Frank Yan

Agentic RL.

[Frank Yan @adler0824]

@Y4tacker 其实训练的时候就根据这prompt适配训练了，否则效果不足

[Frank Yan @adler0824]

@wzenus 累了，我选择Codex

[Frank Yan @adler0824]

"Agentic RL is increasingly an observability problem"

[Mani 🪽 @devxMani]

The infrastructure for self-improving agents still needs three primitives:

step-level trajectories, governed workload capture, and a control plane for deciding when to update weights vs the harness

Agentic RL is increasingly an observability problem.
https://t.co/FBeTFwdAgW

[Frank Yan @adler0824]

@karpathy vibe chatting ?

[Frank Yan @adler0824]

@real_kai42 让kimi code分析一下啊

[Frank Yan @adler0824]

有没有死在2008年的

[那由 @noaayoyaoaya]

北京给我一种整个城市死在了2018年的感觉，上海稍微晚一点撑到了2022年

[Frank Yan @adler0824]

@buaaxhm 😂

[Frank Yan @adler0824]

20刀的Codex额度只够使用两天的，有同感的吗？

[Frank Yan @adler0824]

I switched to GPT-5.6 Sol in Codex during one of the most difficult debugging sessions I’ve had recently.

We were working on R3, a system that records which “experts” an AI model chooses during inference and replays those choices during training. The idea sounds simple, but the data has to survive a long journey: across multiple GPUs, CUDA Graph batches, storage pools, multi-turn agent conversations, sequence truncation, packing, splitting, and finally training replay.

Every part looked correct when tested alone. But in a real distributed experiment, one routing value would occasionally arrive at the wrong token.

GPT-5.6 Sol didn’t just suggest random fixes. It followed the entire data journey, connected logs from different machines and codebases, and helped design an end-to-end canary that gave every token a predictable identity. That let us see exactly where the identity changed.

The real environment then exposed several subtle edge cases: padded CUDA Graph batches, terminal tool-response tokens that never returned to inference, and long conversations truncated in the middle of an active generation. Some failures were bugs in the production pipeline; others were bugs in our verification logic. Sol helped separate the two instead of “fixing” the wrong thing.

After several real multi-GPU experiments, the small validation passed. Then the full-scale validation kept running normally across multiple training steps.

I really sat back in my chair and I was just like, “Oh man! Here it is!” It feels like the Manhattan Project.

That’s what I love about GPT-5.6 Sol in Codex. It doesn’t feel like autocomplete. It feels like debugging with an engineer who can keep the whole system—code, logs, experiments, assumptions, and previous failures—in its head while still reasoning carefully about one mismatched value.

That’s why I switched. @thsottiaux @sama @OpenAI

[Tibo @thsottiaux]

Or… what if we gave you $100 in Codex credits if you tell us what you love about GPT-5.6 Sol or why you switched?

Tweet it, claim your gift, enjoy more usage. First 10k get the free tokens!

https://t.co/8mU93eA13i

[Frank Yan @adler0824]

@wey_gu 笑死

[Frank Yan @adler0824]

用 GTP 5.6 Sol 定位到了一个非常复杂的 Bug。此时此刻我瘫坐在椅子上，仿佛看见了原子弹爆炸。

[Frank Yan @adler0824]

@yan5xu nice work，已Star

[Daniel Han @danielhanchen]

DeepSeek just released DSpark for V4 Flash & Pro, a new speculative decoding method boosting throughput by 51% to 400%!

DS also showed DSpark works well for other models like Gemma & Qwen

Github: https://t.co/EGVYpc1kcK
Paper: https://t.co/TaBMRVlaW9
HF: https://t.co/289jVU2pxh

[Frank Yan @adler0824]

@kfk_ai 学习了

[Frank Yan @adler0824]

@qzhao27 赞！本人做RL，欢迎交流

[Frank Yan @adler0824]

大模型输出的中文内容实在是太难懂了，正常人说不出来这种话，头疼

[Frank Yan @adler0824]

用国内的模型套餐，要绑定手机号、身份证号、实名认证、刷脸认证。关键是这一套搞完之后你告诉我Token套餐卖完了，每天10点抢购？？？

[Frank Yan @adler0824]

"because he thinks this is important though he does not understand why people are not doing it". That's so cool !

[Ying Sheng @ying11231]

I remember when Jiajun told me he wants to push for TITO because he thinks this is important though he does not understand why people are not doing it. It’s great they were able to insist on their own judgement! A long work before this blog.

[Frank Yan @adler0824]

@logiclogic1223 你还是不够卷

[Frank Yan @adler0824]

@sheriyuo 每次看你们这种言论都觉得自己不够卷。。