Raw LLM Responses

The dev loop closing was already evident for the past few months. Most of the info from the article are things Anthropic employees have already discussed, and some metrics (\~427x speedup on a task) are directly lifted from the Opus 4.6 model card (the 427x speedup was for a kernel benchmark). The same model cards tell use that yeah, the models are capable of accelerating the engineering process (Dario estimated a 20% speedup in an interview a few weeks ago) but are uncapable of doing the work of an entry-level engineer, which includes research. But how long will that gap persist? I honestly agree with them that it won't last long, RSI in 2027 is plausible in my opinion. So while the models are clearly not there yet, 2027 is still pretty damn soon.