Policy · 21 September
OpenAI and Anthropic neared a pact to stress-test each other
OpenAI and Anthropic came close to a legally binding agreement under which each would gain access to the other’s API for commercially available models, to search for vulnerabilities, hidden risks and misalignment, The Information reported on Monday. The arrangement would have been the first of its kind between rival frontier labs, turning the public safety consensus of recent weeks into an enforceable technical mechanism. It is unclear why the talks did not conclude. The report follows the two labs’ earlier joint alignment evaluation, in which each ran internal safety tests on the other’s released models and published the findings as a template for cross-lab accountability.
Source: The InformationPrimary: Invezz
Standards · 21 September
OpenAI asks Washington to lead a global standards push
In “Building standards for the next phase of AI”, OpenAI argued that international standards for safety and security practice “may be as important to pacing the frontier as alignment research itself”. The company wants the United States to lead, using the emerging network of national AI safety institutes — Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the United Kingdom — to set common measures for evaluating self-improving systems and common protocols for classifying and reporting incidents. It frames standards as a complement to national law rather than a licence or pre-release approval regime, and as a guard against fragmentation and the concentration of power. The Wall Street Journal noted the awkward timing: President Trump has said he will not “stifle” the industry and that the Justice Department can “rein in things” if needed.
Source: OpenAIPrimary: The Wall Street Journal
Research · 21 September
OpenAI forms a maths advisory group as its model clears 100 problems
OpenAI has established the Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study in Princeton, to give mathematicians a formal voice in its maths research; nine prominent mathematicians are initial members. Alongside the announcement, OpenAI said the same internal model behind its Navier–Stokes result has now resolved more than 100 further open problems across most areas of mathematics. The group is advisory only: OpenAI’s post states it “will not be responsible for advising us on how to pace our internal progress on mathematics”, and the institute says it holds no decision-making power over any AI company. The announcement follows an open letter signed by 25 Fields Medal winners warning that the pace of AI-driven results threatens mathematicians’ work.
Source: TechCrunchPrimary: OpenAI
Product · 21 September
Meta’s Muse beats ChatGPT’s early launch, and Amazon blocks it
Meta’s new AI agent app, Muse, has drawn 1.8 million iOS downloads in the United States and Canada in its first 12 days, against ChatGPT’s 1.3 million over the equivalent window, and 642,000 American daily active users against 231,000, according to estimates from Apptopia. Muse has seen 2.8 million installs globally in its first 12 days and has climbed to number one on the US App Store. The caveat is that Apptopia measures third-party estimates, not Meta’s own figures, and Muse launched in two countries across two platforms. Separately, Amazon has blocked Muse’s agent from using Amazon.com, a step it has taken against shopping bots from Google and OpenAI before; The Information reports OpenAI is building features to counter Grok Bot and weighing a response to Muse.
Source: TechCrunchPrimary: The Information
Hardware · 21 September
Google opens preorders for its $899 Gemini laptop
Google’s AI-native Googlebook is now available to preorder at $899, nine months after it was first shown. The device runs Android OS with a desktop Chrome and threads Gemini through the cursor (Magic Cursor), dictation (Rambler, which turns rambling notes into readable text) and vibe-coded widgets; flagship models come from Acer, ASUS, Dell, HP and Lenovo, with up to 2.8K OLED displays, Intel and Qualcomm processors and dedicated NPUs. The pitch bundles 12 months of Google AI Pro and up to 10 years of updates, and reads as a bid to move the roughly 50 million Chromebooks in schools towards Gemini. It goes on sale on 4 October in the United States and on 5 October in Canada, the United Kingdom, Ireland, France, Germany and Australia.
Source: TechCrunchPrimary: Google
Geopolitics · 21 September
DeepSeek bets on 160,000 Huawei chips to sidestep export controls
DeepSeek plans to deploy at least 160,000 of Huawei’s Ascend accelerators at a new data centre in Inner Mongolia, with the company’s chief executive, Liang Wenfeng, telling investors that deliveries could begin as early as the fourth quarter of 2026, The Information reported. The order is the clearest sign yet of Chinese labs building around American export controls that have restricted the sale of Nvidia’s most advanced data-centre GPUs into China. Huawei is targeting 750,000 Ascend units shipped in 2026, of which ByteDance and DeepSeek alone would account for roughly 510,000. The same equipment controls that pushed DeepSeek towards Huawei also constrain Huawei’s ability to manufacture at that scale.
Source: The InformationPrimary: Value Add VC
Capital · 21 September
Oura seeks $2.2bn as autumn AI listings test the market
The smart-ring maker Oura and some of its backers are seeking to raise as much as $2.2 billion in a US initial public offering, offering 50 million shares at $40 to $44, a top-of-range valuation of about $15.6 billion. It is a test of public appetite for consumer AI hardware: Oura has leaned on its health-data trove for new AI features and reported sharply higher revenue. The listing lands in a crowded autumn window that also includes the data-centre developer SB Energy — building a huge complex for OpenAI — which faces investor scepticism, and an Anthropic IPO that retail investors are reportedly queuing for despite lock-up risks.
Source: BloombergPrimary: The Information