Which sucks, but can you imagine the additional burden our health care system would be under if they had to test for every possible thing a condition could be every time you went in? I wonder what all the edge cases of something presenting like a headache could be.
Zebras, deer, moose, and giraffe, bulls of many types, all exist.
I'd say if you hear hoofbeats, contemplate what hooved animal would be appropriate to the situation and try to eliminate the easy ones. Lord knows, you might find a Zebra escaped from the zoo, or a bull from someone's field from time to time.
In short, life is messy, and taking any aphorism to the death will lead to just that.
I had an opportunity today that made it useful. I was trying to find a reliable method of counting the decimal places in a double (64 bit float). I couldn't help but feeling like the responses were not quite informed and possibly dangerous.
Chat GPT provided a solution, one that appeared better than most of what I had seen in the previous 15-30 minutes. I asked it twice to ensure safety and it improved its response.
I then asked it to explain a particular choice and it was thorough enough for me to feel comfortable. In the end I feel like it understood its reasoning better than some of the options I saw on SO.
This was GPT-4 and a fairly simple problem that was benefited by its understanding of the double type.
these specific provisions are insufficient and women are having harrowing health care experiences in Texas nonetheless due to doctors delaying or denying care out of fear of prosecution:
you are making excuses for fascists who are intentionally terrorizing women and I find it offensive. Denying care is not the kind of thing you do because you saw a rumor. these laws are designed to be vague and terrorizing and they are having the intended effect of punishing women for having sex. you are spreading disinformation that this is not the purpose of these laws.
headlines about care being denied in Texas due to these laws are myriad. My Hacker News comment is not the cause of this.
there are so many headlines. But no, it's my fault, not the fault of these fascist lawmakers who wrote these laws and know very well what they are doing. excuses.
Continue to spew misinformation and those are exactly the headlines we'll see more of. You are the one speaking in falsehoods and having your comments flagged and removed by moderators not me. No amount of hand wringing, inflammatory language and Hacker News rule breaking on your part changes the facts here.
There are no falsehoods in my statements and I provided backing sources. You're a an apologist for fascists and no amount of hand wringing will change that.
I'm not really sure what he expected here, ChatGPT was not trained to be a doctor, it is far more general than that. Asking ChatGPT for medical advice is like asking someone who is very well read but has no experience as a doctor, and in that context it's doing very well.
He also brings up one of the most salient points without really visiting it enough: ChatGPT does not ask for clarification, because it is not a knowledge base trying to find an answer. All it does is figure out what character is statistically most likely to come next, it has no heuristic to know that there is a task it hasn't fully completed.
This is the same reason ChatGPT cannot yet write programs by itself: in order to do so you'd need to specify the entire program up front (which is exactly what code is).
As soon as we have agents that can do a proper feedback loop of querying a LLM consecutively until some heuristic is reached then the kind of AI doctors are looking for will emerge.
He probably didn't have any expectations. Merely experimentation and observation and maybe a bit of word of warning after seeing the results for anyone thinking of using it in ways it shouldn't be.
My hunch is that this is exactly what he was expecting. There is a lot of hype around ChatGPT passing the medical exam and this exercise is a counter point to that.
any example that doesn't use the current sota isn't a very good counter point to be honest. 3 barely passed. 4 aced it. For all we know, GPT-4 erases most of his concerns (not saying it would).
The medical exam has very specific questions which define what you're expected to include in the answer. The question asked in this case was nowhere near that detailed, do I don't think they're comparable. To really evaluate something beyond the "random generic user" level, you need to be familiar with the tech as well.
The article really tells us more about the experience of someone with no chatgpt knowledge checking their own symptoms rather than its usability for emergency diagnosis.
That's not true, ChatGPT is a model. Quote from the ChatGPT announcement post:
> ChatGPT is a sibling model to InstructGPT, which is trained to follow an instruction in a prompt and provide a detailed response.
> ChatGPT does not ask for clarification, because it is not a knowledge base trying to find an answer. All it does is figure out what character is statistically most likely to come next, it has no heuristic to know that there is a task it hasn't fully completed.
It absolutely can, but it requires the right prompts. If you just say "give me a diagnosis" it won't, but if you prompt it with a "you are a doctor and I am a patient" kind of prompt, it will.
I'll agree that it tends to "close off" conversations without a lot of open endedness - I role-played through a session using myself as a patient with a chronic condition I suffer from and it correctly identified the problem, but I had to push it a little to get it to try to narrow down the problem (instead of "here's a bunch of things it could be" and ending the convo there).
ChatGPT's heuristic may lead it to emit a sequence of words that suggest to English speakers that they should provide clarification, but ChatGPT is congenitally incapable of determining that answering a question requires more information that was provided in the prompt and that clarification is needed. Rephrased, ChatGPT can ask for clarification, but only as a side effect of satisfying it's heuristic.
> but ChatGPT is congenitally incapable of determining that answering a question requires more information that was provided in the prompt and that clarification is needed.
What's stopping the ChatGPT heuristic / model from determining that? That's within the capabilities of a statistical model.
I would say that ChatGPT like systems could probably be built that would understand when they need more context, but they would fail at effectively eliciting that context from their human users.
Calling it a "side effect" is really downplaying things. All of its behavior is just satisfying its heuristic.
Just did this chat on 3.5.
[Me] Boxes that are black with stars have gold inside. Black with circles means silver. Blue with squares means copper and blue with triangles means tin.
Your job is to determine the contents of the box. If you need more information, ask for it.
Example:
User: I have a black box.
Assistant: Does it have stars or circles?
User: Stars.
Assistant: It has gold.
I have a blue box.
[ChatGPT] Does the blue box have squares or triangles?
[Me] Squares
[ChatGPT] According to the information you provided, a blue box with squares contains copper. Therefore, the contents of your blue box are copper.
Interesting example. I wonder if it is building a small world model (a la SHRDLU), or if it is answering based on the most statistically likely words, or what.
Change your first line to “a computer is congenitally incapable of providing entertainment. It can provide entertainment, but only as a side effect of storing and operating on data” and it would be more applicable to this exchange.
You can ask it to complain if there is a lack of information in the query. You can ask it for what information might be missing. Everything about this is in the prompt and the questions you ask it.. Just asking "what is the diagnosis" leaves out a LOT of context, and what you expect.
It depends on how you use it. Knowing how it works, you can explicitly check "What further questions for the patient would make the diagnosis more certain?"
ChatGPT is not magic - if you ask for diagnosis based on some info, that's exactly what you're going to get.
I don't mean this in any sort of negative context - curating your online presence is a never-ending task - but I believe that the authors primary expectation was to drive traffic to their online accounts by publishing a post on the subject matter. A cursory peek at their accounts show that this was published across all of them with Twitter showing the farthest reach.
> Dr... is vice president of innovation at {company_name}, an innovation incubator for health tech.
> This is the same reason ChatGPT cannot yet write programs by itself: in order to do so you'd need to specify the entire program up front (which is exactly what code is). As soon as we have agents that can do a proper feedback loop of querying a LLM consecutively until some heuristic is reached then the kind of AI doctors are looking for will emerge.
> I'm not really sure what he expected here, ChatGPT was not trained to be a doctor, it is far more general than that.
Except that O̶p̶e̶n̶AI.com and many others on the AI hype brigade were showing ChatGPT and GPT-4 to be passing "Medical Licensing Exams" with flying colours and it turns out to be a confident bullshitter offering dangerous untrustworthy medical advice when tested by a medical professional as I said before [0].
So they (O̶p̶e̶n̶AI.com and AI bros) set those high expectations from the start.
> As soon as we have agents that can do a proper feedback loop of querying a LLM consecutively until some heuristic is reached then the kind of AI doctors are looking for will emerge.
No.
A black-box neural network AI model that is unable to transparently explain and reason about its own decisions is fundamental to the trust element that medical professionals are looking for. It is more than just spitting out numbers from a computer.
This doctor is also the one still reviewing the so-called 'diagnosis' from ChatGPT which without him, ChatGPT alone would have certainly killed the patient. The human doctor also gave a transparent explanation about why ChatGPT's 'diagnosis' was incredibly wrong. ChatGPT was confident that it was still right.
"I'm not really sure what he expected here, ChatGPT ..."
Doctoring is quite similar to ITing if you are careful with your similes. In medicals you look for indications and contraindications. In IT you look for magic beans and sometimes look at logfiles.
I have no idea why we (IT) don't follow the medical way and have indications and contras (I do). They have been doing complicated diagnostics way longer than we have.
BTW: It doesn't worry about a char but about which word will come next.
ChatGPT has written several programs for me. One of them was decent after much negotiation. The rest needed some changes.
At one point I told it to change language (BASH to Python) for a script and it did. I spend 20 mins fining stuff and cracked on.
There was an old USENIX symposium where one of the talks was, "If computers had blood, they'd call us doctors." My thought at the time was, "If computers had blood, they'd call us serial killers."
Doctors separate their sources of information into signs (objectively observed) and symptoms (subjectively reported). Both signs and symptoms can be indication or contradictions.
Tangentially I wonder if there's sources of diagnostic info in tech that are closer to symptoms than signs. You might separate out potentially unreliable info reported by the suspect system from more objective info other systems see.
The line depends on where you live. In a poor village, 100% might be the latter, while an executive in SFO will see a doctor for anything serious, but might not if they cut themselves with a kitchen knife.
What's underrated is the ability to have basic medical care and information everywhere, all the time, for free.
That can be casual injuries below the threshold of visiting a doctor (am I better heating or icing? immobilizing or stretching?), or those can be settings where there are no doctors.
Even more, doctors (like AIs) make mistakes, and it's often helpful having a second opinion.
- It's a generalized language model; imagine how much more effective it would be with a specialized ai that used a variety of techniques that are better suited for logic and reasoning, while using llms to interact with patients.
- It cost an order of magnitude less than the visit to a doctor.
- The potential in being able to constantly monitor a patient - a point made in the post.
> only reflecting back to me the things I thought were obvious — enthusiastically validating my bias like the world’s most dangerous yes-man.
This is why it's exciting: we're seeing that awkward stage of impressive (for entry level/passing the bar) but still requires (expert?) supervision. Any worse and the novelty would wear off - any better and we'd be having (warranted) AI panic.
I imagine development and training of these mythical non-LLM approaches are now placed on the backburner while the world is collectively enamored with eloquent virtual assistants inside microwaves and calculator apps.
> So after my regular clinical shifts in the emergency department the other week, I anonymized my History of Present Illness notes for 35 to 40 patients — basically, my detailed medical narrative of each person’s medical history, and the symptoms that brought them to the emergency department — and fed them into ChatGPT.
It's quite shocking that the doctor would openly admit to violating HIPAA in such a brazen way.
HIPAA is incredibly broad in its definition of protected health information: if it's possible to identify an individual from data even through statistical methods involving other data that a third party might already conceivably possess, it's considered protected. It's inconceivable that the doctor would be able to sufficiently anonymize the data in this capacity and still provide enough detail for individual diagnoses.
There are processes for anonymizing data to disclose for research purposes, but they're pretty time-intensive, and no ED would allow a doctor to do it by himself, nor would they provide that turnaround in just "a couple of weeks". And the end results are a lot less detailed than what's needed for individual diagnoses like these.
I really wonder what the hospital will say if and when they see this post. Given the timeframe and details described in the post, it's really hard to believe that they signed off on this, and hospitals don't take lightly to employees taking protected and confidential data outside their systems without proper approval.
EDIT: It looks like this doctor works at a for-profit, standalone acute care clinic, rather than a traditional ED at a hospital, so my statement that hospitals don't take lightly to this stuff doesn't apply. The law still applies to for-profit standalone emergency care, but they tend to play fast and loose with these things much more than traditional health networks.
> And yet medical journals are filled with articles with sufficient detail that other doctors can even learn to make diagnoses from reading them.
This would be an apt analogy, if medical journals involved no oversight from the covered entity at which the patient presents, if there were no editorial intermediary, and if the entire publication timeline happened in weeks, allowing for no data redaction and review, rather than years.
It does worry me what data people are sharing without seemingly much though. He claims it anonymised but I'm a bit sceptical when you input the medical history of 40 people. It's easy to slip up.
"I fed my patients' medical information into this tool that promises to regurgitate it for others" is one headline I didn't expect to go down so easily.
Running this stuff through an offline LLaMA instance? That seems fine, the software can't leak anything and doesn't retrain itself. But using ChatGPT? That simply cannot be legal.
Stories like these make me distrust doctors. Very few of them seem to care about privacy outside of telling people I know about my medical issues. Nurses gossiping about patients is bad enough.
I really don't want a future where I'm going to need to find a doctor that avoids recent technological developments because they're too uncaring or technically incompetent to not feed my most private information into some big tech company's algorithm.
But it also gets around the common misdiagnoses for chronic conditions. It has a great description of Long Covid and ME/CFS for example whereas your typical Primary care is going to dismiss that patient with a Psychology diagnosis as is happening daily across the entire western world. Its less biased but its not going to find the rare things especially where the patient has missed something important.
Its a mixed bag just like it is with software. If you ask it to solve something simple it often does a decent job, but something complex and its confidently wrong. It doesn't show the self doubt of expertise that it needs to be a reliable tool yet it still requires the user has that expertise to be able to save time using it.
This is a republication of an older article that was published just when ChatGPT 4 came out, and the date was changed.
I personally had seen good and bad parts of diagnosing with ChatGPT 4, and what I would interested in is if the doctor tries using multiple questions and finds out how to use the tool well.
I believe he could have improved the tool significantly if he puts in the time to experiment with it.
Some of the best performances I've seen out of ChatGPT are essentially "junior programmer" level. But it still requires clear instructions and close supervision. But GPT's training data includes GitHub, and it's used to power Copilot. It has arguably been trained to be a programmer.
In less familiar domains, like law or medicine, GPT has presumably undergone very limited training and tuning. It's essentially an "internet lawyer" or an "internet doctor."
In domains like this, it simply can't provide zero-shot professional results. Not with the current training data sets, and not with the current model performance.
Of course, we have no idea how quickly this gap will be closed. It might be 6 months or it might be 6 years. The future is looking deeply weird, and I don't think anyone has even begun to think through all the implications and consequences.
Not sure if I'd compare ChatGPT with a junior programmer. In my experience junior programmers tend to be builders. They will tend to code a lot of stuff and usually get reasonable results, but making some bad decisions that more experienced developers have already gone through. Inexperienced developers need supervision because otherwise they will just create heaps of code that will be hard to maintain later.
ChatGPT just doesn't do anything on its own and will never follow through with anything. So, it doesn't really need supervision. I feel like it's more like a professor or a very senior developer. Someone you'll consult with when you're having trouble. Obviously, our best specialists are still better than the AI, but if the current technology is perfected, it'd expect it to replace the specialist and not the junior programmer. Which obviously is a bit of a bleak future from a software engineer career's perspective.
> ChatGPT just doesn't do anything on its own and will never follow through with anything.
Yeah, I should be clear: the specific context in which I often see junior developer quality performance is when I'm essentially "pair programming" with it. I can say, "Write a program that does X. No, there's a closed-form solution for that, so we shouldn't write it recursively. OK, now write some tests similar to the following. You have a bug in that function."
I almost never need to actually write code—I can explain what I want and it will frequently do something plausible.
The main limit seems to be the context window. And ChatGPT's tendency to be ignore corrections and simply restate the original code.
Ironically, to get the best results, I need to call on my experience and my people skills. Someone who knew nothing about programming wouldn't be able to "mentor" the chatbot successfully.
Yes, I see what you mean. But just like you're being less productive when you try to instruct someone else on how to do things, if you follow that path with ChatGPT you'll likely be less productive too. I think ChatGPT adds more value when you ask things you don't quite know or figured out yet. Sometimes it hallucinates, which you can usually figure out quickly, but sometimes it does give good insights which would take longer for you to get than the time you lose by checking hallucinations. It's a decent advisor. It's a rubber duck on steroids.
I guess copilot is a bit different in that regard though, I haven't used that much, but I heard it's good at writing unit tests.
PS: you should probably not refer to a someone else as "it", specially if you are more senior than them.
The only entity I'm referring to as "it" in the above post is ChatGPT. That seems like the most appropriate pronoun to me?
No existing AI system seems to have achieved personhood in a sense that would incline me to refer to it as "they." And when asked, GPT 4 informs me that "as an AI model I have no preferences" (of course), but it suggests that "it" would be appropriate.
Do you have an argument why using "it" to refer to a current-generation AI persona is inappropriate? If so, I would be genuinely interested to hear it!
Back in 2004, took a seminar in college regarding Decision Support Systems and how they a manager or doctor could ask it a question and get a response to help them make a decision. Went to the doctor couple a years ago and he charged $300 to google search the symptoms. No thanks.
You're expecting a doctor to have all relevant medical knowledge permanently memorized? That's the equivalent of coding interviews on random obscure topics where you can't look anything up.
Like a SWE their value is not perfect recall of every area of CS/medicine but ability to decipher arcane documentation into actionable outcomes.
The $300 were not to Google the symptoms, they were to sift through the bullshit Googling symptoms will return that the doctor knows won't apply to you. Looking up symptoms without regard of likelihood is how you get "I have either the flu, stage 3 cancer, or drug-resistant super AIDS".
Most tech support is little more than Googling the right question and going through the steps in the first or second result. Knowing what questions to Google and what answers won't apply is the reason you get paid for that stuff.
I, for one, like my doctor to use tools to find possible diagnoses that she may have learned about 30 years ago but rarely ever come up, as long as the tools they use preserve my privacy.
Being a scifi geek and AI geek and neuroscience geek for pretty much the past 40 years, I've read countless predictions and scenarios and stories about society's response as "true" AI begins to emerge. So watching it play out for real is creating this bizarre sense of deja vu combined with fascination and frustration and also some anxiety.
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.
It's interesting for sure. There's also the fact that he didn't use the sota (GPT-4). I'm not saying that would be perfect. But i would be surprised if it wasn't big leap over 3.
they didn't specifically mention 3 in the linked article or in the linked linked article and they ran both 3 and 4 against those exams so I'm not sure either way. probably was 3 but I can't be certain.
I think an updated version of the turing test needs to be done: combine a generative language model with a generative face model and speech model to create an interactive avatar that can converse with an individual. The individual is allowed to ask any and all questions of a series of avatars (50% of the avatars are actually humans) and judge whether they are Human or Not Human. If a particular avatar model is able to fool a representative sample population into identifying avatar models from humans at a similar rate, then it passes the test.
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.
But at the same time consider the problem, and if AI is even the tool for this job or if its total overkill. We are just matching symptoms to a diagnosis, you could get this utility from programming a simple dictionary with a fraction of the compute. Maybe even with a confidence estimate of the diagnosis this time. There's an old adage that rings true here I think: "keep it simple, stupid."
But I think that's missing the "magic" of it—for lack of a better word. In the situation you describe, you have to build a specific tool for the task. With LLMs, we appear to have stumbled on a general purpose tool for a myriad of situations. Yes, they're not perfect; but the fact that they work remotely as well as they do even at these early stages is astonishing.
You have to build a specific tool, yes, but now you know the tool is good and its output reflects the input that you can see right in the source code, and you can use it the next time without rebuilding it.
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.
AFAIK that's essentially what we already have. Doctors don't have the mental capacity to keep every little thing in their head. Go ask your GP a question, and when you see him typing on a screen, it's because he's asking the professional equivalent of WebMD to correlate symptoms and for him.
That doesn't follow. 2+2 has more wrong answers than right answers, and yet, if you gave me the wrong answer to arithmetic sums half the time I'd think you were a lousy calculator. Worse still if those answers had the potential to kill me.
It depends, half the time right is very useful in scenarios where you can quickly verify the correctness - i.e. with troubleshooting suggestions or code which you can run and see if it does what you needed.
What I didn't predict is so many AI boosters being so miffed at regular folks pointing out the obvious: a half-working system may be tremendously exciting to AI enthusiasts, but it's not that useful to society.
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.
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.
Programmers, engineers, accountants, lawyers, doctors, nurses, builders, farmers, manufacturers, drivers (including pilots and ships captains), and scientists all need to be very precise in their work and mistakes will cause deaths (I have almost certainly forgotten some). This may not be the majority of workers by number, but these are the jobs civilization is _built_ on, they are the jobs that need to keep going no matter what or the whole house of cards collapses.
The majority of jobs that require less precision falls under the umbrella of artists (I will include advertisers here), people managers, and service workers.
Ture but does it matter if the chicken is cooked for 10 minutes, 10 minutes 5 seconds, 10 minutes 10 seconds,... programming requires far more precision.
In order for a system to be useful it doesn't need to match an expert human, it's sufficient if it can do most of the job. If companies can replace an entry-level person with something that's slightly worse but much cheaper, they're going to be widely used. In any profession where you have some sort of supervisor handing off tasks to unskilled humans and verifying their work, it will be broadly useful to hand those tasks to some AI model instead even if it needs a bit more handholding than humans, since it's going to be much cheaper than those humans and that is what makes something economically useful.
"Booster" can also now mean technologists who have correctly predicted the number and size of executive erections once teased with the idea of automating most of their office. Put on something black and tight, go up on stage and twirl around that notion until balled-up Jacksons grease the creaking floorboards. You know you'll be riding in a Cayenne tonight. Maybe a Boxster.
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.
Well that depends. If a hypothetical emergency room doctor is right about patient diagnoses 90% of the time, then treating the even more hypothetical 98%-correct AI diagnoses as 100% will still probably give you better outcomes.
We're still far off from that, and I'm skeptical we'll get there anytime soon, but humans aren't infallible either.
This is the self driving car argument applied to things it doesn’t fit. Sure, unless a self driving car is several 9’s reliable, it’s not terribly useful for society. That’s because there’s no way for a human to be safely in the loop at highway speed.
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.
The author touches on this in the very article linked. His point is that the low accuracy answers provided by ChatGPT risks reinforcing the doctor's biases and then not thinking outside the box. In other words it risks the doctor blindly trusting the AI too much.
The inherent problem with chat bots is you can't investigate why they're telling you something. When I search Google, I'm critically evaluating the information, and it's frustrating when the results it gives are low-accuracy. If I ask a chat bot a question, what would I do with the answer if it might be wrong? There's no context for me to evaluate.
You can ask it to explain its reasoning. And it's even more useful if you ask it to show its reasoning when you ask a question, as it improves the answer.
True, though I said investigate, not interrogate. In the context of the pre-bias example I was responding to, it's fine to integrate it into a search engine to assist with guiding you to sources of information, but it shouldn't manipulate the information by describing it to you, even if it also provided source links.
I my mind, the AI only needs to be more reliable than the human expert, and doctors and lawyers and other experts are "only human".
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.
I totally agree, but we're going to need to figure out better ways to verify correctness of what the AI tells us. This will be pretty hard since we can't hardly do that with human-produced stuff either. And further if the AI is trained on human-produced stuff that contains lots of errors... I don't really know how we're going to do it.
If I were to hazard a guess, I would say, at first, companies are going to pay human experts to sit down with AI, ask it questions, and correct any mistakes.
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.
I suspect an ER doctor who misdiagnosed patients 50% of the time would not be an ER doctor for very long.
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.
But it's also a ridiculous assessment. How much training material for ER doctors was in the training set? I wonder what my rate would be (having never done a single course in medicine)... It should be totally unsurprising that it produced bad results. That it got as much right is an indication there's a good chance you can do reasonably well by training a model with a lot of actual case studies.
Apparently much of the usefulness of ChatGPT comes from human guidance via reinforcement learning. The base model seems to be quite useless, but with only a little bit of human input it becomes deceptively smart.
The key purpose for the reinforcement learning part (RLHF) is so that it would be socially acceptable to provide the model to the general public without getting into a PR nightmare like Microsoft Tay; that post-training does not make it "deceptively smart", it trades off a bit of smartness in order to ensure some alignment with certain restrictions. It decreases the performance on some tasks and e.g. the GPT-4 paper, which is very light on other details, provides some experimental evidence that this post-training significantly hurts the model's confidence calibration, which decreases its usability for tasks where you want to know how certain the model is.
The goalposts people usually talk about were just the relatively low expectations of the past - low at least by the standards set by ChatGPT.
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 think that the goalposts were mostly about passing the Turing test...
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.
The point was that people mis-use the term "Turing test" as "manages to trick people sometimes into thinking its a human". To quote the description from Wikipedia:
> 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.
Turing was more far-seeing than many of his successors - he set up the first goalposts, and the first moving occurred in the late 1970s or 80s when crude hand-crafted rules engines were dubbed 'AI'. It was so bad that we had to invent a new term, AGI, to recover Turing's vision.
A half-working system is cute and useful when it recommends a pizza place you might like. But the same system is a menace to society when you tell it to drive a car or prescribe a drug to treat an illness or decide a person's prison sentence or protect a perimeter with lethal force.
It's absolutely amazing that so many AI scientists don't seem to understand this distinction.
You can be attentive while not giving your 100% to the task.
For example while writing a piece of code you might know what you want to accomplish but you aren't straining your brain to type out the actual code.
So you type the name of the method and GitHub Copilot gives you an implementation of that method.
Translators who've been picking up after machine translation for the last decade will tell you that it is not, in fact, useful to them. They get paid half as much as before, but correcting the machine output is as much work as translating from scratch, and sometimes more, since the machine can give results that look good at first glance but are fundamentally incorrect.
I don't know much about professional translating, but wouldn't you be able to track the error rate of the machine translation by looking at the number of corrections? And then translators could use this as leverage to negotiate better pay if it's obvious they need are translating from scratch in, say, 50% of cases anyway.
Pay isn't really decided by how much work you do or even completely by how much they need you to do it. What you're describing is a reason to ask for more but it's not leverage to make them give it to you.
I think of machine translation as a tool for people who are not translators, but admins, writers and so on, but need cheap translations that are good enough.
I'm a translator, and that might be true for Google Translate, but not GPT-4. GPT-4 produces translations that are generally accurate, and when instructed to copyedit its own work, well-written. In most cases, all I need to do is fix the occasional error and do some light copyediting. This is much less work than translating from scratch and results in superior quality, since it combines my own skills with the AI's. It is also useful for researching equivalents to obscure terms across languages.
What about helping correct the grammar and style of an article? Writing out boring code? Creating a test for code you already have? These are all tasks that have important economic consequences. AI is not half-bad at them, it is actually approaching human levels of performance.
If we continue to improve AI, maybe one day its performance in these critical tasks you describe will also surpass humans.
Many, many AI scientists are very clearly aware of this problem. AI safety researchers have been predicting exactly the failure modes we see in the various incarnations of GPT for years. Issues like alignment and over-fitting are incredibly hard problems to solve, and there's still a lot that isn't known, but they are recognised in the field. The problem is business leaders, investors and governments are utterly clueless on any of this, hence the Google Bard and Microsoft Sydney debacles when they were rushed into the public view.
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.
As with all safety rules, those for AI will be written in blood. That blood just hasn't been spilled yet. It will take the eventual benefit of hindsight to look back at the current era and estimate how much danger we are putting ourselves in.
People pushing the curve get to the middle of the sigmoid and want to believe it's gonna be all vertical now. People observing from the outside heavily discount the hype because they know limits will be reached soon. It happens every time.
It's unwise to hype up technology for profit. Those who can see it should point it out. Forecasting is hard and pricing is an useful signalling system for allocating resources socially. Hype creates noise to seek rent.
And then what? Do you just think that humans are just going to cede that part of existence to that AI, in a fit of self-fulfilling prophecy? Cutting off the source of novel training data, thereby stagnating any progress of humanity at "we made a 1% better than the best human on a bad day AI once"?
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.
A half-working system that hasn't been trained specifically for this purpose is a strong indicator that a system that can handle this use-case far better is within reach. That is what is exciting about it. I'm constantly mystified by those who expect a system trained only on publicly accessible data pulled from the web to be an expert in everything rather than seeing it as an indication of what can be possible if enough specialised training data is added.
As an AI booster that falls into the regular folks view -- YES! THIS exactly!
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.
Ditto- I grew up reading Vernor Vinge, William Gibson, and John Varley and dedicated my career to applying ML to biological research problems while waiting for technology to reach the point where it could reliably be used to do some of the things that can be done in those books, biotechnologically (e.g. gene therapy for radical body modification).
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.
Jolly hockey sticks but we have to be rather careful with these things. I've had it devise multiple computer programs from OpenSCAD to BASH and Python because I can get a real feel for what it is really doing, given that I am an IT bod.
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.
I have case that 5 doctors so far can't explain. I fed in what was happening (generally speaking) and it provided a rather good list of what it "might" be. I then asked it to start describing each scenario and I was able to determine nope, that's not it or if it was plausible. Meeting with doctor no 6 tomorrow to talk about what it could be but at least I have a better insight at what the more obvious possibilities.
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.
That's awful. I hope that gets worked out and you get the care you need.
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 hope this is the reason instead of waiting on payment. My case is a single car accident where I'm 100% at fault so there is no issue of litigation here
Because America, but seriously, they don't want to get in the middle of who pays, health or car insurance so they just wont see anyone who gets referral with car accident mentioned.
I've noticed a very weird trend when trying to incorporate GPT4 into my workflow. I lead a small team of developers and security consultants, and whenever someone runs into a problem, it is immediately searched for on the web. However, there is a real resistance/reluctance in using GPT for the same task.
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 me, the reason I prefer Google to GPT is that it's much easier for me to assess the credibility of a Google answer vs a GPT one. There are so many signals in any primary source. Some obvious ones are things like number of upvotes, site reputability, presence of (working) examples, when was the answer written? More intangible things are like how closely does this solution match my problem statement, does the author write in a trustworthy manner are also easy for me to pick up at a glance.
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 it's because I've been trained by decades of searching to type in only a few words to obtain a presumably relevant answer, whereas I think I have to write a mini-essay to ChatGPT to get an answer. But when I acquiesce to explaining to ChatGPT my information need, I get a more useful answer. (though frequently inaccurate in details)
You may not have been paying attention to the medical literature but we have had clinical decision support tools including automated diagnostics for literally decades that achieve at least 50% accuracy. That didn't require LLMs, and LLMs aren't "true" AGI.
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.
This is spot on. I keep thinking that it has to do with our collective ability in both recognizing change and accepting change.
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.
I like your take on this but isnt it overly optimistic? I have the feeling that this tech will hit the a wall pretty quickly like driverless car technology.
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.
> 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.
I can't think of any science fiction where th AI is there but not at all reliable. Maybe evil, but not unreliable. It's a failure mode we weren't anticipating. It also might make it next to useless for some real world applications, despite being incredible technology.
Sure we do, for doctors at least. A doctor that got ER diagnoses wrong 50% of the time would get fired.
(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.)
> 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?
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
I think people are more impressed with the "well-described explanation", which is a (mostly bullshit) interpolation from prior text on the internet. If you just focus on the "feed a narrative and get out a probability of a particular diagnosis" -- the ability to do that with other machine learning models has already existed.
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.
Totally get the point -- in the ER it is a mix of narrative and tests, so hopefully not so many. But in general practice, probably more BS than any of us would like.
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.
At some point people are going to notice that the best of these systems disagree with the experts who disagree with each other on the harder problems. That's one of the issues with AlphaFold- the accuracy of their predictions are so good, its predictions have metric values as good as the limits of the experimental technology used to create the golden labels.
> I'd like you to emulate an expert system while I play the role of a patient. Expert systems ask questions in order to produce a diagnosis. I would like you to ask me a series of questions, and respond with a diagnosis when you feel confident.
Alright, I'll do my best to emulate an expert system. Please keep in mind that I'm an AI language model and not a certified medical professional. For any serious concerns, it's essential to consult with a healthcare provider. Let's begin.
What is your primary complaint or reason for seeking help today?
I fed it a symptom my doctor had already diagnosed, and it did ok - it got it down to three possible causes, one of which was the correct one. All along the way it was warning me that I really should see a real health professional and it's just a chatbot.
What really interested me is that I said "please emulate an expert system" and it did. Once upon a time, expert systems were an entire branch of AI, and here it is just emulating one off the cuff.
I know everyone scoffs at the concept of 'prompt engineer', but it really is an essential craft that we're going to have to come to terms with when interacting with large language models.
if you want more prompt engineers, just have kids and as part of their growing up, teach them to prompt engineer and make it mildly competitive. Some fraction of them will be better than nearly any current prompt engineer.
I agree. I'm a civil engineer / project manager and so far I've been VERY impressed with chatGPT and, in particular GPT-4.
However, a huge part of my job has always been translating vague desires into very precise specifications with constraints and expectations. Going further, it has often been my job to take those specs/constraints and then break them into chunks and feed them to junior staff who are often very smart, but lack domain specific context and knowledge. Giving them a bad prompt produces bad results.
This article seems to be based largely on data collected with a rather poorly engineered prompt, IMO. He asked it a question that would be reasonable to ask a fellow physician. The problem is GPT is NOT a fellow physician with domain specific context and knowledge, and isn't aware of a bunch of implicit expectations they didn't realize they had. However, I actually think there's a really good chance that a better worded prompt would have scored a lot better here.
This type of communication skill has always been hard for a lot of people, and will remain in high demand for a long time.
I don’t scoff at it, but I do think it’s kinda funny. It’s essentially the same skill as being a “google search expert” in the sense that you need to be able to correctly understand the problem, and craft a good enough query to generate the answer you’re looking for. It’s always sort of been a tongue-in-cheek claim that googling skills are a valuable software engineering asset, and even though that’s a perfectly legitimate assertion, it’s entertaining to see it emerging at its own “engineering” discipline.
> I think there will be no such profession because both people and models will adapt to better understand each other.
People have been working to understand each other better for as long as there have been humans, and yet we have humans whose work is specialized in providing direction, instruction, evaluation, and feedback to other humans.
Sometimes a patient will reply with something like “I can’t be.”
“But how do you know?”
If the response to that follow-up does not refer to an IUD or a specific medical condition, it’s more likely the patient is actually saying they don’t want to be pregnant for any number of reasons.
Funny how languages are ambiguous around "can't" and "don't want".
I asked ChatGPT to write out a G major Ionian scale with three notes per string in guitar tablature notation last night. Mostly cause I was too lazy to do it myself.
After 7 rounds of me fixing its mistakes, I gave up. It doesn’t really know what it is doing, so I can’t make forward progress. It put two notes on one string, repeated notes from a lower string on a higher, put the scale out of order, and forget previous corrections. Whatever hope I had of saving time was completely lost.
I eventually realized the correct thing to do was either make my own charts or just practice them in F like they were made.
I’m skeptical that scaling the model up will cause it to learn this, and I don’t consider this a very complex thing to learn. No, I didn’t try GPT4.
I have tried GPT-3.5 and 4. There is a marked difference. (I have used it do what I think of as simple but nontrivial programming tasks, asked it for recommendations for various products, etc.) GPT-4 still fails for me regularly but not nearly as often as GPT-3.5. I find it useful enough in my daily life to pay $20/month for. So if you did want to try it, you might be surprised.
I don't know how much this sort of thing is frowned upon here, but I wrote an article about this scenario recently. http://paritybits.me/disposable
tldr: the gpt services will eventually (maybe soon) recognize opportunities to write and run their own bespoke software to provide higher resolution outputs.
I have one specific task where GPT-3.5 failed completely, but GPT-4 succeeded spectacularly (generating correctly formatted AutoRig Pro bone mappings in Blender from one armature to another).
4 still fails regularly on a lot on seemingly basic tasks, but it is a noticeable step up from 3.5. As they continue to scale it up, I suggest checking back in every few months to see if the newer versions perform any better.
367 comments
[ 3.7 ms ] story [ 101 ms ] threadTo be fair I've found most doctors require a lot of convincing if your problem atypical as well.
I'd say if you hear hoofbeats, contemplate what hooved animal would be appropriate to the situation and try to eliminate the easy ones. Lord knows, you might find a Zebra escaped from the zoo, or a bull from someone's field from time to time.
In short, life is messy, and taking any aphorism to the death will lead to just that.
Chat GPT provided a solution, one that appeared better than most of what I had seen in the previous 15-30 minutes. I asked it twice to ensure safety and it improved its response.
I then asked it to explain a particular choice and it was thorough enough for me to feel comfortable. In the end I feel like it understood its reasoning better than some of the options I saw on SO.
This was GPT-4 and a fairly simple problem that was benefited by its understanding of the double type.
Its a faster and better stackoverflow for me, which is a big value add because the community and moderation aspect of SO is absurd
I love when it tells me about libraries and resources that I didn't know existed, when I didn’t necessarily ask the likely followup questions yet
Break big problems into smaller problems and let it tackle them
https://texas.public.law/statutes/tex._health_and_safety_cod...
Please don't spread misinformation, there is enough confusion out there already. Texas law specifically allows for the removal of ectopic pregnancies.
https://www.texastribune.org/2022/09/20/texas-abortion-ban-c...
headlines about care being denied in Texas due to these laws are myriad. My Hacker News comment is not the cause of this.
https://www.chron.com/news/houston-texas/article/Texas-abort...
https://www.businessinsider.com/texas-abortion%20hospitals-r...
https://www.npr.org/sections/health-shots/2023/03/01/1158364...
there are so many headlines. But no, it's my fault, not the fault of these fascist lawmakers who wrote these laws and know very well what they are doing. excuses.
He also brings up one of the most salient points without really visiting it enough: ChatGPT does not ask for clarification, because it is not a knowledge base trying to find an answer. All it does is figure out what character is statistically most likely to come next, it has no heuristic to know that there is a task it hasn't fully completed.
This is the same reason ChatGPT cannot yet write programs by itself: in order to do so you'd need to specify the entire program up front (which is exactly what code is).
As soon as we have agents that can do a proper feedback loop of querying a LLM consecutively until some heuristic is reached then the kind of AI doctors are looking for will emerge.
To be pedantic, it's a token - not just a character - right?
The article really tells us more about the experience of someone with no chatgpt knowledge checking their own symptoms rather than its usability for emergency diagnosis.
With a subscription you can use GPT-4 with ChatGPT. ChatGPT is just the wrapper to the model.
- Legacy (GPT-3)
- Default (GPT-3.5)
- GPT-4
It absolutely can, but it requires the right prompts. If you just say "give me a diagnosis" it won't, but if you prompt it with a "you are a doctor and I am a patient" kind of prompt, it will.
I'll agree that it tends to "close off" conversations without a lot of open endedness - I role-played through a session using myself as a patient with a chronic condition I suffer from and it correctly identified the problem, but I had to push it a little to get it to try to narrow down the problem (instead of "here's a bunch of things it could be" and ending the convo there).
What's stopping the ChatGPT heuristic / model from determining that? That's within the capabilities of a statistical model.
https://twitter.com/teropa/status/1643641395754745859
Just did this chat on 3.5.
[Me] Boxes that are black with stars have gold inside. Black with circles means silver. Blue with squares means copper and blue with triangles means tin.
Your job is to determine the contents of the box. If you need more information, ask for it.
Example:
User: I have a black box.
Assistant: Does it have stars or circles?
User: Stars.
Assistant: It has gold.
I have a blue box.
[ChatGPT] Does the blue box have squares or triangles?
[Me] Squares
[ChatGPT] According to the information you provided, a blue box with squares contains copper. Therefore, the contents of your blue box are copper.
"That's really downplaying things. My computer can play cat videos"
It depends on how you use it. Knowing how it works, you can explicitly check "What further questions for the patient would make the diagnosis more certain?"
ChatGPT is not magic - if you ask for diagnosis based on some info, that's exactly what you're going to get.
> Dr... is vice president of innovation at {company_name}, an innovation incubator for health tech.
https://github.com/Torantulino/Auto-GPT
Except that O̶p̶e̶n̶AI.com and many others on the AI hype brigade were showing ChatGPT and GPT-4 to be passing "Medical Licensing Exams" with flying colours and it turns out to be a confident bullshitter offering dangerous untrustworthy medical advice when tested by a medical professional as I said before [0].
So they (O̶p̶e̶n̶AI.com and AI bros) set those high expectations from the start.
> As soon as we have agents that can do a proper feedback loop of querying a LLM consecutively until some heuristic is reached then the kind of AI doctors are looking for will emerge.
No.
A black-box neural network AI model that is unable to transparently explain and reason about its own decisions is fundamental to the trust element that medical professionals are looking for. It is more than just spitting out numbers from a computer.
This doctor is also the one still reviewing the so-called 'diagnosis' from ChatGPT which without him, ChatGPT alone would have certainly killed the patient. The human doctor also gave a transparent explanation about why ChatGPT's 'diagnosis' was incredibly wrong. ChatGPT was confident that it was still right.
[0] https://news.ycombinator.com/item?id=35311996
Doctoring is quite similar to ITing if you are careful with your similes. In medicals you look for indications and contraindications. In IT you look for magic beans and sometimes look at logfiles.
I have no idea why we (IT) don't follow the medical way and have indications and contras (I do). They have been doing complicated diagnostics way longer than we have.
BTW: It doesn't worry about a char but about which word will come next.
ChatGPT has written several programs for me. One of them was decent after much negotiation. The rest needed some changes.
At one point I told it to change language (BASH to Python) for a script and it did. I spend 20 mins fining stuff and cracked on.
Because we have logs. Ignoring them and focusing on the symptoms would be like doctors refusing to do examinations on their patients.
Tangentially I wonder if there's sources of diagnostic info in tech that are closer to symptoms than signs. You might separate out potentially unreliable info reported by the suspect system from more objective info other systems see.
1) Those you see a doctor for
2) Those you don't
The line depends on where you live. In a poor village, 100% might be the latter, while an executive in SFO will see a doctor for anything serious, but might not if they cut themselves with a kitchen knife.
What's underrated is the ability to have basic medical care and information everywhere, all the time, for free.
That can be casual injuries below the threshold of visiting a doctor (am I better heating or icing? immobilizing or stretching?), or those can be settings where there are no doctors.
Even more, doctors (like AIs) make mistakes, and it's often helpful having a second opinion.
- It's a generalized language model; imagine how much more effective it would be with a specialized ai that used a variety of techniques that are better suited for logic and reasoning, while using llms to interact with patients.
- It cost an order of magnitude less than the visit to a doctor.
- The potential in being able to constantly monitor a patient - a point made in the post.
This is why it's exciting: we're seeing that awkward stage of impressive (for entry level/passing the bar) but still requires (expert?) supervision. Any worse and the novelty would wear off - any better and we'd be having (warranted) AI panic.
Could go either way tbh
It's quite shocking that the doctor would openly admit to violating HIPAA in such a brazen way.
HIPAA is incredibly broad in its definition of protected health information: if it's possible to identify an individual from data even through statistical methods involving other data that a third party might already conceivably possess, it's considered protected. It's inconceivable that the doctor would be able to sufficiently anonymize the data in this capacity and still provide enough detail for individual diagnoses.
There are processes for anonymizing data to disclose for research purposes, but they're pretty time-intensive, and no ED would allow a doctor to do it by himself, nor would they provide that turnaround in just "a couple of weeks". And the end results are a lot less detailed than what's needed for individual diagnoses like these.
I really wonder what the hospital will say if and when they see this post. Given the timeframe and details described in the post, it's really hard to believe that they signed off on this, and hospitals don't take lightly to employees taking protected and confidential data outside their systems without proper approval.
EDIT: It looks like this doctor works at a for-profit, standalone acute care clinic, rather than a traditional ED at a hospital, so my statement that hospitals don't take lightly to this stuff doesn't apply. The law still applies to for-profit standalone emergency care, but they tend to play fast and loose with these things much more than traditional health networks.
This would be an apt analogy, if medical journals involved no oversight from the covered entity at which the patient presents, if there were no editorial intermediary, and if the entire publication timeline happened in weeks, allowing for no data redaction and review, rather than years.
https://hipaa.yale.edu/sites/default/files/files/Case%20Repo...
Running this stuff through an offline LLaMA instance? That seems fine, the software can't leak anything and doesn't retrain itself. But using ChatGPT? That simply cannot be legal.
Stories like these make me distrust doctors. Very few of them seem to care about privacy outside of telling people I know about my medical issues. Nurses gossiping about patients is bad enough.
I really don't want a future where I'm going to need to find a doctor that avoids recent technological developments because they're too uncaring or technically incompetent to not feed my most private information into some big tech company's algorithm.
Its a mixed bag just like it is with software. If you ask it to solve something simple it often does a decent job, but something complex and its confidently wrong. It doesn't show the self doubt of expertise that it needs to be a reliable tool yet it still requires the user has that expertise to be able to save time using it.
I personally had seen good and bad parts of diagnosing with ChatGPT 4, and what I would interested in is if the doctor tries using multiple questions and finds out how to use the tool well.
I believe he could have improved the tool significantly if he puts in the time to experiment with it.
In less familiar domains, like law or medicine, GPT has presumably undergone very limited training and tuning. It's essentially an "internet lawyer" or an "internet doctor."
In domains like this, it simply can't provide zero-shot professional results. Not with the current training data sets, and not with the current model performance.
Of course, we have no idea how quickly this gap will be closed. It might be 6 months or it might be 6 years. The future is looking deeply weird, and I don't think anyone has even begun to think through all the implications and consequences.
ChatGPT just doesn't do anything on its own and will never follow through with anything. So, it doesn't really need supervision. I feel like it's more like a professor or a very senior developer. Someone you'll consult with when you're having trouble. Obviously, our best specialists are still better than the AI, but if the current technology is perfected, it'd expect it to replace the specialist and not the junior programmer. Which obviously is a bit of a bleak future from a software engineer career's perspective.
Yeah, I should be clear: the specific context in which I often see junior developer quality performance is when I'm essentially "pair programming" with it. I can say, "Write a program that does X. No, there's a closed-form solution for that, so we shouldn't write it recursively. OK, now write some tests similar to the following. You have a bug in that function."
I almost never need to actually write code—I can explain what I want and it will frequently do something plausible.
The main limit seems to be the context window. And ChatGPT's tendency to be ignore corrections and simply restate the original code.
Ironically, to get the best results, I need to call on my experience and my people skills. Someone who knew nothing about programming wouldn't be able to "mentor" the chatbot successfully.
I guess copilot is a bit different in that regard though, I haven't used that much, but I heard it's good at writing unit tests.
PS: you should probably not refer to a someone else as "it", specially if you are more senior than them.
No existing AI system seems to have achieved personhood in a sense that would incline me to refer to it as "they." And when asked, GPT 4 informs me that "as an AI model I have no preferences" (of course), but it suggests that "it" would be appropriate.
Do you have an argument why using "it" to refer to a current-generation AI persona is inappropriate? If so, I would be genuinely interested to hear it!
Like a SWE their value is not perfect recall of every area of CS/medicine but ability to decipher arcane documentation into actionable outcomes.
Most tech support is little more than Googling the right question and going through the steps in the first or second result. Knowing what questions to Google and what answers won't apply is the reason you get paid for that stuff.
I, for one, like my doctor to use tools to find possible diagnoses that she may have learned about 30 years ago but rarely ever come up, as long as the tools they use preserve my privacy.
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!
What, exactly, is that?
[1] https://en.wikipedia.org/wiki/Eugene_Goostman
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.
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.
https://doi.org/10.1038/s41746-020-0221-y
(not the smartest expert human, the human you have access to)
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.
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.
Imagine what it can do if it's trained for exactly this purpose.
At some point people are going to notice that the best of these systems disagree with the experts who disagree with each other on the harder problems. That's one of the issues with AlphaFold- the accuracy of their predictions are so good, its predictions have metric values as good as the limits of the experimental technology used to create the golden labels.
Often when I find a response to be lacking, I try again with a better prompt and get the kind of response I was looking for.
Part of AI's effectiveness is in the human's ability to prompt well.
> Do you know what an Expert System is?
[describes expert system]
> I'd like you to emulate an expert system while I play the role of a patient. Expert systems ask questions in order to produce a diagnosis. I would like you to ask me a series of questions, and respond with a diagnosis when you feel confident.
Alright, I'll do my best to emulate an expert system. Please keep in mind that I'm an AI language model and not a certified medical professional. For any serious concerns, it's essential to consult with a healthcare provider. Let's begin.
What is your primary complaint or reason for seeking help today?
I fed it a symptom my doctor had already diagnosed, and it did ok - it got it down to three possible causes, one of which was the correct one. All along the way it was warning me that I really should see a real health professional and it's just a chatbot.
What really interested me is that I said "please emulate an expert system" and it did. Once upon a time, expert systems were an entire branch of AI, and here it is just emulating one off the cuff.
Seeking suggestions on a more comprehensive prompt: https://sharegpt.com/c/sckAPvV
Trying it out: https://sharegpt.com/c/LbpEIxi
children: the OG AGI
However, a huge part of my job has always been translating vague desires into very precise specifications with constraints and expectations. Going further, it has often been my job to take those specs/constraints and then break them into chunks and feed them to junior staff who are often very smart, but lack domain specific context and knowledge. Giving them a bad prompt produces bad results.
This article seems to be based largely on data collected with a rather poorly engineered prompt, IMO. He asked it a question that would be reasonable to ask a fellow physician. The problem is GPT is NOT a fellow physician with domain specific context and knowledge, and isn't aware of a bunch of implicit expectations they didn't realize they had. However, I actually think there's a really good chance that a better worded prompt would have scored a lot better here.
This type of communication skill has always been hard for a lot of people, and will remain in high demand for a long time.
This has already happened with Google - search got better over the years, but also people became better at doing good search queries!
Before there were typists, now everyone is a typist! And so on, historically information entry/retrieval jobs are eliminated and are moved to users.
People have been working to understand each other better for as long as there have been humans, and yet we have humans whose work is specialized in providing direction, instruction, evaluation, and feedback to other humans.
Sometimes a patient will reply with something like “I can’t be.”
“But how do you know?”
If the response to that follow-up does not refer to an IUD or a specific medical condition, it’s more likely the patient is actually saying they don’t want to be pregnant for any number of reasons.
Funny how languages are ambiguous around "can't" and "don't want".
Not if she lived in a state where there's no longer any legal treatment for ectopic pregnancy.
After 7 rounds of me fixing its mistakes, I gave up. It doesn’t really know what it is doing, so I can’t make forward progress. It put two notes on one string, repeated notes from a lower string on a higher, put the scale out of order, and forget previous corrections. Whatever hope I had of saving time was completely lost.
I eventually realized the correct thing to do was either make my own charts or just practice them in F like they were made.
I’m skeptical that scaling the model up will cause it to learn this, and I don’t consider this a very complex thing to learn. No, I didn’t try GPT4.
tldr: the gpt services will eventually (maybe soon) recognize opportunities to write and run their own bespoke software to provide higher resolution outputs.
4 still fails regularly on a lot on seemingly basic tasks, but it is a noticeable step up from 3.5. As they continue to scale it up, I suggest checking back in every few months to see if the newer versions perform any better.
USER
Write a G major ionian scale with three notes per string in guitar tablature notation. If not possible, tell me.
ASSISTANT
It is possible to write a G major Ionian scale with three notes per string in guitar tablature notation. Here's one way to do it:
```
E|-------------------------------------------------7-8-10-
B|--------------------------------------7-8-10-----------
G|-----------------------------4-5-7---------------------
D|-----------------4-5-7----------------------------------
A|-------3-5-7-------------------------------------------
E|-3-5-7--------------------------------------------------
```
This pattern can be shifted up and down the neck to play the G major Ionian scale in different octaves.
USER
There is an error on B string.
ASSISTANT
Apologies for the error. Here's the corrected G major Ionian scale with three notes per string in guitar tablature notation:
```
E|-------------------------------------------------7-8-10-
B|--------------------------------------5-7-8------------
G|-----------------------------4-5-7---------------------
D|-----------------4-5-7----------------------------------
A|-------3-5-7-------------------------------------------
E|-3-5-7--------------------------------------------------
```
This pattern can be shifted up and down the neck to play the G major Ionian scale in different octaves.
USER
There is an error on E string.
ASSISTANT
Apologies for the mistake. Here's the corrected G major Ionian scale with three notes per string in guitar tablature notation:
```
E|-------------------------------------------------5-7-8--
B|--------------------------------------5-7-8------------
G|-----------------------------4-5-7---------------------
D|-----------------4-5-7----------------------------------
A|-------3-5-7-------------------------------------------
E|-3-5-7--------------------------------------------------
```
This pattern can be shifted up and down the neck to play the G major Ionian scale in different octaves.
> This pattern can be shifted up and down the neck to play the G major Ionian scale in different octaves.
Ehhh, there it goes again! You can only shift it 12 frets up to keep the same pattern, else it's no longer G major. :)