❯ OpenAI releases 722 mathematical manuscripts from an internal model, covering 372 groups of new results, most still awaiting review by mathematicians

722 at onceOpenAI announced on October 6 that it is releasing a batch of new mathematical results produced by an internal frontier model: 722 manuscripts, grouped into 372 families of related results, in a public GitHub repository. According to Runtime Wire, they came from about 4,000 research problems posed to the model, with the average result using compute equivalent to about three hours of ChatGPT Pro thinking. The model that produced them is unreleased and unnamed.

One prompt, one agentAccording to Scientific American, nearly all 372 results came from the same procedure: a single prompt handed to a single AI agent, with some possibly taking multiple attempts. They span number theory, combinatorics, theoretical computer science and mathematical physics. A few results did not follow the standard procedure, including work on the Riemann zeta function and on a case of the Hodge conjecture.

Not all verifiedThe repository includes Lean formalizations for many results, meaning proofs a computer can check step by step, but many manuscripts still lack them. OpenAI itself warns that unformalized results could contain issues and that citations and exposition need more work. It says the release drew on advice from the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, which has no decision-making power.

Unease among mathematiciansWIRED reports that in August OpenAI convened around 40 mathematicians to discuss what to do if AI outpaces humans in the field, and hinted that its models had solved hundreds of longstanding problems. Attendees say the company assured them it would not release everything at once, an assurance an OpenAI spokesperson says the company is “not aware of.” One mathematician told WIRED there is “a perception of mobster behavior” from leading AI companies. With hundreds of proofs arriving together, the work of checking each one’s correctness and value falls to mathematicians.

▮ SIGNALOnce proofs can be produced in bulk, the scarce step in mathematics shifts from finding a proof to confirming one, and that still depends on people.

❯ Mistral releases its trillion-parameter Large 4, which third-party testing calls the strongest model from outside the US and China

Nicknamed "Le Chonk"French AI company Mistral released a public preview of Mistral Large 4 on October 6, officially nicknamed “Le Chonk.” According to AI Weekly, it has 1 trillion total parameters with 49 billion active at a time and supports more than 160 languages, including every official EU language. It became available through the API that day at $1.36 per million input tokens and $4.18 per million output tokens, with weights due on October 27.

Trained in its own European data centersMistral says the model was trained on 3,800 Nvidia Grace Blackwell GPUs in its own data centers in Europe, and that much of the training data was multilingual. Reported scores include about 62% on the coding test DeepSWE and 93% on the security test Cybench. AI Weekly cautions that 62% still trails closed models clustered around 74% and that the figures need third-party validation.

The third-party viewBenchmarking firm Artificial Analysis scores it 38 on its Intelligence Index, says France once again has the most intelligent model from outside the US and China, and puts it on a par with DeepSeek V4.1 Flash. Mistral’s CEO told Reuters the new model is above the Chinese models on certain aspects, including cyber, and that the narrative that Europe cannot compete is not true.

The case for European customersFor European companies and government bodies whose data must stay local, an open-weight model trained in Europe that they can deploy themselves answers questions of compliance and sovereignty, not only performance. Like Reflection’s Beam, the American model released a day earlier, it frames its rivals as China’s open models, not the closed flagships from OpenAI or Anthropic.

▮ SIGNALOpen models from Europe and America are both measuring themselves against Chinese ones, so on the open-weight track there is no longer any dispute about who leads.

❯ Anthropic opens Mythos 5.1 to more vetted security teams and adds a tier for authorized penetration testing

Two programs become oneAnthropic announced on October 6 that it is expanding its Cyber Verification Program and folding the separately run Project Glasswing into it. Verified security professionals can use Claude Mythos 5.1, Opus 5.5 and Sonnet 5.5 with fewer restrictions on security tasks than ordinary users face. Mythos was previously open only to Glasswing members.

Three tiersAccording to SiliconANGLE, the new program has three tiers. The Defense tier covers work such as incident response and malware analysis, is open to individual researchers too, and answers applications within days. The Red Team tier allows penetration testing on systems an organization is authorized to test and accepts organizations only. The top tier has the fewest restrictions, is for organizations cleared to test critical systems such as power grids and telecom networks, and involves US government coordination in vetting.

More than 100,000 vulnerabilities this yearAnthropic also released figures. According to Reuters, 33 Glasswing partners found more than 129,000 verified vulnerabilities between April and July, and Anthropic’s own scanning of open-source software found 5,500 more between April and October, with over 33,000 rated critical or high severity. The company estimates the true impact is at least five times higher.

Looser access, closer watchingDefenders are given stronger tools so they can patch flaws before attackers find them. In exchange, organizations in the program must let Anthropic retain data to monitor for misuse, and even in the Red Team tier, actions such as deploying ransomware are still blocked in real time. Who gets the capability, and how much of it, is decided by Anthropic’s vetting and the US government’s.

▮ SIGNALThe strongest vulnerability-finding model is not sold openly but issued by tier according to who you are, and AI in cybersecurity is developing something like an export-licensing regime for arms.

❯ Google releases image model Nano Banana 2.1 at about half the generation price of its predecessor

Better and cheaperGoogle released its image generation and editing model Nano Banana 2.1 on October 6. Google’s Logan Kilpatrick said on X that the new model produces higher quality images, fixes many bugs from previous models and comes at a new lower price point. It is available in the API, AI Studio and the Gemini app.

Half the priceAccording to The Decoder, Nano Banana 2.1 is based on Gemini 3.6 Flash and replaces Nano Banana 2, released in February. A 1K image costs 3.36 cents, down from 6.70 cents, and a 4K image falls from 15.10 cents to 7.56 cents. The old model will be retired on October 29.

What improvedGoogle says the new model improves on previous versions “across the board”: better visual quality and text rendering, more consistent characters across turns, stronger wide panoramas and infographics, up to 14 reference images, and a connection to Google Search. The Decoder also notes that although it beats the higher-end Nano Banana Pro on some benchmarks, Pro often still produces noticeably better images in practice.

Who cares most about the priceAdvertising, e-commerce and design-tool companies that call image APIs at volume see their costs halve. The model is also wired into AI Mode in Google Search, Google Ads and other products, so Google is using the price cut to make image generation a default feature across its product lines.

▮ SIGNALImage generation is moving from whether a model can draw well to how many cents an image costs, and each price cut makes another set of previously unaffordable uses worthwhile.

❯ Meta and Sierra join Walmart, Stripe and others to launch a Personal Agent Protocol setting rules for AI agents dealing with businesses

An open standardMeta and AI company Sierra announced the Personal Agent Protocol on October 6, developed with Genesys, Instinct, Rocket, Shopify, Stripe and Walmart. According to Sierra’s announcement, it is an open standard anyone can implement, meant to let personal AI agents acting for consumers deal securely with company systems.

The trouble it addressesPersonal agents are AI assistants that book tickets, shop and handle customer service for users. TechCrunch reports that deliberate blocks and anti-bot defenses are getting in their way, leaving consumers caught in the middle. On the business side, the concern is that companies have no reliable way to tell whether a visitor is a person or an agent, and cannot see what an agent does inside their systems.

Authority split between user and businessThe protocol builds on the widely used OAuth authorization standard. An agent can start as a guest or act after the customer signs in. The customer decides whether to grant read-only or write access, and the company decides which capabilities agents may use. Agents can connect through websites, APIs or a company’s own AI agent. A v0.1 specification is due later in October, with extensions such as payments planned.

Who is in the room and who is notMeta has its own personal assistant, Muse, and Instinct also builds a personal agent, so both need businesses to let them in. Walmart, Shopify and Stripe represent merchants and payment providers willing to open up. How far the standard goes depends on whether more large retailers and platforms join. For a business, adopting it adds a sales channel and also hands part of its direct contact with customers to agents.

▮ SIGNALThe first obstacle to agents shopping for people is permission, not technology, and whoever brings merchants into one protocol gets an assistant that can do more.

❯ DeepSeek nears a funding round of at least RMB 80 billion, with Tencent and CATL the largest contributors

Above the original targetAccording to Bloomberg, citing people familiar with the matter, as relayed by The Next Web, DeepSeek is close to securing at least RMB 80 billion (about $12 billion) in a funding round, with Tencent and battery maker CATL the two largest contributors. It originally sought about RMB 50 billion, and based on term sheets already signed, the final total could approach RMB 100 billion.

A lab that came out of a quant fundDeepSeek develops large language models and originated from High-Flyer, founder Liang Wenfeng’s quantitative hedge fund. Its open-source R1 model, released in early 2025, brought it global attention, and it launched the V4 series this year. The report says investor interest exceeded expectations after the reception of its latest model.

The second large raise this yearEarlier this year DeepSeek raised about RMB 50 billion at a valuation of $52 billion to $59 billion, including about RMB 20 billion from Liang himself, about RMB 10 billion from Tencent and about RMB 5 billion from CATL, with JD.com, NetEase and China’s national AI investment fund also taking part. Bloomberg says this round prepares for a listing planned for early 2027. DeepSeek has hired CITIC Securities to prepare a possible Shanghai STAR Market listing, with timing, valuation and size not yet determined.

Why Tencent is adding moreThe Next Web’s reading is that a closer tie to DeepSeek helps Tencent keep up with domestic competitors. Tencent builds its own models and is still a main contributor to both DeepSeek rounds, which suggests it values continued access to the strongest domestic model over necessarily building it in-house.

▮ SIGNALA leading Chinese lab raising tens of billions of yuan twice in a year, mainly from industrial companies and not financial investors, shows the backers are buying model supply as well as equity returns.

❯ Moonshot AI closes its final pre-IPO round at a valuation of about $50 billion and targets a Hong Kong listing in the first quarter of next year

A $50 billion valuationAccording to Bloomberg, citing people familiar with the matter, as carried by Yahoo Finance, Moonshot AI has closed its final private funding round at a valuation of about $50 billion, up from $31.5 billion in the summer. It has confidentially filed for an IPO, targets a Hong Kong listing in the first quarter of 2027, and may raise up to $5 billion.

The company behind KimiMoonshot was founded in early 2023 by Yang Zhilin, who studied at Tsinghua University and worked at Meta and Google. Its main product is the Kimi chatbot, with the latest K3 model launched in July. Investors include Alibaba, Tencent and 5Y Capital.

Fast-growing revenueThe report says Moonshot disclosed annual recurring revenue of $300 million in June, that the figure is now about $1 billion, and that it is expected to reach $2 billion by year-end. Annual recurring revenue projects a year from current subscriptions and contracts and is not realized full-year revenue. Bank of America is overall coordinator, with CICC, Deutsche Bank and Goldman Sachs as sponsors, and early meetings with investors could start this month.

Variables on the way to listingThe report also flags risks: authorities are conducting a data-security probe into Moonshot and DeepSeek, there are allegations about its use of restricted Nvidia chips, and the timeline could shift. Moonshot, DeepSeek and Kling all aim to list in 2027, when Chinese AI companies will face public-market pricing together.

▮ SIGNALRevenue going from $300 million to $1 billion in half a year, and valuation from $31.5 billion to $50 billion, shows Chinese model companies are starting to pitch capital markets on revenue instead of user counts.

❯ Kuaishou’s Kling picks CICC, Goldman Sachs and UBS to prepare a Hong Kong IPO aiming to raise at least $1 billion

Banks selectedAccording to Bloomberg, citing people familiar with the matter, as relayed by AI Weekly, Kling, the AI video business owned by Kuaishou, has picked CICC, Goldman Sachs and UBS to prepare a Hong Kong listing that aims to raise at least $1 billion, as soon as next year. Discussions are ongoing, and size and timing could change.

AI video generationKling is Kuaishou’s AI video generation unit: users enter text or images and the model produces video. Its competitors include ByteDance’s Seedance, Shengshu Technology and PixVerse. The report says Kling entered its commercialization phase after OpenAI’s Sora exited the market.

It raised $2.8 billion in JulyKling closed a $2.8 billion round in July at a pre-money valuation of about $15 billion, with Alibaba, Tencent and Baidu among the investors. Less than three months after that money arrived, IPO preparation is already under way.

Revenue is still a blankAI Weekly points out that Kling has never publicly disclosed revenue, that its Hong Kong filing will carry the first such figure, and that whether the $15 billion valuation holds in public markets is unconfirmed. Who pays for AI video, and how much, is the answer investors most want to see.

▮ SIGNALAI video companies have so far spoken only in valuations, and Kling’s listing will give outsiders their first look at the real accounts of this business.

❯ SpaceX seeks $40 billion of debt, led by Apollo, to buy Nvidia chips

$40 billion for chipsAccording to the Financial Times, citing people familiar with the matter, as relayed by InvestingLive, SpaceX is seeking to raise $40 billion of debt to buy Nvidia AI chips. About $10 billion would be bank loans and about $30 billion investment-grade debt, with Apollo Global Management leading and bond fund Pimco among a small group in talks to provide financing. The transaction is expected to close in 2027.

A rocket company's AI businessSpaceX absorbed Musk’s xAI before its June IPO and runs the Grok model in data centers called Colossus, using Nvidia hardware exclusively. It also rents compute to other AI developers, including Anthropic and Google. The report says Colossus 2 could more than double its Nvidia chip count by December.

AI as a major debt issuerMorgan Stanley estimates AI infrastructure will need $1.5 trillion in external financing by 2028. Large AI-related deals have followed one another in bond markets this week, with Broadcom’s bank group seeking investors for $60 billion of chip financing and Waymo enlarging its own borrowing. After the news, SpaceX shares fell about 1% in extended trading and Nvidia rose about 0.5%.

Nvidia's customers are borrowing to orderThis money ends up as Nvidia’s revenue, but it is repaid from SpaceX’s cash flow from renting compute and running models. Once Apollo distributes the debt, the risk sits with the investors who buy it. If compute rents fall or customers leave, the pressure travels back along that chain.

▮ SIGNALMore and more of Nvidia’s orders are being paid for with customers’ debt, so the strength of chip demand is now tied to how much credit markets are willing to lend.

❯ Waymo raises its first debt financing from more than $3 billion to $5 billion to expand its fleet and pay for compute

Upsized on its first borrowingAccording to Bloomberg, citing people familiar with the matter, as relayed by Crypto Briefing, Waymo, the self-driving company owned by Alphabet, has increased its first debt financing from more than $3 billion to $5 billion. It is an unrated private loan arranged by Goldman Sachs, with Pimco, Blackstone and Sixth Street Partners among the lenders, priced at 5.25 percentage points above the benchmark rate.

500,000 paid rides a weekWaymo operates driverless taxis and currently provides more than 500,000 paid rides a week across more than a dozen US cities. It aims for 1 million a week across 20 cities worldwide by the end of 2026 and plans to launch in London, Tokyo and Munich.

Money for vehicles and computeThe report says the funds will go to expanding the fleet and covering rising AI and compute costs. In February, Waymo closed a $16 billion equity round at a post-money valuation of $126 billion. Until now it had raised money by selling equity.

Why borrow insteadDebt does not dilute existing shareholders, including Alphabet. Waymo already has steady ride revenue, which is why lenders will fund it without a rating, though a spread of 5.25 percentage points shows the money is not cheap. Each doubling of the fleet requires vehicles and compute upfront, and revenue arrives only after the cars are on the road.

▮ SIGNALSelf-driving has reached the stage where future fares can back a loan, which says more than any technical demo about it being a business.

❯ Data center operator DayOne files for a US IPO as first-half revenue more than triples, with nearly 70% coming from one customer

The filing is inSingapore-based data center developer and operator DayOne filed with the US SEC on October 6 to list American depositary shares on Nasdaq under the ticker DODC. According to Mingtiandi, the filing shows first-half revenue of $512 million, up from $151.5 million a year earlier, while its net loss widened from $12.6 million to $77.2 million. The number of shares and the price range are undetermined, and earlier reports put the target at up to $5 billion in proceeds at a $20 billion valuation.

Spun out of GDSDayOne was established in 2022 as the international business of GDS Holdings and took its current name in January 2025, with GDS retaining a 19.4% stake. It builds and operates data centers in Malaysia, Singapore, Indonesia, Thailand, Japan, Finland and elsewhere, leasing space and power to technology companies. In June it closed a $4.5 billion Series C led by Coatue and Hillhouse.

More under construction than in serviceThe company has 962 megawatts in service and 1,328 megawatts under construction, 99.3% of it committed to customers. Completing those projects will cost another $11.4 billion, expected by the end of 2028. Its outstanding debt was $4.4 billion at the end of June. Underwriters include Morgan Stanley, JP Morgan, BofA Securities and Citigroup.

One customer behind nearly 70% of revenueThe biggest risk in the filing is concentration: 69.2% of first-half revenue came from one unnamed technology and short-video company, and the top two customers accounted for 84.3%. Malaysia contributed 87% of revenue. Investors are in effect buying one or two customers’ compute expansion in Southeast Asia, and if those customers change plans, DayOne’s growth and ability to service debt are both affected.

▮ SIGNALIn a data center company’s prospectus, the line most worth reading is often not revenue growth but the share that comes from the largest customer.

❯ Anthropic’s prospectus shows CEO Dario Amodei earned $18 million last year, mostly in stock and options

$18 millionAccording to Reuters, as relayed by The Next Web, Anthropic’s IPO prospectus shows CEO Dario Amodei’s total compensation for 2025 was $18 million, most of it in stock and options. His sister Daniela Amodei, the company’s president, received $16.4 million. Both saw their base salaries double to $1.4 million in July.

Middle of the pack in techThe figure is above Alphabet’s Sundar Pichai ($10.9 million) and Amazon’s Andy Jassy ($2.1 million) and far below the $627.5 million of Oracle co-CEO Clayton Magouyrk. Courtney Yu of executive compensation data firm Equilar said founders “can just live off the wealth of the equity they already own.” The board has also committed to granting Dario Amodei additional restricted stock units, some conditional on tenure and on the company listing.

A pledge to give away 80%The prospectus notes that Amodei and six co-founders have committed to donating 80% of their personal Anthropic shares to charity. Separately, according to The Information, Anthropic reported $540 million in charitable contributions for 2025, against $109 million at BlackRock, the next highest in a comparison with Fortune 500 companies.

What investors look atFor a company targeting a valuation above $2 trillion, executives’ cash pay is not the point. Investors care more about dilution from equity awards and charitable stock matching, and about the pace at which founders sell or donate shares. The prospectus was submitted confidentially in June, with the listing expected this autumn.

▮ SIGNALThe founders’ pay figures are restrained, and the real transfer of wealth is written into the equity and donation arrangements, which will draw more questions after listing than the payroll does.

❯ A developer uses Claude Opus 5.5 to build seven open-source apps modeled on Adobe’s, and says they are far from professional use

Seven repositories in two daysA project called ArtCraft published seven open-source repositories under the GitHub account storytold between September 30 and October 1, each modeled on an Adobe application. According to GIGAZINE, they are PhotoCraft for Photoshop, VectorCraft for Illustrator, FilmCraft for Premiere, and four more corresponding to Lightroom, Acrobat, After Effects and InDesign, all written from scratch in Rust, processing files locally with no subscription.

The developer credits Opus 5.5According to an analysis by OrcaRouter, the project credits Claude Opus 5.5 for its pace, saying about 70 menu items and several major subsystems were completed in a handful of hours using 8 parallel agents. By the morning of October 7, PhotoCraft had 6,691 stars on GitHub.

Claims and realityThe project says all 625 of Photoshop’s menu items are wired to commands, but OrcaRouter points out that a command existing does not mean its output matches Photoshop: of 170 test files, 115 were pixel-matched, and readiness for billable work is estimated at only 25% to 35%. Only PhotoCraft and PrintCraft are in early alpha, with the other five still in development. The project itself states that it is “not yet a Photoshop replacement for daily professional work.”

What it means for AdobeSeven single-app Adobe subscriptions cost about $147 a month in total. One project producing look-alike alternatives in days shows the cost of getting features built is falling fast, while the file compatibility, plugin ecosystem and stability professionals depend on still take long refinement. In the near term, those affected are more likely to be light users who need only basic functions.

▮ SIGNALAI has made it cheap to replicate a program’s interface and menus but has not yet made it cheap to reach professional grade, and that gap is the moat established software companies have for now.