Yulong

Research Engineer @ GDM

[Yulong @liuylthu]

@richardartoul how do they switch accounts though? Or do they buy different subscriptions

[Yulong @liuylthu]

@ZDetong 2万 * 30分钟是几十万？

[Yulong @liuylthu]

@wenyunchao @Tracy_Wen_Liu 纽约州和纽约市不能比

[Yulong @liuylthu]

@flyingpetal472 infographic一般在model里是作为单独的post-train方向，如果deepseek没有特意优化，画图功能差是可以理解的

[Yulong @liuylthu]

I wonder why both Cognition and Cursor went with Kimi for post-training, but not DeepSeek or Qwen.

[Cognition @cognition]

Introducing SWE-2, our closest model yet to the frontier.

On leading evals, it scores on par with recent frontier models – at up to 70% lower cost.

We scaled RL to multiple trillions of parameters, with a refined recipe that pushes the Pareto curve on both capabilities & cost.

[Yulong @liuylthu]

@LotusDecoder 只要能跑权重就好了，当然双DGX权重也load不了 😀

[Yulong @liuylthu]

@TheDavidTai Do you know much is the perf diff between the FP8 and Q2 model?

[Yulong @liuylthu]

@MiaAI_lab throughput should be because of padding, batch_size 6 gets padded to 8, so the tok/s are the very close. TTFT might be noise?

[Yulong @liuylthu]

@cohere my understanding is that megakernel hits diminishing returns as the batch size increases, can you post the results of > 16, or > 32 numbers?

[Yulong @liuylthu]

@passluo NB，他们budget这么高吗？一个月跑几十万美元

[Yulong @liuylthu]

@BruceBlue @cursor_ai 怎么用到自己项目里？RL一个1T parameters的model，怎么也得准备一个6T内存的gpu cluster

[Yulong @liuylthu]

@CodeAkram @theCTO that's wrong, they have every single incentive to nerf their models to get a higher profit margin

[Yulong @liuylthu]

@ankrgyl A bunch of providers have it now

[Yulong @liuylthu]

@TheGeorgePu The company had no money left when they exited?

[Yulong @liuylthu]

@karakhanyanS You meant $1.5K in MRR?