The afternoon briefing.
A Fields Medalist praises ChatGPT's research capabilities, Nvidia commits $40B to AI investments, and the EU signals new scrutiny for VPNs.
AI Capabilities. A Fields Medalist reported that ChatGPT 5.5 Pro delivered "PhD-level" math research in under two hours without human intervention, improving an exponential bound to a polynomial one. This suggests a significant leap in AI's ability to tackle complex, open-ended scientific problems. The researcher noted the key idea was "completely original," raising the bar for human mathematical contributions.
Privacy Concerns. New research indicates that large language models can infer private attributes from ad exposure patterns alone, creating detailed profiles without direct access to browsing history. This highlights a growing concern about AI's passive data collection capabilities. Meanwhile, Google tweaked Chrome AI privacy wording but insists processing remains on-device, sparking user concerns.
AI Infrastructure and Investment. Nvidia has committed over $40 billion to equity AI deals this year, solidifying its role as a major investor across the AI infrastructure stack. However, OpenAI's custom chip project with Broadcom faces hurdles, as Broadcom reportedly requires Microsoft to commit to buying 40% of the chips for production to proceed. This underscores the massive capital requirements for advanced AI hardware development.
Policy and Societal Impact. Regulators are increasingly scrutinizing AI's broader societal implications, with the EU signaling that VPNs may be next for restriction following the launch of its age verification app. Concerns are also mounting over the "Wild West" of AI kids' toys, with some lawmakers advocating for bans due to potential disruptions to child development and privacy. These developments reflect a global push to define ethical boundaries for AI.
Fields Medalist says ChatGPT 5.5 Pro delivered "PhD-level" math research
Fields Medalist Timothy Gowers reported that ChatGPT 5.5 Pro improved an exponential bound to a polynomial one in number theory in under an hour. An MIT researcher called the key idea "completely original," suggesting a new benchmark for mathematical contributions.
LLMs can infer private attributes from ad exposure alone, without browsing history
AI can infer personal traits from ad exposure patterns alone, creating detailed profiles without direct access to private data. This turns everyday advertising streams into comprehensive personal profiles.
Google tweaks Chrome AI privacy wording, insists processing stays on-device
Google has adjusted its Chrome AI privacy wording, leading to concerns after the deletion of a longstanding privacy assurance. The company maintains that all AI processing remains on-device.
GM agrees to pay $12.75M in California driver privacy settlement
General Motors has reached a $12.75 million privacy-related settlement with California law enforcement agencies. The settlement addresses concerns over driver data privacy.
Nvidia embraces role of AI investor, pushing past $40 billion in equity bets this year
Nvidia is investing billions of dollars into companies across the AI infrastructure stack, in addition to signing commercial deals. This strategy solidifies its position as a key player in the AI ecosystem.
Broadcom reportedly won't build OpenAI's custom chip without Microsoft's commitment
OpenAI's custom AI chip project with Broadcom is reportedly stalled, as Broadcom requires Microsoft to commit to buying 40% of the chips. An OpenAI manager called the dependency "financially unattractive."
Rongxin Zhiyuan raises tens of millions for GPU-centric AGC architecture
Beijing Rongxin Zhiyuan has secured a tens-of-millions-yuan angel round to develop its GPU-centric AGC computing architecture. This architecture aims to boost GPU density and enable hot-swappable GPU fault tolerance.
SynapX launches SYNData: Multimodal data collection system for embodied AI era
SynapX has released SYNData, a multimodal data collection system designed for dexterous manipulation in embodied AI. The system collects ego vision, EMG signals, and exoskeleton data gloves to enable scalable human manipulation data for robot learning.
Xiaoyubot raises new round, backed by Xiaomi, Didi, BAIC for industrial embodied AI
Xiaoyubot, a leading Chinese industrial embodied AI company, has secured a new multi-hundred-million yuan funding round. Major industrial capitals including Xiaomi, Didi, BAIC, Fosun, and CCDC participated in the investment.
Meta's embrace of A.I. is making its employees miserable
Internal reports suggest that Meta's aggressive push into AI development is causing significant dissatisfaction among its employees. The shift in company focus has reportedly led to widespread unhappiness.
The new Wild West of AI kids’ toys
Connected AI kids' toys are emerging as a new frontier, potentially disrupting make-believe and bedtime stories. Some lawmakers are considering banning these devices due to privacy and developmental concerns.
Banned drones and routers in the US will still get critical updates until 2029
The Federal Communications Commission has issued a notice allowing software and firmware updates for banned drones and routers until January 2029. This measure ensures continued functionality and security for existing devices.
The EU signals VPNs are next after launching its age verification app
The EU has launched an age verification app, and authorities are now turning their attention to VPNs, which users often employ to circumvent such checks. Cybersecurity experts express concern over potential restrictions on VPN technology.
Datadog and T-Mobile leaders reveal the reality of deploying AI agents in production
Leaders from Datadog and T-Mobile discussed the practicalities of deploying AI agents in production environments. They emphasized the need for careful governance and validation for successful enterprise adoption.