OpenAI Pulls $1M From Caltech Mathathon as 25 Fields Medalists Object

Leaked iOS 27 build shows Siri AI delegating tasks to ChatGPT and Claude

Leaked private frameworks for iOS 27 and macOS Golden Gate show Apple building deep integration between Siri AI and third-party models including OpenAI’s ChatGPT and Anthropic’s Claude, per an X post by pdfu. The frameworks include a model delegation feature where Siri passes requests to another agent, demonstrated by Claude creating a reminder in the Reminders app while Siri acts as intermediary; requests needing proprietary Apple system data stay with Siri AI. A more robust app extension replaces Apple’s Siri AI model entirely with a different LLM: using a model manager service, pdfu ran GPT-5.6 with Siri’s logged information to search emails and send iMessages through the Siri UI and voice. Apple, OpenAI and Anthropic did not immediately comment. The EU’s Digital Markets Act, which targeted Siri AI and Gemini, has shaped how Apple rolls out updates; the revamped assistant was pitched as a key iPhone 18 Pro selling point at Apple’s Surprise and Shine event.

Siri AI Will Have Deep Integration With ChatGPT and Claude Models, Leaked Build Shows →

Microsoft publishes a code of conduct for its MAI models

Microsoft AI published a code of conduct for its MAI models, setting values, behavioral limits and how to handle conflicting goals. The document sits at the top of the rulebook, guiding training, technical controls and evaluation, while operator rules and user requests rank below it; Microsoft does not currently train models on the code. After a six-week public consultation, a revised version is due around the end of 2026 and will guide development starting in 2027. The code applies to Microsoft’s own models, not automatically to third-party models in Microsoft products. The core rule is that human control comes first, and Microsoft says it is willing to give up generality, autonomy or performance to keep it. AI chief Mustafa Suleyman told The Information: „If it isn’t safe we shouldn’t build it.” The code bans „Neuralese” or other unreadable communication in reasoning or between AI systems, and rejects claims that models have consciousness, feelings or rights. It follows calls to slow development from Anthropic’s Dario Amodei, backed by Satya Nadella, OpenAI, xAI and Meta.

Microsoft’s AI rulebook: readable thinking, no inner life, and definitely no rights →

OpenAI withdraws $1M Caltech Mathathon sponsorship after mathematicians object

OpenAI withdrew its $1m sponsorship of the student-run AI maths event Mathathon at Caltech on Thursday 10 September, taking back $1m in credits. Dan Roberts, an OpenAI research lead, wrote on X that the company recognizes AI’s rapid progress in mathematics is disruptive and wants to engage the maths community on integrating the technology. Anthropic put in the other $1m and has not withdrawn. A coalition of current and former Caltech mathematicians circulated an open letter from 8 September, published 10 September with 771 signatories, calling the work slop mathematics and arguing research mathematicians end up verifying and correcting AI companies’ results uncompensated and uncredited. On 11 September, 25 Fields Medallists published a declaration titled A Severe Misalignment of AI in Mathematics, warning that mass production of true or false statements could destroy fertile ground. Organizers, five undergraduates, conceded several concerns are valid, added an arXiv requirement and expanded the educational track, but did not cancel. The event runs 30 October to 1 November.

‘Slop mathematics’: OpenAI walks away from a Caltech AI maths contest →

DeepSeek V4.1 cuts memory footprint 437x at 890 bytes per token

DeepSeek V4.1 achieves a 437-fold reduction in memory usage compared with earlier iterations, operating at 890 bytes per token while supporting a million-token context. The model divides its 40 transformer layers into encoders, which compress the input context into a compact representation, and decoders, which process that compressed representation to minimize redundant memory storage. Two architectural mechanisms underpin the gains: Sliding Window Attention recalculates local context dynamically, removing the need to persistently store intermediate states, and Compressed Sparse Attention (CSA2) lets transformer layers share sparse indices, reducing redundant computation. Memory states are compressed into a 4-bit format using 4-bit Compression (E2M1), with dynamic scaling preserving precision in narrow numeric ranges. The savings come with trade-offs: sliding window attention and CSA2 increase computational overhead, especially during inference, making performance depend on the serving architecture’s ability to handle concurrent requests. Deployment on optimized hardware is required, and independent validation is still needed.

How DeepSeek V4.1 Achieves a 437x Memory Footprint Drop →

AI lion images helped NCA map and remove 200 TikTok smuggling accounts

A network of more than 200 TikTok accounts linked to an Albanian organised crime gang used about 900 videos to promote illegal routes to the UK, reaching thousands of viewers. The network was uncovered after a National Crime Agency investigator spotted a suspicious profile with a Big Ben photo, no posted content and 957 followers. TikTok investigators used the gang’s repeated AI-generated images of lions as a „tell” to map the interconnected accounts; all 200 were removed. Mike Hulett, head of the NCA team combating online organised immigration crime, said the adverts are „a key enabler” of smuggling. Home Office data shows 80 per cent of small boat arrivals interacted with social media posts during their journey, which Hulett said is likely an underestimate because it was recorded 18 months ago. Smugglers now spread content across accounts, misspell words and move conversations to encrypted apps. TikTok removed 10,000 UK videos breaking human exploitation rules in the first half of this year, just 0.17 per cent of content blocked there.

How an AI image of a lion brought down an Albanian smuggling gang →