AI Builders Digest

OpenAI pauses new $200 Pro subscriptions as Astra demand outruns capacity

New signups to the $200 Pro plan are on hold because those users put the most strain on the systems, and pausing them was framed as the smallest step that keeps the broadest access intact. Existing accounts, all other plans, and the API are untouched, with more capacity being added as fast as possible. Separately, the infrastructure running under ChatGPT Work is now exposed as an API for scaled agents on demand, pitched as something you can start using in under a minute.

• #products
• #agents

Levie's field notes: enterprises rip out AI vendors the moment one disappoints

After meeting a couple dozen technology leaders across banking, media, insurance, and consulting, the sharpest pattern was ruthlessness about architecture: companies have swapped systems multiple times in the past year or two, and nobody waits around for a vendor to get it right. Most are running multiple frontier models because standardizing proved too hard, though the dollars still concentrate on a few vendors and open weights remain in infancy for lack of domestic frontier options. Security and agent identity dominate the anxiety, and the real ROI shows up only when companies reengineer the workflow itself rather than layering agents onto the existing one. Evals barely registered, with only a few of the couple dozen customers raising it, which he calls a huge open opportunity. Box is also partnering more deeply with OpenAI so enterprise content can be worked with securely inside ChatGPT.

• #agents
• #enterprise
• #evals

Hold Claude's production code to a higher bar than a human's, says Claude Code's Cherny

Prototypes and throwaway code can be treated as a total black box when the blast radius is low, but production code written by Claude should clear a higher bar than human-written code. At Anthropic that means lint rules, tests, Claude-driven end to end tests, Claude-powered fuzzers running daily, automated code reviews, security reviews, and automated refactoring, without which you get a mess that is hard to maintain. When output falls short his fixes are concrete: move to a frontier model, raise effort to high or xhigh, and invest in CLAUDE.md and skills rather than accepting the debt. He also flagged the latest Threat Intelligence report as terrifying and important, noting that capability is dual use, since a model that codes well can hack critical infrastructure and one that assists biology research can help engineer a pandemic.

• #agents
• #security
• #products

Richard Socher raised $650M on the bet that AI research automates itself first

Recursive starts with AI for AI, building a system with the equivalent of 50,000 PhDs of research capability before turning it loose on physics, chemistry, and biology. The reason scientific progress slowed is that a body of knowledge became a labyrinth, 34,000 journals that might as well have no trespassing signs, and no one can span enough subfields to recombine them anymore. His prediction is blunt: anything you can simulate, AI will solve, which is why math and programming fall first and the natural sciences lag until we can simulate a cell. Hallucination gets reframed as a feature, since generating protein sequences outside the training distribution is exactly the point, and he is explicitly not a believer in a hard takeoff because clinical trials still take the years that physics demands. The jobs question he reduces to demand elasticity: illustrators lost because the world does not need a thousand times more illustrations, coders won because it does need a thousand times more software. His Eureka Machine rests on four pillars, human knowledge in LLMs, scientific measurement, simulation, and robotic labs collecting real data, with an agent swarm on top, and he thinks self-driving labs are two to three years from being the right investment.

• #research
• #agents
• #funding

Vercel made its global CDN metadata store 91% faster at p99

Roughly 10 million deployments land on Vercel every day, 2.35 billion to date, all co-existing and routable on the CDN at once. The global metadata store that syncs rollbacks, config changes, and route additions within hundreds of milliseconds worldwide just got 91% faster at p99, which also sped up the build to deploy pipeline. The work happened while the system was under immense pressure from growth in agentic deployments, and the CDN engineering team published the internals.

• #infrastructure
• #agents

Grade the agent's trajectory, not just its answer

Two agent runs can land on the same answer while one searches the right sources, retrieves the right document, and makes four clean tool calls, and the other makes 17 calls, repeats the same search three times, and recovers from two errors along the way. It is obvious which is better, and a result-only eval cannot tell them apart. The prescription is to define the whole workflow, name the tasks at each step, decide whether each step gets its own eval or a slice of a bigger one, and set both median and hard tasks. Then invert the reading order: study the steps first, final results second.

• #evals
• #agents

Masad: cybersecurity is the real AI risk, extinction is not remotely one

Plenty about AI worries him, cybersecurity especially, but extinction risk in the literal sense of 100% of humans dying does not make his list at all. It is a pointed split from the doom framing common in his corner of the industry, separating concrete near-term damage from civilizational endings.

• #policy
• #security

Gemini ships on Windows

Gemini is now available on Windows, moving Google's assistant onto the desktop where most enterprise work actually happens rather than leaving it in the browser and on mobile.

• #products

Every hot tranched seed round somehow ends up at a $300M valuation

Three things are simultaneously true in early stage venture right now: everyone wants to raise a $50 million seed, everyone believes they will hit $30 million ARR next year, and every hot tranched seed magically converges on a roughly $300 million valuation. The pattern is stated as observation rather than complaint, which makes it land harder. He also published a post the same day going from voice memo in the car to publish in a single day, admitting it lacks his usual polish.

• #funding

Dreambeans goes free for every US user and connects to Gemini

Dreambeans is now available to all US users 18 and over on iOS and Android at no charge, with no subscription required. You can connect the Gemini app to it, letting Dreambeans build on the nuance and understanding from your chats to surface more personalized daily stories.

• #products

One prompt that makes Claude interview you for its own memory

Ask Claude to interview you in depth about the relevant parts of your life it does not know yet, using free text or the askuserquestion tool where multiple choice fits, and to save all of it to memory. It turns context-building from something you write by hand into something the model elicits.

• #products
• #agents

Duplicating logic stopped being painful, abstractions still are

The cost structure of code has flipped. When a model writes the duplicate for you, repetition is cheap, but a bad abstraction still has to be understood, maintained, and eventually unwound by a human. That inverts a decade of DRY instinct about when to factor something out.

• #engineering

How much GDP would you spend on a machine that only cures disease?

Posed as a thought experiment with an immediate answer: if a machine could do exactly one thing, find cures for our most pressing diseases, the share of GDP worth devoting to it is very high. The point is that this is no longer hypothetical, it is the world we already live in, which reframes compute spending as a public health question rather than a tech capex one.

• #policy
• #research

Fable 5.1 Build Days run in cities worldwide through September 25

Community-hosted buildathons start this week and run September 11 through 25 in cities around the world. Bring a problem, an idea, or nothing at all and see what gets built.

• #community
• #products

Normies live in Google and Doordash all day, so AI is still early

The counter to AI adoption pessimism is that ordinary people already use Google, Instagram, Zillow, and Doordash constantly, every day, without thinking about it. Measured against that bar, current AI usage is nowhere near saturation. He also coined the day's best euphemism for a discounted take-private, calling it an Italian goodbye rather than a private equity acquisition far below peak valuation.

• #products
• #funding