Perhaps not immediately obvious to everyone, but the growing popularity of openclaw is directly linked to the development of DeepSeek.
Let me clarify: it was DeepSeek that first demonstrated how reinforcement learning in environments where results can be verified scales and significantly enhances the capabilities of models. By 2024, it can already be said that they were the first to achieve such successes, although no official documentation on this has been released.
It took a whole year for leading laboratories to create truly scalable environments for long-term tasks with extensive context—such as working with code or complex projects.
This, in turn, led to the emergence of Opus 4.5—a model that, thanks to RL, became one of the best agent solutions. It handles long tasks well, can navigate in a tower and maintain the correct course when it needs to return to the truth.
So, besides training scalable pretraining (which is still ongoing), we also have scaling through GRPO / RL with verifiable rewards.
To simplify: over the past year, the speed of “getting smarter” of LLMs has increased by at least twofold—and in reality, this is more of an exponential growth.
Created with n8n:
https://cutt.ly/n8n
Created with syllaby:
https://cutt.ly/syllaby