Recently we've had all these people complaining of moving goalposts, when in fact most people never had goalposts for AI in the first place, and certainly never had goalposts that remotely resembled those that AI experts use to evaluate their own progress.
My mental model seems to clash with that of the boosters. They appear to have little doubt that reliability will emerge from scaling the models up. I think reliability is the lion's share of the problem and probably won't be solved by current architectures. It will be interesting to see who is right.
Have you actually engaged with it in any way? Its free to sign up.
By the way, what on earth is an AI booster? Do you really mean "enthusiast"?
"My mental model seems to clash with that of the boosters." Please tell us about your mental model, we can't read your mind.
A calculator gives you exact results for a very small subset of problems (equations).
Chat-GPT will give you approximate (and occasionally incorrect) solutions to an extremely wide range of problems.
The main problem with AI is that there are only so many problems where an approximate solution is fine, and an incorrect solution is easy to work around. In order for AI to be broadly useful, it has to get to the point where it's correct more often than an expert human. It may well get there within the next couple of years, but this isn't a space LLMs are designed to be able to solve so it's not as guaranteed as a lot of people seem to think.
We humans rely on thousands of things where an approximate solution is fine. Especially when it's about interacting with other humans.
Programming is an exception, but what % of the population are programmers and engineers?
The majority of jobs that require less precision falls under the umbrella of artists (I will include advertisers here), people managers, and service workers.
It actually sounds pretty familiar for those of us in the side shows for other Miracle Technologies That'll Fix Your Business, but it's a hell of a lot easier to set up a convincing model than something like SAP. Also, unlike ERP, GPT is a real thing, and it does real stuff, so that will make it easier to take the money. Real science is happening, which is always more fun, even if Homo Executivus uses it as a marital aid. I guess this makes me a Booster, but one of those old wrinkled Boosters who sounds like a carton of Newports.
A half-working system might actually be worse than useless if people believe it 100% of the time.
We're still far off from that, and I'm skeptical we'll get there anytime soon, but humans aren't infallible either.
Say... you have a broken arm, doc says it is sprained. You say No! and point out why.
AI says this, you say no, and the minumum wage person who works with the AI, shrugs and says the computer says so.
But a LLM assisting a wide rage of professionals at even a low accuracy could be part of an efficient workflow, allowing for be doctor or lawyer to give quality service to more people.
There is only so much a human can read and remember. There is only so much they can be convinced to change their mind in the face of new evidence. They get tired. They get cranky. They have biases, and fears.
Some doctors are just terrible. I think the AI is probably more reliable than them already.
I don't think we are very far away from being able to correct an AI, and for it to remember that correction forever.
It might look like a conversation where a panel of human experts and the AI agree on whats true and whats not. Like scientific consensus.
Eventually the AI will just run its own experiments and test its own hypothesisis.
Also consider that, in this case, the 50% the AI got right was defined such somewhat generously. For each patient, the AI presented several possible diagnoses, only one of which was correct. If the AI were making all the decisions, would the AI also decide to treat the correct ailment in all those cases? In an emergency situation, you may not have the opportunity to go back and try other options if you're wrong the first time.
Before ChatGPT there wasn't anything quite like it, so I think that the goalposts were mostly about passing the Turing test, or at least "not sucking" after 4 or 5 phrases. The leap from whatever existed before it was enormous, and I think comparable to the first iPhone versus, let's say, the Blackberry or Windows CE. Maybe even more than that. Go back to the first iPhone, it had a lot of issues but it offered so much as it was that people wouldn't look back.
It's too much to expect ChatGPT 3 to be better an experienced emergency doctor or neurosurgeon. But the fact that it can be a competent assistant to a doctor instead of a toy assistant - that by itself is amazing.
I mean, if that was ever the goalpost, ChatGPT does not clear it: the Turing test isn't about merely tricking someone who isn't paying attention, or making someone seriously doubt their belief system... it involved a careful discussion with a chat agent by someone who actively was trying to determine whether or not it/they were human. If you sit down with ChatGPT--even 4--and start discussing complex topics about the world with it, you absolutely can quickly get it starting to adamantly defend things about the world which a human would never assert... and, even more damning, if you just ask it "are you human?" it will reply "no", because we actively told it to fail the Turing test.
If we instead merely re-define Turing's test to be "it is able to converse with a human and manages to occasionally trick a human into thinking it is also human" then I had a friend (Roy Keene) who managed to pull that off years ago with a chatbot that was literally just "every time someone messages it, wait a small random period of time and then reply with a random response from the following set: Yes, No, Maybe, I dunno, :), :(, ;P (and like one other thing I forget)". He had it set up to reply when he was not available and when I finally caught him (I asked about a task I gave him the day before which he claimed to have no memory of, so he admitted to me he had this bot) he gave me a log file and OMG I had had long conversations about personal topics with this thing and had never noticed... but only because I wasn't trying.
> Turing proposed that a human evaluator would judge natural language conversations between a human and a machine designed to generate human-like responses. The evaluator would be aware that one of the two partners in conversation was a machine, and all participants would be separated from one another. The conversation would be limited to a text-only channel, such as a computer keyboard and screen, so the result would not depend on the machine's ability to render words as speech. If the evaluator could not reliably tell the machine from the human, the machine would be said to have passed the test. The test results would not depend on the machine's ability to give correct answers to questions, only on how closely its answers resembled those a human would give.
ChatGPT would not manage to pass an actual Turing test anymore than Roy's auto-responder; and, while we might get there frighteningly soon, that goalpost stands.
It's absolutely amazing that so many AI scientists don't seem to understand this distinction.
I'm super curious to learn how you use AI to do that. Thanks!
If we continue to improve AI, maybe one day its performance in these critical tasks you describe will also surpass humans.
Today: "Senior" devs by virtue of knowing the cookbook of, at most, one stack.
2028: "Senior" devs by knowing when GPT-12 output POST when it should have been PUT.
There is also the risk that the recent massive success of LLMs will bring a lot of excited bandwagon seekers into the field that don't have or appreciate the background on AI safety.
This is the central conceit of AI maximalists; and tech maximalists as a whole imo. Someone wants something that fits in a box that everyone else must eventually delegate to. Given capitalism's focus on creating owned systems, and the tendency to centralize and maintain asymmetry of access to rent extract, this is a most concerning direction to keep pushing development in.
We should be building better people; but we've abandoned that in favor of making a better mantrap.
People will use and fund AI tools that have usefulness to them.
AI boosters love to see the capabilities because it unlocks a whole new verdant valley in the continent of useful ideas, but until a street is laid down and a general store is available it remains the domain of early adopters.
LLMs are genuinely impressive. They aren't yet AGI, but they are fantastic tools. More exciting to me as a practitioner is the Segment Anything foundational Model (SAM) since I can put that into immediate use for computer vision in a variety of domains. LLMs still require some baking for broad usefulness (though there are many places they are already useful, its not quite broad enough).
When folks use AI focused on augmenting capabilities rather than replacing people, you get a great return. Until AGI is clear (and we'll know, the AI will tell us without being prompted I reckon), let's treat AI as an exoskeleton rather than a new species.
I was very disappointed to learn when I first started (around '93-94) that neural networks were hard to design, and impossible to train more than a couple layers deep, and the labelled data wasn't there anyway. None of that really changed visibly for a couple decades, but multiple unrelated research projects made the necessary breakthroughs: straightforward NN models (CNNs, etc) that people could adapt, frameworks to train and predict without having to implement the entire ML stack from scratch, absurdly fast computers and acccelerator hardware, and massive labelled data sets.
So far, nothing I've seen is anywhere close to what I consider "true" AI. As we struggle with hard medical problems. But seeing computers handle text with such facility has been exciting because it's the sort of thing that normies can see and appreciate (and even be fooled by).
However, medical problems aren't something that ML is going to magically solve. Because medicine has huge structural issues that are harder to solve than making diagnoses.
It would be nice if James gleick wrote a new version of "What just happened" covering the time between the invention of Transformers and now.
I'm also a Civil Engineering grad and it really cocked up when I asked a question about normal beam strength and I suspect it will get the pressure at depth thing wrong too. Both of these are related to depth only - one dimensional. That's why we have I beams - the flanges are to resist twisting and buckling but the strength is in the depth.
ChatGPT is a tool. It is basically a really good calculator. I do find myself saying please in my requests.
ChatGPT gives you approximate and sometimes-wrong answers to arbitrary questions in nearly any domain.
On the other hand with chatgpt, you have to build a specific prompt, and you don't know if that will lead to a consistent result or what its even pulling from. It introduces uncertainty that you cannot reliably estimate, whereas with the above toolmaking example, you can estimate your uncertainty, you can calculate how much power you have and how much you should have to resolve what you expect, and you can calculate a false positive rate.
When decisions mean money or life, I wouldn't rely on the black box chatbot. I would take a little bit more time and make sure all i's are dotted and all t's crossed.
(not the smartest expert human, the human you have access to)
What, exactly, is that?
This still doesn't really move the needle on any of the important questions about AI, but does hasten the public perception that proving self-awareness, consciousness, or agency exists in humans without depending on subjective experience is probably impossible.
Side note, my ER doctor wrote me a referral to neurologist to who was considered the best in the state and they refused to even talk to me since the referral said car accident. They wouldn't budge and wouldn't reconsider. The car accident was a result of an episode, not the cause of my issues. I hate the US healthcare system.
The thing is, if the healthcare system is shitty and broken, it doesn't necessarily matter how good ChatGPT is. If it's applied with the intent of "streamlining" the medical system by ignoring people's concerns and rejecting their insurance claims in an automated fashion, it might be much harder to even get in front of a doctor.
You could fiddle with your dosage, work with your doctor, etc, and then your insurance company can say, "sorry, ChatGPT has decided that you can do with a lower dosage, so we won't be paying for it. The machine is a black box, so it's not accountable to you or us or anyone else, so you just have to accept it's verdict."
No amount of training data, no architecture, heck no alignment or safety work solves that problem. Because technology doesn't solve social problems.
I think it is a combination of factors. The wrong answers are part of it, but I think another part is the anthropomorphization of it makes it almost like asking another human for help, which is somehow 'less good' than asking a search engine.
For better or worse, LLMs are here to stay.
With GPT, I don't have any of that (or maybe I just need to re-learn it?)
Also, I get a useful answer from most Google queries. GPT performs at a significantly lower bar (at least right now) - it works well for some stuff but not others, and the time it takes to figure out whether it's going to do a good job (and maybe do a couple of rounds of prompt refinement) is much more than just Googling.
for me its wrong answers and long winded english sentences with low information density.
Those existing CDS tools aren't widely used because the data entry is cumbersome and slows down clinician workflow. So LLMs may offer some opportunity there.
I commented a couple of days ago that we feel that the slope of change is relatively flat. As if the current snapshot of AI capabilities is somehow the way it’s going to be for awhile.
But in fact, the slope of technology change is almost vertical. A year or two from now, AI will be a very (perhaps radically) different capability unless something is done to interrupt its development.
It looks great on the surface but it isn't anything like the kind of sci-fi ai I think we're aiming for. Not in anything but a superficial way at least.
It doesnt know when its wrong and just straight up lies constantly inventing all sorts of fantastical stuff.
I asked it for history of a small local area in Ireland and it created a bunch of castles and tribes that never existed.
And then there's ChatGPT querying WolframAlpha:
https://writings.stephenwolfram.com/2023/03/chatgpt-gets-its...
The funny thing is not that we are holding AI to a high standard.
The funny thing is that we normally do not do the same for doctors, marketers, politicians, etc.
(And the AI here didn't even get it right 50% of the time, just 50% of the time it presented the right answer as one of several options. If pressed, it may not have made the one single correct answer for each of those patients.)
we've had this for years (bar the "any language")
they're traditional expert systems with the same flowchart a GP (family doctor) would use to ask questions
and they do a lot better than 50%
If you give me a sample of 100k ER text records and final diagnoses, I would be pretty confident I can estimate a model that gets you +50% accuracy on my laptop using open source tools.
Part of the difficulty with healthcare (that I myself don't have an easy answer to), is that you have very asymmetric error costs. So chest pain, even if heart attack is low probability, you should be risk adverse in many scenarios and still do due diligence to eliminate that as a possibility.
Not saying that smart systems are not possible, just still skeptical that level of complexity is possible on the backbones of LLMs. The smart sounding narratives I am concerned are mostly red herrings though in terms of outcomes we care about, accuracy/making right decisions/etc.
Yet in the end, the correct answers are likely identical to webmd or mayo, and the wrong to reddit or twitter, so all we have is a fancy google search.
So far, all the magic I see is the prompts. The knowledge is just random junk, the validation is just "how many internet pages said it", so you can never train new models again, because you now are training on gpt output.
The author of this article seems to me to be doing a good job of navigating between them.
This article and the comments in this thread are right up that alley. I mean, can you imagine say 1 or 2 years ago saying we'd have a readily accessible system that you could feed it the symptoms a patient is experiencing (in any language!) and out would split a well-described explanation of the diagnosis (in any language!) around half the time? And now it is suddenly here, and so many people are like, "yeah, but it isn't perfect" or "we can't completely trust it" or "we have to fact check it because it can be wrong".
It is all so weird and fascinating and just bigger than I can describe.
But here we are!