Hacker News Reader

Comment by sixtyj | Hacker News Reader

I have tried to rewrite an article with GLM-5.2 and with Sonnet 4.6. Completely different results as LLM is non-deterministic. But GLM-5.2 made a lot of subtle mistakes that needed to be corrected by hand. On the opposite, Sonnet found and corrected all mistakes in the second round.

Similar situation was with planning and coding. GLM-5.2 seems to be good “on paper” but the real usage results was different.

And I am not an attorney for Claude or GLM-5.2… :)

But as I’ve been using LLM models daily since Nov 2022 I have realized that all common tests have to be confirmed in your project - there is no “one model rules them all” - you need to dig out a specific model from that LLM haystack with thousands of models.

Benchmarks help but they start to be similar to fuel consumption specs in car ads - real consumption is different for everybody :)

View on HN

> Completely different results as LLM is non-deterministic.

You'd need to produce this like 20 times by each model and then do 2x20x20 cross comparisons by both models and ultimately distill the 2x20x20 comparison results into two reports of how they differ.

In this non deterministic computing future, everything else is voodoo, feelings and "vibes".

I would expect a model's result each time to be of a similar quality to the other times. There's something wrong if it does a way better or worse job, at the same problem, sometimes. It's possible, but I haven't heard anyone saying that they do.

No, that's not how LLMs work. It's all probabilities, and that issue has only deepened with providers silently falling back to worse models if they suspect you might be distilling their models. If an LLM rolls a bad token that can tip the whole balance of the response into utter nonsense.

People use LLMs to do vulnerability scanning by throwing them repeatedly at a codebase. Depending on the run they return with nothing, with a false positive, with a true vulnerability. These are very different destinations when faced with the same problem, sometimes.

Since GPT2, people have been throwing a ton of crap at the wall just to pick out one nugget that's uncharacteristically more solid than the others. Honestly? It's not just possible—it's core to how they operate. And it always has been.

probably depends a lot on the temperature setting, lower it is, the more similar the quality perhaps. the higher it is the more variance.

I’m just dipping my feet in the water of local models and I really feel this. I had a simple alignment task (align known quality transcript without timestamps with timestamped but lower quality whisper transcriptions) and I went through 12 rounds of testing across 4 generations of 3 models. The results were all over the map even across versions of the same model. Spin out the task to something as big as coding and wow.

If you have any advice/blogs on doing project specific benchmarks I’d love to hear it. I’m trying, but it’s haphazard at the moment

We have found similar when plugging GLM 5.2 into actual benchmarks in our product. The open-source models are really dialled into the public benchmarks, until you try them in context you won't have a solid idea of how they perform (Sonnet is a higher quality model than 5.2, both in prose, reasoning, and alignment).

Most benchmarks notoriously undercount multi-turn instruction following too. And it's where open source models (and Gemini) lose big time.

Probably because benchmarks are leaked. Included in their training or model can cheat by finding answer online.

Reply on news.ycombinator.com