Floating Plants, Amputated Robots, & More

TLDR;

Welcome back! This week, we have a long list of fascinating news, from amputated robots that still move, great models from both the US and China, and even a floating power plant; yes, this newsletter is that varied.

Enjoy!

/

WORLD MODELS

Meta Presents the First Coding World Model

Meta’s Superintelligence Lab, the new team created after the massive Llama 4 failure that led Mark Zuckerberg to panic and poach dozens of top researchers from competing labs, has presented its first model: CWM, standing for Coding World Model.

The model showcases robust metrics for its size (just 32 billion parameters, way smaller than frontier models), competing in many aspects with top models like Gemini 2.5 Pro.

This might sound like just another coding LLM, but it’s different in many aspects. The most significant change stems from the training procedure.

Typically, models are trained using a technique known as ‘imitation learning.’ They are given vast amounts of data (code in this case) and have to learn to imitate it, replicating it in the literal sense. This is akin to teaching kids to write well using dictation exercises. Eventually, the kid learns how to write via rote repetition.

This works, but as you may imagine, it isn’t a particularly great way to create intelligence, in the same way most human learning isn’t about copying texts.

Thus, we then add a ‘post-training phase’ in which the model isn’t guided toward the answer. Instead, we give them a task, the result, and a pat on the back.

Here, the model must find its way to the answer, and we help it by providing intermediate signals to indicate whether the model is heading in the right direction (think of the game ‘hot and cold’). This is called reinforcement learning because we reinforce good choices and penalize bad ones.

This is a superior training method to develop the ‘I’ in ‘AI’ because models must actually figure out how to arrive at the answer instead of simply copying it.

The point I’m trying to get to is that tables are turning, and the latter training method is gaining prevalence as a share of total training.

In other words, the ‘guiding stage’ where models rotely imitate data is less and less important, and the ‘figure it out yourself’ phase is becoming more relevant (Grok 4 was the first model to allocate more compute to RL than to imitation learning).

And what does this mean to the industry?

In my opinion, this means we’re still very early, and much of the progress we’ve seen so far is not only not indicative of future progress (as we’re using a fundamentally different approach), but it could actually be undermining AI’s real potential that could soon be realized.

→ READ RESEARCH PAPER

ADOPTION

How Good Are AIs In Real Problems?

GDPEval is a new evaluation from OpenAI that tests AI models on economically valuable, real-world tasks across 44 occupations, drawn from industries that contribute significantly to GDP.

Tasks are realistic deliverables…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.