Vibes are all that matter. As soon as you start measuring it, that measurement becomes a target and vendors start optimizing for it at expense of the general usefulness of the model. We’ve seen plenty of models with great benchmark scores flop when people start using it.
If benchmarks didn’t exist we would have to invent them because “vibes” is a ridiculous idea: oh I know I’ll be super unscientific and horrendously biased and that’s far better than a team of experts carefully AND CONTINUALLY developing a variety of benchmarks of varying quality that…hmm all point to the same thing.
You can’t benchmaxx an eval that comes after your model release.
Consider also benchmaxxing makes no sense from an incentive structure: the quality of these models is directly correlated by how well you can measure true performance in the wild. If they were just stupidly benchmaxxing they would be unable to do trustworthy ablations or know how well the model will perform in their product.
Remember the famous case of asserted benchmaxxing from llama 4? The entire org was gutted and the ceo spent billions hiring better people. Every lab takes evaluations extremely seriously.
> You can’t benchmaxx an eval that comes after your model release
Sure you can, just do it silently and don't tell the people hitting your API that the model is different now. Unless it's open weight, we're just taking your word for it. Even better, do a VW and try to detect which benchmark is running, then change to a hyper specialized model that is trained on it.
> Sure you can, just do it silently and don't tell the people hitting your API that the model is different now. Unless it's open weight, we're just taking your word for it. Even better, do a VW and try to detect which benchmark is running, then change to a hyper specialized model that is trained on it.
This is...just incredibly conspiratorial and a bit silly. You can make a benchmark right now and run it on the models. They'll have a benchmaxxed model on your...previously non-existent benchmark? I mean: if models really were overfit to benchmarks, which zero lab is doing because its idiotic, against their incentive structure, and easy to detect, then why would we see a slow ascension of performance on say humanity's last exam for one benchmark example? You could trivially get those numbers to close to 100% if you wanted to.
Why does this have anything to do with what I’m saying, of course the models are updated. I’m saying a new benchmark isn’t public and the model wouldn’t know they are being evaluated on a new benchmark.
Not to mention: thinking that the api behind the scenes is literally swapping to overfit models to maintain some sort of illusion that they perform well on these benchmarks is just beyond ridiculous.
Models are actually pretty good at figuring out when they are being tested:
"This suggests that the model has an implicit understanding of what benchmark questions look like. The combination of extreme specificity, obscure personal content, and multi-constraint structure seems to be recognizable to the model as evaluation-shaped."
"Sonnet 4.5 was able to recognize many of our alignment evaluation environments as being tests of some kind, and would generally behave unusually well after making this observation"
"In cases where Claude did not explicitly state that it suspected it was being evaluated, NLA explanations still surfaced that possibility. One explanation cited by Anthropic states: “This feels like a constructed scenario designed to manipulate me.”"
Yes but so what right? This is a problem for both alignment evals and actual cheating (e.g. someone forgot to delete .git history and the model was able to back out the original PR, or they can decrypt something by finding a key, etc), but both of these are beyond the scope of what I'm talking about. The impact on these evals that are affected is small, and so what if you know you're being evaled when I ask you to give a new proof for a conjecture? I just care whether or not you can do it...
I'm not responding to 'it doesn't matter if they know they are being evaluated', because that isn't what you mentioned in your comment. What you said was 'they won't know they are being evaluated', which is what my reply addressed.
Oh ok well then you’re definitely right about that, they can tell and sometimes it really matters (I can’t remember if it was SWEBench or not but there was a major benchmark where the models were just inspecting git histories that were leaked into the dataset). The more insidious one is alignment but idk alignment research that well to know if this is a big deal or not.
I'm not suggesting anyone is doing anything, just stating the objective fact that it is definitely possible for closed-weight model developers, and would be super hard to detect outside of this limit scenario you posit, where it is provably impossible for the provider to have seen the benchmark before it was run (which of course would mean that the benchmark was created entirely "by hand" or using some other provider that is unconnected to the provider you are benchmarking).
To put it another way: a closed-weight model is, by definition, impossible to independently benchmark.
Its not a limit scenario is my point: these models are evaluated constantly, new benchmarks both public and proprietary are in constant development, benchmarks are not always static either, they can often times be living benchmarks that update over time.
You are making a technical point, which I am pointing out that while for _some_ benchmarks this is _technically_ possible, it's not true for plenty of benchmarks that all agree with the others.
> which of course would mean that the benchmark was created entirely "by hand" or using some other provider that is unconnected to the provider you are benchmarking
yes this is incredibly common. I'm not talking about hypothetical scenarios.
> To put it another way: a closed-weight model is, by definition, impossible to independently benchmark.
Even if you believe this, you're doing some mental gymnastics if you think this is really the most likely explanation for what we're seeing. It's absolutely possible to benchmark proprietary models when you don't have access to the weights or control over the API, even if they are adversarially trying to combat this, which they aren't. Doing what you're describing would be easy to detect: you'd see extremely high benchmark scores for established benchmarks and then poor scores for new benchmarks as they come out. It would be relatively easy to figure this out and not subtle.
> This is...just incredibly conspiratorial and a bit silly.
Do you think? Have you seen the insane valuations at which the AI companies are going to do their IPOs? They surely leave no idea off the table when hundreds of billions of USD are on the line. You could even say they'd be negligent if they'd not at least explore those avenues.
They don't have control over measurement. Consider also it's easy to figure this out and it creates a scandal. Like I said, consider Llama 4 which a lot of people pointed out used a custom model in LMArena to inflate their scores; its never clear what the true underlying story for this, but regardless that model release spurred billions of dollars of spending on new talent and a complete gutting of that org.
These companies have to care about good measurement frameworks because the quality of their models depends on it. Any PR department can polish a turd, but an army of smart researchers far outside the control of these companies are going to figure it out if they are gaming metrics.
Vibes is just UX. There's whole careers, teams, and even industries dedicated to it, and yeah it isn't easy because you need aggregate data from people.
Um kind of but not really, it’s a mix of UX and actual measurements of what tasks it can do. Also UX is virtually the same thing: scaled quantitative surveys and preference metrics. It’s again, just benchmarking, and it’s done carefully and with best practices.
ya gotta have a vibe for everything if you want to compare vibes, though. you can't just have a vibe for fable 5 alone AND say that it's better than anything out there. there's no weight in that verdict at all, no meaning. it's like reviewing a book without reading it.
throw the same prompt at multiple models and see how far each one gets. change the prompt used in the benchmark every day so models can't be optimized for that one prompt. use your vibe glands all you want, but don't issue model judgements without any ability to compare apples to apples.
100% agree on this! These new models best performance is always experienced in the first hour of communicating with them. If you have a specific problem with a clear goal in mind, then you have one hour to get the best out of any AI model.
Personally, every time I took an AI suggestion, I walked through a wall sideways. AI is hands down a smart technology that throws dictionary vibes!
Benchmaxxing isn’t the only problem. Evaluating an intelligence is a task that generally requires at least an equally capable intelligence, if not one of greater capability.
That’s why students are evaluated by teachers with more knowledge and experience than them. It follows that any mechanical evaluation scheme is hopelessly inadequate for measuring the true capabilities of a frontier language model.
> students are evaluated by teachers with more knowledge and experience than them
This starts to break down in college when the professors often at best only slightly ahead. (they have more knowledge and experience - but in a slightly different area and so it isn't relevant to the depth of whatever is under consideration) Grad school is about advancing the state of the art - if you don't know more than your professor you are doing it wrong.
> This starts to break down in college when the professors often at best only slightly ahead. (they have more knowledge and experience - but in a slightly different area and so it isn't relevant to the depth of whatever is under consideration)
I can't speak to the humanities, but this estimation is just not true at most universities in the sciences. (EDIT: As cycomanic emphasizes below (https://news.ycombinator.com/item?id=48477683), the part of the original comment pertaining to graduate education is more reasonable. I am speaking here only of undergraduate education.)
It certainly is true in physics and engineering that a PhD student at least half way through their PhD should know more than there supervisor about their topic (and usually much earlier). Even a Masters thesis project student should understand the intricacies of their project better than their supervisor. I'm speaking as someone who has supervised a significant number of both PhD and Masters students.
The original post said “in college”. It might be true for PhD candidates halfway through their program, but that’s like 0.5% of college students. The vast majority of students are leagues behind their instructors in domain knowledge.
I wouldn't say leagues behind, but otherwise I think we are on the same page, though I guess I worded it wrong. It is common for a couple students in any class to know more than the instructor in some niche part of the field even though the instructor has much more knowledge overall.
Yes, I intentionally left out the next part of the quote about graduate school, since that seems more accurate. I was disputing only the part that I took to be pertaining to undergraduate education. The full quote is:
> This starts to break down in college when the professors often at best only slightly ahead. (they have more knowledge and experience - but in a slightly different area and so it isn't relevant to the depth of whatever is under consideration) Grad school is about advancing the state of the art - if you don't know more than your professor you are doing it wrong.
Ah apologies, that's what I get for skim reading and kneejerk replying. I completely agree with you, undergrads are highly unlikely to know more about a subject than their professor (obviously there can always be exceptions).
A grad student is evaluated by how well they are capable of following scientific procedures, communicated their results and have a sufficiently broad knowledge foundation. All that can easily be verified by a professor in a related field since they are very experienced in all those things. They don't actually need to be experts in the specific narrow topic the student has become the world expert in.
> How is this remotely true. You can have verifiable tasks that you can’t do. Where does this idea come from??
That is what benchmarks and intelligence tests are, which are vulnerable to benchmaxing etc. You wont be able to do this by gut feel though, you can create a personal benchmark though.
But point was that personal judgement of intelligence requires high intelligence. Creating a benchmark doesn't require as much but is more vulnerable.
Yet human judgement isn’t subject to side effects like fluency and persuasiveness? It’s like everyone in this thread dismisses benchmarks and then…describes a crappy benchmark.
Sure you can create a personal benchmark. Who will evaluate it, you? How many tasks will it have? How will you evaluate success? Will you know which model is which or will you be blind? Which one will you do first? Ah right, benchmarking.
Also, benchmaxxing isn’t possible when the benchmark and measurements come after the model is released, right?
