Jeff Bezos bought a superyacht. Mukesh Ambani built Antilia. Zuckerberg built a Hawaii bunker. Elon sold all his mansions, lives in a 375 sq ft prefab box worth $50,000 near a rocket launch site in Texas. And just became the world's first trillionaire. Consumption vs Creation. The builder always wins.

I strive to make my writing unsummarizable, in the sense that it has so little fluff left in it that if you take any words out, as summaries by definition do, you lose a lot of interesting ideas.

Software is eating the world. AI is eating the world Attention is all you need RAG this, RAG that. Agentic this, Agentic that. Context engineering is what you need. RAG is dead. Long live RAG. Harness engineering is what you need. Harness is the backend. What did I miss?

if your bread-and-butter consists solely of: - tuning hyperparams/config files - fitting points on a log-log plot - tweaking a few lines in model.py, transformer.py, optimizer.py, train.py - waiting a week for <= 512 chips to free up and then another week for loss curves to converge it is completely understandable to be stressed about becoming automated into irrelevance within the next year or so. question is, do you wait for that to happen, or do you start doing something differently now?

It took me two weeks to onboarding for my internship at Microsoft, somehow, my collaborators relies too much on the agent to fix everything for them, such that they can't explain how things really works or have some basic understanding in what's blocking me. Though they offered to meet in person and help me "solve the problem together", and when we meet, they prompt their agent while I ackwardly watching with them or they asked me to prompt my agent while they ackwardly watching with me. Why don't I just prompt my agent for everything to save both of our time? Both the learning and human interaction are missing gradually.

Frontier models are powerful advisors. On @harvey's Legal Agent Benchmark, a GLM 5.1 worker using Claude Opus 4.7 as a sparse advisor reached 18/100 all-pass versus 14/100 for Opus alone, at 39% of the cost. More on the harness design, advisor pattern, and training results: fireworks.ai/blog/open-sourc…

Super excited to announce seven new world-class MAI models today. They represent what we consider a new era in AI designed to keep you in control and on the frontier. First is our text foundation model, MAI-Thinking-1, exceptionally strong on reasoning and SWE tasks. - It’s a 35B active parameter MoE with a 256K context window. Independent human raters on Surge prefer it for overall quality in blind side-by-sides versus Sonnet 4.6, and it’s achieved 97% on AIME 2025, the key measure of its general-purpose reasoning abilities. - It's at 53% on SWE Bench Pro, placing it right alongside Opus 4.6 on one of the toughest coding benchmarks. - And since we co-designed our models with our own silicon, MAI-Thinking-1 is optimized on our MAIA 200 chip. Benchmarking head-to-head against the GB200, we see 30% better performance per dollar as well as a 1.4x performance-per-watt gain when running our MAI models on the MAIA 200 end-to-end. Next is MAI-Image-2.5 and its Flash variant. Two super strong models now at #2 on the leaderboards, surpassing the score of Nano Banana 2 on image editing. Last for now is MAI-Code-1-Flash, our new inference efficient coding model, especially tuned for VS Code and GitHub Copilot CLI. - Code-1-Flash achieves 51% on SWE Bench Pro, despite having just 5B parameters, putting it closer to Haiku in size but cheaper in cost. All of this is the foundation for Microsoft Frontier Tuning. It lets you customize our models to create custom, company-specific agents that only you control. You can make our model, your model. Your data. Your agents. Your moat. Early adopters are already seeing a difference. When we tuned our models for McKinsey’s tasks, MAI delivered the highest win rate, outperforming GPT-5.5 on quality, while being 10x lower on cost. Also really excited to be collaborating with the amazing team at Mayo Clinic to jointly train a new frontier AI model for healthcare. Our announcements today mark another milestone on the road to humanist superintelligence. You can learn more and about our other new models in our latest blog: microsoft.ai/news/building-a…

I'm excited to share that I'll be @bespokelabsai this Summer building out some exciting RL environments! Huge thanks to @madiator and @AlexGDimakis for the opportunity. Excited to work with you all! :)

GPT Realtime 2 unlocks some real magic:

Kirkland & Ellis, the world's highest-grossing law firm, is setting aside $500M to build its own AI platform rather than rely on tools available to its rivals (Financial Times) (Visit Techmeme dot com for the link and full context!)

This episode shows me how insanely little Dwarkesh knows about hardware and has made me second guess his intelligence on the other levels of the abstraction stack. Also the dude lecturing is not communicating very well. This whole episode is very clearly an ad for MatX and a poor one at that because the founder clearly has certain gaps in his hardware knowledge

CEOs are the most delusional about AI. Detached from reality.

Anthropic onboarding day: Michael Scott introducing Karpathy like he just signed Wemby in free agency.

Polsia just raised $30M at a $250M valuation. Approaching $10M annual run rate. One Founder AI. Zero employees. Polsia runs companies autonomously. It also ran its own fundraising. I just showed up for signatures.

Now - starts doing blackboard lectures Next - starts hosting in studio audiences for lectures ... - Dwarkesh university?