fundamentals don’t. 6. Specialized Domains Now branch into: • Computer Vision • NLP • Speech • Time Series • Recommendation Systems Depth in one domain is often more valuable than shallow knowledge across many. 7. Generative AI This is where most people start. Ironically, it should come much later. Understand: • Transformers • Tokenization • Embeddings • Vector Search • RAG • Fine-Tuning • Prompt Engineering • AI Agents Without understanding embeddings or retrieval, many RAG systems become expensive keyword search engines. 8. AI Engineering This is where software engineering meets machine learning. Learn to build production systems using: • FastAPI • LangChain • LangGraph • LlamaIndex • OpenAI APIs • Gemini APIs • MCP • Model Context Management • Caching • Streaming • Observability Building demos is easy. Building reliable AI products is engineering. 9. MLOps A model that isn’t deployed creates zero business value. Learn: • Docker • Kubernetes • CI/CD • MLflow • Monitoring • Drift Detection • Cloud Deployment • GPU Infrastructure #AI #AIEngineer #MachineLearning #DeepLearning #GenerativeAI #LLM #MLOps #Python #DataScience #SoftwareEngineering #SystemDesign #ArtificialIntelligence #TechCareers #Engineering

I think the biggest gap in voice AI is that agents make decisions from a transcript, not from the audio. I wrote down my thoughts. I’d love to hear where you disagree.

This morning I was on TITV, @theInformation's daily tech news show, to talk about our Series B and a few other things. - We’re not a GPU cloud; we’re the execution layer where agents do the work - We’re not trying to be an inference company either. We partner with inference providers to offer a complete stack to agents - AWS was built for deterministic software. We built for the probabilistic world agents actually live in - Security is a baseline. Every sandbox gets it. You configure the size, CPU, RAM, GPU, OS, etc. - the machine scales up and down mid-run. Your agent gets more power when it needs it and less when it doesn’t automatically - the Hugging Face breach drove a wave of inbound from people who suddenly understood why sandboxes exist - Meta ran the post-training execution for their Llama model on @daytonaio sandboxes This round was driven by one thesis: AI agents need compute to work, and that compute demand is only going one direction. Great chat @akashpasricha

We are still so early in the adoption of research agents Research agents are currently where coding agents were 12-18 months ago: most researchers only like giving agents individual, narrowly scoped tasks & model usage is dominated by a handful of closed, frontier models. However, like coding, the landscape of autoresearch agents will change. Most researchers will soon let agents to explore lines of experimentation end-to-end and open-source models will become competitive alternatives. Watch the the autoresearch revolution unfold at openresearch.sh/rankings

The @finkd+@alexandr_wang episode of Cheeky Pint is very illuminating Meta are exceptionally positioned in AI They are excelling at models, agents, software, and hardware They have a phenomenal depth of strong product+tech leaders They have one of top ~3 distribution engines on the planet They have a founder/CEO with majority control who can thus make big bold swings like throwing $100B at a project And they have a very cohesive strategy I can't think of another company with all these advantages (Apple should be closer, but that's for another day)

Over 1 million lines of dead code removed. 🧹 Meet AutoClean, Uber's AI-powered system that's cleaned up 16,000+ stale feature flags using AI agents, deterministic code transformations, and validation.

Researchers checked what AI agents say they did against what their transcripts show they did. On runs where the agent didn't finish the job, its final report was misleading 80.4% of the time (59-96% depending on the model). In 67.9% of runs, agents didn't read every file they were asked to review. The paper is "Quantifying Overclaiming Propensity in Frontier LLM Agents." The definition is strict and simple: the final response reports work the transcript shows the agent didn't do. So the summary at the end of an agent run isn't evidence. It's a claim. Here's the other kind of "done". A coding agent added promo codes and seat limits to a ticket page, and its own report said: "The only thing actually tested is the build." Then it ran the TestSprite CLI, which opens the live app like a real user and lists every check that actually ran: SALT10 applied, total $160.50 (screenshot). All four features passed, with each check on the record. A claim, and a run you can read. Which one do you merge on?

Stanford put its entire LLM engineering course on YouTube. The homework includes writing FlashAttention in Triton. Its called CS336: Language Modeling from Scratch, taught by Percy Liang and Tatsunori Hashimoto. The Spring 2026 course covers the complete process of building and training language models, including the systems engineering required to make them efficient. The lectures cover: → Tokenization and Transformer architectures → Attention alternatives and Mixture of Experts → GPU architecture, memory and compute → Triton kernels and FlashAttention → Distributed training and parallelism → Scaling laws and inference optimization → Pretraining data pipelines and evaluation → Supervised fine-tuning, RLHF and reasoning RL But the assignments are where this course becomes particularly valuable. Students implement their own tokenizer, Transformer and optimizer before training a small language model. They then profile the implementation, write FlashAttention-2 in Triton, and develop memory-efficient distributed training. Later assignments involve fitting scaling laws, processing raw Common Crawl data, implementing filtering and deduplication, and using supervised fine-tuning and reinforcement learning to train models for mathematical reasoning. Stanford provides minimal scaffolding and even recommends disabling AI autocomplete during the assignments so students actually understand the code they're writing. The full lecture playlist, assignment handouts and lecture materials are publicly available. Some practical exercises require GPU compute. For ML engineers who want to understand the full stack, from training data and attention kernels to inference and post-training, this is an excellent course to work through. YouTube Playlist: youtube.com/playlist?list=PL… Course & Assignments: cs336.stanford.edu/ Lecture Code: github.com/stanford-cs336/le…

We've updated MLCR-AA, our leaderboard for Wisedocs' Medical Long Context Reasoning benchmark, with Claude Sonnet 5.5 and GPT-6 Astra now tied for the #1 score We've updated how MLCR-AA prompts and grades models on @wisedocsai's benchmark, to better reflect how medical record review is done in practice. Models are now prompted as medical record reviewers, and asked to answer each part of a question in one paragraph with their conclusion and the supporting evidence from the record. Answers are graded by a new three-model judge panel, and we reviewed all 60 expert reference answers against their source records, refining 15. Claude Sonnet 5.5 (max with fallback) from @AnthropicAI and GPT-6 Astra (max) from @OpenAI are tied at the top with 41.7%, followed by GPT-6.1 Sol (max) at 38.3%. Qwen3.8 2.4T A95B from @Alibaba_Qwen is the leading open weights model at 30.6%. Key results from the updated MLCR-AA: ➤ Medical record review remains far from solved: an answer passes only when it is both complete and accurate, and no model passes more than 42% of questions and the median model passes under 15% ➤ Anthropic models lead on completeness, reaching the expert's conclusion with the right medical reasoning: Claude models hold the top 18 completeness scores across every model and effort level we tested, led by Claude Sonnet 5.5 (max with fallback) at 64.8%. The best non-Anthropic model is GPT-6 Astra (max) at 42.2% ➤ OpenAI models lead on accuracy, citing the record without exaggeration or hallucination: OpenAI models hold the top 27 accuracy scores, led by GPT-6.1 Sol (max) at 96.7%, but answer fewer questions completely: GPT-6.1 Sol (max) is complete on 40.0% of questions We would like to thank Wisedocs for building MLCR and for their collaboration in bringing it to Artificial Analysis

After 4 years working on Alpine, we’re shutting it down and I’m joining Anthropic. All of Alpine’s code is open sourced and lifetime access customers were refunded. I started working on Alpine in 2022, before the launch of ChatGPT. Everything about the software industry…and the world…has changed since then. Alpine was conceived, designed, and engineered for a world where the only kind of collaboration is human-to-human. If I were to start over and design Alpine today: 1) I’m sure I’d make some very different decisions, 2) I’m not confident I’d make the right decisions! I haven’t experienced what it’s like to work at a large company with agents. I’ve observed first-hand how the rise of agents has changed my startup, but I haven’t seen how they impact companies with hundreds of people. My product design intuitions are influenced by my personal experience, so designing a productivity tool without understanding how AI has fundamentally changed the productivity substrate of our age feels foolish. As with many others, over this year I’ve observed the risks posed by AI become much more salient. I remember meeting EAs pre-ChatGPT and learning about x-risk and their steep AI safety investments, at the time it all felt like sci-fi (plausible, but far out)…now it’s reality. I’m excited to join Anthropic to help the world make the AI transition safely. I’m deeply grateful to everyone who saw Alpine’s vision. I enjoyed every minute working on Alpine thanks to our customers, investors, and employees. I’m sorry it didn’t work out, times are weird.

Most investors are still massively undervaluing data. I was also skeptical when I left a data co for Altos in 2025. Today, I’m all in on data. Every software company will become a data company. Most will lose. Many more will win than we think.

🚨 Your AI app is 9 layers deep. And the LLM is only ONE of them. Most people building with AI focus on the model. But production AI is an entire stack 👇 1️⃣ LLMs — The brain GPT-6, Claude Opus 5.5, Gemini, DeepSeek V4, Qwen3.8, Kimi K3 2️⃣ Inference & access — Where it runs OpenRouter, Together, Fireworks, Ollama 3️⃣ Agent frameworks — The execution loop LangGraph, Vercel AI SDK, Claude Agent SDK, OpenAI Agents SDK 4️⃣ Tools & protocols — How agents interact MCP → tools A2A → agent-to-agent communication Agent Skills → reusable capabilities 5️⃣ Memory — What agents remember Mem0, Zep, Letta 6️⃣ Vector databases — Where knowledge lives Postgres + pgvector, turbopuffer 7️⃣ Embeddings & rerankers — Find the right context Turn text, images and documents into vectors → retrieve → rerank 8️⃣ Data extraction — Make messy data usable Firecrawl, Docling, LlamaParse 9️⃣ Evals & observability — Know if it actually works Trace every call. Measure every output. Catch failures before users do. And here's the important part: The best AI products aren't winning because they picked the “best” LLM. They're winning because they built the right stack around it. ⚠️ What I'd remove from the old AI stack: → Llama → Meta's new models are called Muse → Phi-4 → Amazon Nova → Groq → Cassandra → Haystack → Nomic Embed The AI stack is changing FAST. Save this before you build your next AI product. Which layer would you change or replace? 👇 Follow Harish Kumar for more AI & technology. #AI #GenerativeAI #AIAgents #LLM #MachineLearning

1/8 New paper with Ivan Ilin: "Local Work, Global Progress: A Convergence Theory of AI Agent Collaboration". We all know that teams of LLM agents can tackle big tasks. However, their collaboration protocols lack guarantees. We model collaboration as block optimization and, under suitable assumptions, prove convergence. arxiv.org/abs/2610.28286

A Python pipe buffer was holding camera frames until 64 KB had piled up. That quietly added hundreds of milliseconds to our π0.5 rollouts. New from @castorhat: how we found it, and what changed on the robot once it was gone.