flambreaux

The argument here has some big flaws: 1. It points that that *at the moment* the AI-improving-AI feedback loop is much to weak to explode. But then erroneously infers that it probably won't explode in the future. The jump from GPT-2 to GPT-3 increased AI researcher productivity by a *tiny tiny* amount. Probably much less than 0.1%. Ramez would have concluded that there was absolutely no chance of an intelligence explosion. The feedback loop would need to increase in size by >1000X! But the jump from GPT-5 to GPT-6 increased AI researcher productivity by much much more. By >1% and plausibly by >10%. Ramez concludes there won't be an intelligence explosion bc the feedback loop would have to increase in strength by 5X. But it's already increased in strength by a lot! When AI becomes more capable over the next year, it will increase in strength much more. 2. Ramez claims that "Keeping Up the Pace Takes Exponentially More Resources", as if this means progress is unsustainable. But "keeping up pace" is delivering exponentially growing revenues. Exponential growth in inputs -> exponential growth in outputs. That's a feedback loop that can totally explode. Tbc, @ramez argument is made in good faith and is in many ways well informed and well argued. But I think his conclusion is wrong!