Hacker News Reader

Comment by otabdeveloper4 | Hacker News Reader

Somewhat of a hot take. Normal classification algorithms output a confidence weight, LLM's don't. This means you can only use LLM classification if you don't care at all about managing your precision and recall metrics.

This means LLMs are useless for 90 percent of real business requirements. (Business actually wants some sort of binomial regressor, not a plausible-looking hallucination. Plausible-looking hallucinations are the domain of spam and copywriting.)

View on HN

The reality is LLMs aren’t competing against normal classification algorithms, as the tasks have no algorithm that approaches the observed precision and recall of LLMs. They’re competing against humans, who also have no metrics other than observed precision and recall. When you observe their performance vs humans at these sorts of tasks they’re generally superior, especially as the quality of the human work is very low - poorly trained, poorly paid, and poorly incentivized people doing very boring work do very badly. LLMs do very well.

The savings can be substantial even tho the people are poorly paid and trained, because you need so many of them, and need a complex organization built around them to source, train, manage, and observe all the humans - of which there can be many thousands with high turnover. Their function is actually pretty important and the high failure rates cause material impact to the business in terms of loss and risk. Improving these dimensions substantially upgrades the business over all.

The truth though is we are really early in the development cycle. These sorts of changes will play out over a decade. In that time I expect the technology to improve in many dimensions - power, capability, ease of integration, ease of observing, ease of fine tuning/alignment to task, applicability to other functions, etc.

You can fine tune an LLM to output a particular token for category X. This way, you get 1) fast inference (just need to generate a single token) and 2) confidence metrics (just use the logprobs for the token that was chosen). Gets me to 93% accuracy with a 7B model, trained with only 15 examples per category (around 10-15 categories in total, I think).

Auto Label can do it even with closed-source LLMs: https://www.refuel.ai/blog-posts/labeling-with-confidence

Also, your 90% number is a 'plausible-looking hallucination', which is mildly funny.

>Explicitly prompting the LLM to output a confidence score is the least accurate

They don't show the whole prompt they use for this so I am sceptical of the comparison that recommends their service which does the token level accuracy. In my short testing you can drill down some acceptable accuracy for specific domains if you lead it over some general "ask yourself" questions then "finally, output a % confidence"

I couldn't care less about the specific implementation.

My parent comment said "LLMs are useless for classification, because they can't give confidence scores". I think i have shown that to be demonstrably false.

My understanding is these are more “confidence in the tokens selected being appropriate” rather than the more fundamental “confidence the classification chosen is accurate given training data,” where training data is samples of the distribution being classified rather than language corpus. They seem similar at first blush but the first is more about appropriateness of phrasing and how esoteric the semantics are, and the latter more descriptive of some statistically observable process - which is something LLMs are literally incapable of reasoning about directly.

Reply on news.ycombinator.com