Jtor

ex crypto founders: crypto is supposedly 'so back' now, so please gtfo of AI

[Jtor @jtorrence325]

@greg16676935420 @JoeBiden L

[Jtor @jtorrence325]

@MLStreetTalk @cohere @aidangomez Can someone share the clip where hallucination is discussed?

[Jtor @jtorrence325]

@nearcyan They did, it just didn't work out as well

[Jtor @jtorrence325]

@nearcyan @arpitingle Source?

[Jtor @jtorrence325]

@itsandrewgao Bro liked his own post

[Jtor @jtorrence325]

@1vnzh IRON YOUR CLOTHES

[Jtor @jtorrence325]

@lqiao Any possibility for a dataset release as well?

[Jtor @jtorrence325]

@hoyeon_chang What's more interesting here is probably forgetting: when a model forgets a factual info, does it hallucinate or does it admit to have forgotten? How do we steer this behavior better?

[Jtor @jtorrence325]

@soniajoseph_ Name and shame

[Jtor @jtorrence325]

@cloneofsimo Are they? I thought MoEs are mostly a quick hack that orgs have since moved away from

[Jtor @jtorrence325]

@dejavucoder I've never heard anyone call it underrated, it's the defacto resource for a lot of researchers

[Jtor @jtorrence325]

@andyzou_jiaming Really cool work but wouldnt short circuiting still suffer from the same issue with degraded performance?

This is especially tricky since a lot of user prompts are not intentionally harmful but the short circuit logic is never perfect.

[Jtor @jtorrence325]

@eugeneyalt You can't just annouce this and not share the alphas

[Jtor @jtorrence325]

@giffmana I met these guys irl. They were selling their torch fsdp implementation on 10 4090s as something that can rival OpenAI's H200s.

They even got second place on a hackathon.

Should tell you enough who they are lol.

[Jtor @jtorrence325]

@hingeloss I met these guys irl. They were selling their torch fsdp implementation on 10 4090s as something that can rival OpenAI's H200s. Should tell you enough who they are lol.

[Jtor @jtorrence325]

@realAnthonix @karpathy @Yuchenj_UW If this works on 4x 7900, why not try it on 4x 4090

[Jtor @jtorrence325]

@hingeloss @maksym_andr 4o voice isn't even out and people are already writing papers claiming jailbreaks lol

[Jtor @jtorrence325]

@Dynamoblaze @signulll Philosophers are semantic jugglers, just very skillful ones

[Jtor @jtorrence325]

@carmguti What a slop ass tweet

[Jtor @jtorrence325]

@yacineMTB Which robotics company is building tomagachis