granzymes

We call things AI until they start working. See also: robots (your washing machine is a robot, but it works so you don’t think of it that way).

The “overlap” is that all machine learning is AI, but not all AI is machine learning.

The proposal would also involve other US AI companies giving a similar stake to the government, the FT reported, although it is not clear yet whether companies such as Anthropic, Google and Meta would agree to the plan.

I can't see Google or Meta shareholders agreeing to this? That said, Google, Meta, and SpaceX are all still founder-controlled using supervoting shares.

To be clear, that is not "Google's memo". It's a memo by a guy who happened to work at Google. There is a diversity of opinions at a company that employs 180,000 people.

HN is of course not a monolith, but from my recollection most of the frustration and criticism w/r/t Mozilla is about its product strategy and executive team, not its corporate structure. Some other examples of nonprofits with for-profit subsidiaries include National Geographic, the AARP, and most research universities.

On Altman, the trial showed that Altman does not have any equity in the for-profit. He does have some indirect exposure through his investments in YC, since YC has a small position in OpenAI.

I'm curious why you think that creating a for-profit subsidiary is objectionable, since it is extremely common for large nonprofits. A good example for this forum would be Mozilla, but many more were mentioned during the trial.

Also curious what conflicts of interest you have in mind.

Elon's purchase offer was in 2025, after the success of ChatGPT showed that OpenAI's IP (much if not most of it developed after 2019) could be commercially valuable. I think it is also debatable whether Elon's purchase offer was in good faith.

It was not clear in 2019 that OpenAI's IP would ultimately be worth billions. That was well before the current AI boom.

I for one will only buy Chinese RAM from now on, no matter the prices.

Memory is a commodity, so I think you will be very lonely in your quest.

I think it would be helpful for you to clarify which part of the chain you found objectionable:

- in 2015, OpenAI was founded as a Delaware nonprofit

- in 2017, OpenAI discovered the scaling laws and realized they needed far more compute (and thus money) than they had initially anticipated

- that discovery precipitated a series of negotiations between the founders on how to restructure OpenAI to raise more money for compute, ultimately resulting in Musk’s departure when the other founders would not give him control

- in 2018, OpenAI attempted to dramatically increase its fundraising despite Elon ending his contributions, but raised only $50M of its $100M goal

- in 2019, OpenAI created a capped-profit subsidiary in order to attract funding from commercial entities

- the nonprofit hired an independent assessor to value its IP, and then transferred that IP to the for-profit for fair value (around $60 million in 2019)

- the OpenAI nonprofit received a right to 100x capped return on its IP investment, or $6B, once the for-profit began making a profit. The nonprofit also received the right to the residual profit after all future investors reached their caps

- in 2019, OpenAI’s capped-profit received $1B in investment from Microsoft. OpenAI later received $2B from Microsoft in 2021 and $10B in 2023 as compute scaling continued

- Microsoft received a cap of 20x on its $1B investment, and 6x on its $2B and $10B investments, for a total of $92B target redemption

- in 2025, OpenAI’s for-profit entity recapitalized from a capped-profit entity with residuals flowing to the nonprofit to a traditional public benefit corporation with traditional equity

- in exchange for the residual (and 100x profit cap on the original $60M transfer) the nonprofit received a 26% equity stake in the for-profit. That stake is currently valued at around $200B

All of the above is from the record in Musk v. Altman, thanks to which we now have all the details. The upshot for the nonprofit is that it transferred IP worth around $60M in 2019 for rights to $6B in future profit, and then ended up with $200B in equity after the recapitalization. I see a lot of people in this thread assuming that the nonprofit no longer exists, which is not true.

Thanks, dang! The blog post[1] might be a better source than the twitter thread. Also I regret my typo above (lab -> labs) but too late now!

[1] https://cursor.com/blog/composer-2-5

The intellectual property (code, data) was transferred from the nonprofit to the for-profit for about $60M, which is what an independent firm hired to assess the value said the IP was worth in late 2018 / early 2019. The nonprofit itself was never converted to a for-profit, and indeed remains a nonprofit to this day.

The $60M in IP has grown to about a $200B stake in the OpenAI for-profit.

This is not correct. jongjong is correct that a nonprofit does not have owners in the sense that a for-profit has owners. Nonprofits are dedicated to their mission, and are run by a board of directors.

You cannot have a % ownership in a nonprofit because its resources must be used exclusively to carry out its mission. You could have a % control in its decision making process.

Correct, Musk based his claims on the 2023 Microsoft deal.

The 2025 recapitalization was discussed at trial, but it was ancillary since all that changed was the existing for-profit changed from a capped-profit with weird cash flow mechanics to a traditional public benefit corp with ordinary equity.

Surprised this got pushed off the front page so quickly! It’s exciting to see what the Cursor team has been able to do with significantly fewer resources than the frontier labs.

I do wish they weren’t joining xAI. Something tells me there will be a contingent of researchers that departs Cursor if that merger is consummated.

Evidence at trial showed that Musk attempted to pursue AGI at Tesla starting in 2017 before he left the board of OpenAI. He was unsuccessful in that endeavor and later restarted his efforts in xAI after the success of ChatGPT.

You are correct that which statute of limitations applies is a question of law. If facts are undisputed, that is the end of the issue. In this case, the facts were disputed, and the jury found for the defendants.

The jury instructions are public and the final jury form will be published, likely later this week.

I can tell you that the instructions told the jury to decide whether Musk could have brought his case before 2021.

The statute of limitations is not a trivial issue. Defendants have rights just as much as plaintiffs do, and our justice system does not allow plaintiffs to unreasonably delay in bringing their claims.

This case sets no precedent one way or the other on that question. The IP was transferred to the for-profit for fair value in 2019, and Musk never argued otherwise in this case.

The attorneys general of California and Delaware could challenge that 2019 IP transfer if they so wished on behalf of the public.

The statute of limitations takes into account when the plaintiff discovered or with reasonable diligence should have discovered their injury.

In this case, the jury found that Musk knew or should have known of his alleged injury prior to 2021.

Whether the claim accrued before the statute of limitations expired is a question of fact, and is therefore reserved for the fact-finder which in this case was the jury.

The statute of limitations exists to prevent unreasonable delay, to protect defendants from prejudice due to loss of evidence to the passage of time, and to recognize that people who are injured tend to complain immediately and not sit on their claims.

This case demonstrates why. Musk only complained after OpenAI was commercially successful with ChatGPT and after he started a competing effort. He repeatedly said “I do not know” and “I do not recall” on the stand, and argued that the passage of time made it hard for him to remember facts that would have been helpful for OpenAI.

My own thoughts:

If I had been on the jury, I would have found against Musk on every point.

His lawyers created a “3 phases of doubt” to try and sidestep the statute of limitations, but it was clearly bogus and he was on notice of OpenAI creating a for-profit in 2019.

Musk was perfectly happy to have OpenAI be a for-profit, a non-profit with an attached for-profit (the current structure), or even just absorbed into Tesla. His complaints fell flat for me given the number of emails where he said that a non-profit was likely a mistake.

This is technical, but Musk clearly never created a charitable trust, which was a precondition for his claims. His funds were donated for general use by OpenAI, not for any specific use that would allow him to claim breach of charitable trust. Also, all of his funds were spent by no later than 2020 which is before his alleged breach in 2023.

Musk unreasonably delayed bringing this case until the success of ChatGPT and starting a competing AI company, and he had unclean hands because he attempted to sabotage OpenAI repeatedly by poaching its key staff while on the board.

Because no one has commented yet on the legal significance:

Musk lost today because the jury found that he waited too long to bring his claims. The jury answers only yes/no questions, so we do not know their exact thoughts, but it is likely they determined that the 2019 and 2021 Microsoft deals were too similar to the 2023 Microsoft deal that was the centerpiece of Musk’s lawsuit. Musk could have brought the same lawsuit in 2019 or 2021, meaning his claims were untimely for the 3 year statute of limitations.

Because the statute of limitations is a precondition, the jury was not asked to find any other facts. They may tell the press what they thought on other issues, or they may not.

The judge was prepared to immediately accept the jury’s finding, and said she agreed that the jury’s decision was supported by the evidence.

It is possible for Musk to appeal, but success is vanishingly unlikely. Whether Musk’s claims are barred by the statute of limitations is a quintessential question of fact, and appellate courts are extraordinarily deferential to factual findings by juries so as a practical matter it’s almost impossible to appeal this verdict.

Power companies are a classic example of a natural monopoly because they require a ton of extremely expensive physical infrastructure to connect every house to the grid that would be wasteful to duplicate for every competitor.

The whole point of airplanes is that they require no physical infrastructure between point A and point B.

You can have competing power companies generating the power if the grid is owned by the state (or a regulated monopoly). Coincidentally, that is a good mental model for airlines because airports are often state-owned or if not are highly regulated.

This is a truly incredible exchange.

Savitt asked Musk if he’d raised any objection to the structure then, when he’d received the documents. Musk said that he didn’t read beyond that first box.

>Musk: I didn’t read the fine print.. We’re going into the fine print of this document.

>Savitt: It’s a four-page document.

https://archive.ph/rjkoa

Political violence is not acceptable in a democracy.

Full stop, no "but". That's all that needs to be said on this thread.

Comes impressively close to GPT 5.4 / Gemini 3.1 Pro / Opus 4.6! Mostly behind OpenAI on coding/agentic benchmarks, behind Google on text reasoning, behind Anthropic on Humanity's Last Exam with tools (surprisingly the only benchmark where Anthropic leads currently).

Meta hasn’t fully caught up, but they came close and I think can solidly claim to be a frontier lab again. I’d call it a 3.5 horse race right now, and hopefully their next model improves. More model competition is good!

Poor Grok 4.2 should probably be dropped from the table.

You are commenting on the article that contains data that backs that up.

The irony of this webpage being terribly optimized for mobile clients…

Nested scrollbars! Horizontal and vertical scroll!

They’re pretty upfront in their release post that they took an open source model and improved it with their own coding data. They mention “continued pretraining” (on top of the base model) and RL. Cursor never claimed to have done a full pretraining run.

More to the point, beating Opus 4.6 at coding and coming within striking distance of gpt-5.4 is impressive! The benchmarks outperform raw Kimi K2.5.

It’s particularly impressive given larger labs like Meta are struggling to catch up to OpenAI/Anthropic.