Google confirms Gemini 4 is in early post-training as it readies to send TPUs to space
量子位 · wechat · 2026-09-26
Google DeepMind's new chief Koray Kavukcuoglu has confirmed that Gemini 4 has entered early post-training, with a goal of shipping fast. The once-planned Gemini 3.5 Pro was quietly shelved while the team pivots to Gemini 4, and internal safety and testing pipelines are already running on Google's internal coding tools.
Community sleuths have been spotting suspected Gemini 4 checkpoints on Arena — a "pelican on a bicycle" prompt rendered in two very different styles across checkpoints — fueling talk that Google is closing the gap with OpenAI and Anthropic, though skeptics say it can't yet face off with GPT-6 SOL or Anthropic's Mythos lineup.
On hardware, Google is now selling TPU chips directly to challenge Nvidia, and launched Project Suncatcher: a prototype satellite carrying 4 TPUs will launch on a Falcon 9 with Planet to test in-orbit operation. Thermal limits mean chips can run only 15 minutes at a stretch.
Meanwhile, DeepMind chip-tooling engineer Robert O'Callahan resigned — not because AI is moving too slowly, but too fast. He fears cheaper, more capable AI, and says many colleagues privately share safety concerns but don't speak up.
More from Infra
- Google's $15B India AI datacenter draws farmer protests over confiscated land — nordicinst · 2026-09-26
- Splitting DeepSeek-V4.1-Flash across an M5 Ultra and 2× RTX PRO 6000 — harrythunder · 2026-09-26
- Nvidia's SoL-Pi cuts coding agent token usage by up to 49% by optimizing the harness — The Decoder · 2026-09-26
- Running Qwen3.8 Locally on 6x3090s with exllamav3 Hits 80-120 tok/s — takoulseum · 2026-09-26
- Gemma 4 31B UD-Q8_K_XL Suddenly Hits "Exceeds Shared Memory" at 68k Context — Few_Professional6859 · 2026-09-26
- Meta Open-Sources Muse Glimmer: 30B Agentic Model That Fits a 24GB Consumer GPU — bibryam · 2026-09-26