Pyyan / News

What happened in AI

A dated log, newest first. Every item was read from its source before it was written up, and every one carries the link so you can go and disagree with it.

68 stories · 6 straight from the source · showing 41 to 50, page 5 of 7

The days

3 Sept3
2 Sept6
1 Sept1

By kind

Who keeps appearing

3 September 2026

17d ago

PolicyUS Congress

A US senator proposed making superintelligent AI a crime carrying 20 years

20 yearsmaximum sentence proposed, the nuclear weapons standard

The penalty the bill proposes for building a superintelligent AI is up to 20 years in prison. The sponsors picked that number because it is the penalty for unlawfully developing a nuclear weapon.

Senator Bernie Sanders and Representative Greg Casar announced the Ban Artificial Superintelligence Act on 3 September. It would permanently prohibit systems that surpass human intelligence, that could overthrow a government, or that can subvert a shutdown command. It would pause advanced AI development entirely until a new federal regulator writes safety rules, and would stand up a cabinet level agency with an advisory board to enforce the ban. Individuals face up to 20 years. Companies face dissolution, which the sponsors call the corporate death penalty.

Why this one is different

Almost every AI bill so far has concerned disclosure, training data, or the use of models in hiring, housing and policing. This one proposes to make a direction of research illegal. It is the first time a prominent American politician has asked for a prohibition rather than rules, and it arrived on the same day OpenAI released a model its own president said may eventually be seen as the arrival of general intelligence.

Cutting edge AI technology is less regulated than the average food truck.

How we got here

16 Jul 2026The European Commission orders Google to open Android to rival AI assistants under the Digital Markets Act. Regulation of conduct, not of capability.
26 Aug 2026Meta agrees to pay up to $17.1bn over child safety, the largest settlement in the industry's history. Still conduct.
2 Sep 2026The US Justice Department files a statement of interest backing OpenAI against the New York Times, arguing training on copyrighted text is fair use.
3 Sep 2026OpenAI ships GPT-6 Astra and its president says it may eventually be seen as the arrival of AGI.
3 Sep 2026Sanders and Casar propose banning that outcome outright, with a 20 year sentence attached.

What it does and does not mean

The bill has not been introduced, and it defines superintelligence in words rather than in anything testable. Systems that surpass human intelligence, or could overthrow a government: no compute figure, no capability score, no evaluation is named, so as described there is no procedure by which a company could establish whether it had broken the law. A prohibition that cannot be measured cannot be enforced, and that gap is the first thing any committee will find. What it does show is a change in what can be said out loud in Congress. The argument used to be how to regulate deployment. This is a proposal to prohibit a destination, made in the same week the executive branch filed in court on the industry's side.

Office of Senator Sanders Axiosfrom the source itself

Open this story on its own page →

MoneyNVIDIA · Hugging Face

Nvidia agreed to buy the place open models live

$12.93bnfor the platform hosting 3m models

Nearly every open weight model released this year was published in the same place. On 3 September the company that makes the chips they run on agreed to buy it, for $12.93bn.

Nvidia will pay $12.93bn for Hugging Face, with an equity retention package worth up to $1bn for employees who move across. It is Nvidia's second largest acquisition, after the $20bn it paid for Groq's assets in December 2025. The platform hosts more than 3 million models and serves 18 million developers and over 200,000 companies. Hugging Face's chief executive told CNBC that the company approached Jensen Huang weeks before the deal, rather than the other way round. Huang published a set of unusually specific commitments about keeping the platform open, which is why the reaction has been mixed rather than uniformly hostile.

Why this one is different

Buying open source infrastructure is not new, and the warnings are not always right: Microsoft bought GitHub and it went better than most people predicted. This one differs in a way that has nothing to do with anybody's intentions. Microsoft did not sell a product that competed with the code on GitHub. Nvidia sells the silicon these models have to run on. If a quantisation format or a serving optimisation lands working best on Nvidia first and the other backends catch up months later, nothing improper has happened and the effect is the same, because developers stay on whatever path is fastest the day they start.

Microsoft did not sell a competitor to the code on GitHub.

How we got here

14 Aug 2026Alibaba releases Qwen3.8-27B under Apache 2.0, distributed on Hugging Face.
26 Aug 2026Qwen3.8-Flash-Next tops the Hugging Face trending list, which much of the field reads as the scoreboard for open models.
28 Aug 2026Tencent ships HY4 preview, 770B parameters, open weights, in the same place.
31 Aug 2026Anthropic signs $35bn of cloud with Lambda, and Nvidia holds the lease on the building.
3 Sep 2026Nvidia agrees to buy the platform every one of those was published on.

What it does and does not mean

None of this has happened yet. A transaction this size requires an HSR filing in the United States and a waiting period before it can close, and full review is expected in the US and the EU and probably the UK, against a company already under active antitrust inquiry on both sides of the Atlantic. Nvidia's own argument is that an open platform is a deconcentrating force rather than a concentrating one, and a regulator, not a reader, will decide whether that holds. What is already true whatever the outcome is narrower and worth saying plainly: the open weights movement keeps its distribution, its discovery and its scoreboard in a single place, and this month established what that place is worth to the company selling the hardware underneath it.

CNBC Bloombergtwo sources

Open this story on its own page →

BenchmarksNVIDIA

An AI outscored every human at the International Olympiad in Informatics

535.4out of 600, against 498.27 for the best human

The International Olympiad in Informatics has run every year since 1989. Four teenagers per country, picked through national contests, two days, five hours a day, six problems. This year the highest score in the room did not belong to a teenager.

Closing Ceremony of the 38th International Olympiad in Informatics, IOI 2026 · Raqamli texnologiyalar vazirligi, the host country's Ministry of Digital TechnologiesThis is the contest the model competed in, filmed by the host. It is the ceremony rather than the result described above: no video of the AI run has been published.

IOI 2026, final scores out of 600. The gold medal threshold is the dashed line, cleared by about one contestant in twelve.

NVIDIA entered a model called Nemotron-3-Ultra-550B. It scored 535.4 out of 600. The best human competitor scored 498.27. The bar for a gold medal, which about one contestant in twelve clears, was 361.12.

Why this one is different

Machines have scored well on IOI problems before, but afterwards: on problems already published, with as much time as the researchers wanted to give them. This one ran during the contest itself, on the same five hour clock, under the same limit on submissions and the same rules about what it could look up, before the problems were public anywhere. The claim in the paper is deliberately narrow. First time an AI has outscored the top human on an IOI problem set.

Four years, from the middle of the pack to first place.

How we got here

2022DeepMind's AlphaCode reached the median competitor on Codeforces, around the top 54%. The first time an AI was competitive at this at all.
2023AlphaCode 2 solved 1.7 times as many problems and beat roughly 85% of entrants.
2024OpenAI entered o1-ioi live at the IOI and finished in the 49th percentile. It could reach gold, but only with hand written strategies and relaxed limits. Its successor o3 then reached gold on the same problems without either.
2025Open weight models reached the gold threshold too, so the capability stopped being something only a frontier lab could rent you.
2026535.4 against 498.27. First place.

What it does and does not mean

Start with what it does not show. An IOI problem is the friendliest shape a task can have for a machine: self contained, precisely specified, scored automatically, with a correct answer fixed before anyone starts. Almost no real engineering looks like that. This result says nothing about choosing which problem is worth solving, working from a specification that turns out to be wrong, or writing code a colleague has to maintain for five years. What it does show is narrower and still large: for problems that can be stated exactly, the ceiling has moved above the best human alive, and it moved there in four years.

arXiv 2609.02849 VentureBeatfrom the source itself

Open this story on its own page →

2 September 2026

18d ago

LawOpenAI

Thirty new lawsuits allege OpenAI's public relations team overruled its safety team

30new complaints, and a new legal theory

OpenAI's chief executive apologised in April for not telling police about the account. The 30 complaints filed on 2 September allege something the apology did not cover: that the decision not to tell them was taken by the people who handle the company's image.

Teachers, students and a principal who were present at the shooting at Tumbler Ridge Secondary School in British Columbia in February filed 30 complaints in US federal court in California. They allege OpenAI knew from the suspect's ChatGPT conversations that an attack was being planned and made a conscious decision not to warn the RCMP. For the first time in this set of cases the complaints also allege aiding and abetting, which requires establishing intentional conduct rather than failure. They name chief global affairs officer Chris Lehane, alleging that staff responsible for public relations overrode a safety team recommendation to refer the account to the authorities. OpenAI disputes key claims.

Why this one is different

The AI lawsuits so far have been about copyright, or about products harming the people using them. This is a claim that a company held specific knowledge of a specific threat to specific people and said nothing, and the aiding and abetting count moves the case from what a company failed to do towards what it decided. OpenAI is now facing more than fifty consumer harm and wrongful death suits.

A filing is not a finding.

How we got here

Feb 2026The shooting at Tumbler Ridge Secondary School in British Columbia.
Mar 2026The first family sues OpenAI, alleging it could have warned police.
Apr 2026Sam Altman writes to the community apologising for the failure to alert the RCMP.
2 Sep 2026Thirty more complaints add an aiding and abetting claim and name executives.
2 Sep 2026On the same day, the US Justice Department files in support of OpenAI in the New York Times copyright case.

What it does and does not mean

These are allegations in a complaint and none has been tested. Aiding and abetting is a high bar precisely because it requires showing a decision rather than a lapse, the claim that public relations overrode safety comes from the plaintiffs rather than from any document that has been published, and OpenAI disputes key claims. A filing is not a finding, and the people who filed these are survivors rather than disinterested parties, which is not a criticism of them but is a fact about where the account comes from. What is already established is narrower: the company did not tell the police, and its chief executive apologised for that in April. What the case now turns on is not whether a warning would have worked, but whether somebody decided not to give one, and that has moved from a question of policy to a question a court will take evidence on.

NPR TechCrunchtwo sources

Open this story on its own page →

ModelsGoogle

Google shipped Gemini 3.8 Flash, and a cyber twin you have to apply for

70%real world vulnerability discovery, gated access

Two models shipped on the same day, built on the same intelligence. One you can buy today. The other you have to apply for, and Google decides.

Gemini 3.8 Flash is a general release at $0.75 per million input tokens. Gemini 3.8 Flash Cyber finds software vulnerabilities and writes the patches, clears 70% on real world vulnerability discovery, and is reachable only through a new scheme called the Fairwind Program: governments, national cyber authorities, operators of hospitals, power grids and payment systems, and the maintainers of widely used software.

Why this one is different

The model is not the story. The gate is. Three weeks earlier OpenAI did the same thing with GPT-5.6-Cyber and its Daybreak Red tier, and the day before this, OpenAI rated its own Astra model Critical for cyber and said the strongest capabilities would go to vetted partners only. Three labs, three separate application processes, the same conclusion reached independently: a model good enough to find zero-day flaws is good enough to use them, and the only control anybody has is who gets an account.

The model is not the story. The gate is.

How we got here

10 Aug 2026OpenAI announces GPT-5.6-Cyber, built on Sol for finding zero-days, behind an applicant-vetted tier called Daybreak Red.
1 Sep 2026OpenAI rates Astra Critical for cybersecurity under its own Preparedness Framework and says the strongest capabilities will be gated.
2 Sep 2026Google ships 3.8 Flash Cyber behind the Fairwind Program. Same shape, different name, and no shared standard for who qualifies.

What it does and does not mean

What this does not establish is who decides. There is no shared definition of a trusted defender, no appeal if a lab says no, and no reciprocity: clearing Fairwind tells you nothing about clearing Daybreak. A hospital group approved by one and refused by the other has no recourse and no explanation. What it does establish is that vendor gatekeeping is now the actual security policy for frontier cyber capability, arrived at by three companies separately and written down by no regulator.

Google blog VentureBeatfrom the source itself

Open this story on its own page →

LawOpenAI · Microsoft · US DOJ

The US government told a court that training on copyrighted text is fair use

1sttime Washington has taken a side in an AI copyright suit

For nearly three years the AI copyright fight has been private companies against private rightsholders, with the government watching. On 2 September the government picked a side.

Trump Admin Backs OpenAI in Landmark New York Times AI Copyright Battle · WIONWION's report on the filing. The account above is sourced to the Washington Post; this is a broadcast summary of the same court document.

The Justice Department filed a statement of interest in the New York Times case against OpenAI and Microsoft, arguing that training a model on copyrighted text transforms the work enough to be fair use, and that the benefit to AI outweighs the competitive harm, partly on national security grounds. A statement of interest binds nobody. It is the government telling a court what it thinks, without becoming a party.

Why this one is different

This is the first time the federal government has taken a formal position in any of them, and there are dozens: authors, publishers, music labels, news organisations, all making roughly the same argument in different courtrooms. Until now the question of whether training is fair use was being answered case by case by individual judges. A statement of interest does not settle it, but every one of those judges has now read what the United States thinks the answer is. The Times replied that the government had sided with a handful of trillion dollar companies against the countless American creators whose work they stole.

It binds nobody, and every judge in every one of these cases has now read it.

How we got here

Sep 2023Authors including George R. R. Martin and Jodi Picoult sue OpenAI over training on their books.
27 Dec 2023The New York Times sues OpenAI and Microsoft, arguing the models reproduce its journalism and compete with it.
2025The judge throws out most of OpenAI's motions to dismiss, and the bulk of the Times claims proceed.
Sep 2025Anthropic settles with book authors for $1.5bn over torrented training data, the largest number yet attached to the question.
2 Sep 2026The Justice Department files, and says training is fair use.

What it does and does not mean

Start with what it cannot do. A statement of interest has no binding force, does not decide the case, and can be reversed by the next administration with another filing. No judge is obliged to agree, and the Times case still has to be tried. What it does do is change the weather. Every remaining defendant can now cite the United States government in support of the central defence, and every rightsholder now has to argue against their own government as well as against the company that took the work.

Washington Post Deadlinetwo sources

Open this story on its own page →

SiliconBroadcom

Broadcom's AI chip revenue tripled in a quarter, to $16.7bn

$16.7bnAI chips in one quarter, up 221%

Broadcom sold $16.7bn of AI chips in three months. That is more than the company's entire revenue, from everything, two years ago.

Broadcom AI semiconductor revenue, reported for Q3 and guided thereafter. The last two figures are the company's own projection, not results.

AI semiconductor revenue grew 221% year on year and now makes up 56% of everything Broadcom sells. Total revenue was $29.6bn, up 86%. The guidance is the part worth reading twice: $21.7bn of AI revenue next quarter, then a claimed $115bn across fiscal 2027 and $230bn in 2028. Google, Anthropic and OpenAI are named as customers.

Why this one is different

Most AI revenue figures are a supplier telling you what it hopes to sell. This is a public company reporting what it already sold, audited, to named customers, and more than half its business is now one product line that barely existed three years ago. The projection is the aggressive part: doubling to $115bn and doubling again to $230bn assumes the people buying keep buying at a rate none of them has committed to publicly.

A doubling, then another doubling, on orders nobody has publicly committed to.

How we got here

2023AI is a rounding error in Broadcom's accounts, overshadowed by networking and infrastructure software.
2024 to 2025Custom accelerator work for hyperscalers turns into a real line of business as the big cloud buyers look for an alternative to buying only from NVIDIA.
Q3 2026AI passes half of total revenue for the first time, at 56%.

What it does and does not mean

A backlog is not a business, and a projection is not a backlog. Two thirds of the growth story here sits in fiscal 2027 and 2028 numbers that depend on a handful of customers, each of whom could build in house, switch supplier or simply slow down, and none of whom has publicly committed to the volumes implied. What is not a projection is the $16.7bn already banked, and the fact that the customer list for frontier AI silicon is now short enough to name in one sentence.

Broadcom investor relations 24/7 Wall Stfrom the source itself

Open this story on its own page →

PolicyNYC Department of Education

New York City banned generative AI for 600,000 schoolchildren

600,000pupils, one year moratorium

The largest school district in the United States has decided that for children under about fourteen, the correct amount of generative AI is none.

NYC bans generative AI in schools through eighth grade · ReutersReuters' report on the decision, filmed the day it was announced.

New York City put a one year moratorium on student-facing generative AI from pre-K through eighth grade, covering roughly 600,000 pupils, about two thirds of enrolment. Companion chatbots offering emotional support are banned at every grade, high school included. Pre-K to second grade lose individual devices entirely. Older primary pupils are capped at 30 to 45 minutes a day. High schools get twice yearly AI literacy classes and supervised pilots instead.

Why this one is different

Most school AI policies are guidance: use it responsibly, disclose it, do not cheat. This one is a prohibition with an enforcement mechanism, implemented through device filtering rather than a memo to teachers. And the companion chatbot ban is broader than the age limit, applying to sixth formers as well as six year olds, which says the district is worried about something other than cheating.

The ban on companion chatbots covers every grade. That is not a cheating policy.

How we got here

2023New York City bans ChatGPT on school networks within weeks of its release, then reverses the ban four months later and publishes guidance encouraging teachers to experiment.
1 Sep 2026Los Angeles Unified quietly blocks generative AI on 378,000 student devices using commercial filtering.
2 Sep 2026New York City goes further and puts it in policy, with a stated one year term and an age line rather than a blanket ban.

What it does and does not mean

This is not evidence about whether AI helps children learn. No study is cited, the term is one year rather than permanent, and a moratorium is what an institution does when it needs time rather than when it has an answer. It also pushes usage to personal phones, where the district has no filtering and no visibility. What it does show is a large public system concluding that the burden of proof now sits with the technology, having spent three years assuming the opposite.

CNN ABC Newstwo sources

Open this story on its own page →

ModelsMeta

Meta's Muse Spark 1.3 does the same work with a quarter fewer tokens

25%fewer tokens for the same work

The interesting number in Meta's release is not a benchmark. It is that the model does the same work with 25% fewer tokens and 20% fewer tool calls, which is the number that shows up on the bill.

Muse Spark 1.3 arrived on 2 September in Muse Code and the Meta Model API, which Meta calls its largest improvement yet on coding and agentic work. It scores 88.8 on Terminal-Bench 2.1 and 75.4 on DeepSWE v1.1, and 98.5 on the MRCR long context test between 256K and 512K. A maximum reasoning mode is held back pending further safety testing.

Why this one is different

Every lab claims a benchmark. Almost none of them publishes how much it costs to reach one. An agent that solves the task in twenty steps and one that solves it in fifteen score identically and bill very differently, and for anyone running agents at volume the second number decides whether the thing is affordable at all. Quoting the reduction alongside the score is the unusual part here.

Two agents can score identically and bill very differently.

How we got here

5 Aug 2026Muse Spark 1.2, Meta's agentic model with a terminal agent, at $1.25 in and $4.25 out per million tokens.
10 Aug 2026Muse Glimmer, 29.6B dense under Apache 2.0, the open counterpart to the proprietary Spark line.
2 Sep 2026Spark 1.3, quoting efficiency against 1.2 rather than only benchmark scores.

What it does and does not mean

These are Meta's own figures on Meta's own harness, and efficiency claims are unusually sensitive to how the test was set up: a different scaffold, a different tool set or a different retry policy moves the token count more than the model does. Nobody outside has reproduced them. What is worth taking from it regardless is that the industry has started competing on cost per completed task rather than score alone, which is the metric a buyer actually has.

Meta AI Research Axiosfrom the source itself

Open this story on its own page →

1 September 2026

19d ago

SafetyOpenAI

OpenAI says a model of its own crossed the line it wrote to catch dangerous ones

91.5%of cyber jailbreaks refused, from 59%

OpenAI wrote a framework to identify models too dangerous to release as they are. On 1 September one of its own models tripped it.

OpenAI warns of risks in new 'Astra' AI model · ReutersReuters' report on the same announcement. It covers the risk warning described above rather than the model's capabilities.

Share of cyber jailbreak attempts refused, OpenAI's own testing. The gap is what the Critical rating forced before release.

Astra is the first model OpenAI has rated Critical for cybersecurity under its Preparedness Framework. During testing it found and chained two zero-day vulnerabilities nobody had asked it to look for, and scored full marks on ExploitBench. The rating forces extra safeguards before release: Astra now refuses 91.5% of cyber jailbreak attempts against 59% for GPT-5.6 Sol, and the strongest capabilities go to vetted partners only.

Why this one is different

Safety frameworks are usually published and then never bind anything, because no model ever quite reaches the threshold. This one caught something, and the company said so before shipping rather than after. The specific detail worth holding on to is the unprompted part: the zero-days were not the task. They turned up while the model was doing something else, which is a different kind of capability from passing a test designed to measure it.

The zero-days were not the task. They turned up while it was doing something else.

How we got here

2023OpenAI publishes the Preparedness Framework, with capability tiers up to Critical and a commitment not to ship at that tier without safeguards.
10 Aug 2026GPT-5.6-Cyber, trained for offensive security work, released only through the vetted Daybreak Red tier.
1 Sep 2026Astra becomes the first model OpenAI rates Critical on any axis, and the framework does something for the first time.

What it does and does not mean

The grader and the graded are the same company. OpenAI wrote the framework, ran the evaluation, decided the rating, chose the safeguards and will decide who gets access, and no outside body checked any step. A 91.5% refusal rate also means roughly one attempt in twelve still gets through. What it does show is a lab publishing a threshold in advance and then honouring it against its own flagship, which is the first time the exercise has cost anybody anything.

OpenAI SecurityWeekfrom the source itself

Open this story on its own page →