Saptashwa Mandal

[Saptashwa Mandal @percydamson]

@dev___shreya wow all the best

[Saptashwa Mandal @percydamson]

@dev___shreya same 🫪

[Saptashwa Mandal @percydamson]

@HarveenChadha doesn't gpt-4 cost more than gpt-5 per MTok? I assumed it was improved optimisation wrt model and also compute allocations

[Saptashwa Mandal @percydamson]

@thsottiaux @sandislonjsak real. that's all it takes to reach limits.

[Saptashwa Mandal @percydamson]

@0xSero @KingBootoshi one interesting thing tho if you tell the model its in an eval setting and that cheating is allowed, for some reason it cheats less

[Saptashwa Mandal @percydamson]

@0xSero @KingBootoshi You have to either make the right path easier to choose or hard path harder to choose. There has been work done in RL post training to reduce reward hacking but nothing promising until basically an ideal reward-punishment system is created

[Saptashwa Mandal @percydamson]

isn't the claim still true? Because from what I understand llm didn't come up with a "new idea" but made use of existing ideas + proofs + axioms and moved and rearranged them in a way that would serve as a "new" proof for the problem.

[ℏεsam @Hesamation]

I often think about professors coming up with these claims.

[Saptashwa Mandal @percydamson]

@_blast_furnace_ @djcows ooo I thought all of them used byte pair encoding including the frontier labs like anthropic and gpt so imcurious why is this annoying?

[Saptashwa Mandal @percydamson]

@Shirooooo______ @melhpine I should've specified but it cheats less even if its told that the 'bad' trajectory is also a valid path in the eval setting. So it's essentially told that it's allowed to cheat and somehow cheats less.

[Saptashwa Mandal @percydamson]

@_blast_furnace_ @djcows its a tokenizer, doesn't every model do the same?

[Saptashwa Mandal @percydamson]

Am I the only one who thinks this is cool or has value? Too expensive but useful

[Mark™ @thinkwithmark]

We raised $1M dollars to reinvent how people read. Introducing Mark II - a $159 AI bookmark. Thread below

[Mark™ @thinkwithmark]

We raised $1M dollars to reinvent how people read. Introducing Mark II - a $159 AI bookmark. Thread below

[Saptashwa Mandal @percydamson]

@karpathy no way

[Saptashwa Mandal @percydamson]

Is the idea to crop and repost a tweet specifically or just attach an image in general

[Saptashwa Mandal @percydamson]

@RaeAlisa_ @lucent_ai Codex Pro

[Saptashwa Mandal @percydamson]

genuinely insane. what's next?

[Ryan Fleury @rfleury]

Presented without comment

[Saptashwa Mandal @percydamson]

@thsottiaux Anything for Indian remote intern who wants to be involved in any way possible

[Saptashwa Mandal @percydamson]

How is codex app working for you guys in a mid tier windows laptop?
I have performance issues so running TUI instead using either codex or Pi

[Saptashwa Mandal @percydamson]

@sama 👺