AI / Compute
Machine learning, inference hardware, open-weight models — and the reality check of their promises.
Topic weight 28%
Last week 24%
UP
Reason High vote and prediction activity around open-weight models.
Articles
3-
Open-weight 7B reaches frontier level on math eval
Distillation paper reports 71% on the MATH benchmark — peers contest the evaluation harness.
Reality 46Hype 71 -
Edge inference on 2nm node at < 0.4 W
New mixed-precision engine runs 7B inference locally on an ARM chip below 0.4 watts.
Reality 68Hype 38 -
Agentic benchmarks diverge sharply by eval harness
Three independent setups produce results that differ by up to 28 points for the same model.
Reality 58Hype 42