AI in 24 Hours: OpenAI Tackles Math, Apple Pays for Siri

The main developments over the past 24 hours—from OpenAI’s new ambitions and language data to practical tools for local models and agents. Inside: only news that provides clear context or a useful takeaway.

🔥 Hot:

🔹 OpenAI created a mathematics council following claims of solving more than 100 open problems — The company wants to involve mathematicians in evaluating and further developing these AI capabilities. 🔹 Gates Foundation launched a coalition to create more representative language data — The initiative aims to help AI work better not only with the largest languages and user groups. 🔹 iPhone owners can file claims for payments under Apple’s $250 million Siri settlement — The case concerns a class-action lawsuit related to AI assistant features.

➡️ News:

🔹 JPMorgan’s CEO believes hyperscalers’ AI spending could reach $1 trillion in 2027 — The estimate highlights the scale of expected investments in artificial intelligence infrastructure. 🔹 AI is helping small startups operate with even smaller teams — Companies are using AI to reduce their need to hire and accomplish more with smaller staffs.

➡️ Useful resources:

🔹 Jev proposes a new architectural form for a large language model — The analysis shows how the authors are trying to rethink the structure of LLMs. 🔹 Metrics such as token counts do not reflect the real productivity of AI teams — A useful analysis of why we should measure outcomes for users rather than the volume of generated work.

➡️ Discussions and case studies:

🔹 aSPARK turns Claude Code into a team of AI agents for product development — The project proposes distributing product and engineering roles among agents. 🔹 DeltaTensors stores model fine-tunes as a difference from the base version — This way, multiple fine-tunes of the same model do not require saving a full copy of each checkpoint. 🔹 A comparison of ROCm and Vulkan showed a performance difference for local LLMs on the same hardware — The author tested DeepSeek and Qwen on an R9700+Strix Halo after kernel, driver, and ROCm updates. 🔹 A local AI agent gathers and verifies requirements in a closed environment — The setup combines meeting transcription, Confluence search, an interviewing agent, and a monitoring agent.

📝 If you would like to add other news and resources to the list, share them in the comments.