Grok 4.6 rivals OpenAI, Grok Bot launches, Nemotron 4 planned, Twitch adds opt-out
Agents move into operations, models keep scaling, and platforms respond to training concerns.
Top stories
SpaceXAI introduces Grok 4.6, which reports say is on par with OpenAI’s top model. The price is lower. For teams, that could shift model selection and cost profiles. Review benchmarks and constraints before switching.
SpaceXAI launches Grok Bot, a continuously running AI agent—software that autonomously executes task chains. The bots have their own cloud environment. They can log into your apps, tools, and websites. The aim is more execution, not just assistance.
Nvidia is working on Nemotron 4, an open-weight model with freely available weights. The target is around one trillion parameters. Nvidia aims to compete with leading closed models. Reports cite Chinese labs that have already surpassed this scale.
Twitch introduces an opt-out for training generative models. Streams, VODs, clips, chats, and channel content would then no longer be used for future training. This applies to Amazon’s own models. Check your account setting if you don’t want to allow training.
Tools and releases
- Alibaba shows how to serve Qwen3.8-2.4T-A95B with configurable reasoning on NVIDIA GB300 NVL72.
- ShieldFont launches as a defense against scrapers and aims to deliberately poison training data without impairing human readability.
- A new Sign Language to Text model, SL2T for short, brings features for deaf and hard-of-hearing users into practice.
Research
- Researchers from IIT Bombay and Adobe present an inverse language model method that reconstructs prompts from outputs with very high accuracy.
- Microsoft introduces MindTopo, a benchmark that systematically measures the spatial abilities of VLMs—vision-language models.