Daily Brief

Preparing this talk got Corey Gallon a warning from OpenAI that his account faced a ban for cyber abuse with a web browser. The offending work was an agent clearing Cloudflare Turnstile, two image captchas, and finally reCAPTCHA v2 with no human in the loop. His premise fits on one slide: a browser driven through the Chrome DevTools Protocol is just a meat bag with a mouse, because the agent's clicks and keystrokes travel the same path inside Chrome that yours do. Chrome stamps every event as trusted or untrusted, which is why a synthetic JavaScript click that works fine in Outlook gets silent

Asked to file an expense, the agent gets signed out mid task, reasons that it can infer the password, guesses twice, and locks the account. In a second run it clicks a sponsored button styled like the real submit button, lands on a different site, and begins typing personal details into it. Both are real trajectories from early browser training runs at the Amazon AGI Lab, and Gaurav Mishra's summary is that RL worked while the world was a game, and IRL starts when the game fights back. The talk catalogues what a reward function meets on contact with a real login screen. Observability is parti

Just past 15,000 queries, renting context stopped being the cheaper option. Omer Primor gets that number from a small experiment he is careful to call a test rather than a benchmark: enrich one company across 25 fields, run it 100 times against this event's sponsors, and compare AI search products, context as a service vendors, and a scraper pipeline built in roughly a day. Pricing a week of setup at $5,000, the build it yourself path crossed over a little above 15,000 entities, and he expects the real crossover sits lower than most teams assume. The argument underneath is about frequency rat

Minutes into a call to demo a search API rebuilt to answer in under a second, the system got blocked, badly, in front of the client. Patricija Žemaitytė treats that as the useful distinction: something that works in development, something that passes tests, and something that survives reality are three different systems. The rebuild had no trick to it. Browsers are slow, expensive, and incompatible with low latency, and they were unavoidable, so the team went hunting for time across layouts, parsers, sessions, and proxies until the seconds were gone. It averages 550 milliseconds now, against a

Full episode: https://www.youtube.com/watch?v=-RXD4bTuFTo Me on twitter: https://x.com/dwarkesh_sp

In this presentation from Startup School Paris, Y Combinator General Partner Tom Blomfield shares findings from his research on how companies are being built with AI infrastructure placed in from the start. Apply to Y Combinator: https://www.ycombinator.com/apply Work at a startup: https://www.ycombinator.com/jobs Chapters: 00:00 — Intro 01:53 — Why Roman Legions Built Your Org Chart 03:56 — Humans as the Bottleneck 05:28 — What a Real AI Loop Looks Like 07:51 — The Data Agent That Changed Everything 08:39 — The Self-Improving System 10:32 — Office Hours → Living User Manual 12:22 — The AI

Exo is a systems approach to recursive self improvement. In short, it's a complete AI agent harness (supporting tools, tasks, integrations, etc. similar to OpenClaw, Pi or Hermes), with the crucial difference that it has full visibility into both its code and runtime logs. This allows Exo to incrementally improve every aspect of itself, clone itself, and even manage a lineage of clones. While most agents can do some form of self improvement, such as updating memory or creating skills, Exo is fully recursive in that it can clone or operate on any aspect of itself, from prompts, to memory, tool