Thyndd

[Thyndd @Thyndd]

@scaling01 @apestein_dev Keep in mind investors are usually not technical people. Sure, the pieces were there, but without venture capitalism we wouldn't have this race. OpenAI did greatly contribute to kick it all off.

[Thyndd @Thyndd]

@yacineMTB Don't get me wrong, I love RL and actor critic methods, but PPO is quite literally retarded. In general our best RL algos are so horribly sample inefficient it's not even funny.

[Thyndd @Thyndd]

@Rufus87078959 @scaling01 Well, It's not like he threw a tantrum because of the criticism. I think it's fair to not want to see it constantly on your timeline. Nothing wrong with wanting to make the cesspit that is social media a better experience for yourself.

[Thyndd @Thyndd]

@yacineMTB I'm not unemployed enough to dig up old tweets of yours claiming the opposite, back when models were considerably worse.
Did you arrive at the conclusion that saying they're awful generates more engagement these days?

[Thyndd @Thyndd]

@DeryaTR_ @sama Well, that might be true, but I think often what happens is that they think long term. They don't care about the "move fast and break things" of Silicon Valley. They want to establish the foundations of a scientific field, and that takes time.

[Thyndd @Thyndd]

@nhegde610 @JFPuget Raw intelligence has nothing to do with "emotions and feelings".
Can it understand and solve the problem? Can it generate plausible hypothesis? That's what matters. I don't think Hinton is claiming LLMs feel teenage angst.

[Thyndd @Thyndd]

@YurifyEveryting @ChibiReviews If I received one tenth of the targeted harassment from the alphabet people Chibi gets on the daily, you don't want to know what slurs I'd throw their way. Chibi is a fucking saint.

[Thyndd @Thyndd]

@NHydraTwit Don't threaten me with a good time, troon.

[Thyndd @Thyndd]

@beffjezos Well, one could argue that if we build AGI, no matter how energetically inefficiency, you could task it with optimizing its own energy cost. There is a limit to how expensive it can be, of course, but it would be quite large.

[Thyndd @Thyndd]

@MoonL88537 The reason it's hard to know how smart a model really is is that there's no human that has all of the internet memorized to compare it to. Patterns that might look incredibly complicated for you might become trivial. We're constantly re-inventing the wheel, LLMs are not.

[Thyndd @Thyndd]

@redtachyon Maybe they shouldn't have been fearmongering like there's no tomorrow.
But still, they just need to release Mythos under a different name. At least until a clear ban on all Mythos-tier models arrives.

[Thyndd @Thyndd]

@VictorTaelin If any lab, let alone Anthropic, open sources a Mythos-tier model, then I would have no choice but to believe AGI has been achieved internally.

[Thyndd @Thyndd]

@baalzamon35

[Thyndd @Thyndd]

@jeremyphoward That's great marketing for Anthropic, so I doubt they will mind.

[Thyndd @Thyndd]

@Hitchslap1 @NYMag By doing "research level math" do you mean actually making a significant contribution to the field, or just holding a PhD? Because if it's the former, yeah, I wouldn't expect anyone but the brightest. But doing a PhD... I think that might be achievable with enough time and effort

[Thyndd @Thyndd]

@cmuratori @imadr_ @Danukeru The question is at what point does the lossy compression become lossy enough that you wouldn't consider it piracy.

[Thyndd @Thyndd]

@nekohive86 A wise friend once told me, "if I'm gonna be staring at a character for dozens of hours, I at least want it to be pleasing the eyes".

[Thyndd @Thyndd]

@fchollet Everything is a bubble when you look at it hard enough

[Thyndd @Thyndd]

@yifever I don't know, I think a paying customer is owned at least the peace of mind that the product they are paying for won't silently and deliberately sabotage them.

[Thyndd @Thyndd]

@gfodor Just use ONLY their subscription service and make sure you spend every last token.
That hurts them more than it helps them.