On the other hand, a lot of people who ran into dead ends with physicians found incredible insights into their health problems using AI.
I have an issue I’ve had for years, and ran into dead ends with physicians and medical professionals. The AI said it could be a rare condition, then I asked it for what physicians were near me who could treat it. It found me a physician, I made an appt and they did some tests to verify the diagnosis the AI gave.
AI can hallucinate. Nothing new about that.
Maybe your AI gives great medical advice 99% of the time. But what happens that one time it gets it wrong?
Don’t put your health in the care of your computer. Go see your doctor.
As the others pointed out, AI is like a bus that can you take you in the right direction, just don’t treat it like a taxi that gets you to the precise address.
I once used AI for legal matters and it was highly useful. When I fed it a 500-word summary of my situation, it brought up issues like “promissory estoppel” and “detrimental reliance” that I would never have had a clue about on my own. It wasn’t 100% accurate, but it made me aware of much more.
Just saying:
Dutch lawyers increasingly have to convince clients that they can’t rely on AI-generated legal advice because chatbots are often inaccurate, the Financieele Dagblad (FD) found when speaking to several lawfirms. A recent survey by Deloitte showed that 60 percent of lawfirms see clients trying to perform simple legal tasks with AI tools, hoping to achieve a faster turnaround or lower fees.
Right. I wouldn’t advise to treat AI as something that’s fully precise, but it can steer you in the right direction - and then you search more for the actual technical details on your own.
It also helps to use multiple AIs - ChatGPT vs Grok vs Claude vs Gemini, etc. If they all agree, then you MIGHT have something good. If they mutually contradict, then check and re-check much more.
Have you seen how many things humans are wrong about? Humans are not perfect, infallible experts. If we were, we wouldn’t need AI.
Also you act like AI we have now is the end goal. Its just another step in the progress, AI in 2029 will be vastly better and more accurate.
On top of that, the AI has infinite patience and can crunch far more data in one go than a doctor can. A human doctor is likely to be jaded, fatigued, you are his 17th patient out of 20 that day, and just wants to go home, while the AI can calmly process a 3,000-word long medical history that you send to it in one massive paragraph and answer a hundred follow-up questions with no loss of patience.
I don’t disagree that the ultimate end point of the process should be a visit to the doctor. But to say that you shouldn’t use AI because it occasionally gets things wrong discounts the handful of times I’ve seen a doctor get something wildly wrong too.
To give an AI a list of symptoms and then ask “what’s wrong with me” is probably not its best use. But you can certainly have enlightening conversations about what the next steps could be, what other questions to ask, things to watch out for, etc.
I think the main issue is if the AI gets it wrong there is no backup. You are now following bad advice.
A doctor can get things wrong too but there are systems surrounding the doctor that have a better chance at catching a mistake.
Certainly Google/AI whatever is ailing you but also go see your doctor. The AI is not a replacement for seeing the doctor. Not yet anyway.
And yet I have seen those supposed ‘systems’ either fail catastrophically or simply not exist.
But at any rate, I’m not saying that AI is a replacement for proper medical treatment, but rather that it’s an excellent additional tool that can give you more information and insight to frame better questions and challenge assumptions, either your own or the doctor’s.
Think about that.
You are saying that you still need a doctor to diagnose and resolve whatever it is that ails you.
How do you think this works?
Things that never happen in real life:
Doctor: I think you are suffering from X.
Patient: I asked my AI and it suggested this, that and another thing might be wrong with me.
Doctor: OMG!!! You are so right! I never considered this, that and the other thing. It is great you asked your AI for medical advice. If you hadn’t you’d be dead.
ETA: To be clear - Of course you (anyone) will Google any malady you have been told you have. And it makes perfect sense. What can you expect, how do you live with the problem, are there support groups and so on. All well and good. But the diagnosis and treatment have to come from your doctor.
A diffuse set of self observed symptoms is not going to lead to a useful diagnosis or treatment for the majority of important illnesses or conditions.
So the value to ordinary folk is always going to be limited. AI for some parts of medical diagnosis has a long history, going back to Mycin. Something that has the ability to store and correlate the knowledge of rare, unusual and weird medical conditions and presentations will always help. However any general practitioner will tell you that half of their work is done listening and gauging the person, not the symptoms.
Perhaps in places where medical care is out of reach for some people AI systems can help. But if you have available and affordable care, it isn’t going to be especially useful. Particularly if the damned thing cannot distinguish pseudoscience and pop science medicine from solid medical science.
It isn’t as if the AI can write a script or order tests. And going to a doctor with an AI diagnosis in your hand and demanding a script the AI has advised isn’t going to go down well.
Depends where you are. My advice to anyone reaching middle age is to find a trustworthy sensible general practitioner and stick with them. Not everyone can do this, which is unfortunate. But continuity of care is important in many ways. You won’t realise it until you need it.
In total, they visited 17 different doctors over three years. But Alex still had no diagnosis that explained all his symptoms. An exhausted and frustrated Courtney signed up for ChatGPT and began entering his medical information, hoping to find a diagnosis.
“I went line by line of everything that was in his (MRI notes) and plugged it into ChatGPT,” she says. “I put the note in there about … how he wouldn’t sit crisscross applesauce. To me, that was a huge trigger (that) a structural thing could be wrong.”
She eventually found tethered cord syndrome and joined a Facebook group for families of children with it. Their stories sounded like Alex’s. She scheduled an appointment with a new neurosurgeon and told her she suspected Alex had tethered cord syndrome. The doctor looked at his MRI images and knew exactly what was wrong with Alex.
Way to oversimplify a valid use case all the way down to absurdity.
I am saying that in my direct personal experience, AI can be a valuable partner in organizing symptoms and observations into into a coherent story that gives a harried doctor better odds at coming to a reasonable conclusion. At the same time it can can also challenge assumptions, provide the patient with a line of perceptive questions, and suggest particular specialists to engage.
All too often in the past couple of years dealing with my wife’s significant, sometimes life-threatening, definitely complex medical conditions, we’ve been left with fragmentary information and occasionally contradictory opinions. She sees three oncologists, a rheumatologist, an endocrinologist, and a pulmonologist. After speaking with those six doctors, we will sometimes come away with eight opinions. There is no one doctor that plays referee here. So AI has given us tools to know when to ask more questions and to push back, and to whom.
Would that I have had such a tool when caring for my father in his last years of battling MS. More than once, I found that he had been prescribed incompatible medications that no one caught because different doctors in different record keeping platforms were sending prescriptions to different pharmacies. What didn’t I catch? Who knows.
I am not bagging on doctors here. But realistically, we live in a system where medical attention is rationed even to those who can pay and the specialities have become so fragmented in some cases that it feels like the blind men examining an elephant. AI is a new tool in the toolbox, one that I welcome.
I’ll tell you an actual exchange I had with my doctor a few weeks ago.
Me: I have this issue. I looked up several treatments on AI for it. What do you think of these?
Her: (Pulls out her own AI app) two of them are OTC and harmless to try, but the effectiveness is unclear. The one with the best reported results has too small of a sample size to be meaningful.
Me. ok thanks.
Whatever you think AI can or can’t replace, your professionals are already using it too. It’s a massively effective way to save time and surface new ideas for people who are already skilled in problem-solving and analysis. There are some things where human input will always be needed, but in a few years there will be few knowledge-intensive trades that aren’t served by a hybrid AI-human team.
If you work in knowledge-intensive trades, you will use AI more heavily than you thought possible, or you will be out-competed by those who do. In that way it will be like the internet revolution. At the time many professionals saw it as a toy. Some adopted it early. Practically everybody eventually adopted it out of necessity because it simply wasn’t possible to work without having mastered what quickly became standard practice.
Does it have flaws? Of course it does. But the tools are pretty damned good. And for all their flaws, they are currently the worst they’ll ever be.
I’m an AI skeptic. Or I was a really strong skeptic for a long time. My employer markets an AI tooling brand that you’ve heard of, so I’ve been forced to use these tools in my job for several years now. Initially the tools were bad, and I was bad at using them, but both of those things have changed over the past few years.
Last night I was using the new 4.6 model of Claude Sonnet for something like literary analysis, and its insights and self-awareness really blew me away. It’s the first time its capabilities ever really made me apprehensive about it surpassing my abilities, or at least appearing to surpass my abilities in ways that might convince a prospective employer to use it instead of me. The future has already crossed the threshold and is lounging in your foyer.
It is fine for a professional to use various resources to find an answer. They have the training and experience to assess the advice given. Most of us are not doctors and we are unable to assess if an AI is hallucinating when giving medical advice. Given how crucial such advice can be (and possibly dangerous) we should be asking our doctors and not our computers for medical advice.
This is probably a very difficult challenge (for anything/anyone), unless someone has already reverse engineered it (did you check Github?). Otherwise, you might have to figure out the binary encoding first — there are people who do that for fun or for a living, reverse engineering programs and their file formats.
I am fairly confident LLMs can’t do this yet. Binary encodings are not natively “readable” by LLMs, so would have to go through some sort of parsing/decoding first.
It’s not something you could just throw money at unless (maybe) you had a sufficiently large training dataset of them (tens of thousands? maybe more?) to train a new LLM as though the binary were a language of its own, i.e., training it to understand that binary encoding as just another language. That might be doable if the binary were a simple bit-for-bit representation of the underlying data, as opposed to some sort of compression or encoding scheme. It would not be easy.
This should be more doable. Can you post some samples (both the files themselves, and the layout they are supposed to represent)? Or DM me with them?
LLMs might be able to do them either out of the box (if it’s a format they’ve seen before or can relatively easily reason about). Otherwise, with enough sample inputs & outputs, it can at least help write converters and tests and then iterate through them (with agentic workflows) until it gets them right and the outputs can pass some validation testing (that it can also help write). It might take a lot of prompting, trial and error, and a lot of notes for the LLM though (and probably not through a simple chatbot interface, but something more like Claude Code or another agentic workflow).
Anyway, feel free to post/send some of those if that’s a possibility and I can look into it for you in ChatGPT, Claude, and Gemini.
Or you can try it yourself with coding agents (like Claude Code, OpenAI Codex, Google Antigravity, Cursor, Windsurf, etc.). Those are good for more complex coding tasks like this if the basic chatbots don’t work well enough.
Philosophically, computer programming is something that the LLMs excel at, both because they’re just a form of language translation, and also because they have immediately verifiable outputs (in a way that medical, legal, and most other questions don’t).
They still do frequently hallucinate while coding, but the agentic workflows are not just pure LLMs. They can call other tools, write tests, run those tests, etc., again and again until the output actually verifiably compiles/builds and actually works. It’s less of a Q&A type chat interface and more of an orchestration system for agents and sub-agents that prompt each other, test their outputs, repeat until it works, etc.
So it’s not really the same problem domain as in most other fields. Here you’re not asking them to find truthiness, merely build & rebuild code until it becomes verifiably correct/functional. Programming is probably the single most powerful use for LLMs right now, which is why there are so many companies and models for agentic coding all competing each other.
As far as models go, I think Claude and Gemini are currently in the lead, but that changes every few weeks — there are numerous benchmarks available you can look up if you really care. But your prompting and context notes will make a big difference in something like this, maybe even more so than raw model performance. Even if nothing can “one-shot” that particular format, you can probably get a working converter written with a few hours of effort (mostly waiting + occasional prompting).
FWIW, I know five people who do coding for major tech companies (3 or 4 are household names and the other 1 or 2 are still quite recognizable) and they all use Claude. Also they all speak rather positively of it which was a little surprising since I’ve heard a number of “It was so bad, I spent more time fixing it…” gripes from random corners of the internet. In fact, one guy was recently saying how surprised he was at how good it was because he assumed it would be trash. I’d almost say he was apologetic about it – “Yeah, I know, AI sucks and all but damn this is actually really good…”
For personal use, I don’t use LLMs often but I’ve done a small assortment of stuff with them:
Gaming: As suggested upthread, I’ve used an LLM to give suggestions on some side content and encounters. I wouldn’t run of a game off it or have it develop the main story but it’s fine (IMO) for side stuff. In a game like Twilight: 2000 which is basically a big hex crawl it’s nice to have an additional source of encounter suggestions. I’ve also used it for making location battlemats, character art and other flavor material via various Stable Diffusion programs (mainly local, not LLM)
Music: I’ve used it for music suggestions, giving it a list of 7-10 songs I’ve enjoyed recently and asking it to suggest similar music with some constraints such as songs since 2020, songs by artists with fewer than 100k Spotify followers or songs that haven’t been featured in a movie, TV or commercial. The first time I tried this was probably the best but I’ve gotten some good suggestions since then. I used ChatGPT at the time but might try Gemini. Major complaint was that it made up Spotify links and did have some song name hallucinations.
Product Help: I used it to help me narrow down suggestions for a new turntable cartridge. Google was pretty useless for this, mainly pointing me to Reddit threads plagued by tribalism, elitism and very little actionable information. Gemini took my technical info and price ranges and gave me a couple of options with some pros and cons. I used that to confirm suitability and was happy with the results.
Work Help: Not too much but I have used it to help identify and decipher some construction plan symbols that are usually outside my scope. Being able to just upload a photo or PDF and say “What’s this?” and get a response is super handy. Again, used that info to confirm accuracy rather than assuming the LLM was infallible but it was accurate and helped out a lot.
Yeah, but that changes frequently. Every once in a while, OpenAI pulls ahead for a few weeks, and then Gemini takes the crown for a month, then it’s back to Claude, etc. (This is specifically in the context of coding ability and benchmarks.)
I work for a tiny tech company that you’ve never heard of, and a few months ago we were all using ChatGPT, then Gemini, then Claude for now. Thankfully the subscriptions are monthly so you can just pause and restart the plans as the leaders change.
There’s frequent and rapid incremental advancements in both the LLM models themselves and the productization around them, such as all the agentic stuff. It’s way too early to declare Claude (or anyone) a winner yet. Chances are most of those companies will be bankrupt in a few years anyway, or be absorbed into the remaining giants.
It’s a mad gold rush right now and the only clear winners are Nvidia and TSMC, the shovel vendors.
Right now, I’m deep into one of my other serious use cases for Generative AI: Editing YouTube videos.
I just finished a video explaining in-ear-monitors for worship teams and bands, with demos of latency and so on.
After I was happy with all of my edits, I used a thing called “whisper” to create a transcript from the audio. I uploaded that to ChatGPT and explained the premise of the video.
It then had it QC my transcript, looking for goofs and ambiguities. It spotted a metaphor I repeated too often, and gave me several places where a lower-third with text would be helpful.
I then asked it to give me some recommended chapters (it’s not quite so good at this, but it helps).
Then once I had my chapter headings figured out, I had it tighten them and harmonize them, so they look like the same person wrote them (tense, capitalization, tone)…it does that in spades.
I then had it work on titles and description and hashtags.
Once that’s all done and I have locked in my choices, I tell it “give me a youtube.md file with this stuff” and it does, with everything I need to include when I upload it.