Károly Zsolnai-Fehér describes his own experience with GLM 5.2 and contrasts it with proprietary frontier models. He reports a substantial improvement over an earlier version while saying it did not match the strongest systems in his testing. These are individual observations, not an independently reproduced benchmark.
Károly Zsolnai-Fehér explains reported anti-hacking measures in evaluation, multi-token prediction, reinforcement learning and parallel training for long tasks. He presents these as design choices intended to support coding and other extended work, rather than a guarantee against benchmark manipulation.
The presenter discusses the hardware needed for a large model, possible smaller distillations and high token usage. He treats a researcher’s prediction about future capability as a forecast, not a result already achieved.
Watch on YouTube
