4 ms·
4 Charts That Show Why AI Progress Is Unlikely to Slow Down
- hospitalJail 3y agoGiven how the market sold Google after their Bard press release had incorrect information, I'm quite skeptical of everyone's opinion on AI since its a bunch of people who are new to the market and enjoying the hype. I wrote my first AI program Random Forest (from scratch) in 2019, and it helped me incredibly understand what AI is. Read what I'm about to say as if I'm an old grumpy man yelling at the kids/bandwagon tech journalists on my lawn: These graphs literally don't matter. >First Graph showing competence They skyrocket then flatten. Just as I expect to happen with LLMs and AI Art. >Amount of compute for AI Probably the only valid graph here. Lots of people are using it now, so I expect further investment. Although we still might be nearing end game for LLMs, now its going to be all about application. >Algorithm Efficiency and Data Points graphs Garbage data is the real issue. More isnt necessarily better.
- cubefox 3y agoIf you look at the massive AI progress we had in the last year, last five years, last ten years ... doesn't that suggest that progress will continue at a similar pace for at least some time?
- hospitalJail 3y agoThe low hanging fruit are gone. With LLMs, everything except math problems have been solved. Real AI is a monster compared to ML.
- cubefox 3y agoI see no evidence for this.
- Havoc 3y agoProbably right directionally but I’m a little sceptical of their scale showing various markers above human 100%. Even simple things a kid would manage still trip them up. Eg if it takes 5 shirts an hour to dry in the sun how long would 10 shirts take? I’ve yet to see one get that even remotely right. Admittedly a bit of a trick question (for a machine) but that to mean means they’re not yet at 100%
- cubefox 3y agoMany humans would get that question wrong, too. 100% just means average human. Also, Bing precise mode gets it right for me: https://abload.de/img/screenshot_20230804-19fffj.jpg https://abload.de/img/screenshot_20230804-19fffj.jpg
- Havoc 3y agoThat's impressive @ Bing
- sliken 3y agoI've seen LLM reviews that totally nail it. 2 hours, if space contrained to 5 shirts, 1 hours if space is not a constraint. Things do seem to be improving quickly, but if the inputs are largely human written content I do wonder if it's feasible to be much better than humans.