What the people actually building AI said today. One page — a 3-min read.
Evals were the through line today: two separate builders argued that what enterprises can measure, not what models can do, is the real ceiling on AI adoption. OpenAI's Codex team went public with the specific bugs behind its rate limit complaints and promised a full usage reset. And a resurfaced conversation with Microsoft's CTO makes the case that the gap between model capability and shipped product is now bigger than the gap between models.
OpenAI names three Codex bugs behind rate limit burn, resets all paid usage
Codex was quietly eating usage in three places: images in long sessions with multiple compactions, high p95 usage for Computer History, and a conversation title generator that cost more than intended. Fixes ship tomorrow alongside a full usage reset for every paid subscription. A separate and unrelated efficiency approach is queued for next week.
Kevin Scott: the bottleneck is now capability overhang, not scaling
Model reasoning has run ahead of what products actually deliver, and closing that gap is the industry's real work right now. The fix is an agentic web with open protocols doing what HTTP and HTML did for the internet, MCP and NL web being the first credible glimmers, plus agent identity and entitlement systems so an agent can ask for exactly the permissions a task needs. Expect the interaction model to shift from synchronous prompting to agents that go away, make many calls, and come back much later. On the craft objection to agentic coding: "If you value the process more than you value the outcome, sometimes you'll make different decisions than people who value the outcome more than the process."
The real eval market is one company's workflows, not model leaderboards
Public model evals only tell you the shape of general progress and how models rank against each other. The far larger space is evals on the specific workflows enterprises run, down to the quirks of an individual company. You cannot automate what you cannot measure progress on, and enterprises will not deploy on vibes.
Hill climbing on evals: pick one dimension that matters and grind it
Part six of a running eval series argues the work reduces to better harnesses and model selection through prompt engineering, context engineering, memory, post training, and plain deterministic code. Your failure mode taxonomy tells you where to aim: if tool calling is the top failure, you probably stuffed 20 tools into context when the task needed 3 to 5. On cost, launch with the best model, get quality high, then hill climb toward the same quality on something smaller and cheaper once you know users love it.
AI gives a solo builder 10x, and the same person inside a big company 20%
The leverage AI grants an individual mostly evaporates inside a large organization, where the same talented person gains at most 20 percent and sometimes goes backward. That asymmetry is why more strong people are leaving big companies, with top labs like OpenAI and Anthropic as the likely exceptions. Also worth pinning up: everyone who is actually ahead on AI thinks they are behind.
Use a coding agent to audit which apps still hold your Google data
Open your Google connections page in Chrome, point Codex or Claude Code at the open tab, then have it revoke the ones you no longer want. Doing this disconnected roughly half of the apps that still had access. Separately, the Instinct team shipped external data deletion after public pressure, so 36 Gmail records could finally be removed from the Data Privacy section.
Ragebaiting investors then sending SAFE docs is not a fundraising strategy
A cold DM that should not have cleared the bar became a public spectacle, and the takeaway aimed at young founders is blunt: provoking investors and then asking them to wire money is not how investing works. Doxxing the sender was not the response, but choosing who you take advice from is the part that matters. Talented people with real substance keep burning themselves on these games.
Source data comes from the open-source project follow-builders by zarazhangrui, released under the MIT license. Summaries are generated by an LLM from that project's public feeds, and the summarization prompts are adapted from it. Every item above links to its original source.
Summaries generated automatically. Read the original before relying on any claim.