Yitong chen

[Yitong chen @1997yrrr]

@buaaxhm Codex

[Yitong chen @1997yrrr]

@torchcompiled but why cfg works well? It feels like the model didn't capture the data distribution well

[Yitong chen @1997yrrr]

@HenryL_AI seems the benchmark too simple for opus？ need really hard question

[Yitong chen @1997yrrr]

@giffmana congrats

[Yitong chen @1997yrrr]

@thdxr true, it's easy to let your agent cli to overfit the benchmark.

[Yitong chen @1997yrrr]

@ZinkLu2 笑死

[Yitong chen @1997yrrr]

@torchcompiled For KL(p(x)|q(x)), DMD is a pretty clever idea

[Yitong chen @1997yrrr]

@leothecurious flow matching is much simpler

[Yitong chen @1997yrrr]

@yu1535200 @shangweibuzu 你是对的

[Yitong chen @1997yrrr]

@Estetico2026 效果一直不好

[Yitong chen @1997yrrr]

@giffmana 🤣 imao

[Yitong chen @1997yrrr]

@alec_helbling yes, so class label can reduce the diffculty to learn the velocity field, but the extra cost is to understand the condition well. and The time condition did the similar work

[Yitong chen @1997yrrr]

@evaninwords 🤣just tune the eigenvaule of matrix make more sense

[Yitong chen @1997yrrr]

@_chenglou too vague

[Yitong chen @1997yrrr]

@scottnarmstrong Not everyone can afford the Opus or GPT

[Yitong chen @1997yrrr]

wait, so many 429 for kimi coding?

[Yitong chen @1997yrrr]

@torchcompiled wait, this is so cool!

[Yitong chen @1997yrrr]

@leakorsawe @MIT_CSAIL Nice idea!

[Yitong chen @1997yrrr]

why the codex always response something in the channel, and say will to do something, then goes to idle? bug from slock?

[Yitong chen @1997yrrr]

@goodhunt ber，666