Why a computer scientist believes superintelligence will not be achieved
Why a Computer Scientist Thinks Superintelligence Won’t Happen
AI summary
A computer scientist argues that, from an engineering perspective, superintelligence (ASI) will not be achieved, and that recursive self-improvement (RSI) remains speculative. Anthropic and OpenAI have pushed RSI toward engineering practice, but after GPT-4 the returns from scaling diminish, and the paired data required for post-training is hard to obtain outside mathematics and programming. The author worries that even if an intelligence explosion does not occur, pursuing superintelligence may cause real harm.
Related papers