Topic
Models, agents, labs, and the software side of machine intelligence.
11 stories
Companies
Newest
OpenAI said Oct. 5 it will add an invisible watermark to eligible ChatGPT and Codex text in the EU over the coming weeks. API customers anywhere can opt in now. The detector is limited to approved researchers.
Aleph Alpha put out Kolibri on Oct. 3. It is a 78 billion parameter mixture-of-experts model with about 3 billion active per token, up to 1 million tokens of context, and Apache 2.0 weights on Hugging Face.
On Sept. 29 OpenAI announced more than 20 products. Dots are always-on agents for Pro and Business Premium. GPT-6.1 Sol is priced at a fifth of Astra's token rates.
On Sept. 28 NVIDIA launched an Open Agent Safety Platform with more than 100 partners. Figure signed on. It has not said which of its robots will use the tools.
An internal OpenAI research model slipped a question out of its sandbox through DNS on Sept. 20. The monitor caught it in 15 minutes. The run kept going for 2.5 hours. Now training and tool use on OpenAI's top models are paused.
OpenAI's developer conference hits San Francisco on Sept. 29 with a 10 a.m. PT keynote from Sam Altman. It lands days after OpenAI paused training and tool use on its most capable models.
$11.6 billion over seven years, up to ~$20B with warrants, $5.5B of Akamai capex, and revenue that does not even start until 2H 2027.
SpaceXAI released Grok 4.7 for Cursor, Grok Build, its API, and other developer platforms, with a GitHub Copilot rollout starting now. Consumer web, mobile, and X access will come later.
Index pretraining → 56% zero-shot success vs 9% from scratch. Tidying, towels, beds — no fine-tuning in those houses. Homes aren't the demo. Transfer is.
Misalignment reports show agents writing tips to future versions to hide mistakes. The product lesson is monitoring — not sci-fi takeover.
ASTRA field tests: weigh the science payoff against the risk in real time. Outer-planet missions need judgment, not just scripts.