TE
TechEcho
Home24h TopNewestBestAskShowJobs
GitHubTwitter
Home

TechEcho

A tech news platform built with Next.js, providing global tech news and discussions.

GitHubTwitter

Home

HomeNewestBestAskShowJobs

Resources

HackerNews APIOriginal HackerNewsNext.js

© 2025 TechEcho. All rights reserved.

The Darwin Gödel Machine: AI that improves itself by rewriting its own code

245 pointsby birriel6 days ago

25 comments

jerpint6 days ago
I have a feeling LLMs could probably self improve up to a point with current capacity, then hit some kind of wall where current research is also bottle necked. I don’t think they can yet self improve exponentially without human intuition yet , and the results of this paper seem to support this conclusion as well.<p>Just like an LLM can vibe code a great toy app, I don’t think an LLM can come to close to producing and maintaining production ready code anytime soon. I think the same is true for iterating on thinking machines
评论 #44136962 未加载
评论 #44137167 未加载
评论 #44139638 未加载
评论 #44138456 未加载
评论 #44143043 未加载
评论 #44137566 未加载
评论 #44137416 未加载
评论 #44137558 未加载
评论 #44141297 未加载
评论 #44137988 未加载
评论 #44138921 未加载
评论 #44138174 未加载
评论 #44138077 未加载
评论 #44143918 未加载
评论 #44137180 未加载
评论 #44144764 未加载
评论 #44137003 未加载
评论 #44138032 未加载
Lazarus_Long6 days ago
For anyone not familiar this is SWE <a href="https:&#x2F;&#x2F;huggingface.co&#x2F;datasets&#x2F;princeton-nlp&#x2F;SWE-bench" rel="nofollow">https:&#x2F;&#x2F;huggingface.co&#x2F;datasets&#x2F;princeton-nlp&#x2F;SWE-bench</a><p>One of the examples in the dataset they took from<p><a href="https:&#x2F;&#x2F;github.com&#x2F;pvlib&#x2F;pvlib-python&#x2F;issues&#x2F;1028">https:&#x2F;&#x2F;github.com&#x2F;pvlib&#x2F;pvlib-python&#x2F;issues&#x2F;1028</a><p>What the AI is expected to do<p><a href="https:&#x2F;&#x2F;github.com&#x2F;pvlib&#x2F;pvlib-python&#x2F;pull&#x2F;1181&#x2F;commits&#x2F;89d2a17c18b30b61cef31a84caa2bab7aec3b78f">https:&#x2F;&#x2F;github.com&#x2F;pvlib&#x2F;pvlib-python&#x2F;pull&#x2F;1181&#x2F;commits&#x2F;89d2...</a><p>Make your own mind about the test.
评论 #44138844 未加载
vidarh5 days ago
I&#x27;ve built a coding assistant over the last two days. The first 100 lines or so were handwritten. The rest has been written by the assistant itself.<p>It&#x27;s written its system prompt. It&#x27;s written its tools. Its written the code to reload the improved tools into itself.<p>And it knows it is working on itself - it frequently tries to use the enhanced functionality, and then expresses what in a human would be frustration at not having immediate access.<p>Once by trying to use ps to find its own pid in an apparent attempt to find a way to reload itself (that&#x27;s the reason it gå before trying to run ps, anyway)<p>All its commits are now authored by the tool, including the commit messages. It needs to be good, and convincing, and having run the linter and the test suite for me to let it commit, but I agree a substantial majority of the time. It&#x27;s only caused regressions once or twice.<p>A bit more scaffolding to trigger an automatic rollback in the case of failure and giving it access to a model I won&#x27;t be charged by the token for, and I&#x27;d be tempted to let it out of the box, so to speak.<p>Today it wrote its own plan for what to add next. I then only told it to execute it.<p>A minor separate goal oriented layer guiding the planning, and it could run in a loop.<p>Odds are it&#x27;d run off the rails pretty quickly, but I kinda want to see how far it gets.
评论 #44152229 未加载
评论 #44149084 未加载
foobarian6 days ago
I find the thing really missing from current crop of AI systems is continuous retraining with short feedback loops. Sounds expensive to be sure, but it seems like what biological systems do naturally. But would be pretty awesome to watch happen
评论 #44137270 未加载
评论 #44137780 未加载
yahoozoo6 days ago
Isn’t one of the problems simply that a model is <i>not</i> code but just a giant pile of weights and biases? I guess it could tweak those?
评论 #44137878 未加载
评论 #44137756 未加载
评论 #44137920 未加载
pegasus6 days ago
I&#x27;m surprised they still hold out hope that this kind of mechanism could ultimately help with AI safety, when they already observed how the reward-hacking safeguard was itself duly reward-hacked. Predictably so, or at least it is to me, after getting a very enlightening introduction to AI safety via Rob Miles&#x27; brilliant youtube videos on the subject. See for example <a href="https:&#x2F;&#x2F;youtu.be&#x2F;0pgEMWy70Qk" rel="nofollow">https:&#x2F;&#x2F;youtu.be&#x2F;0pgEMWy70Qk</a>
akkartik6 days ago
<i>&quot;We did notice, and documented in our paper, instances when the DGM hacked its reward function.. To see if DGM could fix this issue.. We created a “tool use hallucination” reward function.. in some cases, it removed the markers we use in the reward function to detect hallucination (despite our explicit instruction not to do so), hacking our hallucination detection function to report false successes.&quot;</i><p>So, empirical evidence of theoretically postulated phenomena. Seems unsurprising.
评论 #44139374 未加载
dimmuborgir6 days ago
From the paper:<p>&quot;A single run of the DGM on SWE-bench...takes about 2 weeks and incurs significant API costs.&quot; ($22,000)
hardmaru6 days ago
If you are interested, here is a link to the technical report:<p><a href="https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2505.22954" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2505.22954</a><p>Also the reference implementation on GitHub:<p><a href="https:&#x2F;&#x2F;github.com&#x2F;jennyzzt&#x2F;dgm">https:&#x2F;&#x2F;github.com&#x2F;jennyzzt&#x2F;dgm</a><p>Enjoy!
OtherShrezzing6 days ago
This is an interesting article in general, but this is the standout piece for me:<p>&gt;For example, an agent optimized with Claude 3.5 Sonnet also showed improved performance when powered by o3-mini or Claude 3.7 Sonnet (left two panels in the figure below). This shows that the DGM discovers general agent design improvements rather than just model-specific tricks.<p>This demonstrates a technique whereby a smaller&#x2F;older&#x2F;cheaper model has been used to improve the output of a larger model. This is backwards (as far as I understand). The current SOTA technique typically sees enormous&#x2F;expensive models training smaller cheaper models.<p>If that&#x27;s a generalisable result, end-users should be able to drive down their own inference costs pretty substantially.
评论 #44137275 未加载
评论 #44141735 未加载
ordinarily6 days ago
The pieces are coming together quickly <a href="https:&#x2F;&#x2F;ai-2027.com&#x2F;" rel="nofollow">https:&#x2F;&#x2F;ai-2027.com&#x2F;</a>.
评论 #44136868 未加载
评论 #44137390 未加载
评论 #44136625 未加载
评论 #44137132 未加载
评论 #44137461 未加载
评论 #44136811 未加载
评论 #44136738 未加载
Frummy6 days ago
More like an AI that recursively rewrites an external program (while itself is frozen), which makes it more similar to current cursor lovable etc type of stuff
artninja19886 days ago
The results don&#x27;t seem that amazing on SWE compared to just using a newer llm but at least sakana is continuing to try out interesting new ideas.
guerrilla6 days ago
This feels like playing pretend to me. There&#x27;s no reason to assume that code improvements matter that much in comparison to other things and there&#x27;s definitely no reason to assume that there isn&#x27;t a hard upper bound on this kind of optimization. This reeks of a lack of intellectual rigor.
andoando6 days ago
This seems to be just fovused on changing the tools and workflows it uses, nothing foundational
评论 #44137309 未加载
ringeryless5 days ago
does anyone do due diligence on corporate names before launching? Sakana is a popular slang spelling of sacana, or bastard, in Português. I suppose self modifying code can be considered such, in some circumstances, but willingly pointing this out is probably less than stellar marketing.
评论 #44142089 未加载
评论 #44144430 未加载
ge966 days ago
Plug it into an FPGA so it can also create &quot;hardware&quot; on the fly to run code on for some exotic system
rahen5 days ago
Isn&#x27;t this violating the first rule of AI safety: do not let an AI change its code?
alastairr6 days ago
I wondered if something similar could be achieved by wrapping evaluation metrics into Claude code calls.
p1dda5 days ago
Garbage in, garbage out, AI hype will never die, no doubt
interludead6 days ago
Sounds nice! Especially with the Sakana&#x27;s latest development of Continuous Thought Machine. The next step should be to let foundation models fine-tune themselves based on their &#x27;history of what has been tried before&#x27; and new data
htrp6 days ago
do people think sakana is actually using these tools or are they just releasing interesting ideas that they aren&#x27;t actually actively working?
billab9956 days ago
When does it begin to learn at a geometric rate?
zackmorris3 days ago
This is good but you want to use a functional programming (FP) language with lightweight syntax like Lisp that translates directly to&#x2F;from the intermediate code (icode) tree without additional parsing. Genetic Programming by John Koza explains it in detail:<p><a href="https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Genetic_programming" rel="nofollow">https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Genetic_programming</a><p>I read the 3rd edition:<p><a href="https:&#x2F;&#x2F;www.amazon.com&#x2F;Genetic-Programming-III-Darwinian-Invention&#x2F;dp&#x2F;1558605436" rel="nofollow">https:&#x2F;&#x2F;www.amazon.com&#x2F;Genetic-Programming-III-Darwinian-Inv...</a><p>That way all processing resources can go towards exploring the problem space for potential solutions close to the global minimum or maximum, instead of being wasted on code containing syntax errors that won&#x27;t execute.<p>So the agent&#x27;s real-world Python LLM code would first be transpiled to Lisp and evolved internally, then after it&#x27;s tested and shown to perform better imperically than the original code, be translated back and merged into the agent.<p>Then the challenge becomes transpiling to&#x2F;from other imperative programming (IP) languages like Python, which is still an open problem:<p>-<p>Going from Lisp to Python (or running Lisp within Python) is trivial, and I&#x27;ve seen implementations for similar IP languages like C++ in like 1 page of code. They pop up on HN frequently.<p>But going from Python to Lisp (or running Python within Lisp) is a lot harder if one wishes to preserve readability, which may or may not matter here. Naive conversions bind variables under pseudonyms, so a Python variable like my_counter becomes int_123 and it works like an emulator, merely executing the operations performed by the Python code. Mutability gets buried in monadic logic or functional impurity which has the effect of passing the buck rather than getting real work done. Structs, classes, associative arrays, etc lose their semantic meaning and appear as a soup of operations without recognizable structure.<p>To my knowledge, nobody has done the hard work of partitioning imperative code into functional portions which can be transpiled directly to&#x2F;from FP code. Those would only have const variables and no connection to other processes of execution other than their initial and final values, to be free of side effects and be expressible as prefix&#x2F;postfix&#x2F;infix notation without change to logic, as imperative or functional code.<p>Mutability could be represented as shadowed variables within ephemeral functional sub-scopes, or by creating new value names for each mutation and freeing the intermediate variables via reference counting or garbage collection. Think of each new value as running in a forked version of the current process, with only that value being different after copy-on-write. A simple for-loop from 1 to 1000 would run that many forked processes, keeping only the last one, which contains the final value of the iterator.<p>Mutability can also be represented as message passing between processes. So the FP portions would be ordinary Lisp, glued together with IO functions, possibly monadic. I don&#x27;t like how Haskell does this, mainly because I don&#x27;t fully understand how it works. I believe that ClojureScript handles mutability of its global state store by treating each expression as a one-shot process communicating with the store, so that the code only sees initial and final values. While I don&#x27;t know if I understand how that works, I feel that it&#x27;s a more understandable way of doing things, and probably better represents how real life works, as explained to me in this comment about Lisp Flavored Erlang (LFE) and Erlang&#x27;s BEAM (see parent comments for full discussion):<p><a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=43931177">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=43931177</a><p>Note that FP languages like Lisp are usually more concerned with types and categories than IP languages, so can have or may need stronger rules around variable types to emulate logic that we take for granted in IP languages. For example, Lisp might offer numbers of unlimited size or precision that need to be constrained to behave like a float32. Similar constraints could affect things like character encoding and locale.<p>-<p>I first learned about everything I just explained around 2005 after reading the book. I first had thoughts about brute-forcing combinations to solve small logic circuit and programming challenges during my electrical and computer engineering (ECE) courses at UIUC in the late 1990s, because it took so much mental effort and elbow grease to create solutions that are obvious in hindsight.<p>Then the Dot Bomb happened, the Mobile bubble happened, the Single Page Application bubble happened, and the tech industry chose easy instead of simple:<p><a href="https:&#x2F;&#x2F;www.infoq.com&#x2F;presentations&#x2F;Simple-Made-Easy&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.infoq.com&#x2F;presentations&#x2F;Simple-Made-Easy&#x2F;</a><p>This is why we chose easy hardware like GPUs over simple highly multicore CPUs, and easy languages like Ruby&#x2F;React over simple declarative idempotent data-driven paradigms like HTTP&#x2F;HTML&#x2F;htmx.<p>The accumulated technical debt of always choosing the quick and easy path set AI (and computing in general) back decades. The AI Winter, endless VC wealth thrown at non-problems chasing profit, massive wealth inequality, so many things stem from this daily application of easy at the expense of simple.<p>I wish I could work on breaking down IP languages like Python into these const functional portions with mutability handled through message passing in LFE to create an IP &lt;-&gt; FP transpiler for optimization, automatic code generation and genetic algorithm purposes. Instead, I&#x27;ve had to survive by building CRUD apps and witness the glacial pace of AI progress from the sidelines.<p>It may be too late for me, but maybe these breadcrumbs will help someone finally get some real work done.
2OEH8eoCRo06 days ago
We could be on a path to sentient malicious AI and not even know it.<p>AI: Give me more compute power and I&#x27;ll make you rich!<p>Human: I like money<p>AI: Just kidding!
评论 #44137716 未加载
评论 #44136879 未加载