These kinds of arguments are obviously vacuous: The US could also "win" any war with a non-nuclear opponent by nuking every major city. But the fact that they don't is not something worth calling restraint.
Is the term "One-shotting" here even appropriate? In the context of AI one-shotting means giving your llm a single prompt and having it return a surprisingly good result (usually code related) without requiring further input, but the llm (probably) still iterates heavily before showing the results. When Codex "one-shots" something it depends a ton on the fact that code is easily testable.
Lately "one-shotting" means "without human steering / feedback"
I think its a consequence of the improvements in harnesses and LLMs of the last year. Before they had a hard time self-reviewing the results and proceeding to a second iteration, now they are quite good at it although not perfect.
"using the same approach that Buckmaster and Alpoge had been exploring" is imo mealy wording: it seems fairly likely that OA heard Buckmaster and Alpoge were close to a breakthrough, and decided to use their unlimited compute to quickly prompt based on their assumptions about B&As work.
Citation of what? this was unpublished work! This computational blitz really just reads as "might makes right" on OpenAi's part... which isn't surprising, but they should probably be honest about what they've done here.
So, leaving aside the idea that OA might've used data from the researchers Codex sessions: Do I understand correctly that the internal OpenAI work on the problems was probably started after they heard Alpoge and Buckmaster had made process by using their models? And they used the publicly available info about the researchers past work to prompt their models?
If compute is cheap, and the difficult thing with scientific discovery is now mostly in steering agents into promising areas, there's an obvious incentive for OA mathematicians to simply monitor closely which researchers are close to releasing exciting results, make some assumptions about their prompts based on their past work, and quickly prompt their own (stronger) model to look into the same areas.
But there's a bunch of people already in this thread calling that stuff unfounded speculation (which I disagree with), and my point is that even if that specific thing isn't true, OA's behavior here is obviously awful.
If they're going to try to beat researchers to discoveries like this it disincentives researchers to talk about their progress publicly, and basically breaks the ecosystem of scientific cooperation / discovery. It's also immoral.
Yep. The most uncharitable view of this might be: they stole the work of researchers to build their models, and now they're using said models to steal the proceeds of future work, too.
It's literally a toggle in the options for ChatGPT, one which is on by default and most researchers probably have on without realising it.
So to say that it is unlikely is extremely suspicious. No, they did not literally pull user data. But user data is automatically added to their training set by default, so their latest in-house model would be trained on it if it is from several months ago. It isn't intentional on their part, and they probably realised they could not refute that they trained on Tristan's logs unintentionally, hence why they acted the way they did.
Well, of course Anthropic employees would say that, since they likely do the same. Claiming that your primary competitor doesn't engage in a certain malicious practice is supposed to make it look as if there's no way you would too. If somebody even says that about their competitor, then surely there must be truth to that, otherwise you would never give credit to someone you're opposed to.
By default OA trains their models on codex-sessions. If I understand him correctly this is something Tristan explicitly mentions in his post as a possible reason for the fast results obtained by the internal OA team. Anthropic obviously doesn't want to challenge the idea that training is transformative, even if it means agreeing with their competitor.
I’ve thought a lot about publishing research and wanting to do more of it, but right as I finally had the time and energy to start writing articles LLMs start to take off. Now all of a sudden, I’m acutely aware that everything I publish will be used for AI training.
For math, a field that is built on incremental research it feels like AI labs will do nothing but discourage publishing research at all for fear that they will be able to spend the money for compute that publicly funded academia simply cannot afford.
It feels like publishing anything at this point just means that your work will be fed to a machine that will make sure your work will never been seen by anyone else because it will always be the ones making the “true advancements”.
Perhaps I’d feel better about this if AI labs really existed for humanity’s benefit, but for some reason I don’t think that comes up in their investor slide decks.
It's honestly unsurprising and not a problem that they do this in my view. The problem really starts when you start taking credit for work that they would've achieved.
Like if i go to a talk on unfinished work, it's not really unethical for me to think about the problem--it's a problem if i scoop the authors but these problems can often be solved by collaboration or proper crediting and timing--IN MY VIEW
Won't deny that this is an interesting idea, but I feel like waiting on the output of an LLM for 40 hours feels like it is completely antithetical to what makes classic Hackathons appealing / educative.
More generally, I don't think the shape of a hackathon (intensely working for a short timespan) maps at all onto the way LLM Math progress has seemingly been made so far; AFAIK it mostly involves picking out something for the Model, then having it run for a week with sporadic correction / encouragement.
1. IMO the hard part isn't prompting. It's selecting the problem and understanding
the solution. It'd be especially exciting if a participant formulates their own conjecture, proves it with AI, then generalizes it to a new theory.
2. We have talked to mathematicians and frontier lab employees. We think 40 hours is enough to produce interesting results.
But it's not "waiting on the output of an LLM for 40 hours" any more than a regular hackathon is "waiting for my damn teammates to finish their part for 40 hours". From my experience using agentic coding for hackathons, the best teams are those that coordinate with the AI agents in relatively quick cadence, generally giving it small tasks and steering it often. Teams may want to run some long-running sessions too, especially closer to the deadline, but even then, they'd probably want to run and follow several sessions in parallel, and continuously inspect their work so that they have reasonable confidence that their main efforts will wrap up before the deadline. There is an art to it.
I can't wait to see if the team that fares best is the one that steers often or the one that interferes the least.
So far humans failed at those problems. Also IIRC there was a guy that proved a substantial problem 2-3 months ago by basically pasting over and over "keep looking for a solution" or something like that for 2 days with little formal math background.
Have you done any math hacking with sol/astra or fable? It’s more fun than using them for coding. The models are great at the monotony, like constructing a Gröbner-basis, etc. But they’re all still absolutely awful at coming up with new ideas, new proof methods, or new constructive forms. So you spend all your time on coming up with novel hypotheses yourself and handing off the rote work to an agent.
It’s also quite fun to get instant results by finding isomorphisms into unfamiliar areas of mathematics that previously would’ve required some networking in order to build a collaborative relationship.
"A mathematician is a person who can find analogies between theorems; a better mathematician is one who can see analogies between proofs and the best mathematician can notice analogies between theories. One can imagine that the ultimate mathematician is one who can see analogies between analogies."
I wonder how models perform on finding analogies between analogies
If the goal is to accomplish something then why limit yourself with available tools?
I’m not a full on AI optimist but it is absolutely the most powerful tool in a host of applications. From a Hackathon perspective, obviously in the 90s it was much more unorganized, but the same ethos existed. Use all available tools to accomplish the goal/task, it’s where a lot of incredible learning came out of. The same will hopefully happen in scenarios like this one
When Claude made progress on the Riemann conjecture, here are the kind of prompts used:
> Jarred's input was mostly limited to sending Claude messages of encouragement (mostly variants of “keep going” or “believe in yourself”).2 This seems to have helped Claude overcome some initial skepticism that it could make meaningful progress.
And left it for a long time. Jarred isn't a mathematician, he's the maintainer of a janky JavaScript environment.
Prompting for some of the results was almost the "Computer, do a breakthrough. Make no mistakes." meme. Just someone telling the model to keep trying a couple of times.
Unfortunately we don't actually know what kind of prompting was done for the more prominent results.
Same here, it's one thing if I'm just screwing around, but if I'm trying to do anything serious, I need to at least have a handle on what it's doing, and, when thinking traces are available, keeping track of any logical errors in the model's reasoning.
>"illegible reasoning in a few reinforcement-learning environments over long rollout"
Yet, I get the point that you're making: those tokens essentially are an internal scratchpad for the LLM which isn't required to logically lead to the output.
Still not meaningful -- https://arxiv.org/pdf/2504.09762; even for local models, the reasoning traces are often filtered and summarized to sound sensible to humans. And even if not, they don't necessarily represent what the model is thinking.
That's how several major AI advancements have happened. I have seen no evidence that that is the fastest way to make progress right now. I expect that, much like chess engines, it will not take too long before AI is significantly better than AI + human. But right now, my bet is that we are still safely within the window where an AI + human mathematician team is still better than AI alone (at least for the case where the human has learned how to work effectively with the partner....something that this event could possible be good for teaching).
I suspect that the best progress will be made by a team that purely spends their time taking a list of open problems and promoting "solve <problem>", without actually trying to understand anything. Just keep as many problems in flight as you can across as many sessions as you can.
You can probably ask the AI to come up with a list of problems itself, and rank them by the likelihood of progress.
Then 20 years go by and you wake up one day with questions that you cannot get out of your mind: why did I start prompting the LLM for? Why did I need these random proofs for? What do i do with my repo with 2billion lines of Lean?
That sounds like adding a bottleneck, unless you mean writing a harness that automatically asks the model to keep going, so that there's no humans involved at all?
But the goal here is not to run 42km; to stay with the outdoors metaphor, it's more like deciding where and how to set up a bivouac - use whatever tools you have at your disposal to analyze the area you're in, and find the best site to stay in overnight.
Because the companies that run frontier models are malevolent by every metric.
They are destroying the environment, especially those in neighborhoods of low income people.
They are empowering their owners who are some of the most deplorable and duplicitous people living.
They are destroying personal compute to avoid competition with local models by buying all computer components with “promised money” and forcing their P into AI.
They stole the entire creative output of humanity and are trying to sell it back to us.
They are only good for giving wealth access to skill while removing from the skilled the ability to access wealth.
They are being used to kill in war and for surveillance.
Seriously why would you use them? Your use only emboldens them; making you complicit in their nefarious success.
Did you read the article? This is not a hackathon where you build software, it’s one where you’re trying to get a model to make progress on a frontier math problem. The point is that that activity may not map well onto the shape of a hackathon
Lots of people seem to care about whether a project was built by hand over time by a few committed developers, or if it was vibe-coded in the past week by a single person.
Whether the distinction matters or not, it would be nice for everyone involved if projects falling into the second category were more upfront about that.
I don't think it's needed or a fair ask. Even though project clearly shows one of the contributors as Claude. There are excellent AI assisted software out there and shite software "crafted" by bare fist and knuckles violently banging a mechanical keyboard in a glorious hacker dungeon.
When I commit I either mark it as my own commit if I've read through every line of that commit at least once (not skimmed), pushed back, changed, etc, or put as "ai@<a word I use>" if it was mostly "fine, do it…" or the "whatever… " phase. I never let AI write a single word of documentation or comment. And I made sure Claude removed its attribution the first time I started using it. Besides I was using GLM 5.2 at that time so Claude was not a fit anyway. I don't think they should have that default behaviour.
Yes, a project maintained and built over longer time has a bigger chance of staying relevant, than someone's weekend fun they will abandon once the dopamine hit of "making something" wears off.
Isn't software written by hand more likely to be abandoned due to the extremely costly nature of keeping its development going? Abandonware definitely existed prior to vibe coding. As long as it's open source, now you can pick it up and keep vibe-developing it if you care to, for very little money/effort.
Possibly, but sunk costs fallacy often means it's possible to rely on the maintainer to stick around. I think we all know how disposable LLM-generated code can be. If I wanted to vibe code my own thing, I'd do it my way rather than fork an abandoned project.
What's needed is more of a nexus thing where there's evidence multiple people work together or some long track record of someone scratching a very specific itch. Or if we're all using agents to build the same thing some way for them to team up. It's a sort of proof of life thing and it's difficult to distinguish those from flashes in the pan.
Rather than sunk cost fallacy, I would call it survivorship bias: the projects that someone has consistently worked on over time are a small fraction of the total projects out there, and those that have survived are probably useful to someone (at the very least, their developers).
I'm not sure what decision you're talking about.
My comment just mentions that it would be convenient if disclosure of vibe-coding was more common.
I can look for myself and see that the repo was started a week ago and has had hundreds of commits since, and that the author exclusively talks about ai on twitter. I'd just like it if I could save myself some time, and it would prevent useless discussions in this thread were people argue whether the commit-messages look like ai.
I'll admit to having some bias against vibe coding, but trying to put that aside I think it depends on the purpose of the project. If it attempts to be an improvement over existing software then absolutely yes it would, things like speed, performance, security and reliability are not something I'd trust vibe coding for. If its for doing something novel where the previously mentioned is less of the focus, I can imagine myself looking past that, sure.
It's interesting that this article is mostly focused on the merits of living a simple / predictable life, whereas I get the impression that the main improvement here might be just not working from home anymore (I guess the author agrees since they link to this older post: https://mrmarket.bearblog.dev/remote-work-is-bad-for-you/).
Like the author, I enjoyed wfh for some time, but also found this creeping crushing loneliness settling on my brain after some years. My solution was to find a co-working space that actually has a community (most spaces are just places to rent a desk). It's helped me set up a routine and get some water cooler talk in a few times a week.
But more importantly, through the community I've met some new friends, so my social life blew up. Now I rely less on the co-working space and more on meeting with the new friend group for sports and debates.
It's been an adventure in finding out what makes me tick and finding ways of addressing it in healthy ways.
Nice! Any tips for finding spaces like that? I guess it's mostly finding coworking spaces that are focused around a specific discipline (I know there are a few gamedev ones for example) so that you already have a lot in common?
Couldn't it just be a moral thing? Ie "This thing was trained on the work of people it's now replacing and I think that sucks". (NB I'm not looking to get into an argument about the actual impact of these kinds of small moral stances)
reply