However this migration is still in its early stages, and current prices have not yet fully reflected the logic of robotics as a foundational infrastructure.
Robots are likely undervalued now, seen as hardware rather than next gen AI infrastructure.
It isn't just an execution terminal, but a generator of real world data, simulation, control, and feedback loops. Its strategic position is far more foundational than realized.
To be more precise, it’s not that the market is completely unaware, but rather that those who see it have yet to reach a consensus on pricing.
You can already see capital shifting from pure computing power toward the underlying physical AI layers like chips, power etc
some good counter-arguments on why llms can generate novel ideas:
- rlvr (verifiable rewards) can encourage novel ideas given a good goal. i think this can work, but i am not sure if this can produce a model that generalize this capability to other domains without verifiable rewards. same logic for alpha go, alpha zero, they can do well in one domain, but can’t generalize.
- it depends on the question, if a new way to ask questions is discovered, it might trigger the model to give novel ideas (new model trajectory that was possible but unexplored). i think this is possible for very large models, but i’ve yet to this happen.
- combination or synthesis of existing ideas can produce novel ideas. i agree with this approach, but i don’t think it’s that simple. there’s more to it before one comes up with a genuinely novel idea, such as thinking deeply for days, talking to a friend, taking a walk, visiting places, getting a good sleep and other personal experiences. i don’t think current llms can mimic those human experiences yet.
also two more arguments against llms that emerged during the discussion:
- humans who came up with genuinely good ideas were often ridiculed and rejected by the society before. we can expect the same behavior during training where they were treated as part of loss during scoring, hence never make it to the final checkpoint.
- llms have all of human knowledge compressed in the weight, the knowledge can act as a barrier to new discoveries as the model defaults to applying existing knowledge and rules, instead of questioning them. but questioning existing rules is how genuinely novel ideas comes about.
thanks to everyone who participated in the discussion in good faith. i learnt a lot!
The alignment mechanism hides within microscopic, high-dimensional Calabi-Yau manifolds. The geometry of these extra dimensions dictates how strings vibrate, ultimately aligning and creating the speed of light, gravity, and spacetime itself.
AI training requires massive amounts of human text, essentially reverse engineering the laws of the universe through human lens. The real world is made of information, but even info needs a stable generative source. Where does this source align with reality?
If we decode how this source is constructed, humanity could build simulated universes. According to string theory, this "rendering engine" is 1D energy strings, generating matter & gravity via varying vibration frequencies (info).
The AI Scientist: Towards Fully Automated AI Research, Now Published in Nature
Nature: https://t.co/nNfpSV5e5I
Blog: https://t.co/i6h8LVQOdl
When we first introduced The AI Scientist, we shared an ambitious vision of an agent powered by foundation models capable of executing the entire machine learning research lifecycle.
From inventing ideas and writing code to executing experiments and drafting the manuscript, the system demonstrated that end-to-end automation of the scientific process is possible.
Soon after, we shared a historic update: the improved AI Scientist-v2 produced the first fully AI-generated paper to pass a rigorous human peer-review process.
Today, we are happy to announce that “The AI Scientist: Towards Fully Automated AI Research,” our paper describing all of this work, along with fresh new insights, has been published in @Nature!
This Nature publication consolidates these milestones and details the underlying foundation model orchestration. It also introduces our Automated Reviewer, which matches human review judgments and actually exceeds standard inter-human agreement.
Crucially, by using this reviewer to grade papers generated by different foundation models, we discovered a clear scaling law of science. As the underlying foundation models improve, the quality of the generated scientific papers increases correspondingly. This implies that as compute costs decrease and model capabilities continue to exponentially increase, future versions of The AI Scientist will be substantially more capable.
Building upon our previous open-source releases (https://t.co/H1tBT14Yx8), this open-access Nature publication comprehensively details our system's architecture, outlines several new scaling results, and discusses the promise and challenges of AI-generated science.
This substantial milestone is the result of a close and fruitful collaboration between researchers at Sakana AI, the University of British Columbia (UBC) and the Vector Institute, and the University of Oxford. Congrats to the team!
@_chris_lu_@cong_ml@RobertTLange@_yutaroyamada@shengranhu@j_foerst@hardmaru@jeffclune