Hacker News new | past | comments | ask | show | jobs | submit
Claude got really good at ML Engineering recently. If they can't get the current models to RSI, they aren't trying.

How much juice is left in that tank after 5 years of intense focus on optimizing LLMs is another matter. If we are on top of a local architecture optimum, speeding up the hill-climbing will deliver underwhelming results until someone discovers the next hill.