高发量码农Jan

Deep learning | Reinforcement learning | Algorithm | Computational bioinformatics & love Animal and Peace

[高发量码农Jan @smallstone626]

@thsottiaux Codex was waiting the running CI runner by loop checking, several percentage week buffer is ated. what's wrong? why don't use event drived monitor?

[Massimo @Rainmaker1973]

Look at the baby's interaction with the horses, it's so nice how the horses greet him gently

[Nature is Amazing ☘️ @AMAZlNGNATURE]

It is International Cat Day today

Celebrated every year on August 8, the day celebrates cats and raises awareness about their welfare, care and protection

[高发量码农Jan @smallstone626]

@thsottiaux So, don't you think it is a good moment to reset for celebrating?🤣🤣🤣

[Connor Davis @connordavis_ai]

We accidentally built self-improving AI systems. This paper from University of Oxford proves it.

Most people assume model improvements come from bigger architectures or carefully designed reinforcement learning pipelines.

This work shows something more subtle and more unsettling.

If you deploy a model, let users interact with it, filter out the failures, and fine-tune only on the successful traces, the model starts improving its planning abilities on its own.

No explicit rewards, hand-crafted curriculum and no external planner.

Just iteration.

The authors call this iterative deployment, and they test it in controlled planning environments like Blocksworld, Rovers, and Sokoban.

The setup is simple:

1. Deploy an LLM on planning tasks
2. Keep only the plans that actually work
3. Fine-tune the next version on those valid traces

Repeat

After just five generations, planning performance more than doubles across all domains. In some cases it improves by 4 to 5x. Even more interesting, later generations discover much longer plans than the base model, showing real out-of-distribution generalization, not just formatting tricks or prompt compliance.

Here is the key insight.

The paper proves that this process is mathematically equivalent to reinforcement learning with a binary reward signal.

But the reward function is never written down.

It is implicitly defined by user behavior and curation.

Supervised fine-tuning on “only the good outputs” turns out to be REINFORCE in disguise.

That has two big implications.

First, iterative deployment is a powerful alternative to explicit RL for improving reasoning and planning. It works even when rewards are hard to define, as long as you can validate outcomes.

Second, and more worrying, the reward function shaping future models is opaque. User preferences, platform incentives, and validation biases silently become training signals. Over time, these signals can override or conflict with alignment objectives set during pretraining.

In other words, models do not just learn during training.

They keep learning after release.

And they learn whatever the world rewards them for.

This paper reframes deployment itself as a training loop. Once you see it, you cannot unsee it.

Read the full paper: https://t.co/6NfyJ7v91s

[高发量码农Jan @smallstone626]

Definitely, pickle is necessary for cooking shichuan food.

[高发量码农Jan @smallstone626]

聪明人的失败大部分源于对自己的高估…

[高发量码农Jan @smallstone626]

@Charleswcqca 百度的很便宜

[Nature is Amazing ☘️ @AMAZlNGNATURE]

If you're having a bad day, watch this

[Elon Musk @elonmusk]

💫

[高发量码农Jan @smallstone626]

@afra_art56 congatulations

[DogeDesigner @cb_doge]

Grok 4 is smarter than almost all graduate students in all disciplines simultaneously.

[Buitengebieden @buitengebieden]

Sharing is caring.. 😊

[TESLA CARS ONLY⚡️ @teslacarsonly]

Nigerian man grilling meat on Cybertruck.

[高发量码农Jan @smallstone626]

@CoachStanley95 哪里可以看到

[高发量码农Jan @smallstone626]

@Rustallintsla 老马本来就是互联网出生，这个很容易解决，自己搭建团队就可以。

[高发量码农Jan @smallstone626]

@CoachStanley95 无方向盘的车的市场流动性较差...除非fsd能处理世界上所有开车的情况，这是很长的一段时间是不可能的，人也达不到这种水准。

[高发量码农Jan @smallstone626]

@kc_lin20427 我其实更倾向于有方向盘的车，这样才有市场流动性

[高发量码农Jan @smallstone626]

@elonmusk @Tesla_AI Congratulations

[高发量码农Jan @smallstone626]

test