| ▲ | zug_zug a day ago |
| I mean they're obviously not stochastic parrots, I think most of us engineers have known this for a couple years now, though I do admit I saw a pretty obnoxiously vocal minority on the point here on HN. But at the same time, where's the RSI? If the RSI is there then make the company profitable already [though I suppose that means letting go of all the staff? But if it's improving itself, not sure what humans are in the loop for]. Genuinely curious when AI will be able to solve real-world problems, what would 20,000 hours of AI say on the topic of whether Epstein killed himself? |
|
| ▲ | saulpw a day ago | parent | next [-] |
| RSI seems to be here? It may not be the glorious 100% autonomous RSI that sci-fi authors predicted, but 90% autonomous RSI seems to be what's creating Claude Code and putting inference models on FPGAs and whatever happens when I say "Claude, plan it. Claude, implement it. Claude, fix it. Claude, make it faster." I mean it's not perfect but it's remarkable how much you can just...tell it to make itself better? |
| |
| ▲ | zug_zug a day ago | parent [-] | | I mean it's definitely remarkable. But I don't think it's RSI. My mental model of RSI is essentially isn't quite asymptotic but exponential. If a system can make itself 5% smarter every week then it would be 12x as smart in one year at which point it's not clear to me if we even need governments or markets because we have a being that can direct everything. In short I don't believe there's any world where there's true RSI intelligence and anthropic returns green paper to "shareholders" so they can buy pretty cars as status symbols. | | |
| ▲ | saulpw a day ago | parent [-] | | Perhaps your mental model of [the effects of] RSI is mistaken, though? It's not an inherent part of RSI that there will be consistent 5% improvements to be made. Even with domains that have been 'solved' by AI, like Go (board game) or OCR, the improvements hit a point of diminishing returns. Like at some point there aren't any 5% improvements left, so you make 1% improvements, and then .1% improvements, etc. And that is also just providing an explanation for the sake of accepting your premise, that we're not experiencing exponential improvements. In a lot of ways it sure seems like like a radical acceleration. I mean Claude Opus was only released exactly 1 year ago, and now long-standing math problems are being solved every other week. I mean if it were RSI, what would literally be different today? | | |
| ▲ | zug_zug 19 hours ago | parent [-] | | Yeah the 5% rate is somewhat arbitrarily picked, but the reason I expect it not to look like go is because the machine is improving itself, so presumably it unlocks a higher "iq" which in turn will in turn mean that the types of optimizations it will be able to find will grow. Of course there is a raw hardware limit, but it's my intuition that we are under-utilizing the hardware by multiple degrees of magnitude. The whole pattern of using token prediction and realizing intelligence falls out the other end has been more or less stumbled upon by us, there's no reason to think it's optimal. |
|
|
|
|
| ▲ | wolfinterval 20 hours ago | parent | prev | next [-] |
| Stochastic parrot was a brilliant, catchy label to make a point. Arguing over whether a language model is really a stochastic parrot mistakes a metaphor for an actual claim. Like arguing back and forth if a paper tiger is made out of printer paper or newspaper. |
|
| ▲ | abletonlive a day ago | parent | prev [-] |
| [flagged] |