I counted something like 30 different instances of run-on exclamation marks ("!!!!!!!!!!!") and weird mannerisms ("Waitwaitwaitwait.") in just one GLM 5.3 Flash session. Our token budgets are getting eaten up by this stuff...
I expect it's actually not wasted and there's meaning behind what seems like nonsense to us in helping it achieve it's goal. Which is mildly chilling but not unexpected.
I think this is a known phenomenon: even in non-reasoning models, adding useless/filler tokens before an answer improves task performance. The model is doing some computation during the filler. See: https://arxiv.org/html/2404.15758v1
It is. Processing tokens is the only time model has to do computation, and if you ask it a tough problem, there is some minimal amount of computation it needs to perform to process and solve it - pre-CoT in particular you could guarantee failure by forcing model to be concise, and thus giving it less computational budget than necessary to compute the answer.
(This is I think where people parroting out "stochastic parrot" are stuck even today - not realizing that "predicting next tokens" is hiding arbitrary computation underneath, with token stream acting as input and clock signal...)
if one were to remove the expressions of excitement from the previous messages would it the model continue to demonstrate that same excitement scaling?
Yes this sounds just like the effect where people new to coding AI negatively berate it like a person and it continues to get worse and make more mistakes because that’s what those tokens are related to.
100%, back when it was Ox Alpha I had a little fun trying to guess what it might be by looking at the reasoning and I consistently laughed at how excited it got
> Algorithmic mitigation is being explored, but scientists describe it as “embryonic” — potentially demanding computing power that rivals the science processing itself.
It sounds like this would work? Just be more costly, at least in the short term.
> Engineering fixes are the real answer
What is the fix and tradeoff? The article doesn't seem to cover it at all.
You can still do all that. The only thing going is that in some cases the human isn't going to be able to claim that they made the key insights that first solved the problem.
A lot of this would be public funding.
If AI is reducing the need for mathematicians in an area it makes sense to fund it less and fund something else that isn't solved.
Surely it is. Ask it to keep a list of all the promising sub paths, reprompt the collections of agents again on these after the main problem has been addressed.
Or even release a list of them and let others investigate.
I feel like most of the latest big Chinese lab models are good enough now. If I had to pick a difference it feels like sol goes further towards 1 shotting things and needs less babysitting. But if I am actively prompting and reading the results can get just as far on a lesser model.
