Post ATdxVTbOPP147NtO1A by jmb@mstdn.party
(DIR) More posts by jmb@mstdn.party
(DIR) Post #ATdpifwf1tlmWxchHs by TedUnderwood@sigmoid.social
2023-03-15T13:46:56Z
0 likes, 0 repeats
This could be totally wrong, but I expect deep learning to slow down as it approaches the performance of the best human solutions ā just because we don't have rich training data above that level.
(DIR) Post #ATdppi5FTdhdBqziCm by afamiglietti79@mastodon.social
2023-03-15T13:48:12Z
0 likes, 0 repeats
@TedUnderwood I like mine better. It's got that MS Paint Authenticity that the 2020s are gonna crave.
(DIR) Post #ATdqTOOo5D2Jit44Dw by budgibson@me.dm
2023-03-15T13:55:23Z
0 likes, 0 repeats
@TedUnderwood It strikes me that the machine could train itself, much like game playing models so as not to be constrained by the limits of human data. But there might still be limits on how far it can optimize due to the structure of the environment.
(DIR) Post #ATdqwHgo6mcFJMGEQy by adamgurri@mastodon.social
2023-03-15T14:00:28Z
0 likes, 0 repeats
@TedUnderwood @brendannyhan if it actually followed the trajectory of the pictured graph it'd be revolutionary. It doesn't have to be better than the average human solution because it scales vastly larger than any human (or really any number of humans). I'm skeptical it will follow a trajectory as good as the graph though.
(DIR) Post #ATdsTtJNLqBdVKtDvM by Scmbradley@mathstodon.xyz
2023-03-15T14:17:48Z
0 likes, 0 repeats
@TedUnderwood what about cases like alpha go where we can train it against itself? That seems like an important caveat. But those are cases where we have clear rules...
(DIR) Post #ATdsx4bgR0huvyMt72 by matthewmaybe@sigmoid.social
2023-03-15T14:23:09Z
0 likes, 0 repeats
@TedUnderwood I am just a guy on the Internet, but the way around this would be if combinations and/or juxtapositions of lessons from existing training data performed unreasonably well on whatever the target task or metric is, and were consequently rewarded in RLHF loops.
(DIR) Post #ATdu7UKZkQtzZ8tcoa by MrCheeze@wetdry.world
2023-03-15T14:36:14Z
0 likes, 0 repeats
@TedUnderwood One hopes so! There is a more worrying possibility, though, which is that it actually speeds up once "amount of human-quality data" is no longer a chokepoint.
(DIR) Post #ATduyVZnqbA0mEIhOq by TedUnderwood@sigmoid.social
2023-03-15T14:45:49Z
0 likes, 0 repeats
@adamgurri @brendannyhan I agree about the first part. Crossing the dotted line is a big deal. I feel like we're crossing it in some areas right now; GPT-4 is much better at explaining things than most people who aren't professionals.
(DIR) Post #ATdvL314CS2erKnhLc by adamgurri@mastodon.social
2023-03-15T14:49:51Z
0 likes, 0 repeats
@TedUnderwood @brendannyhan I dunno. I remain pretty skeptical. For me the proof is not in the argument but in the putting: when this stuff is successfully operationalized and it's as invisibly part of the infrastructure as normal search engines are, I'll see it (or at least once we're on the path there, whereas now we are still perpetually doing proof of concepts)
(DIR) Post #ATdvpqwV8hHKOHM1PE by adamgurri@mastodon.social
2023-03-15T14:52:57Z
0 likes, 0 repeats
@Jonathanglick @TedUnderwood @brendannyhan It would be kind of funny if the only economically significant outcome of this for the next 10-15 years was largely as a substitute for human programmers (rather than the writers and chattering class as people keep predicting)
(DIR) Post #ATdvprTp8nvw3cwdLU by TedUnderwood@sigmoid.social
2023-03-15T14:55:28Z
0 likes, 0 repeats
@adamgurri @Jonathanglick @brendannyhan As someone who does both of those things professionally, they feel much less different than our discourse about them would suggest. I'm regularly leaning on GPT now for aid in both areas. It exceeds me as a programmer, I'd guess, more often than as a writer. But I've had more practice writing!
(DIR) Post #ATdvw9h6KmHkHk9kBc by adamgurri@mastodon.social
2023-03-15T14:56:36Z
0 likes, 0 repeats
@TedUnderwood @Jonathanglick @brendannyhan A functional programmer who did a little crash course at a startup I worked for had a saying that stuck with me, "good coding is good taste."
(DIR) Post #ATdw1VHNVaygg6ftBY by Jonathanglick@mstdn.social
2023-03-15T14:57:33Z
0 likes, 0 repeats
@TedUnderwood @adamgurri The differences probably have more to do with the culture and function of those vocations than the quality of the output of the AI.
(DIR) Post #ATdxVTbOPP147NtO1A by jmb@mstdn.party
2023-03-15T15:14:10Z
0 likes, 0 repeats
@TedUnderwood "... past 1T parameters (20T tokens), training data collection would naturally have to rely on alternative text-based and multimodal content. ...Fundamentally, it should not be an incredibly onerous process to collect petabytes of high-quality and filtered multimodal data (converted to text), though that task has not yet been accomplished by any AI lab to date (Jun/2022). - https://lifearchitect.ai/chinchilla/
(DIR) Post #ATdy26ckrT8YPR5XVI by chengyjonathan@sigmoid.social
2023-03-15T15:20:06Z
0 likes, 0 repeats
Iām thinking about a paper from 2022 about how RL found faster matrix multiplication algorithms than the best human solutions š I suspect RL will, surprisingly, have a comeback to bridge that exact gap. *takes off tin foil hat*
(DIR) Post #ATe1nbRIPOx1SC4J3Q by rf@mas.to
2023-03-15T16:02:18Z
0 likes, 0 repeats
@TedUnderwood and models are built to *predict what humans will say* which doesn't fit well with getting superhuman behavior out of them. good chess NN's aren't just models of how people will move!you need ways to nudge output further in the correct direction analogous to self-play. i suspect RLHF will not end up enough because you can't get enough HF
(DIR) Post #ATeG6QdoB9tcp9nAbw by lobrien@sigmoid.social
2023-03-15T18:42:34Z
0 likes, 0 repeats
@TedUnderwood I think Francois Chollet argues similarly: intelligence is embedded in a huge context that argues against the prospect of runaway improvements.
(DIR) Post #ATeM0G3ZDc9bz5HVAm by TedUnderwood@sigmoid.social
2023-03-15T19:48:40Z
0 likes, 0 repeats
@lobrien Interesting ā I'll keep an eye peeled for his take. There are a lot of people skeptical about the "superintelligence" thesis.
(DIR) Post #ATeZgbk2HdLGo0fbA8 by Csosorchid@universeodon.com
2023-03-15T22:21:59Z
0 likes, 0 repeats
@TedUnderwood @adamgurri @brendannyhan So AI is better at explaining things than people that do not specialize in the topic?
(DIR) Post #ATebrqMYShmPFLAkC0 by TedUnderwood@sigmoid.social
2023-03-15T22:46:26Z
0 likes, 0 repeats
@Csosorchid @adamgurri Or even those who do but don't specialize in writing. I think it tends to be better at explaining, e.g., science than people who aren't professional science writers.
(DIR) Post #ATec21LIUiheCAR9Ky by Csosorchid@universeodon.com
2023-03-15T22:48:16Z
0 likes, 0 repeats
@TedUnderwood @adamgurri But is it right?
(DIR) Post #ATedX2IB75rI4WKadU by TedUnderwood@sigmoid.social
2023-03-15T23:05:05Z
0 likes, 0 repeats
@Csosorchid @adamgurri Usually! :) Not always, but GPT-4 is notably, definitely better than ChatGPT was. I ran into hallucinations a lot with the old version and have seen none (that I noticed!!) with this one. To approach 100% they'll need to bolt on an error-checker of some kind ā but even then we won't be 100%, because humans are not 100%. Truth is elusive.