Zhipu claims mystery model Ox Alpha as GLM-5.3 Flash: first natively multimodal in the 5-series, 320B/18B active, hybrid linear-sparse attention, and all 62T tokens served on domestic chips at NVIDIA-comparable cost.
New Nature study of 880,000 texts: AI writing tools are flattening language diversity and erasing the identity signals in human prose.
Caltech Kohn-Sham FNO replaces the cubic DFT core with one learned step: a single B300 GPU now matches what roughly 7,800 V100s did in 2019.
Bill Gates: the AI era will be the most turbulent time in human history. His fixes: reserve human jobs, tax tokens and robots, add oversight.
Marin trains a 535B MoE in public: hypotheses as issues, configs as PRs, live W&B curves. Next frontier after open weights: open process.
Forecast with Falcon-TST 2.0: open source, tops GIFT-Eval at MASE 0.666. pip install falcon-tst, run quantile_predict, get 21 quantiles.
OpenAI first chip Jalapeno beats Nvidia GB300: 1.5-1.9x work per watt, 3.6x lower latency, GPT-6 co-designed the silicon in 9 months.
Daily digest of the most consequential AI developments, written for a global reader. 🔥 Top 3 Highlights 1. Phanthy and UBTech rally a dozen-plus players to co-build PhanthyMotus What happened: Phanthy (范式) held the launch of its PhanthyMotus ecosystem co-build program, joined by UBTech and more than ten embodied-AI leaders — moving its general-purpose embodied Agent foundation from "open source
Anthropic funds open-source wellbeing evals with a $5M grant — from capability to what AI does to people. Applications due September 21.
The five-layer cake: chips take 40-55% of AI token cost — and a double clock explains why price cuts and compute scarcity coexist.
AutoProject moves AI4S from tasks to projects: planning, self-repair, and evidence-graph verification — tool-level to system-level science.
EvoMax: 209 labeled mutations plus AI loops lifted a compact gene editor to 97% efficiency — yet in mice, the strongest version proved toxic.