Tag
Articles about AI benchmarks from podcasts, articles and papers, with links to the original sources.
RoboCurve reports that GPT-6 Astra placed a block in a bowl in 19 of 20 trials, up from 8 of 20 for its predecessor. But it completed only two puzzle insertions—the same as the older model—in a 120-trial evaluation that separates basic manipulation gains from reliable precision.
A controlled comparison of 2019–2025 datasets and training recipes reported compute-efficiency gains of 12-fold from data improvements versus 3.7-fold from model recipes. The Moonshots panel explored the business opportunity—and used BloombergGPT’s reportedly short-lived advantage to question whether owning unique data is enough.
DeepSeek says V4.1-Flash cuts the high-bandwidth memory needed for its cache to a quarter of its predecessor’s requirement. The Moonshots panel sees an opening for companies to build on efficient open-weight models—but a smaller cache is not the same as a small model.