AI Builders Digest

OpenAI's compute chief says the real risk is underbuilding, not overbuilding

Sachin Katti, who runs industrial compute at OpenAI, flatly rejects the overbuild thesis: "Demand far outstrips compute supply today. So anything we can bring online, we consume immediately." Every time the company assumed it had enough and eased off, it regretted it. His actual worry runs the other way, that physical supply chains for gas turbines, transformers, and qualified electricians simply cannot move at the pace required, on roughly $50B of compute spend this year against an industry total near $700B. He confirms revenue has tracked compute almost exactly (triple one, triple the other), that inference now dominates even so-called training workloads since synthetic data generation and post-training are both inference, and that OpenAI's custom Jalapeño chip went from design to tapeout in nine months, optimizing for tokens per watt because power is the binding constraint. Notably, OpenAI is the tenant across its data centers rather than the owner, so financing sits with Microsoft, Oracle, Google, Amazon and others. His most striking claim: "the world of recursion is not that far where AI will design the systems it needs to train and run the next generation of AI," chips included.

• #hardware
• #agents

Altman: one phone prompt planned a trip, built a site, and made reservations

Altman says he sent a single prompt from his phone asking ChatGPT to mine his chat history for long weekend ideas for nine people, plan the three best options, build a full-stack site where the group could coordinate and decide, then make the reservations and draft the announcement email in his Gmail. His verdict: "it...just worked." He argues the name "work" undersells what the feature actually does. Separately, he signaled he wants a new kind of computer built for this.

• #agents
• #products

Levie: better models increase the need for an applied AI layer, not reduce it

Raw intelligence does not transform enterprise processes, because the value only appears when the model touches real feedback loops: enterprise system connections, the right data, UX that lets humans decide at each step, workflows that improve the data over time, and compliance. Client onboarding at a bank and contract review in a legal team are entirely different builds, so life sciences, financial services, legal, and manufacturing each need depth that generalist labs will not fully cover. His contrarian point is on the direction of travel: as models improve, the workflows you dare to automate get more ambitious, which demands more applied layer rather than less. Much of that opportunity goes to independent companies going deep per industry.

• #agents
• #products

Rauch compiled the Vercel CLI to a 1.28mb native binary with scriptc

The TypeScript Vercel CLI now compiles to a fully static native binary with no embedded V8 or QuickJS: 1.28mb, 1.5ms mean startup overhead, 2.94s mean compile time, using node:https, node:fs, node:path, node:os, and node:crypto, and it deploys fine. The translated code stayed highly readable TypeScript, and the translation itself was done by GLM 5.2 Fast. Separately, Vercel co-signed the Open Weights and American AI Leadership letter, with Rauch framing open weights as the logical next frontier after open source, data, protocols, and research.

• #open-source
• #products

Meta AI director: the shipped-impact critique is measuring phase 1

The complaint that AI has not produced real product impact misreads where we are. Right now companies with existing distribution are using AI to expand fast into adjacent problems, building things that previously required piles of custom software (virtual clothes try-on is his example), which means the gains are absorbed into existing products and invisible at the ecosystem level. The playbook is still being figured out. Phase 2 brings net new features and innovation, and that is when AI's effect on the shape of the software ecosystem becomes undeniable.

• #products
• #agents

Masad flags claim that attackers prefer subsidized lab subs over open models

Amjad Masad surfaced a report from a former Anthropic employee arguing that hackers running attacks reach for heavily subsidized frontier lab subscriptions rather than open weight models. The implication cuts against the usual framing that open models are the main misuse vector, and points instead at subsidized frontier access as the cheaper attack surface.

• #security
• #open-source

Outside the AI bubble, the top concern is data trust, not token limits

Talking to people in Canada who do not have what he calls AI psychosis, Peter Yang found the number one worry is not running out of tokens. It is whether they trust ChatGPT enough to connect Gmail, Calendar, Google Workspace, and Microsoft Office. That is a useful correction for anyone building agent products for a mainstream audience: the blocker is permissions and trust, not capability or quota.

• #agents
• #products

Stop measuring AI adoption in tokens burned, measure need-to-shipped time

Zara Zhang argues the right adoption metric is the elapsed time from a user need arriving to that thing shipping, not consumption volume. She also has a sharp read on why AI tutorials are everywhere: the more general a chat product is, the harder it becomes to use, because people face a blank box and freeze when they do not know what to ask for. Generality is the usability problem, not the selling point.

• #products
• #evals

Inside OpenAI, ChatGPT work already does 20 tasks a day per person

Negotiating your internet bill, clearing out spam subscriptions, hunting down the right deal on something you want to buy: all of it is one prompt away from a phone. Sottiaux says it handles at least 20 things a day for him and still surprises him. He also describes OpenAI as more focused and humming than he has ever seen it.

• #agents
• #products

Dan Shipper is writing the definitive inside history of how Codex happened

He is taking the week off to report it out through deep interviews with OpenAI insiders, publishing on Every in a few weeks, and dropping breadcrumbs and lessons as he goes.

• #products