A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis
lynchlab.net505 points by sungam 13 hours ago
505 points by sungam 13 hours ago
This is an example of why I'm a little stumped about why people are excited about Jev-style decision models. For any application at scale it'd be way more accurate to have a dedicated classifier using the right architecture (neural net, bayesian, etc.) LLMs drive the cost of making those models down to near zero.
Just use the LLM to make a thing that's actually good at what you need it to do.
Jev's launch was the best AI era marketing campaign I've seen so far. It was executed and timed so perfectly that Jev became synonymous with decision models overnight.
I think it reveals how many people were only familiar with LLMs as the generic AI tool for everything. It only took one ambitious marketing campaign for someone to come along and present a simpler technology as a new breakthrough to audiences that hadn't considered anything else.
The largest LLMs are really amazing for how much they can accomplish on unseen tasks. That is undeniably remarkable. The mistake that gets made everywhere is thinking that they're the best tool for every task, because LLMs are the only AI tool in most developer's AI toolbelt right now.
>many people were only familiar with LLMs as the generic AI tool for everything
That’s exactly how the biggest vendors have been marketing them, it’s no wonder non-experts think of them that way. I’m not an expert myself but have a passing familiarity with theory as a SWE and grad student who did some research for machine translation and currently am in a computing-adjacent field. I use LLMs heavily for work and have never come across any coder-oriented tutorial, guide, etc. encouraging use of LLMs as decision models specifically, even if they are being used that way “beneath the hood.” Jev’s announcement seemed sort of new functionality to me but even in the first post about it on HN (where I heard about it), iirc people were pointing out that LLMs do the same thing, at least to some extent.
I think you have a very narrow and specific experience with people claiming to be "devs".
I don't know of any devs in real life who would loudly and publicly get excited about anything unless it was their own completed project to show off.
The tools are a footnote. It's precisely because it's so hard to get a project that far along that they would be familiar with many more tools than they would ever speak about. I suppose some people have a low bar. I wouldn't really call them "devs", especially now in this "AI era".
> I don't know of any devs in real life
I think you're coming at this from a very narrow and specific experience with a few people you know.
Jev really went from the announcement to being discussed all over overnight. We're discussing how popular things are in general, not among your narrow and specific group of people you work with.
I agree with what you said above, but not this. Jev was discussed "all over" by people chasing hype. I think that is actually a narrow set of people.
"No you!" :-)
I think it's clear we're at the tail end of this cycle. I'm very glad to see it.
There are a lot of people working in software that have limited knowledge of information theory or machine learning.
Unfortunately LLMs are magic to them.
When you lack a first principles understanding of how something really works, you will bolt on hacks and duct tape on the wrong approach.
it's also the unfortunate practice of trying to use the new shiny thing everywhere and anywhere to signal that you're keeping with the times. It's funny how i've seen newer devs brush off things as 'traditional ml'
Which is nuts because one thing LLMs are actually pretty good up is generating "traditional ml" harnesses. You just have to describe the problem you are trying to solve and ask the LLM for algorithm suggestions.
It's also nuts because "traditional ml" generally requires a fraction of the resources of an LLM. If you need to auto label, tag, categorize, it's never been easier to setup that sort of system in any project.
It seems to me that hype cycle of jev made it harder to understand rather than easier.
Universal classifier just seems nonsensical. It seems clear that it can't have accurate probability distributions on any possible question you might ask it. Moreover, giving only "maybe" answers to everything adds a layer of equivocation that makes it harder to catch it in a clear mistake.
That said, I've encountered plenty of fuzzy judgment questions now (ie is this case more like example X or example Y?) that reasoning models struggle with. I could see how, in lieu of a dedicated classifier, a model like jev, ostensibly trained to guesstimate the relative likelihood of natural language hypotheses, may do well.
But it still feels wrong to call it a classifier in the traditional sense, let alone a universal classifier.
I think there are a lot of practical applications where something that works few-shot and doesn't require training is quite appealing if the cost of inference is low. That's the sweet spot for Jev and similar models. If you have time and an eval set to train it, clearly a "real" special-purpose classifier should have much better performance (and for most applications cheaper cost/faster inference), but Jev and friends are way better than using a full-fat transformer LLM for a classification task.
As an example, I've been using winnow running locally as a decision model in openchamber to select between llms for doing various tasks. I don't have an eval set of different coding and other tasks that I could use to train a proper classifier. This I just download off huggingface and it works out of the box and has cut my token spend to a fraction of what it was. So it's good enough to be beneficial.
> Just use the LLM to make a thing that's actually good at what you need it to do.
This is the end state of most imagined use cases for “agentic” software IMO. Use the AI to build software, with a classifier or LLM as the expansion joint in awkward spots.
The frustrating part about an LLM is at a fundamental level it's designed to do the wrong thing, it's designed to provide an answer for things it can't do, e.g. one argument I've been in recently at work is someone using it to summarise numerical data. They've a raw data spreadsheet they're putting into ChatGPT and asking it to produce a report in a particular style. So the LLM is doing its job, predicting the tokens to emit, and not actually doing mathematical analysis of the underlying data.
I had a team proudly show me how they were feeding their dashboard data into an LLM to predict growth needs, instead of doing basic linear regression or any number of pre-existing basic statistical analysis. The part that depressed me the most on that was it was a bunch of engineers that did it, the people you'd most hope would understand what an LLM is/does.
For numerical work, there’s no reason to assume the computation is happening in the LLM. All frontier models could handily write a Python script to do linear regression then report the result, so the quality of the responses here depends entirely on the prompt + harness.
+1. I noticed ChatGPT now pulls out a calculator or writes a script instead of hallucinating numbers. Initially it was terrible, but recently it seems pretty good... that said, if you're doing serious work with this you should probably double check. And if you're doing repeated work, obviously get it to write a script or formula instead of doing 1-offs.
It makes sense to use an LLM to write the queries when exploring data, but if you are building a dashboard that’s going to be used repeatedly, you should save the queries so they can be validated and improved over time. The LLM shouldn’t be rewriting a new query every time you open the dashboard.
Definitely. I’ve found Cursor + Grok to be the best budget models for this type of tool use, although I’ve yet to try out the latest Anthropic ones.
You can go “here’s a spreadsheet of some data, do a thing to it” and it’ll go through, have it plan out the type of calculations you want to do, and it’ll write you a program to parse your sheet, calculate what you want, then output an result sheet.
The key to good results with minimal effort seems to be the planning modes to me, combined with the /goal command. A little bit of validation before implementation goes a long way.
They’re very much trained to do this, in fact. Even Step 3.5 Flash, a medium sized open model from way earlier this year is trained to use python to calculate answers on the fly :)
I've seen so many "truthy" results coming back that suggest it's not even remotely as smart as you might hope.
I'm mostly starting to have some success telling TPMs and engineers to ask chatgpt to give them appropriate code that generates more accurate answers.
These days I wouldn't be surprised if the models are smart enough to get the analysis correct.
However sometimes I think that AI use shows a revealed preference around things. In this case, the actual data just ... maybe never mattered to them at all? People wanted a shiny chart and they got a shiny chart.
Either chatgpt or claude are building scripts to do those and has been for a few months at least now.
This is exactly what we’ve been doing.
For example, a client want a classifier to decide to make a go/no go decision on roofing jobs based on weather. The end result doesn’t use LLMs at all, but does use a lot of other techniques including human feedback when it makes the wrong call.
(Now the work crew gets the decision at 5 AM instead of 5:30 and the guy making the decision doesn’t have to wake up at all.)
(Presumably you mean the guy affected by the decision?)
I think it means the person making the decision does not have to get up early anymore in order to make a decision. The people receiving the decision still receive the decision, and a bit earlier.
A complaint from the crews was not getting the go/no-go at the same time consistently. This fixes that problem.
I do agree with you, but I do think you're also making the classic "curse of knowledge" mistake.
You have to know the right words to ask the LLM to make and train a classifier. You and I already know the words. But they are not actually that trivial to learn for an average joe, and people are lazy.
The tooling around LLMs is still nascent and, in my opinion, quite poor. You may be able to shotgun easily-falsifiable mathematical proofs with raw compute, but get into the messy world of real data, uncertainty quantification, and a host of other things that go into building real world models, and I think that an LLM plus tooling dedicated to the task plus a human in the loop will open up capabilities not available to an LLM in a chat or development centric software environment.
I don't believe this is true with current frontier coding agents. Or at least, it's becoming increasingly less true, such that I think it will effectively be nullified in ~6 months.
I'm gonna cite myself here:
«We came full circle on the problem that also exists with Google Search, and why so many people don't get good results out of it.» [0] It is not that access to all information was hidden to anyone before, it was harder to find it, sure. Now it can be found easily and aggregated, however, it is not as simple as saying "Do X" and believe that X is correct. Recall the infamous Swing & Product Management meme? [1] The problem is the absence of information you provide, the LLM will have to make a guess of what you mean, much like the human. Likelihoods of this guess being correct depends on the domain and background. So yes, it'll likely build whatever you ask it (I do think we're there already), but if that matches what you intended to actually get is vastly different. You'll be settling with "Good enough" though and declare it working.
Others already pointed out that 6 months seem to be the constant of distance, irrespective of the date.
[0] - https://news.ycombinator.com/item?id=50063427#50063744
[1] - https://www.productftw.com/productftw-2-the-tire-swing-carto...
I'm done waiting 6 months. Just let me know the night before when I don't have to come in to work tomorrow.
Its a bummer because that is also exactly when self-driving cars are going to work. You'll be able to sleep through the commute but won't have to go.
Self driving cars have been working for years... And waymo launched this year in Denver.
https://waymo.com/blog/2025/10/creating-an-all-weather-drive...
LLMs are the perfect tool for "finding the right words". They are literally translation machines.
Just ask the LLM for these words, obviously !
It's funny you mention this, because I've been thinking lately that all the focus on code quality benchmarks for LLMs seems to miss something. In my experience, LLMs aren't actually all that good at prompting other LLMs, and that's exactly the skill that matters for proper subagent usage.
The corpus of code available to train LLMs on is rich in both volume and quality. The same can't be said for LLM prompts. Prompts only came into existence alongside LLMs themselves, so most of the prompts out there are actually LLM generated, and training on them causes every model to overfit to the same patterns.
LLMs can give clear and concise directives to other LLMs, but only for basic tasks. Beyond that, they frequently overstuff their prompts with too much prose and too much direction, and worse, they often fill them with subtle contradictions.
Of course, there's a spectrum here. Some models turn out to be better at this than others, and I suspect this is what differentiates Fable-type models from the rest, and why vendors have started keeping their models' reasoning confidential.
Quick tip here: Where possible try to trick the model into thinking whatever task it is writing will / could be read and performed by one of your coworkers rather than another agent. It will quickly fall into line and stop writing prompts meant for dumb historic low intelligence LLMs.
The "agent factories" available to employees at the labs and hyperscalers clearly prove that agent delegation does work. Today, maybe that capacity lies with Fable class models, as you pointed out, but there is no doubt with opus 5.5 that this capability will be available to cheaper models as they get better.
That's what I've been doing in areas I lack knowledge tbh. Instead of a half assed attempt at describing what I need, I ask an llm to take my base idea and write it up using industry standard jargon as if I were commissioning a team to do the work. Has been working pretty well so far!
Okay so I just asked it "What question should I be asking you but I don't because I'm not smart enough to think of it?" Response was kind of interesting I have to say.
It gave me a very helpful answer, my project will be improving just because I asked.
I did this just now, ironically before I even saw this discussion. I ended up questioning the AI about humanism.
So what? The average joe shouldn't be trying to diagnose people with cancer.
Healthcare is more than having medical knowledge.
Ten years ago, in France, two doctors (one in a professional setting and the other during a private consultation) asked me about a wound I had on my shoulder; obviously I couldn't identify what it was, and they offered no suggestions, so I supposed it was not something important.
It took a visit to a cardiologist—for an entirely different reason—for him to suggest that I rapidly consult a dermatologist.
It was a carcinoma, and surgery removed it. If ChatGPT had existed back then, this carcinoma would have been removed several years earlier.
This sounds like a competency problem on the part of the GPs. They’re supposed to make these calls.
Maybe if chatgpt existed back then you'd have had your carcinoma extracted along with 10 other benign anomalies.
Automation doesn't make up for carelessness.
Skin cancer is easily treated with excisions if it’s caught early. Better to have ten small false positives than a single miss.
Even if addressing each false positive would carrying low but real risk of death or complications?
Is your "low but real" risk comparable to the risk of driving a motorcycle, of riding in a car, or of using the bathroom?
Reasonable treatments for small, non-aggressive skin cancers (or suspected cancers) are extremely low risk. And reasonable practitioners will often suggest that you merely keep an eye on an abnormal spot - because it doesn't look worth the bother of any sort of treatment for now.
Not sure why this is getting downvoted. In its specific case that's quite obviously not a good example because removing benign spots isn't harmful, but overlooking malignant ones is, so there is no downside.
But if we think a little further and generalize then there is great risk (and by extension, harm) to delegate diagnosis and treatment (of health issues in general, including potentially or supposedly fatal ones) to ChatGPT and similar tech. That we'd end up with a net positive is far from obvious.
Obviously, error rate of ChatGPT should be measured by comparing to human diagnoses and when it plateaus, we should investigate a possibility that the number is driven mostly by human errors.
I still maintain my point that it isn't harmless to blindly address benign anomalies. Every medical procedure carries risk.
Your point doesn’t really make a lot of sense. Even with ChatGPT existing and making 90% bad calls, a doctor isn’t likely to just “blindly” be cutting out anomalies.