> If this benchmark becomes popular, then presumably to avoid such embarrassments synthetic data is eventually added to training sets to make sure even esolangs are somewhat more in-distro
"After the paper was finalized, we ran agentic systems that mimic how humans would learn to solve problems in esoteric languages. We supplied our agents with a custom harness + tools on the same benchmark. They absolutely crushed the benchmark. Stay tuned"
Zooming out a little, all the ai companies invested a lot of resources into safety research and guardrails, but none of that prevented a "straightforward" misalignment. I'm not sure how to reconcile this, maybe we shouldn't be so confident in our predictions about the future? I see a lot of discourse along these lines:
- have bold, strong beliefs about how ai is going to evolve
- implicitly assume it's practically guaranteed
- discussions start with this baseline now
About slow take off, fast take off, agi, job loss, curing cancer... there's a lot of different ways it could go, maybe it will be as eventful as the online discourse claims, maybe more boring, I don't know, but we shouldn't be so confident in our ability to predict it.
I thought I was the only person going crazy by the new default behavior not showing the file names! Please don't expect users to understand your product details and config options in such detail, it was working well before, let it remain. Or at least show some message like "to view file names, do xyz" in the ui for a few days after such a change.
While we're here, another thing that's annoying: the token counter. While claude is working, it read some files, makes an edit, let's say token counter is at 2k tokens, I accept the edit, now it starts counting very fast from 0 to 2k and then shows normal inference speed changes to 2.1k, 2.3k etc. So wanted to confirm: is that just some UI decision and not actually using 2k tokens again? If so, it would be nice to have it off, just continue counting where you left off.
Another thing: is it possible to turn off the words like finagling and similar (I can't remember the spelling of any of them) ?
I initially planned to run a chess tournament for LLMs but they are not good: besides obvious mistakes, they output incorrect moves, get stuck in loops by repeating the same moves and the smaller models fail to output valid json frequently. I thought the reasoning models like o3 mini might be good, but they are an incremental improvement in chess.
Feedback and suggestions for other games to explore welcome.
The article mentions, he is going to run the marathon, looking forward to what he can do in that distance. I feel it's only a matter of time until someone breaks the 2 hour barrier in an official race. Lot of people thought it would be Kelvin Kiptum, unfortunately he passed away in an accident.
The impact this organization had was incredible. I doubt they would have been able to do this work if they were based out of any other country, which makes me wonder how the US legal system, regulators and law enforcement in general are not extremely corrupt. What reasons or incentives make the system work in the US? Of course there are many instances of corruption and injustice, but in comparison to almost any other country, it seems to work surprisingly well.
Great work! When I use models like o1, they work better than sonnet and 4o for tasks that require some thinking but the output is often very verbose. Is it possible to get the best of both worlds? The thinking takes place resulting in better performance but the output is straightforward to work with like with sonnet and 4o. Did you observe similar behaviour with the 1B and 3B models? How does the model behaviour change when used for normal tasks that don't require thinking?
Also how well do these models work to extract structured output? Eg- perform ocr on some hand written text with math, convert to html and format formulas correctly etc. Single shot prompting doesn't work well with such problems but splitting the steps into consecutive api calls works well.
https://x.com/lossfunk/status/2034637505916792886
"After the paper was finalized, we ran agentic systems that mimic how humans would learn to solve problems in esoteric languages. We supplied our agents with a custom harness + tools on the same benchmark. They absolutely crushed the benchmark. Stay tuned"
A little harness engineering was enough!