ByteDance reveals a training run roughly ten times the scale of GPT-4-era models; the man who built Google's AI era is starting something new.
Sources say ByteDance's next foundation model is being trained at up to 10 trillion parameters — larger than any publicly disclosed training run and roughly an order of magnitude above the GPT-4 range. The model is expected to power future Doubao and enterprise AI products.
Jeff Dean — Google Brain co-founder, architect of TensorFlow, TPUs, and the scaling era of transformers — is departing along with other senior researchers to launch a new AI company.
AMD acquired Taalas, a startup that etches model weights directly into processor die rather than loading them from memory bandwidth-constrained DRAM, to reduce inference latency and power draw.
OpenAI rolled out reasoning improvements to GPT-5.6 Sol in ChatGPT and extended access to GPT-5.6 Luna — its faster, lighter model — to free-tier users who previously topped out at older models.
Google DeepMind's WeatherNext model achieved state-of-the-art accuracy forecasting tropical cyclone tracks up to 10 days out, outperforming operational numerical weather prediction on intensity as well as track guidance.
Google Maps is rolling out agentic capabilities that let users complete multi-step real-world tasks — restaurant orders, hotel reservations, venue bookings — without leaving the app.
Anthropic is hiring hardware engineers to design custom silicon for AI inference, with the goal of reducing dependence on Nvidia and controlling its own compute stack.