AI News for May 25, 2026 — 7min.ai

[RESEARCH] Huawei unveils Tau scaling law, targets 1.4nm-equivalent chips by 2031: Huawei presented a new Tau (τ) scaling law designed to deliver 1.4-nanometer-equivalent performance by 2031 without cutting-edge EUV equipment.
[RESEARCH] HBM now 63% of AI chip component costs, up from 52% in early 2024: Epoch AI estimates high-bandwidth memory grew from 52% to 63% of total AI chip component spending between Q1 2024 and Q4 2025, weighted across Nvidia, AMD, Google, and Amazon designs.
[RESEARCH] Lancet study finds AI-fabricated citations in biomedical papers grew 12x in three years: A Columbia-led audit of 2.5M PubMed Central papers and 97M citations found more than 4,000 fabricated references across nearly 3,000 papers, with the rate accelerating after AI tools entered widespread research use in 2024.
[RESEARCH] AutoTTS lets Claude Code discover test-time scaling algorithms that beat human-designed ones: Researchers from UMD, UVA, WUSTL, UNC, Google, and Meta built AutoTTS, an offline environment where a coding agent searches for better test-time scaling control algorithms instead of humans hand-crafting them.
[NEWS] White House approves Anthropic–NSA Mythos deal with no-American-data clause, as template for future AI contracts: White House Chief of Staff Susie Wiles personally approved an arrangement letting Anthropic continue delivering Claude Mythos to the NSA, despite the Pentagon's "supply chain risk" label.
[NEWS] Meta, Google, Amazon, and OpenAI held Vatican lobbying blitz before papal AI encyclical: Months before Pope Leo XIV issued his first encyclical on AI and human dignity, tech giants ran a coordinated lobbying campaign at the Vatican.
[RESEARCH] Xiaomi's MiMo-V2.5-Pro tops Artificial Analysis open-source agentic ranking: Artificial Analysis ranked Xiaomi's MiMo-V2.5-Pro as the world's top open-source model for agentic capabilities.
[OPINION] Uber COO says AI token spending no longer justifies its productivity gains: Uber's operations chief said it is getting harder to justify the company's AI spend because token usage isn't translating into proportionally more shipped consumer features.
[RESEARCH] ByteDance's MMProLong shows Q&A pairs beat OCR for training long-document multimodal models: ByteDance Seed and HKUST researchers found that training multimodal models to transcribe page text actually hurts long-document performance, while question-answer pairs whose answers are buried across pages work far better.
[OPINION] BofA pitches 10x AI productivity boom while admitting current effect is 0.1%: Bank of America's global economics team published a report arguing AI is more powerful than electricity or the internet and could deliver productivity gains 10x larger than what's currently visible.

About 7min.ai · Subscribe to newsletter