Remix.run Logo
▲ aabdi 5 hours ago

Contrarian take. I think everyone should use LLMs and the most expensive possible, unless they know how to run evals (both on line and off line).

Yes its expensive, ugly, etc. But...

I've seen enough IRL that people just create really bad evals that do not reflect real world. Then the product is just bad. Outright broken for the p90 customer.

People don't know statistics. Offline Evals are hard. Sampling the entire population distribution is hard. Sampling the fat tail is harder. Online evals are big ops projects that are giant balls of death.

▲abhgh 4 hours ago | parent | next [-]

Maybe I should write up a post about this, but about evals ... I have feelings. Yes you're right that writing evals is hard. And it comes down to the simple fact that the thing that you're testing - a LLM-driven feature that presumably suffers from variance and inadequate accuracy - and your measuring device - an LLM judge - suffer from the fundamental problems. But the lure of natural language is such that instead of seeing this pipeline of LLM-driven things as a path of cascading errors, it is easy to view it as one set of problems correcting for another set.

In this world, the LLM-driven feature is a probabilistic entity, and the judge layered on top of that is another probabilistic entity. The real work is to look at the judge's response and infer the performance of the feature. Good evals make the inference step "pass-through", i.e., the judge outcomes have a 1:1 mapping with feature outcomes. But this kind of reliability often doesn't naturally appear. I think of reliability as being composed of three attributes: (a) consistency: do we get the same result for an eval run on the same llm with the same input (b) inter-llm agreement: do outcomes agree over multiple trials of running an eval over multiple llms with the same input (c) accuracy: does the eval agree with human judgement. You need extra work to guarantee these properties. And even when you can, you won't get 100% reliability, but say high 90s, and you still to make a judgement call as to call a feature "functional".

▲aabdi 3 hours ago | parent [-]

Two things, my problem isn't with the probabilistic nature, its:

1. with the quality of the data that customers use to eval.

2. with the lack of rigor/ops/testing around accuracy for deployment and regression.

I'll give you examples for failures IRL:

Offline:

- sample distribution for test data didn't have p99 customers in it, failed in production when it launched for those customers.

- test data sets were all synthetic, launch to real customers fails because data set was not representative of real.

Online:

- yeah we used LLM as a judge in prod/online eval, but the judge sampling failed to measured against our japanese customers, our bilingual customers, customers with media devices etc.

- the LLM as a judge was used, but there was no measure of human accept rate measures/alarms for various use cases (lightbulbs turning off, music playing, etc) so the lightbulbs started failing.

- there was no consistent human sampling of online data, so for example samsung TVs stopped working, but no one caught it until 3 months later when some VP was watching his TV and it broke and was accelerated despite it affecting 2-3% of customers.

▲mrbungie 3 hours ago | parent | prev | next [-]

If you are not getting feedback from those p90 then your bad evals are the least of your problems.

Also, unless those p90 customers are bringing a big portion of your income (again, if that happens you'd KNOW that your product is broken for them) I'd not pay for the most expensive LLM to "avoid doing evals".

▲aabdi 2 hours ago | parent [-]

>If you are not getting feedback from those p90 then your bad evals are the least of your problems.

the feedback is part of your online evals... if you have bad feedback you have bad evals

> Also, unless those p90 customers are bringing a big portion of your income (again, if that happens you'd KNOW that your product is broken for them) I'd not pay for the most expensive LLM to "avoid doing evals".

they are but that's besides the point.

Two ways:

- logically ROI of accuracy degradation as compared to cost has to be measured, sure.

- emotionally, morally sub 10% groups in our customer distribution include such things as disabled people, minorities, and the elderly. In my view you should love all your customers.

▲dieselgate 4 hours ago | parent | prev | next [-]

Can you elaborate on what you mean by Eval? All I can think of is eval() in js...

I hope you agree people should use the "right tool for the job" meaning a calculator for arithmetic rather than an LLM for everything.

▲aabdi 3 hours ago | parent | next [-]

Evals are tests of model accuracy. For example here are 100 people's sample houses, go test if the agent can turn off the lightbulbs correctly.

Right job sure. But, If you're replacing the LLM with a jev like as in the article, or switching with a smaller classifier, then you better have good data/evals otherwise you're reducing accuracy without you even knowing.

▲dominotw 3 hours ago | parent | prev [-]

its basically assertEquals(x,y) where function is calling llm to check equals

▲sick_of_slop 3 hours ago | parent | prev [-]

[dead]