Why eval startups fail (2025)(thomasliao.com) |
Why eval startups fail (2025)(thomasliao.com) |
The market's being split into
1. Longitudinal LLM observability tooling
Most eval startups have gone down the route of something more like being an observability platform for LLM inference. They want to be in your stack and running the inference to collect data on performance of it.
They collect things like how often a model returns JSON that's out of spec or returns values that aren't expected as well as general timing and cost info.
2. Safety Limiting / Pentesting
Say you're doing something in the medical field or that's sensitive in some way and you want to figure out what model has the best outputs for your task that won't fly off the guardrails.
3. Simple cost + performance + quality swapping
This is what my tool does, basically lets you test if you _really_ need to be running that frontier model in a loop across a million records or if you'd be better with an older model or something else.
Example eval: https://giyd8stidy.evvl.io
It's probably not gonna be exactly glorious work, but designing expert evals settings and collecting and crunching the data for quality assurance and control is going to be needed.
AI is not deterministic like regular code, so imagine you use it for "search" (RAG) or for summarizing or for classifying emails etc. How do you know it is giving you the right results? In this context, AI evals are an important idea and very often neglected.
You can use an initial "dataset" to evaluate your prompt and AI calls + code (think test cases), this dataset will of-course be curated by humans. But as the software is used, you want to incorporate, real production data as well and run the evaluation pre and post launch. Sounds simple, but can get complicated specially since this area is new and as the post mentioned there are too many players and options out there (since everyone thought this is a money maker).
I do agree that the author does not do a good job of introducing the term.
What kind of stupid business is this. Though nothing can beat SEO in that spirit.
For example, a simple eval is a dataset of multiple-choice questions, which each have one correct answer, and scored by accuracy. An example of this kind of eval is the Massive Multitask Language Understanding benchmark (2020) (https://arxiv.org/abs/2009.03300).
A more complex eval is FrontierCode (2026). Questions in FrontierCode represent coding tasks needed for real-world repos and are evaluated against rubrics scoring for correctness, code quality, cleanliness, and other factors. https://cognition.com/blog/frontier-code.
¹Note that this is a slightly different definition we used in [0], which used a definition of a fixed input-output correspondence pairs combined with a metric. What's different from 2021 is: models are now given more open-ended inputs (prompts like "find the bug" and a codebase rather than a simple question), have freeform generation (rather than choosing a fixed answer), and are graded in a more complex manner (e.g. beyond correctness, one might care for a coding eval also to grade adherence to coding guidelines, test coverage, etc).
[0] Liao, T., Taori, R., Raji, I. D., & Schmidt, L. (2021, January). Are we learning yet? a meta review of evaluation failures across machine learning. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2). https://thomasliao.com/are_we_learning_yet.pdf
Therefore your knowledge is better used in training than letting users be slightly better at the token casino. Which is mentioned in this post as well, eval startup people either go to work at frontier labs or finetune startups.
Aha.
Personally, I agree with the Goodhart problem, but isn't the reason Eval startups fail because they try to sell an 'evaluation service' rather than a 'verification toolchain'? The problem, it seems, is that AI verification toolchains require a model in the end, because they internalize AI and sell it under the name of a 'harness.'
So an AI verification(eval) toolchain would have to be structurally different. Verifying AI code isn't about whether it compiles. AI code can always be made to compile. The issue involves various semantic criticisms, such as overfitting to existing designs and tests. To catch those issues, you ultimately need to build an AI. But building that AI is expensive. So in the end, AI verification companies depend on external model providers for the core components of their verification engine. I think this is a bad business decision
> built the tools to verify, deploy, and build them, such as CI/CD, static analysis tools, and testing frameworks.
Curious. Which company made money with testing frameworks?
A key point I want to make is that the notion of "frontier" is somewhat fictive in the sense that a model which dominates all others on a given eval is not guaranteed to be number one on another eval, even if both evals are ostensibly for the same task.
For example, the best publicly-available model (i.e. excluding Claude Mythos and Fable) on DeepSWE[0] is gpt-5.5-xhigh at 67%, which is soundly better than claude-opus-4.8-max at 59%. I would say an 8pp gap on a benchmark is quite large. But on FrontierCode[1], claude-opus-4.8-xhigh is the best, at a score of 13.4% compared to gpt-5.5-medium at 6.3%.
That's quite a significant reversal!
Now, one might wish to claim that either DeepSWE or FrontierCode is poorly constructed and that the other is more accurate. But I think you'll find that the degree to which eval-design considerations in this case affect measurement is probably of no less magnitude than user-specific considerations affect measurement in general.
[0] https://deepswe.datacurve.ai/ [1] https://cognition.com/blog/frontier-code