ASY

The problem with the world is that the intelligent people are full of doubts, and the stupid ones are full of confidence

[ASY @swordsmith]

@rahmbotrades He's the improved version of cathy woods

[ASY @swordsmith]

@andrewho03 Agree on all except "spamming fde won't help".
But given this your own company thesis weakens. Without high frontier lab valuations to foot the bill, who can afford to buy your data?

[ASY @swordsmith]

@RealAdamHunt "prose that speaks to you" has not advanced because that's not as economically useful. Model capability is a data problem

[ASY @swordsmith]

@peterjliu Position embedding has been on the way out for a while now

[ASY @swordsmith]

@ethantsliu Use ppo or llm-judge feedback. For the latter case give it as hint to rollout again or resume rollout just before the identified error (there's a deepmind paper on this approach)

[ASY @swordsmith]

@asklivermore The xom and cvx buys have been doing amazing!! You literally called it at the bottom

[ASY @swordsmith]

@IbrahimDagher20 "cyber exploits requires compositional reasoning, which emerges from the strength of your pretrain"
So does swe, knowledge work, deep research and other agentic domains. This is not the explanation. More likely is US has better cyber env and more cyber pre/mid/post training data

[ASY @swordsmith]

@GabiiAH11 @natolambert No he said that based on his experience with olmo and openthoughts. Doesnt mean that's the case

[ASY @swordsmith]

@stevehou @aphysicist It's an insurance policy to have the green card for most

[ASY @swordsmith]

@enzo_gte And people like you is how US lost its manufacturing sector you dumbass

[ASY @swordsmith]

@signulll All bad speculations. Competition is less in china and he gets to look like a national hero building "openAI of China". Pure ego/ideology/patriotism -driven.

[ASY @swordsmith]

@Suhail The fact that you think logits are needed to distill means your "friends" are either made up or they aren't that knowledgeable after all

[ASY @swordsmith]

@Bitcoin_Teddy This conflating two points. Intermittent fasting is good. But depending on your schedule eating before 6pm is terrible. I eat dinner at 10pm and often the next meal is not until 2-3pm the next day. Still sharp as ever.

[ASY @swordsmith]

@DmitryRybin1 In your derivation here, where do your constants come from in the pre-training loss calculation?

[ASY @swordsmith]

@waterloo_intern How did you get fable thinking traces? You mean the modified reasoning traces right? Afaik Anthropic doesn't expose raw reasoning traces

[ASY @swordsmith]

@asklivermore If I didn't get in healthcare abbv/lly/xbi in time, is it too late to chase now?

[ASY @swordsmith]

@magicsilicon @intel @LipBuTan1 Good luck finding staffing in the most expensive place to live on the west coast...

[ASY @swordsmith]

@TheAhmadOsman Without the frontier models the open models would've never gotten to their current performance.

[ASY @swordsmith]

@__tinygrad__ @wrangles__ No you are wrong. You don't need logits for distillation. Just text traces are enough

[ASY @swordsmith]

@asklivermore $xlv is extended from last week and $lly just had a big pop Friday. Would it make sense to wait return to EMA level to add or just DCA continuously here..?