AI Updates on 2025-09-26

OpenAI launches GPT-5 Pro which is generating nontrivial new mathematics and solving problems earlier models couldn't handle, with Mark Chen noting it can automate months of student work for physicists and mathematicians @a16z

Anthropic reports dramatic revenue growth from $87 million at the start of 2024 to over $5 billion run-rate in August 2025, with 80% of consumer Claude usage coming from outside the United States, particularly strong in South Korea and Australia @AndrewCurran_
China bars its major tech companies from buying NVIDIA chips, signaling sufficient progress in domestic semiconductors to break away from US dependence, with DeepSeek-R1-Safe model trained on 1000 Huawei Ascend chips demonstrating system-level design approach @AndrewYNg
Developer reports "wasting tokens" on a problem in standup, highlighting how AI cost considerations are becoming part of everyday development workflow and decision-making @GergelyOrosz
Perplexity Search API claims superiority over Google for LLM use cases, scoring higher on Simple QA/HLE benchmarks since Google optimizes for ad/link click ranking rather than utility as search snippets for AI @AravSrinivas
Rumors suggest OpenAI and Google are both launching "AI native" browsers soon, with owning the primary computer app being critical for distribution, data, and easy-to-use automations @deedydas
Data center capacity demand projected to increase more than 3x globally by 2030 according to McKinsey research @a16z

AI Now Institute advocates for industry-independent scrutiny of AI benefits and risks claims, and for a people-centered AI sovereignty agenda at UN Global Dialogue on AI Governance @AINowInstitute
François Chollet predicts 2026 will be the year companies market their products as "AI free" following the 2023 trend of "AI powered" marketing @fchollet
Gergely Orosz criticizes the vision behind Vibes product launch, describing it as promoting people glued to phones scrolling through AI-generated content infused with ads as a "terrible future" @GergelyOrosz
Simon Willison reports classic prompt injection exfiltration attack against Salesforce Agentforce, now fixed with Trusted URL allowlists enforcement starting September 8, 2025 @simonw
MIT Technology Review reports US investigators are using AI to detect child abuse images made by AI @techreview

NVIDIA and ParaboleAI achieve 1,000x speedup in industrial optimization, reducing processing time from 10 hours to under 1 minute using causal AI on NVIDIA GH200 Grace Hopper with Gurobi @NVIDIAAI
Exelon and Deloitte built OptoAI autonomous drone solution for grid asset inspection powered by NVIDIA Jetson and Omniverse, achieving 100x increase in operational efficiency and expedited defect identification @NVIDIAAI
Perplexity launches Comet shopping agent that can handle requests like "Buy me three books recommended by Druckenmiller" and execute the purchase automatically @AravSrinivas
Google expands agentic capabilities in AI Mode for finding restaurant reservations to all users opted into Labs in the US @rmstein
MIT develops photonic processor chip that performs deep learning at the speed of light, potentially giving edge devices new capabilities for real-time data analysis @MIT

OpenAI releases GDPVal benchmark measuring AI performance on tasks that make up everyday jobs across the entire economy, with models approaching parity with humans on expert-level tasks averaging 7 hours of work @emollick
Research paper demonstrates inadequacy of older public benchmarks for medical AI, showing models are memorizing or using heuristics for answers rather than genuine understanding @emollick
OpenAI confirms their models solved ICCP programming challenges using code execution sandbox but no internet access, clarifying the tools available during competition @simonw
Alexandr Wang clarifies SweBench Verified number refers to TTS pass@1 performance metrics in response to questions about benchmark results @alexandr_wang