yzlnew

[yzlnew @menlzy]

看到 ai slop 文本会感觉上不来气，何意味

[yzlnew @menlzy]

感觉是强化信号的问题，我没有证据

[songkeys 🐿️🦋@song.work @songkeys]

gpt 5.6 最让我受不了的就是：
我让它做一盘番茄炒蛋，它往里还加了东坡肉。
我说有必要加东坡肉吗？它说你说得对，然后把东坡肉去掉。
我说好，你提 PR 吧。再一看，它 PR 写着「番茄炒蛋（无东坡肉）」并且注释里会写一大堆为什么本道菜不需要加东坡肉。

[yzlnew @menlzy]

@GenAI_is_real 我宣布这是更 ai native 的梗图更新

[yzlnew @menlzy]

@Alibaba_Qwen simply goat

[yzlnew @menlzy]

@latepostnews 在 GPT4 时代几乎没人讨论训那些 GPT4 合成的、在 HuggingFace 开源的 SFT 数据，是邪恶的「蒸馏」。

[yzlnew @menlzy]

@q25601858 @kotekjedi_ml b.1 - b.3 的方法可以用这种 pair 来验证；b.4 的实验结论也不太符合预期

[yzlnew @menlzy]

@teortaxesTex run some tests before, recent models are more likely to regard itself as Claude in coding context, which I think may result from the massive co-author records in Github. https://t.co/6vYcckQ5Ju

[yzlnew @menlzy]

我只用了一个月的年付 AI Pro 能退款就好了 :)

[Logan Kilpatrick @OfficialLoganK]

We have started our most ambitious pre-training run yet, for Gemini 4, and are excited by the progress : )

[yzlnew @menlzy]

@yetone https://t.co/G28KncfJMP

[yzlnew @menlzy]

@dviolettchan @paperpaper886 优化器 paper 里面全是 sweeps

[yzlnew @menlzy]

https://t.co/KSTdJaMRo1

[yzlnew @menlzy]

plot like openai (skill reversed by cc)

[yzlnew @menlzy]

@m0d8ye Kimi K2 系列还是 DSV3 架构，比较容易训练。用 DSA 之后会有一些训练上的挑战。

[yzlnew @menlzy]

Training across continents is a pretty wild.

[Cognition @cognition]

Introducing SWE-1.7, the most capable model we’ve trained yet.

It scores within a few points of the strongest frontier models at a fraction of the cost, and is now available at 1000 tok/s.

RL is not hitting its limit: after refining our recipe, we keep seeing gains as we scale

[yzlnew @menlzy]

如果每天都可以 happy happy 誰想要sad:))

[yzlnew @menlzy]

@Xianbao_QIAN Is it just the same thing as test-time scaling?

[yzlnew @menlzy]

@midjourney cool

[yzlnew @menlzy]

@jietang congrats, impressive indeed!

[vivek @itsreallyvivek]

https://t.co/ZhOgyq7Vgn