[chongli]

New models literally do get worse after launch, due to optimization. If you charted performance over time, it'd look like a sawtooth, with a regular performance drop during each optimization period.

That's the dirty secret with all of this stuff: "state of the art" models are unprofitable due to high cost of inference before optimization. After optimization they still perform okay, but way below SOTA. It's like a knife that's been sharpened until razor sharp, then dulled shortly after.

[girvo]

> If you charted performance over time, it'd look like a sawtooth

People have, though, and it doesn't show that. I think it's more people getting hit by the placebo effect, the novelty effect, followed by the models by-definition non-determinism leading people to say things like "the model got worse".

[gobdovan]

Is this insider info? The 'charted performance' caught my eye instantly. Couple things I find odd tho: why sawtooth? it would likely be square waves, as I'd imagine they roll down the cost-saving version quite fast per cohort. Also, aren't they unprofitable either way? Why would they do it for 'profitability'?

[bonoboTP]

It's rumors based on vibes. There are attempts to track and quantify this with repeated model evaluations multiple times per day, this but no sawtooth pattern has emerged as far as I know.

[chongli]

I don't want to go too far down the conspiracy rabbit hole, but the vendors know everyone's prompts so it would be trivial for them to track the trackers and spoof the results. We already know that they substitute different models as a cost-saving measure, so substituting models to fool the repeated evaluations would be trivial.

We also already know that they actively seek out viral examples of poor performance on certain prompts (e.g. counting Rs in strawberry) and then monkey-patch them out with targeted training. How can we be sure they're not trying to spoof researchers who are tracking model performance? Heck, they might as well just call it "regression testing."

If their whole gig is an "emperor's new clothes" bubble situation, then we can expect them to try to uphold the masquerade as long as possible.

[eudamoniac]

If a claim is unfalsifiable, it contains no information

[chongli]

What I said is quite far from unfalsifiable. Any number of insiders could step forward and set the record straight.

[eudamoniac]

Not really. Only in the affirmative. If an insider said they don't do that, you'd still think they might do that.

[chongli]

It's not insider info, it's common knowledge in the industry (Google model optimization). I think they are unprofitable either way, but unoptimized models burn runway a lot faster than optimized ones.

The reason it's not a square wave is because new optimization techniques are always in development, so you can't apply everything immediately after training the new model. I also think there's a marketing reason: if the performance of a brand new model declines rapidly after release then people are going to notice much more readily than with a gradual decline. The gradual decline is thus engineered by applying different optimizations gradually.

It also has the side benefit that the future next-gen model may be compared favourably with the current-gen optimized (degraded) model, setting up a rigged benchmark. If no one has access to the original pre-optimized current-gen model, no one can perform the "proper" comparison to be able to gauge the actual performance improvement.

Lastly, I would point out that vendors like OpenAI are already known to substitute previous-gen models if they determine your prompt is "simple." You should also count this as a (rather crude) optimization technique because it's going to degrade performance any time your prompt is falsely flagged as simple (false positive).