[HN Gopher] Darwin Godel Machine: Open-Ended Evolution of Self-I...
___________________________________________________________________
Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents
Author : tzury
Score : 26 points
Date : 2025-06-11 18:07 UTC (4 hours ago)
(HTM) web link (arxiv.org)
(TXT) w3m dump (arxiv.org)
| jinay wrote:
| I recently did a deep dive on open-endedness, and my favorite
| example of its power is Picbreeder from 2008 [1]. It was a simple
| website where users could somewhat arbitrarily combine pictures
| created by super simple NNs. Most images were garbage, but a few
| resembled real objects. The best part is that attempts to
| replicate these by a traditional hill-climbing method would
| result in drastically more complicated solutions or even no
| solution at all.
|
| It's a helpful analogy to understand the contrast between today's
| gradient descent vs open-ended exploration.
|
| [1] First half of https://www.youtube.com/watch?v=T08wc4xD3KA
|
| More notes from my deep dive:
| https://x.com/jinaycodes/status/1932078206166749392
| publicdaniel wrote:
| Did you see their recent paper building on this? Throwback to
| Picbreeder!
|
| https://x.com/kenneth0stanley/status/1924650124829196370
| jinay wrote:
| Ooh I haven't, but this is exactly the kind of follow-up I
| was looking for. Thanks for sharing!
| jinay wrote:
| Timestamped link to the YouTube video:
| https://youtu.be/T08wc4xD3KA?t=124
| bwest87 wrote:
| This video was fascinating. I didn't know about "open
| endedness" as a concept but now that I see it, of course it's
| an approach.
|
| One thought... in the video, Ken makes the observation that it
| takes way more complexity and steps to find a given shape with
| SGD vs. open-endedness. Which is certainly fascinating.
| However...
|
| Intuitively, this feels like a similar dynamic is at play with
| the "birthday paradox". That's where if you take a room of just
| 23 people, there is a greater than 50% chance that two of them
| have the same birthday. This is very surprising to most people.
| It seems like you should need way more people (365 in fact!).
| The paradox is resolved when you realize that your intuition is
| asking how many people it takes to have _your_ birthday. But
| the situation with a room of 23 people is implicitly asking for
| just one connection among _any two_ people. Thus you don 't
| have 23 chances, you have 23 ^ 2 = 529 chances.
|
| I think the same thing is at work here. With the open-ended
| approach, humans can find _any_ pattern at _any_ generation.
| With the SGD approach, you can only look for _one_ pattern. So
| it 's just not an apples to apples comparison and sort of
| misleading / unfair to say that open-endedness is way more
| "efficient", because you aren't asking it to do the same task.
|
| Said another way, I think with the open-endedness, it seems
| like you are looking for thousands (or even millions) of shapes
| simultaneously. With SGD, you're kinda flipping that around,
| and looking for exactly 1 shape, but giving it thousands of
| generations to achieve it.
| yodon wrote:
| Is this essentially genetic algorithms for the LLM era?
| mountainriver wrote:
| Yep, the interesting thing is genetic algorithms previously
| were mostly good at course search and less good at fine search.
|
| They also often converge to a local minima, and are costly.
|
| It'll be interesting to see if LLMs change that, or whether we
| are just approximating something a gradient could do better
| clayhacks wrote:
| Earlier discussion: A deep dive into self-improving AI and the
| Darwin-Godel Machine
| https://news.ycombinator.com/item?id=44174856
| seu wrote:
| Yes, seems interesting, but honestly, an abstract that includes
| sentences such as "accelerate AI development and allow us to reap
| its benefits much sooner" and "paths that unfold into endless
| innovation" sounds like written by the marketing team of a AI
| company.
| behnamoh wrote:
| So it's basically "throw spaghetti at the wall and see what
| sticks". It works in evolution because evolution doesn't have an
| end goal to achieve in a certain amount of time, but for AI we
| want to know how long it takes to go from performance A to B.
| Then again, this paper might be yet another validation of the
| bitter truth of machine learning.
| darepublic wrote:
| In the abstract the reference to 'safety' gave me pause. For one
| it seems doubtful that the AI could ever improve enough to cause
| serious trouble, unless of course you equipped it with things
| that just about any piece of software could create trouble with
| --elevated permissions, internet access, network endpoints etc.
|
| They mention putting it in a sandbox which I assume to just mean
| something like a VM or docker container. I wonder if that would
| be sufficient if the AI truly reached singularity level
| intelligence. Could it figure out some kind of exploit to break
| free of its sandbox, and transmit its code over the internet for
| further replication?
| whattheheckheck wrote:
| It already has and its controlling humans to do it!!!
| Teever wrote:
| You may be interested in this link[0]. Someone posted it in
| another thread yesterday.
|
| [0] https://www.aisi.gov.uk/work/replibench-measuring-
| autonomous...
___________________________________________________________________
(page generated 2025-06-11 23:00 UTC)