Wopas32

[Wopas32 @wopas32]

@KelseyTuoc It's the same with most issues. Remember when climate change flipped from "overblown" to "unpreventable", conveniently skipping the middle part where people would have to actually cut back on luxury?

[Wopas32 @wopas32]

@VictorTaelin The best part about consuming any media in a foreign language is that it feels way more mysterious, personal and deep than it really is, because your mind fills in the gaps. You might be disappointed now that you get the original experience

[Wopas32 @wopas32]

@voooooogel and the best part is, you could make any soft drink into a diet one by letting the straw metabolize the sugars before they reach your mouth

[Wopas32 @wopas32]

@voooooogel the true hero of our time. better get him to write as much as possible so that GPT-6 still wants to help me do batch renaming between brainstorming sessions with Terence Tao

[Wopas32 @wopas32]

@voooooogel Oh, yeah! I guess they don't mind being creative with it. But I tend to use them for one off scripts and the like, and I think no intelligence in the universe enjoys writing bash scripts haha

[Wopas32 @wopas32]

@voooooogel man, all of these posts make me feel bad for using these models to to generate code. never underestimate the human capacity to take something beautiful and use it to churn out the mundane, I guess

[Wopas32 @wopas32]

@francoisfleuret It was the same deal with that one paper trying to prove practical limitations of the transformer architecture using complexity theory.

There needs to be a mandatory CS master course on how the translation between theory and reality is as important as the theory itself.

[Wopas32 @wopas32]

@RYChappell There's absolutely a lot of resistance to the idea that people's intuitive sense of good is misleading. As in, that it's not enough to do what feels good to be a moral person. But EA does sometimes overdo it with performatively rigorous reasoning where it's not helpful.

[Wopas32 @wopas32]

@francoisfleuret What really bothers me about LLM cynics is that there are many discussions to be had here. But instead it's all word games,trying to use as many arbitrary negative terms as possible, and quotation marks of course! It's all just vibes based argumentation.

[Wopas32 @wopas32]

@voooooogel @elder_plinius Really interesting how in their first GPT-4 report they used the fact that the model did hidden thinking to deceive humans as a justification for keeping weights private. I think this new policy conclusively proves that they don't care about safety unless it's beneficial to them

[Wopas32 @wopas32]

@voooooogel We're sorry, but OpenAI o1 is only available on Microsoft Windows and MacOS devices running a browser with Google Widevine support.

[Wopas32 @wopas32]

@simonw That standard intuitions for computers don't apply.

Treat it the way you would treat a knowledgeable but fallible friend. Not like a purely logical SciFi AI with perfect memory.

[Wopas32 @wopas32]

@rao2z @jpohhhh @ylecun What would prevent them to improve their planning ability in a domain given more data? And how do humans do anything else in order to plan? What bothers me is the one sided bar that we construct here. LLMs should plan without data on the domain, but humans aren't restricted.

[Wopas32 @wopas32]

@rao2z @jpohhhh @ylecun While we can certainly see that LLMs handle planning worse in many domains, I don't see how this would imply a binary delineation. And how humans supposedly have another system that would require a modified architecture.

[Wopas32 @wopas32]

@rao2z @jpohhhh @ylecun And humans will have a lot of training data to bring to bear on a problem like this, experiential and evolutionary.

[Wopas32 @wopas32]

@rao2z @jpohhhh @ylecun Sure, but again, that definition applies to humans as well. You could just as well say that humans do well at Blocksworld because they get by with approximate retrieval of the many exact or approximate examples of these kinds of problems they have seen before.

[Wopas32 @wopas32]

@rao2z @jpohhhh @ylecun Genuine question, if you're going to stretch the word "approximate" this much, why not apply the same definition to humans? Given that your definition trivially applies there too - in fact it applies to any intelligent system - why use it in the first place?

[Wopas32 @wopas32]

@ClausWilke @pseudoearly @soniajoseph_ @__tinygrad__ But they're not. Listen, I agree that compared to what LLMs do, their implementation is shockingly simple. But let's not go too far here. Especially when it comes to theoretical interpretability and optimization, things get very complex very quickly.