Are we sure these exams are not present in the training data? (ability to recall information is not impressive for a computer)
Still I'm terrible at many many tasks e.g., drawing from description and the models widen significantly types of problems that I can even try (where results can be verified easily, and no precision is required)
That's probably true, which is why human most knowledge workers aren't going away any time soon.
That said, I have better luck with a different approach: I use LLM's to learn things that I don't already understand well. This forces me to actively understand and validate the output, rather than consume it passively. With an LLM, I can easily ask questions, drill down, and try different ideas, like I'm working with a tutor. I find this to be much more effective than traditional learning techniques alone (e.g. textbooks, videos, blog posts, etc.).
Well written textbooks are consumable on their own for some people, but most are not written for that.
To be more precise can you please give a topic you know well and your % guess how often the answers are wrong on the topic?
Automatic proof generation is a massive open problem in all of computer science and not close to be solved. It’s true LLMs aren’t great at it and more is required for example as with the geometry system Deepmind progresses on.
On the other hand they can be very useful to explain concepts and allow interactive questioning to drill down and help build understanding of complex mathematical concepts, all during a morning commute via the voice interface.
Yesterday I was looking for some help on an issue with the unshare command; it repeatedly made bad assumptions about the nature of the error even I provided it with the full error message and one could already guess the initial cause by looking at that.
I guess such errors can be frighteningly common once you get outside of typical web development.
But a common situation is that with code generation it will fail to understand the context of where the code belongs and so it's a function that will compile but makes no sense.
Try getting it to properly select crystalloids with proper additives for a patient with a given history and lab results and watch in horror as it confidently gives instructions that would kill the patient.
What is even more irritating is that I had gpt4 debate me on things that it was completely wrong about and it was only when I responded with a stern rebuke that it hit me with the usual "Apologies for the misunderstanding..."
Medical problems are highly contextual, so you are not going to get much valuable information at the level of what a doctor is thinking from the first page of Google. That doesn't mean it isn't a simple within our area of expertise.
It is like a calculator that only worked in one digit, and now it works on 2, the improvement is immense but its still nowhere close to replacing mathematicians since it isn't even working on the same kind of problems.
Edit: In several years we might have a perfect calculator that is better than any human at such tasks, but it still doesn't beat humans at stuff unrelated to calculations. Or in the case of LLMs pattern matching texts, humans don't pattern match texts to plan or mentally simulate scenarios etc, that part isn't covered by LLMs. Human level planning with todays LLM level pattern matching on text would be really useful, we see a lot of humans work that way by using the LLM as a pattern matcher, but there is no progress on automating human level planning so far, LLMs aren't it.
GPT-3.5 was released in March 2022. We are now in June 2024. Over 2 years later.
And on average GPT-4 is about 40% more accurate.
For me, LLMs are very much like self-driving cars. On the journey towards perfect accuracy it gets progressively harder to make advancements.
And for it to replace the status quo it really does need to be perfect. And there is no evidence or research that this is possible.
Ppl dont want to hear that, but you see less and less offers and not only for junior positions.
Hard truth is that like with any tool/automation - the higher performance improves, the less ppl are needed for this kind of work.
Just look at how some parts of manual labor were made redundant.
Why ppl think it wont be the same with mental work is beyond me.
E.g. I had it autocompleting a set of 20 variable#s today Something like output.blah=tostring(input[blah]). The kind of work you give to a regex.
In the middle of the list, it decides to go output.blah=some long weitd piece of code, completely unexpected and syntactically invalid.
I am still in my AI evaluation phase, and sometimes I am impressed with what it does. But just as possible is an unexpected total failure. As long is it does that, I can't trust it.
Is it generally because the LLM was not trained on that data, therefore have no knowledge of it or because it can't reason well enough?
I've heard that claim many times, but never is there any specific follow-up on which topics they mean. Of course, there are areas like math and programming where LLMs might not perform as well as a senior programmer or mathematician, sometimes producing programs that do not compile or incorrect calculations/ideas. However, this isn't exactly "garbage" as some suggest. At worst, it's more like a freshman-level answer, and at best, it can be a perfectly valid and correct response.
That is garbage.
> I hope you don't hold a teaching position at a university then.
You think teachers shouldn't have growth mindset for students? I think students can grow from producing garbage answers to good answers, that is what they are there for. An LLM however doesn't grow, so while such students are worth teaching even though they produce garbage answers the LLM isn't.
I think many students including freshman have interesting and sometimes thought provoking ideas. And they come up with creative solutions, which is based on their previous experience in life. I would never describe that as garbage.
But they aren't meaningful for anything other than humans since the correlations between abilities which make them reasonable proxies are not the same.
The idea that these kind of test results prove anything (other than the utility of the tested LLM for humans cheating on the exam) is only valid if you assume not only that the LLM is actually an AGI, but that it's an AGI that is indistinguishable, psychometrically, from a human.
(Which makes a nice circular argument, since these test results are often cited to prove that the LLMs are, or are approaching, AGI.)
I've noticed one thing that LLMs seem to have trouble with is going "off task".
There are often very structured evaluation scenarios, with a structured set of items and possible responses (even if defined in a an abstract sense). Performance in those settings is often ok to excellent, but when the test scenario changes, the LLM seems to not be able to recognize it, or fails miserably.
The Obama pictures were a good example of that. Humans could recognize what was going on when the task frame changed, but the AI started to fail miserably.
Me and my friends, similarly, often trick LLMs in interactive tasks by starting to go "off script", where the "script" is some assumption that we're acting in good faith with regard to the task. My guess is humans would have a "WTF?" response, or start to recognize what was happening, but a LLM does not.
In the human realm there's an extra-test world, like you're saying, but for the LLM there's always a test world, and nothing more.
If I'm being honest with myself my guess is a lot of these gaps will be filled over the next decade or so, but there will always be some model boundaries, defined not by the data using to estimate the model, but by the framework the model exists within.
The problem isn’t the LLMs per se, it’s what we want to do with them. And, being human, it becomes difficult to separate the two.
Also, they seem to attract people who get real aggressive about defending them and seem to attach part of their identity onto them, which is weird.
> data from a recent July administration of the same exam suggests GPT-4’s overall UBE percentile was below the 69th percentile, and 48th percentile on essays. Third, examining official NCBE data and using several conservative statistical assumptions, GPT-4’s performance against first-time test takers is estimated to be 62nd percentile, including 42nd percentile on essays. Fourth, when examining only those who passed the exam (i.e. licensed or license-pending attorneys), GPT-4’s performance is estimated to drop to 48th percentile overall, and 15th percentile on essays.
At the end of the day (a) LLMs aren't accurate enough for many use cases and (b) there is far more to knowledge worker's jobs than simply generating text.
That's called a programming language. It's nothing new.
Edit: LLMs biggest feat is being a natural language interpreter, so it can run natural language scripts. It is far from perfect at it, but that is still programming.
Add some obnoxious pseudo-intellectual windbags building a cult around it and people would be down right turned off.
Hype is also taken as a strong contrarian indicator by most scientific and engineering types. A lot of hype means it’s snake oil. This heuristic is actually correct more often than it’s not, but it is occasionally wrong.
The fact I can talk to the computer and it responds to me idiomatically and understands my semantic intent well enough to be nearly indistinguishable from a human being is breath taking. Anyone who views it as anything less in 2024 and asserts with a straight face they wouldn’t have said the same thing in 2020 is lying.
I do however find the paper really useful in contextualizing the scoring with a much finer grain. Personally I didn’t take the 96 percentile score to be anything other than “among the mass who take the test,” and have enough experience with professional licensing exams to know a huge percentage of test takers fail and are repeat test takers. Placing the goal posts quantitatively for the next levels of achievement is a useful exercise. But the profusion of jaded nerds makes me sad.