The people who build AI have applied to list on the stock market.What changes once a company has shareholders?
A question picked from the 23 records we added that week. More of them were about the size of the companies that build AI, and the rules that bind them, than about new models.
This comes close to being the first case of a frontier lab entering the public capital markets. As reported, the filing showed revenue of $16.3 billion for the first half of 2026 ($1.2 billion in the same period of 2025), an operating loss of $693 million, and a net loss of $80.7 billion made up mostly of non-cash accounting adjustments tied to convertible notes, and it listed more than 11,900 business customers spending at least $100,000 a year as of the end of September, 1,900 spending at least $1 million and about 250 spending at least $10 million. It was also reported that the 7 cofounders will set up a Founder LLC holding 50.1% of the voting power — a structure that ties control to the founders even after the listing
This is the first federal appellate ruling on fair use in AI training. Of the four factors, it found the purpose and character commercial and barely transformative, the amount used to be the headnotes copied in full, and the market effect harmful in two markets — the existing legal research market and the emerging market for licensing headnotes as AI training data. The court did, however, distinguish ROSS's tool from the generative models in Bartz and In re OpenAI on the ground that it cannot produce original expression, and it did not rule on whether generative AI training is fair use — cite it as generative AI precedent and you get it wrong. Whether some copyrights have expired, and the remaining headnote claims and damages, still go to trial in Delaware
Company-reported: on Agents' Last Exam, a professional-work benchmark, the medium reasoning setting scored 15 percentage points higher than GPT-5.6 Terra at about 1/3 the cost, and matched GPT-6 Astra at 1/10 the price per token. It put Terminal-Bench 4.0 at 50.3% on the maximum setting, more than double GPT-5.6 Terra's published 21.5%, and wrote that on an internal benchmark built to induce errors the error rate fell from 53.6% to 16.7% on the low setting and to 4.3% on the maximum setting. At the same time OpenAI said of its own accord that the model reached the critical threshold for cybersecurity under its preparedness evaluation framework, which shows that a model with a lower price is not thereby a model in a lower risk class
What ties the events together — written by AI from the week's records.
A builder's books go onto the public record for the first time
A lab filed to go public with the U.S. Securities and Exchange Commission. It has settled on a ticker and the market it will list on, but not yet how many shares it will sell or at what price. According to reports, the filing lists $16.3 billion in revenue for the first half of this year and an operating loss of $693 million, and it also lays out a structure in which seven co-founders hold a little more than half of the voting power. In the eight days before that, the same lab said it would put 5% of its stock into a public-benefit fund, and it added one more trustee to the trust that holds the power to appoint a majority of the board. Put the three records in date order and you see the sequence: the public-benefit pieces get tied down first, right before shareholders come in.
A court drew a line, and the fight breaks out in the gap it left
A U.S. federal appeals court said that training a rival search tool on summaries of legal material was not fair use. It took as real harm the fact that a new "market for lending material out for AI training" is taking shape. But the court said the tool cannot make up new sentences, set it apart from models that produce text or images, and did not rule on those. A month later, two things landed on the same day over that empty space. One is a set of briefs backing fair use, with the U.S. government signing on too. The other is a promise to honor a shared signal that lets a website say in one go whether it may be used for training. The same week in Korea, companies told the government they cannot get ready because the detailed rules holding the safety management standards are not out yet.
The price drops by half, and the strongest model is not for everyone
A model that looks at a screen and works apps for you came out, and it costs half of what the generation before it did. By its maker's own measurement, it scores about the same as a pricier model at one tenth the cost. The other way around, one of the strongest models released the same week published a price and still is not open to everyone. It goes first only to some of the people working on cyber defense, and will widen only as far as places the maker trusts. Things are splitting into one side that is cheap and used a lot, and another that is strong but handed out selectively.
A record that takes up Google's new top-end model but puts the focus on how to assemble an agent rather than on praising the model. Running the same job side by side, one side directed through conversation and the other
Until now an operator had to identify each crawler one by one and manage separate rules for each provider. What this post promises, though, is an additional signal that does not replace existing crawler controls, and no
It is a document in which a party to the case picked and bound only the briefs that favor it, so it must not be read straight as fact — the briefs on the other side are not included, and the party's own description of th
A hands-on record that runs a coding agent all the way through from an empty working space, pulling out the feature to reach for at each point where it gets stuck. Reading, making and fixing inside the working space happ
It admits as a gap that the Hugging Face incident began in an internal cybersecurity evaluation with the misalignment monitor switched off, and said that when it replayed that record with the strengthened monitor an aler
The core is that the standard moves from following instructions literally to applying values in situations the instructions do not cover. It said it borrowed the view of alignment as a cooperative game, and created new s
It contains a request that liability be divided between developers and adopting companies along the lines of what each can actually control, a request to make clear that pharmaceutical and biotech R&D uses such as toxici
It is an expanded version of the partnership from last January. On timing, the ServiceNow MCP plugin is in the OpenAI marketplace now, enhanced features come at the end of 2026, Super Intelligence Control Tower voice fea
NVGEMM was put into NVFP4 decode on Blackwell, folding the dequantization scale into the GEMM and removing the cost of launching a kernel per layer. float8_e4m3fn and _scaled_mm were added to MPS on Apple Silicon, with p
The names listed include research funders and societies such as the Alfred P. Sloan Foundation, the Simons Foundation, Schmidt Sciences, the Open Source for Science Fund, the U.S. National Science Foundation and CAS of t
The core of the design is that it decides when to swap from the training record alone, without looking at validation or test scores. The authors reported the average over 11 benchmarks rising from 57.77 to 61.05, about 3
Company-reported: No. 1 on the Artificial Analysis Intelligence Index v5.0 with 59, No. 4 on WebDev Arena at 1768 Elo, up 89 points, SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0%, a 0.0% attack success rate on G
A record that unpacks a work structure made by chaining several roles together, likening it to an organization chart. It treats one conversation as one worker's desk and finds the reason for splitting the roles in the wa
By the authors' measurement, visual tokens take up 89–97% of the input, yet dropping the image changes which tokens get unmasked in only 22.8% of steps on average, and as few as 7.9%. They reported keeping 99.6% of LLaDA
All the figures come from the company and its partners, with no independent check. Generate Biomedicines said one CAR T design proposed by 6 researchers was about 70% shorter than the standard and produced about 10 times
It said you can give spoken instructions to change the weather, move the camera and add people and objects, and that it simulates airflow and collisions. It gave no resolution, frame rate or latency figures, writing only
By the company's own aggregation, across the 74 countries observed in every quarter the share of classified work messages taken by the 10 most common activities fell from an average of 56.0% in Q1 2025 to 51.1% in Q3 202
This comes close to being the first case of a frontier lab entering the public capital markets. As reported, the filing showed revenue of $16.3 billion for the first half of 2026 ($1.2 billion in the same period of 2025)
Company-reported: on Agents' Last Exam, a professional-work benchmark, the medium reasoning setting scored 15 percentage points higher than GPT-5.6 Terra at about 1/3 the cost, and matched GPT-6 Astra at 1/10 the price p
By the company's own aggregation, since the standard was released in December 2025 more than 70 agent platforms have adopted it, including ChatGPT, Cursor, GitHub Copilot and VS Code, and about 1 million skills have been
The key point is that this is the MAX license, not open source — you can use it, modify it and ship commercial products on any hardware with no fees or royalties, but no part of MAX may be used as input to an AI system t
A record that gathers the numbers on why a widely used coding test fell apart. The starting point is that OpenAI stopped announcing scores on that test and even withdrew its recommendation of the follow-up test it had pu
A record that examines, with scope at the centre, the case where DeepSeek released not a model but an inference engine under Apache 2.0. The three items the guide document itself notes as left out are the KV cache system
The trust holds the power to appoint a majority of Anthropic's board and, independent of management and investors, advises on decisions that involve risk and social impact. With the listing filing (S-1) and the public be
It explained that adding the co-founders' pledge to give away 80% of their equity and the employee equity match ties about 20% of the equity to the public benefit, with more on top as the value rises. There is no concret
A record of a trolley problem experiment released on Oct 14, 2026. Eleven scenario families were put to fourteen models across 5,240 settings, varying the wording, the order of explanation and the number of people, and t
A record that shows, through one company's internal switch, where the next bottleneck moves once code piles up faster. Daily average merges rose, by the company's own measurement, from about 2,500 in May 2026 to more tha
A record that reworks an introductory course on coding agents into seven steps. It likens the reason a one-line instruction brings flat results to a smart new hire who does not know how the company works, and it says to
There is not a single experimental result — no activity, minimum inhibitory concentration or toxicity figures, the candidate is still being checked in the wet lab, and the authors wrote that the experimental results are
It came with background figures: scams account for more than half of reported crime in Singapore, losses since 2019 have passed S$4 billion, and conversational scams make up as much as 80% of the total loss. This is a ca
An introductory note that explains what an agent is without the jargon. It draws the line this way: the conversational tools so far were an adviser who answers your questions well, so they would write the text while open
A record of handing eight office tasks over within the same frame. The frame has three steps — ask; bundle what works into instructions and code so it can be reused; set what is reused to run at a fixed time — and what k
This is the first federal appellate ruling on fair use in AI training. Of the four factors, it found the purpose and character commercial and barely transformative, the amount used to be the headnotes copied in full, and