Daily Brief

Full episode: https://www.youtube.com/watch?v=OS1NZLgKM2c Me on twitter: https://x.com/dwarkesh_sp

Jerry Tworek led reasoning at OpenAI, convinced that scaling reinforcement learning was the path to AGI. Rohan Anil co-led Gemini pre-training and built the Shampoo optimizer. Now they've teamed up at Core Automation on a contrarian premise: the transformer has carried us as far as it can, and the bottleneck to smarter systems is no longer scale — it's the architecture itself. The missing capability is continual learning, models that adapt at test time, which transformers can't do. In-context learning taps out fast (Codex needs compacting after ~20 minutes) and fine-tuning invites catastrophic

Since skills were open sourced, Yogendra Miraje's team at FactSet stopped thinking about shipping features and started thinking about shipping skills. A skill is a capability you hand the agent, and its heart is a short skill.md whose name and description are really routing signals: get them distinct and the agent triggers the right one, blur them and it fires the wrong skill or none at all. He walks through a minimal skill registry, progressive disclosure so the agent only loads what it needs, and trigger words, like asking for a PDF versus an HTML report, that decide which skill runs. The h

Nubank serves 135 million customers, so an AI agent that mishandles a support conversation fails at scale. The talk opens with the result: five agents in production, higher customer satisfaction, and roughly 20 times faster shipping. Shreya Rajpal, CEO of Snowglobe, argues the thing that unlocked that pace was evals, and specifically simulated data standing in for real conversations. Good agent evals are hard because the data is multi turn and stateful, not single turn question and answer, and hand curating it and waiting on production to confirm can take forever. Snowglobe points at the agen

Ask a frontier model how to make a rental property more profitable and it will tell you, with total confidence, to raise a good tenant's rent by 5% to 10%. Udi Menkes calls this the fluent bluff: advice that reads perfectly and is quietly wrong, because the model has fluency but no experience. He points to a Princeton study where models handed a million dollars to run a business barely turned a profit, and to his own tests where off the shelf advice would have cost real customers real money. His argument is that context is not experience, and the fix is grounding. At Intuit he turns real busi

A team ran about a thousand people through a market research survey, then had LLM agents replay the same questions, and the agents matched the humans closely while carrying less noise than the humans did themselves. That is exactly the tell: a synthetic respondent smooths over the messiness that makes a real population real. Ishan Anand walks through where it breaks. Nudge one variable in the prompt template and purchase probability swings, because the model infers latent confounders nobody stated, a little like it is playing improv. So how you ask matters as much as which model you pick, and

A card gets declined and no one, including the customer, can say exactly why. That gray zone is where Divakar Kumar points his agents. In a payments and fraud system, a rule based engine and an ML model already score most transactions cleanly; the hard cases are the ambiguous ones that neither can resolve. His approach adds an agentic layer on top of an existing event sourced architecture rather than replacing it, so the bounded contexts already in the system, transaction, device, and account, become the context the agents reason over. Events flow through change feeds into projections and a s

When a company is about to spend $100 million, Shawn Chan is the person in the room deciding whether it happens, and after fifteen years and roughly 200 investment committees he can spot the flaw fast. Now some of the decks in front of him are AI generated: cleaner formatting, never defensive, and confidently wrong. His frame is the difference between a demo and a memo. A demo is a beautiful confident slice built to impress; a memo is what survives a room whose entire job is to find what breaks. One marketing demo with a single wrong sentence, he reminds you, once erased enormous value the mom

The slowest part of shipping a production finance agent is not the model or the GPUs, it is you, the developer in the loop. Ramana Siddanth Emani's point is that the same agent harnesses you use to build products can automate your own developer loop. Coding agents can multiply how much you ship; run an army of them across separate git worktrees and they clear tasks in parallel, with skills making sure each one uses the right patterns. The tasks come from where they already live, QA reports, Jira tickets, GitHub pull requests, and a sub agent pulls the traces and logs, writes and runs end to e

Humanity compressed the road from the enlightenment to the moon landing into a few hundred years, and Richard Socher's wager is that automating research compresses it again. He frames it through open ended evolution and Popper: science advances by trying things, finding the shortcomings, and fixing them, and an agent swarm can run that loop across medicine, economics, astrophysics, and more without any single person bottlenecking a field. He calls the goal a Eureka machine, and argues that rethinking the tools around it, web search that returns usable context instead of ten blue links, browser