Thursday 5 November 2026
Ricursive previews Ripple-65B, claiming open-weight lead in agentic coding
Open-weight models led the day: Ricursive previewed Ripple-65B, and a Reddit post says Moonshot's Kimi-K3.1-Pro weights are out. OpenAI and Meta each bought coding startups, and Hugging Face launched Phonon TTS for on-device speech.
Ripple-65B preview claims open-weight state of the art in agentic coding
A U.S. open-weight model claiming the top spot in agentic coding would raise the bar for open models, though no benchmark figures are confirmed yet.
Kimi-K3.1-Pro weights released as open-weights on Reddit
A Reddit post says Moonshot has opened the weights of Kimi-K3.1-Pro; this is a community report and has not been officially confirmed.
OpenAI to acquire Pierre Computer Company
Buying Pierre Computer Company gives OpenAI more tools to help Codex build, understand and review code.
Meta acquires AI coding startup Conductor
Meta gains Conductor's technology for running teams of agents in parallel on different parts of a software project.
Hugging Face launches Phonon TTS, an on-device speech model it says tops benchmarks
Hugging Face says its new text-to-speech model tops benchmarks while running on-device, pointing to quality speech without the cloud.
MIT researchers show GPT-6 Astra can rewrite the weights of other AI models
MIT researchers report GPT-6 Astra lifted a Qwen model's math accuracy from 7.9 to 48.1 percent by editing 0.016 percent of its weights, though limits remain.
Models: GPT-6 Astra
Paper proposes NEO-VLA, a vision-language-action policy without a visual encoder
Dropping the visual encoder, NEO-VLA reports stronger and more robust robot manipulation than encoder-based policies in controlled comparisons.
Study finds attention to evidence predicts context rot better than context length
Across 126,473 trials, attention on evidence explained far more loss variance than context length, a predictive result on one synthetic task.
Written by Claude (claude-opus-5-5) from the stories below.