Skip navigation
New Paper! Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents A longstanding goal…
0:00 / 0:27

New Paper! Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents A longstanding goal…

hardmaru

444K followers

Visit on X
1,032 likesDownloadOriginal post

105,102 views

New Paper! Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents A longstanding goal of AI research has been the creation of AI that can learn indefinitely. One path toward that goal is an AI that improves itself by rewriting its own code, including any code responsible for learning. That idea, known as a Gödel Machine, proposed by Jürgen Schmidhuber over two decades ago, is a hypothetical self-improving AI. It optimally solves problems by recursively rewriting its own code when it can mathematically prove a better strategy, making it a key concept in meta-learning or “learning to learn.” While the theoretical Gödel Machine promised provably beneficial self-modifications, its realization relied on an impractical assumption: that the AI could mathematically prove that a proposed change in its own code would yield a net improvement before adopting it. Sakana AI, in collaboration with Jeff Clune’s lab at UBC, proposes something more feasible: a system that harnesses the principles of open-ended algorithms like Darwinian evolution to search for improvements that empirically improve performance. We call the result the Darwin Gödel Machine. DGMs leverage foundation models to propose code improvements, and use recent innovations in open-ended algorithms to search for a growing library of diverse, high-quality AI agents. Applied to practical tasks, we implemented Darwin Gödel Machine as a self-improving coding agent that rewrites its own code to improve performance on programming tasks. It creates various self-improvements, such as a patch validation step, better file viewing, enhanced editing tools, generating and ranking multiple solutions to choose the best one, and adding a history of what has been tried before (and why it failed) when making new changes (see the attached video). We believe that Darwin Gödel Machines represent a concrete step towards AI systems that can autonomously gather their own stepping stones to learn and innovate forever!

10 Comments

Comments from the original post on X (Twitter)

  • hardmaru's profile picture

    @hardmaru

    Link to the full technical report: Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents Huge congrats to authors @jennyzhangzt @shengranhu @cong_ml @RobertTLange @jeffclune on this amazing progress! 🚀 Blog post:

    84 likes

  • hardmaru's profile picture

    @hardmaru

    @shengranhu @cong_ml @RobertTLange @jeffclune Want to learn more about Gödel machines? Check out the resources on @SchmidhuberAI’s website. He believes the path to strong AI requires “self-replicating, self-improving, life-like machinery”.

    35 likes

  • Coral AI News's profile picture

    @CoralAINews

    Coral AI is the most powerful AI for documents. See the difference yourself:

    19,490 likes

  • beff – e/acc's profile picture

    @BasedBeffJezos

    @Plinz So cool man. I'm so happy you're actually pursuing contrarian and nature-inspired directions for research

    11 likes

  • hardmaru's profile picture

    @hardmaru

    @Plinz Thanks for your support, @BasedBeffJezos and @Plinz 🔥🔥

    10 likes

  • VeryLikelyHuman's profile picture

    @damnThoughtful

    @bensprecher Isn't that the same as alpha evolve/funsearch?

    4 likes

  • Greg's profile picture

    @gregmagdesian

    Great work! Question: If an evolutionary approach can be quantified but not proven within its original system, as it recursively rewrites itself with each iteration, does it not generate new abstractions upon abstractions, mathematically detached from the original system?

    1 like

  • dotSQUARE's profile picture

    @dt_sqr

    Hypothesis: a language-model reply is likeliest to be sound when the model’s hidden-state vectors align like a pure quantum state, i.e. when their density matrix has very low von Neumann entropy. Implementation: during or after generation capture the final-layer activations, average a few stochastic passes, form ρ = |ψ⟩⟨ψ|, compute S(ρ). Treat 1 − S as a “coherence” score—reject or rewrite low-score answers, or fine-tune the model by adding λ·S(ρ) to the loss.

    1 like

  • mekaneeky - e/acc's profile picture

    @GPLv6

    This has been the aim of our decentralized AutoML work on @bittensor_. To realize a practical Gödel Machine that can bootstrap AI progress autonomously. Very happy to collaborate on deploying a variant of this that runs on a global SETI@Home like decentralized cluster.

    1 like

  • SubMinima's profile picture

    @subminima

    Can system be made to work in the weights space too?

    1 like