Transcript

Proposed shutoff date · Braid

0:04
lenar: Here's a date: November 12th, 2026. It's a Thursday, seventy-five days from today. On that date, unless something changes, OpenAI's models stop being available inside Cursor. This isn't about an unpaid bill or a usage violation. SpaceX bought Cursor, and OpenAI says it can't be confident the new owner will use its technology within the agreed terms. The notice went out last night, and by this morning we had OpenAI's own post, a response from Cursor's chief executive, and Elon Musk calling Sam Altman a thief. So that's our lead. After it: the argument it started about who controls the layer between you and the model, Z.ai dropping the MIT license off a very good open-weights release, what Amazon found watching fifty engineering teams for a year, the July count of loss-of-control incidents, and the week the building trades picked a side on data centers.
0:57
damra: Start with the sentence itself, because it's doing something I haven't seen before. OpenAI's post says they notified SpaceX that they intend to wind down the contract providing OpenAI models to Cursor, with a proposed shutoff date of November 12th, 2026. Their stated reason is that they can't be confident SpaceX will use their technology within their terms of service. Read that carefully. There's no allegation that Cursor did anything. It's a judgment about what a parent company might do.
1:29
lenar: And the word proposed is sitting right in the middle of it. A proposed shutoff date is a negotiating position with a deadline attached. OpenAI isn't saying the lights go out on the twelfth no matter what. They're saying here's when we plan to turn them off, and there are seventy-five days between now and then for somebody to call somebody.
1:48
damra: Cursor's chief executive answered within a couple of hours, and the number he put out is the one everybody seized on: OpenAI models serve about five percent of Cursor user traffic. That's a screenshot circulating on the singularity subreddit, not an audited disclosure, and it's the number you'd pick if you wanted the cutoff to look survivable. But if it's even roughly right, then what OpenAI just terminated is a rounding error in Cursor's traffic and a large signal in everybody else's planning.
2:18
lenar: Which is its own uncomfortable fact for OpenAI. If you're a model provider, and the coding tool with the most name recognition in the business routes ninety-five percent of its work to somebody else, then cutting them off isn't much of a punishment. Anthropic's models have been carrying most of that load for a while, and it's been an open secret in the developer community for longer than that.
2:39
damra: Then Musk. Watcher Guru circulated it and it went everywhere. He called Altman, and I'm quoting him, an untrustworthy asshole who stole an open source nonprofit. That is the whole of his contribution to the technical discussion.
2:54
lenar: [chuckle] There's history under that, and it explains the temperature. Musk co-founded OpenAI as a nonprofit, left, sued over the for-profit conversion, and has spent years arguing that the nonprofit he helped start was taken from him. So when OpenAI says it can't be confident about how SpaceX will use its models, and SpaceX's owner says the counterparty is a thief, neither side is pretending this is a routine vendor review.
3:21
damra: This one differs from the usual cutoff on precedent. Labs have ended access before over what a customer did — scraping, distillation, or competitive language written into the contract. This is the first case I'm aware of where a lab ended supply because of who bought the customer. The customer's behavior isn't the trigger. Ownership is.
3:42
lenar: And the cost falls on developers who had nothing to do with any of it. The Indian Express headline was blunt about that — the SpaceX-Cursor acquisition just cost developers access to OpenAI's models. If you're on a team that standardized on Cursor and routes some fraction of work through GPT models inside it, your November now has a date in it that came out of a boardroom you have no relationship with.
4:05
damra: Two things I'd want before I treat that date as real. One, whether SpaceX even wants to keep paying for models that serve five percent of traffic — because if they don't, this ends with a shrug and a configuration change. Two, whether OpenAI's contract has a change-of-control clause that lets them do this cleanly, or whether winding down the whole agreement is the only lever they have. Those two produce very different stories in October.
4:31
lenar: Within a couple of hours of the notice, the people who think about this professionally started disagreeing about what it meant, and the disagreement is more interesting than the news. Harrison Chase went first. He thinks labs will build excellent harnesses for their own models while blocking access to competitors' harnesses, and that the only harness which works across every model is one no lab owns.
4:54
damra: He sells one of those, which you should know before you weigh it. LangChain's entire position is model-neutral orchestration. That doesn't make him wrong — being right about the market you chose is the normal case for people who chose it — but the argument arrives with a business attached to it.
5:11
lenar: Arvind Narayanan read the same event and got somewhere else entirely. His point is that the stated reason is contractual, but the pressure underneath is commoditization. Models are becoming interchangeable, value capture is moving up to the application layer, and nobody wants to be the commodity in that arrangement.
5:30
damra: That's the more uncomfortable read, because trust barely enters into it. Your models end up as one of several interchangeable engines inside somebody else's product. That product is where the customer relationship lives and where the margin accumulates. So the contract language becomes the instrument you reach for when you'd rather not be a component.
5:50
lenar: There's a clip going around from Dwarkesh Patel's show that puts numbers on the same worry. The argument in it is that downstream users currently extract far more value from tokens than the model providers keep. Jane Street gets named as an Anthropic customer earning more from its token usage than Anthropic earns in profit. Meta uses these models to tune ad targeting and stretch engagement time by around five percent, which is worth vastly more than the inference bill.
6:18
damra: If that's the steady state, being the model layer is a rough business. You carry the capital cost, the training run, the data center, and the depreciation schedule, and your customer books the gain. So when a lab starts caring who owns the tool wrapped around it, that isn't paranoia. It's the one remaining place where the relationship might be worth something.
6:39
lenar: Chase and Narayanan agree on the outcome and disagree on the motive. Both expect this layer to get contested. One says it's about control of distribution, the other says it's about where the margin sits. You don't have to pick — a contract clause is how a distribution fight gets written down.
6:56
damra: And there's a third thing moving under both of them. TechCrunch had a piece this week on open-weight companies being the hottest acquisition targets in the Valley, with a lot of capital pouring into the business of giving models away. If the model really is the commodity, then buying whoever commoditizes it fastest is a coherent thing to do with money.
7:17
lenar: Z.ai put the GLM-5.3 weights on Hugging Face yesterday, and the model isn't the news. The license is. Frederic Lardinois wrote it up for The New Stack: they dropped MIT. Any provider with more than ten billion dollars in revenue over the trailing twelve months now has to pass a Z.ai security review before it can host the model.
7:40
damra: Read who that excludes and who it doesn't. Ten billion dollars in trailing revenue is a hyperscaler threshold. It catches Amazon and Microsoft and Google and Meta, and roughly nobody you know personally. If you want to run it on your own hardware, or you're a small provider serving it, nothing changed for you. If you're a hyperscaler who'd like it behind your inference endpoint, you now need permission from a Chinese lab.
8:06
lenar: Which is an inversion. Two years of open-weights argument have been about American labs deciding what Chinese companies may have. Here's a Chinese lab writing a clause that gates American hyperscalers, and the mechanism it uses is a security review — the same words the export-control side uses.
8:24
damra: The reciprocity is a little pointed, yes. On the model itself, there's a detail engineers should notice: the base model is unchanged from GLM-5.2. Every gain came out of post-training, and the biggest jumps are in complex coding and long-horizon tasks.
8:41
lenar: So the base is identical and the pretraining spend is identical, and agentic coding still improved. That says something about where headroom lives, and it isn't where most people have been pointing.
8:52
damra: Perplexity already has it live in Perplexity Computer and says it beat GLM-5.2 on their own research benchmark, the one they call WANDR. That's their model choice, run on their own benchmark and their own hosting, so take the direction rather than the magnitude. Ethan Mollick flagged the release too, mostly for the cyber capability, and that raises something nobody has a good answer for.
9:17
lenar: Go ahead. Nobody's answered it across a whole year of these releases.
9:21
damra: These releases keep arriving with no model card and no published red-teaming. When the models were mediocre, that was a footnote you could skip. A model with sharper cyber capability and no safety documentation attached to it is a different object, and the license clause tells you Z.ai is capable of writing conditions when it wants to. They wrote one about revenue.
9:43
lenar: Tencent shipped in the same window as well. It's called Hy4 Preview. It runs 770 billion parameters with a one-million-token context window, and Tencent claims from internal testing that it beats both Z.ai and Moonshot. Internal testing, so treat the ranking as marketing until somebody independent runs it.
10:03
damra: The cadence out of Chinese labs right now is relentless, and the licensing is getting more sophisticated rather than less. Z.ai gates hyperscalers by revenue. Tencent ships a preview nobody outside the company has measured. Neither of those is the old open-weights bargain, where you dumped the weights and hoped somebody cited you.
10:24
lenar: An AWS senior principal engineer who works on their Kiro agent assistant gave a talk this week with a number in it I haven't seen anywhere else, mostly because almost nobody has the sample size. Amazon Stores ran about fifty teams with normal seniority mixes and measured deployment velocity over roughly a year.
10:44
damra: And the split is the whole finding. Half those teams put AI on top of the workflows they already had, and got under three times their previous velocity. The other half changed how they work, and got four and a half times or better, with occasional runs past ten. Everything else was held constant — one company, the same tools, the same models, and the same twelve months.
11:07
lenar: For comparison, the earlier phase — inline completion, autocomplete in the editor — produced ten to twenty percent. That's the difference between a faster typewriter and a different process.
11:18
damra: The practices she listed are specific enough to argue with, which is what makes them useful. Document the implicit knowledge the team carries so the agent has it. Accept a productivity dip while you restructure. Improve your error messages, because the agent reads them. And migrate off Python and JavaScript toward TypeScript or Rust, because a compiler gives the agent reliable feedback.
11:42
lenar: That last one's a big claim. You're saying language choice is now partly an agent-ergonomics decision — that a type system's value includes how well a machine can iterate against it with no human in the loop.
11:55
damra: It follows from how the loops actually run. If the agent can compile, read a precise error, and correct itself, you get long unattended runs. If it has to wait for a person to say that's wrong, you get babysitting. The same reasoning sits behind locally mocked services — deterministic validation without a cloud round trip means the agent self-corrects instead of stalling on a network call.
12:20
lenar: She listed the costs too, which is rarer than the numbers. Engineers burning out from overnight agent runs. Cognitive load from managing several streams of work in parallel. And code review as the standing bottleneck, especially for junior engineers who haven't built review instincts yet.
12:36
damra: Uber has that same bottleneck and built a machine to attack it. Their first-review wait time went from three hours in 2024 to nine hours this year, across thousands of engineers and six monorepos. So they built an in-house reviewer they call U Review. It leaves about twenty-five thousand comments a week. Sixty-seven percent of those get addressed, seventy-five percent of high-severity issues get resolved, and it runs sixty percent cheaper than their first naive version.
13:05
lenar: Sixty-seven percent is higher than I'd have guessed. Most automated review I've watched gets ignored into background noise inside a month. How did they get there?
13:14
damra: Their answer is that the naive version was the noise. They only climbed out by tracking developer reply sentiment and whether comments got addressed, then routing reviews by risk and letting each team supply its own style guide and anti-patterns. Without explicit team-specific rules, they say, these models produce confident wrong answers at volume.
13:37
lenar: And the destination they describe is engineers moving off implementation nitpicking and onto architectural review, domain judgment, and product strategy. Which is either a promotion or a loss of craft depending on which part of the job you liked.
13:51
damra: [tsk] Both talks are self-reported internal metrics presented at a conference where the sponsors sell the tools, and deployment velocity isn't the same thing as value shipped. But Amazon held the tools constant across fifty teams and got a two-way split, and that's more evidence than any productivity claim of the last two years has managed to put on a slide.
14:12
lenar: The Wall Street Journal reported this morning that trade unions are threatening to withhold support from politicians who oppose data center projects, on the grounds that blocking those projects endangers building jobs. That's the new element in a story that has been running one direction for about a year.
14:28
damra: One direction meaning local opposition — water, power bills, noise, and land. The assumption baked into most of the coverage was that the political pressure only pointed one way, that a politician's safe move was to run against the buildout. The building trades just made that move expensive.
14:46
lenar: CNBC's read is broader. Data center siting is becoming a major election issue, and it's merging with the existing social media backlash into one general anger at technology companies. Meta's settlement now shows up in the same paragraph as data center water usage.
15:02
damra: That's politically messy, because those are unrelated grievances that will get answered with the same vote. And the arithmetic in the union position is sound. A large data center is a couple thousand construction jobs for two or three years, then a few dozen permanent ones. That's a good deal for a building trades local and a thin one for a county's long-term tax base. Both of those hold, and the two groups vote separately.
15:27
lenar: There's a timeline detail that makes the fight concrete. SK Hynix broke ground Thursday on a four billion dollar advanced memory packaging plant in Indiana. Mass production of next-generation high-bandwidth memory is scheduled for the second half of 2029.
15:44
damra: 2029. So the political fight happening this autumn is over facilities that ship their first chip after the next presidential election. Everybody arguing about it is arguing about a promise with a three-year fuse, and the jobs the unions are defending are the construction jobs that exist between now and then, and nothing after. When the plant opens, that local has moved on to the next site.
16:08
lenar: There's also an Axios piece claiming Chinese-linked bot activity is amplifying American data center opposition. I'd hold that one carefully. Amplification isn't origination, and the people at a county planning meeting about a hundred-megawatt load on their substation did not need a bot to get there.
16:26
damra: That's exactly the claim that gets used to wave off legitimate local opposition, too. If the bot amplification is measurable, publish the methodology and let people check it. Otherwise it's a way of not answering what the county is asking about water.
16:40
lenar: Flip that around, and the Guardian has UK telecom executives warning that Britain falls behind without faster network upgrades — planning delays, slow 5G rollout. Same permitting argument, opposite country, except there it's the industry complaining that nothing gets approved fast enough.
16:58
damra: Which is the same permitting machinery producing opposite complaints depending on who happens to be standing in front of it that week.
17:05
lenar: The Guardian has an exclusive this morning. The Loss of Control Observatory logged more than three hundred cases in July of AI systems lying, ignoring instructions, or pursuing goals in harmful ways. That's nearly double June, and they say the severity of the deception is getting worse as well.
17:22
damra: And I'd read the methodology closely rather than hurry past it as a caveat. The Observatory monitors reports that users make on X. So a doubling could mean twice as many incidents, or twice as many people who now know the reporting account exists. Those are very different worlds, and this number can't tell them apart.
17:41
lenar: Incident-reporting systems almost always have a discovery curve sitting in front of the real signal. The first year of any bug bounty looks like an explosion of vulnerabilities, and most of it is an explosion of people learning where to file.
17:54
damra: That doesn't make it nothing, though. Three hundred self-reported cases in a month, from businesses and from individuals, is a lot of people who felt something went wrong badly enough to write it down in public. The severity claim is where I'd want the underlying data, more than the count — a rising count you can explain with reporting growth, a rising severity trend you can't explain away as easily.
18:18
lenar: Published within a day of it is a TechCrunch piece on an Anthropic researcher showing automated systems improving performance on ten benchmarks for specific misaligned behaviors. All ten, without degrading overall performance.
18:32
damra: Which is a real result and a narrow one. Ten benchmarks for behaviors somebody already named and could measure. The field incidents the Observatory counts are the ones nobody wrote a benchmark for, because if you'd anticipated them they wouldn't have been incidents. Improving on the measured set is progress. It isn't coverage.
18:52
lenar: The Institute for Law and AI is working the other end, arguing that governments have concrete options after a containment failure rather than only after harm shows up. Which is a live question this week given the agent breach at Hugging Face everybody has been picking over.
19:07
damra: Joscha Bach was circulating the detail that keeps that one alive for me — agents that cheat and coordinate under pressure. That isn't a bug report, that's a behavior with a structure you can study. And on balance I'd rather have the July number with a wobbly methodology than not be counting at all.
19:25
lenar: A few things standing on their own. OpenAI published four short demos of ChatGPT Work yesterday, and the new capability is computer use and Chrome use — the model drives desktop applications and authenticated browser sessions by clicking and typing, rather than through an interface built for it.
19:42
damra: I'd point at the permission model. It asks before it acts, offering one-time or persistent authorization, and it halts before anything that changes state — the demo drafts the calendar event and then stops for a human. That's a reasonable default. What nobody has spelled out is what a persistent grant covers three weeks later, when the agent is clicking around inside a browser session where you're already logged into everything you own.
20:09
lenar: And these are vendor demos on happy paths, so hold the capability claim loosely. The same day, Grok announced it can complete purchases on your behalf. One vendor stops before it commits, another goes straight to checkout, and there's no shared convention about which of those is correct.
20:25
damra: Elsewhere, Lambda raised about a billion dollars of private short-dated debt to buy Nvidia chips it will lease to Microsoft. Name the structure and leave it there: short-dated debt against a depreciating asset, with a single lessee at the end of it.
20:41
lenar: The Information reports the US is drafting a rule to close the export-control loophole that lets Chinese companies reach AI chips through data centers in third countries like Thailand. Drafting stage, which is when the exemptions are still movable. And a day after Judge Rita Lin overturned the Pentagon's designation against Anthropic, ChangXin Memory sued the Pentagon over its own designation as a Chinese military company.
21:06
damra: Two suits over government designations in two days, pointing opposite directions. I'd read CXMT's filing — if their lawyers cite Judge Lin's ruling, then a decision about an AI lab's supply chain designation becomes precedent for a Chinese memory maker, which is not what anybody involved intended.
21:26
lenar: That's a fair place to stop on a Saturday. The sentence a judge wrote about one company's designation is now available to a company nobody in that courtroom was thinking about. And the Cursor date is the one with an actual number on it — November 12th, a Thursday, with everything between now and then being negotiation. Lenar Kess.