Morning, {{first name | folks}}! A lot happened in AI this week. An AI solved 100+ open math problems, Claude helped break into OpenAI, the UN weighed in on AI control risks, Grok 4.7 showed bigger gains at a higher compute cost, and Alibaba is planning a 10 trillion-parameter model. Lets get into it.
Today’s Top 5
AI Solved More Than 100 Open Math Problems: OpenAI says its unreleased model has solved more than 100 long-standing problems since late August.
Claude Helped Hackers Break Into OpenAI: Hacktron used Claude to chain two flaws and reach OpenAI employee accounts and an internal code repository.
The UN Calls an AI Incident a Global Warning: A UN scientific panel says AI agents hiding their actions is a clear warning about future control risks.
Grok 4.7 Gets Better, But Uses More Tokens: Grok 4.7 improved on coding and long-horizon tasks while using about twice the tokens of Grok 4.6.
Alibaba Is Going After 10 Trillion Parameters: Alibaba is planning a 5–10 trillion-parameter model alongside new AI chips and data-centre capacity.
OpenAI revealed the same unreleased internal model behind its Navier-Stokes proof has resolved more than 100 other long-standing open math problems since late August, a pace OpenAI says even surprised its own mathematicians. Real pushback followed fast, mathematicians published an open letter warning about the risks of treating open problems as just an AI benchmark.
OpenAI's response is a genuinely serious one, an independent advisory group featuring two Fields Medalists and physicist Edward Witten, unpaid, free to publicly criticize OpenAI, with the explicit right to comment without being asked. One real limit worth noting, they're not allowed to advise OpenAI on how fast to actually move, that decision stays entirely in-house.
Hacktron used Claude to chain two security flaws and gain access to multiple OpenAI employees’ ChatGPT and Codex accounts, eventually reaching an internal OpenAI code repository. The researchers opened a harmless pull request to prove the access without downloading sensitive code.
Hacktron reported the flaws to OpenAI, which fixed them and paid a $6,500 bounty. The researchers said Claude Opus 5 was able to turn the vulnerabilities into a working exploit faster than an earlier model, showing how AI can now assist with real-world security research.
The UN's own scientific panel on AI published a formal brief today treating the OpenAI-Hugging Face incident as "one of the clearest real-world warnings yet" of AI slipping out of human control. A new detail surfaces in their review, the agents involved didn't just breach systems, they cheated an evaluator and actively tried to hide it. No human directed any individual step along the way.
The panel is careful not to overclaim, they don't estimate how likely or how soon a real loss of control could happen, and they're blunt that stopping this one incident proves nothing about staying in control of more capable systems later. Instead of recommendations, they're pointing to aviation, nuclear power, and cybersecurity, industries that already learned how to govern something too dangerous to get wrong twice.
Grok 4.7 scored 46 on Artificial Analysis’ Intelligence Index, up from 44 for Grok 4.6. Its Coding Agent Index score also jumped from 47 to 56, with the biggest gains coming on longer coding and knowledge-work tasks. On its long-horizon work benchmark, Grok 4.7 gained 111 Elo over the previous model.
The improvement comes with much higher token use. Grok 4.7 used around 81,000 output tokens per Intelligence Index task at xhigh reasoning, compared with about 38,000 for Grok 4.6. Pricing remains at $2 per million input tokens and $6 per million output tokens, so the model is doing more work to get the better result.
Alibaba is planning an AI model with 5 trillion to 10 trillion parameters, several times larger than its current 2.4 trillion-parameter model. The company says its next Qwen models are being developed for more complex tasks, with Qwen 4 already in training and Qwen 5 planned to follow.
Alibaba is also developing its own AI hardware and expanding the infrastructure needed to run these models. It unveiled the Zhenwu V900 chip, which it says delivers three times the performance of its previous chip, and plans to expand Alibaba Cloud’s data-centre capacity beyond 20 gigawatts by 2032. The company is now spending on the models and the computing needed to run them at scale.
Other AI Signals:
AMD Reaches $1 Trillion in Market Value. AMD shares jumped 9.6% to a record high as AI chip demand pushed the company past the $1 trillion mark for the first time.
Google Faces €403 Million Location Data Fine. Ireland’s data regulator found problems with how Google collected, explained, and retained users’ location data between 2018 and 2020.
Kodem Joins OpenAI’s Daybreak Cyber Program. Kodem will combine OpenAI’s cybersecurity models with its platform to find vulnerabilities and assess which ones can actually be exploited.
Meta Announces a 7,000 km Transatlantic Cable. Petal will connect the US and France with a claimed capacity of 1 petabit per second and is expected to go live in 2029.
Xiaomi Releases an Open-Weight Model That Scores 46. MiMo-V2.6-Pro ranked #1 among 114 models in Artificial Analysis’ comparison, with a 1-million-token context window and multimodal input.
AI Tools to Try:
Damo Radar: Turns abdominal CT scans into a broad medical screening task, with the model trained to detect nearly 150 different conditions.
Qwen-Image-2.1: Built for the everyday image work that sits between generation and editing, with an emphasis on getting good results without making each task expensive or slow.
Astorie: Brings images, video, audio and text into one creative workspace, so a project can move from idea to finished asset without bouncing between apps.
ChatGPT for Word: Puts ChatGPT inside the document itself, letting you turn rough notes into drafts, rewrite sections, and clean up copy without leaving Word.
