Model Training

Tutor • Series B

Tutor, which is training a frontier robot foundation model and builds AI warehouse robots, raised a $75M Series B led by a16z, taking total funding to $117M.

MIT researchers demonstrated that OpenAI's GPT-6 Astra can read and rewrite the weights of other AI models without traditional training, boosting a Qwen model's math accuracy from 7.9% to 48.1% by changing just 0.016% of its weights. The approach could simplify AI optimization but the article notes it still hits clear limits.

Research • The Decoder

DeepSeek reportedly plans to train its next-generation V5 model on Huawei Ascend chips, according to STAR Market Daily citing industry sources. The reported plan would extend Ascend's role from inference to the compute-intensive training of a new model.

Infrastructure • TechNode

Ricursive Intelligence released Ripple-65B Preview, which it says sets a new U.S. open-weight state of the art on agentic coding. The lab trained the model from scratch, optimizing architecture for its own training and serving platforms.

Models • Hacker News

A new study from NYU and UT Austin identifies six linguistic patterns that frontier AI models systematically overuse, making their text exhausting for humans to read. The researchers attribute this to training on code and math, which prioritizes correctness over readability.

AI Research • The Decoder