Hacker News Reader

Comment by Aurornis | Hacker News Reader

I doubt that would change the perception. Every model release is followed by accusations of nerfing.

There are several projects that repeat benchmarks on published models. None has ever found significant fluctations

Here's one example https://marginlab.ai/trackers/claude-code/

Fluctuations of a few percentage points are to be expected and should not surprise anyone who knows how LLMs work.

This Twitter analysis of Fable 5 is not that at all. They analyzed their coding sessions and blamed all of the fluctuations on Fable changing. They then compared to ARC-AGI-2 questions as the benchmark for thinking tokens and tried to stir up anger that coding turns don't produce as many thinking tokens as the ARC-AGI-2 problems.

View on HN

This page has been in 'New model — collecting baseline data. Degradation detection paused.' state for months now. It seems to never say 'degraded'.

If you look at the graphs, the latest benchmarks are showing a pretty significant dip, and they match pretty well with some horrible experiences I've had in recent weeks. You can see token usage steadily going down, matching exactly what the author measured on his own.

> This page has been in 'New model — collecting baseline data. Degradation detection paused.' state for months now. It seems to never say 'degraded'.

Click the part at the end that says "View historical performance". They wait to collect more data about a new model before adding it to the overall charts.

The overall solution rate continues to climb when new models are considered.

> If you look at the graphs, the latest benchmarks are showing a pretty significant dip, and they match pretty well with some horrible experiences I've had in recent weeks. You can see token usage steadily going down, matching exactly what the author measured on his own.

The y-axis is amplified to make differences look larger than they are.

Hover over the dots to see the confidence interval. A 1-2% change means nothing.

The page is currently tracking Opus 5, which was released July 24 and has not seen any updates since.

30-day average 83% [71-91], last result is 79% [66-88], and it dipped to 75% [61-85] a week ago.

The page says

> We always use the latest available Claude Code release and the SOTA model

They've stepped up the benchmark for each new model. Presumably they're gathering Fable data now.

They also link to Anthropic's public blog post about some degradations, their cause, and how they fixed them. The time period sounds like the "recent weeks" you experienced: https://www.anthropic.com/engineering/a-postmortem-of-three-...

Fable 5.1 was released 20 days ago. That's a lot of time to gather data. Since they're still tracking Opus 5 anyway, there is no obvious reason the perf delta section would be disabled. Are you involved with the project?

The Anthropic post points to the latest fix on Sept 12, and the issues mentioned also only affected Sonnet 4 and Haiku. Opus was misbehaving just last week. You are choosing to not see the evidence of degration, 85% -> 75% is a generational dip in intelligence.

Not involved, no. Just guessing. It doesn't look frequently updated.

A lot of these daily-benchmarking sites popped up earlier this year. Most of them have faded away after they all failed to produce the smoking gun that everyone expected. This site survived because it kind of caught a dip one day, maybe.

The results are really rather flat and daily benchmark runs are expensive, so most of these projects give up after a while.

It would change my perception but only if there were a competent and stringent administration in place. I didn't used to have to wonder if the ground beef I was buying was actually 1lb because there were inspections and repercussions, but stuff is kind of chronically underweight these days.

A properly run OWM enables you to stop wondering if you're being ripped off and that's what AI needs because I think it's incredibly easy to just assume we're being ripped off because these companies are all built on a foundation of wonton theft. (Not that I really care about that — I think all information should be free, but still.)

If you go to https://marginlab.ai/trackers/claude-code-historical-perform... there is a very clear downwards trend in the two weeks before Opus 4.7 release. Then a sudden and dramatic drop seven days before Opus 4.8. And now we seem to have entered another decline in the last ten days, beyond the usual noise of Opus 5 scores

Reply on news.ycombinator.com