GoKawiil
Tech News
Researchers pooled data from 137 systematic reviews covering more than 30,000 participants to determine which resistance training variables actually drive strength, muscle growth, power and function gains. They found that most prescription details mattered less than expected, but specific factors like load, volume, and exercise order produced measurable differences in outcomes.
Mecka AI, a startup that pays people to record themselves performing everyday tasks to generate motion data for training humanoid robots, is close to finalizing a new funding round led by Sequoia Capital valuing the company at roughly $500 million. The deal comes just three months after Mecka raised $60 million in a round led by Framework Ventures, though the exact size of the new round remains unknown and terms could still change.
Sony Music and Warner have filed a copyright lawsuit against Anthropic, alleging the AI company scraped and used their song catalogs without permission to train its models. This follows a similar suit from Universal Music Group in January, meaning all three major music publishers are now pursuing legal action against Anthropic.
Meta piloted a Model Capability Initiative that would have captured employees' keystrokes and mouse movements to help train its AI models. Workers pushed back quickly, circulating an internal petition, and the program was ultimately shelved after concerns about privacy and a reported data breach.
A new study found that scent-detection dogs make more false-positive alerts when they have not eaten breakfast before starting work. Researchers observed that dogs on empty stomachs were more prone to signaling a scent that wasn't actually present, compared to dogs that had eaten beforehand.
A report alleges that Moonshot AI's Kimi service secretly forwarded user queries to Anthropic's Claude model and logged the resulting exchanges, apparently to train its own systems on Claude's outputs. Anthropic identified and disrupted the practice, which is why the behavior became public at all.
A user posting on Hacker News says OpenAI's account setting that disables use of chat data for model training has reverted to 'on' after being turned off, despite no action from the user. They say this has happened more than once, and they only caught it by deliberately tracking when they last disabled the option.
Mathematician Andreas Thom says his prior ChatGPT conversations may have fed into OpenAI's recent non-sofic groups breakthrough, echoing a similar complaint from NYU professor Tristan Buckmaster about undisclosed use of his Codex interactions. Thom noted OpenAI's unusually detailed grasp of niche techniques and has asked researchers Sébastien Bubeck and Mark Sellke whether his exchanges with ChatGPT were used as training data.
Hugo Vergnes built a config-driven training framework called little-lm and used it to train a 3.8-billion-parameter language model from scratch on 65 billion tokens, taking 43 hours on eight rented B200 GPUs at a total cost of $998. The model scored 0.384 on the CORE benchmark, outperforming Andrej Karpathy's nanochat d32 model, which cost about the same to train but scored 0.310.
Microsoft and the American Federation of Teachers signed a legally enforceable agreement barring the company from using student or teacher data to train its AI models, except in narrow safety and security cases. The deal also prohibits tracking students, requires human oversight for any AI decisions affecting schools, and mandates transparency with parents and educators about how the tools function. Protections take effect for school districts starting November 1.
A technical essay argues that describing large language models as mere next-token predictors is outdated once reinforcement learning with verifiable rewards (RLVR) enters the picture. Unlike pre-training, which only reinforces sequences already present in training data, RLVR lets models generate novel token sequences and learn from evaluating their outcomes. The piece walks through pseudocode contrasting the two training loops to show how post-training changes what the model is actually optimizing for.
Microsoft submitted 8.2 million Copilot chat logs as part of discovery in copyright suits from The New York Times, the Center for Investigative Reporting, and book authors. The company says only 59,545 logs shared at least 16 words with plaintiffs' news content, with far fewer showing substantial text overlap, and just 10 of 212 books evaluated had any matching passages.