Researcher Notes Issues With AI Agent Models

Alexander Doria, online handle of Pierre-Carl Langlais, co-founder and CTO of Pleias, posted on X that it is not the greatest period for agents. He stated that Opus and Fable show 2-bit quant crashout. He added that GPT models are so RL overcooked they are hardly usable. He also wrote that Chinese models remain under the US capability ceiling. The post is presented as his assessment of recent developments in the field.