Google Is Testing Its Medical AI Chatbot At the Mayo Clinic 38
According to the Wall Street Journal, Google is testing its Med-PaLM 2 AI chat technology at the Mayo Clinic and other hospitals. It's based on the company's PaLM 2 large language model (LLM) that underpins Bard, Google's ChatGPT rival. Engadget reports: Unlike the base model, Med-PaLM-2 has been trained on questions and answer from medical licensing exams, along with a curated set of medical expert demonstrations. That gives it expertise in answering health-related questions, and it can also do labor-intensive tasks like summarizing documents and organizing research data, according to the report. During I/O, Google released (PDF) a paper detailing its work on Med-PaLM2. On the positive side, it demonstrated features like "alignment with medical consensus," reasoning ability, and even the ability to generate answers that were preferred by respondents over physician-generated responses. More negatively, it showed the same accuracy problems we've seen on other Chat AI models.
In an internal email seen by the WSJ, Google said it believes the updated model could "be of tremendous value in countries that have more limited access to doctors." Still, Google has admitted that the technology is still in its early stages. "I don't feel that this kind of technology is yet at a place where I would want it in my family's healthcare journey," said Google senior research director Greg Corrado. However, he added that the tech "takes the places in healthcare where AI can be beneficial and expands them by 10-fold."
In an internal email seen by the WSJ, Google said it believes the updated model could "be of tremendous value in countries that have more limited access to doctors." Still, Google has admitted that the technology is still in its early stages. "I don't feel that this kind of technology is yet at a place where I would want it in my family's healthcare journey," said Google senior research director Greg Corrado. However, he added that the tech "takes the places in healthcare where AI can be beneficial and expands them by 10-fold."
Reposting my old comment again (skip if u read it) (Score:2)
I can provide numerous examples of AI providing totally incorrect or slightly wrong information -- even in critical areas such as medicine. (In this example chatGPT, but I bet it could be applicable to Google's chatbot.) I'm not gonna list all the examples here, just one randomly. Here's the scenario: Let's assume you have a doctor that has grown to rely on chatGPT. Now what if he gets a patient with syphilus and barely sort of recalls that he can treat syphilus with Penicillin G, but he forgets which type
Re: (Score:2)
First, they're not using ChatGPT, this is specially-trained AI using data sources much more reliable than "the internet."
Second, humans also often provide "totally incorrect or slightly wrong" information. My bet is that, with enough iterations of training, AI will do so much less often than humans.
Re: (Score:3)
Re: (Score:2)
Parroting pieces of text that statistically might have come from a medical textbook is a totally shit way of practicing medicine.
And yet this is pretty much how medicine is practiced. How else? I don't see a whole lot of difference between this and your "alternate" method.
Re: (Score:3)
Precisely this. Actually, there is one difference. AI can spot trends that humans haven't actually included in their prescriptive texts. To me, this makes AI a better bet than just textbook knowledge.
Re: (Score:2)
Parroting pieces of text that statistically might have come from a medical textbook is a totally shit way of practicing medicine.
That's your own statement, not backed up by anything in the article or by any source anywhere.
And I don't really see a distinction between your "expert systems" that are hand-coded, and predictive models, that use pattern analysis. Well, there is one difference: AI can find patterns that weren't necessarily spotted by humans.
Whether the doctor consults a textbook or manual, or AI, the doctor needs to know what they are doing. This is true with coding too. I use ChatGPT regularly to help me find ways to code
Re: (Score:3)
this is specially-trained AI using data sources much more reliable than "the internet."
It's based on Google's PaLM 2, which absolutely was trained on massive amounts of data from the internet.
Second, humans also often provide "totally incorrect or slightly wrong" information.
Humans are capable of reviewing their work to identify and correct errors. Language models are not.
My bet is that, with enough iterations of training, AI will do so much less often than humans.
That is a bet you will lose. Over training will degrade model performance. This isn't a problem that can be fixed. The nature of the technology makes this inevitable.
Re: (Score:2)
> It's based on Google's PaLM 2, which absolutely was trained on massive amounts of data from the internet.
PALM2 is a combination of multiple AIs like Gato that has been previously been told to be possible solution to AGI. Major difference to for example ChatGPT is that you can train an individual AI to do a specific task and then attach it to the system. Much like you could give a human a general education, then attach a specific education to inspect X-ray images. So it is not an LLM. It is LLM + other
Re: (Score:2)
that has been previously been told to be possible solution to AGI.
"told"? By whom? What a ridiculous thing to say.
This is why some people already think that AGI is solved
Some people think the earth is flat. The world is full of stupid people.
AGI is science fiction. We are so far from a solution that we don't even know what questions to ask. Anyone making any claims about AGI is either a crackpot or a conman.
Re: (Score:2)
It's based on Google's PaLM 2, which absolutely was trained on massive amounts of data from the internet.
PaLM 2 is a technology, not a training method. The training is separate. And I challenge you to show me a source that says Google's Med PaLM 2 was trained randomly from "the internet."
Humans are capable of reviewing their work to identify and correct errors. Language models are not.
Humans train language models, they don't train themselves.
If you use AI as a tool, you don't get to turn your brain off. I use ChatGPT for coding ideas. But when I get a suggestion, I still have to know whether it does what I want it to do, or if it's a viable solution. Likewise in medicine, the doctor still has to know what t
Re: (Score:2)
PaLM 2 is a technology, not a training method.
I didn't say it was, troll. Why tell such an obvious lie?
I challenge you to show me a source that says Google's Med PaLM 2 was trained randomly from "the internet."
I didn't say "randomly" you lying troll. As for a source than claims they used data from the internet, that's trivial [ai.google] If you don't know anything about the subject, you shouldn't post.
Humans train language models, they don't train themselves.
I never claimed that they did, you lying piece of shit troll.
Don't waste my time with bullshit like this.
Re: (Score:2)
You are conflating PaLM 2 with PaLM 2 Med. Your link does not mention PaLM 2 Med at all. The two are based on the same technology, but different training sets.
Clearly, you have run out of actual arguments, because you have started resorting to insults.
Re: (Score:2)
Quiet troll. All you do is tell obvious lies. Fuck right off.
Re: (Score:2)
Just like I said, you have only insults, no logic.
Re:Reposting my old comment again (skip if u read (Score:4, Insightful)
Many years ago, some classical AI expert systems claimed to outperform doctors in real-world tests. Never gained much traction. Explanations varied.
Personally, I am all for encoding medical knowledge in a machine-readable, open format knowledge base. I might even support requiring every journal article to provide input to it. Computers are designed for retrieving information and applying rules such as formal logic and Bayesian inference to filter and aggregate results. People are notorious for memory or logic lapses, especially under fatigue and other stress.
I can imagine an LLM-frontend might be vastly more doctor friendly than writing formal predicates for input or interpreting raw numbers for output. It could also provide a natural prompting system that recommends tests and warns of rare but serious alternative diagnoses.
Present AI tools can also look at data to suggest correlations that may merit research into improved treatment protocols.
I am strongly opposed to the idea of present-gen AI being used to supplant a formal knowledge base by "somehow" learning "better" than expert humans through unsupervised processing of existing texts and data. Such hubris seems endemic to AI hype.
Re: (Score:2)
I am strongly opposed to the idea of present-gen AI being used to supplant a formal knowledge base
This is like the relationship between Stack Overflow, and ChatGPT. You can't have ChatGPT spit out answers to a lot of technical questions, without it first consulting the Stack Overflow "knowledge base." They work together. ChatGPT doesn't replace the knowledge base, it just digests it.
Re: (Score:1)
Re: asking the wrong question (Score:2)
Asking "What is the difference between benzathine penicillin g and procaine penicillin g?" was the wrong question in this case. It might have been appropriate to ask "what would be the best implementation of penicillin for treating syphilis". If the answer to that was wrong the machine would more clearly be at fault.
And a smart answer would have been to ask for more detail about the specific form of syphilis or other relevant details just like a human would. I expect the machines can do that.
FACE PALM (Score:2)
Future Anomalous & Catastrophic Events (caused by) Physician Assistant Language Model
For the Love of God (Score:3)
Re: For the Love of God (Score:2)
Which field are you working in? I do CS research, and grants are carefully read. You usually get very detailled review from NSF panels.
Re: (Score:2)
I do CS research
Of course you do. Let us know if you find anything.
I would totally want this in my own treatment (Score:3)
I think human doctors are often arrogant or poorly trained. They work shifts that are too long, and are often overly motivated to opt for more treatments, and more expensive treatments, because they get paid for each procedure.
Note that this is not ChatGPT, which was trained on "the internet." This is AI trained on specially curated data sets. I would expect much higher quality than ChatGPT, and with enough iterations and training, I'd literally bet my life that it will have a better track record than human doctors.
Re: (Score:2)
[...]and are often overly motivated to opt for more treatments, and more expensive treatments, because they get paid for each procedure.
Re: (Score:2)
Re: (Score:2)
The American Healthcare System (if it can be called a system at all) is all about profit, and pushing expensive procedures, or even better, expensive forever treatments like continual, perpetual drugs, carefully balanced with other drugs to knock back the side effects, carefully balanced with still more drugs to knock back those side effects, and so on and so on. We even got clockers watching the doctors and nurse practitioners (since the clockers said doctors spent too much time with patients) to make sure
Re: (Score:2)
As Winston Churchill famously said, "Democracy is the worst possible form of government that anyone could imagine, except for all the others."
Healthcare is kind of like that. Yes, American healthcare is expensive. But it's also effective. If you need treatment, you don't have to wait for months, as you do in many countries where the government pays for everything.
It's also not totally one-sided for the "owner class" as you say. Providers still have to work within the bounds of what insurance companies will
Re: (Score:2)
As Winston Churchill famously said, "Democracy is the worst possible form of government that anyone could imagine, except for all the others."
Healthcare is kind of like that. Yes, American healthcare is expensive. But it's also effective. If you need treatment, you don't have to wait for months, as you do in many countries where the government pays for everything.
It's only effective if you can afford it. And that waiting months thing, if you actually research it, doesn't appear to happen in the upper-tier "government pays for everything" countries any more than it does here in America. Especially these last two to three years, where everything seems to take for freakin' ever, if you can dare to pay.
It's also not totally one-sided for the "owner class" as you say. Providers still have to work within the bounds of what insurance companies will pay, and that in turn is limited to what employees and employers are willing to pay. It's not perfect, and it is expensive, and there are losers in the system. But in government-run healthcare systems, there are also losers, particularly those who can't afford to pay for private care if they want it. Money pays for nice things, and no amount of regulation can change that.
Yeah, the insurance companies are part of the owner class in America. They are massively profitable companies in their own right.
And while you can brag about how great it
Re: (Score:2)
Yes, this is true. But in countries where doctors are paid only a salary, they are still motivated by factors not related to the quality of patient care. Even government-sponsored healthcare systems want to control costs. So they might be motivated NOT to perform procedures they really should, in the name of cost savings.
There are certainly problems with the American healthcare system, but Europeans haven't solved all their healthcare problems either.
Re: (Score:1)
Let me suggest that you perhaps investigate the range of different payment models present in medical settings wherever you live. Your comment strongly suggests that you're only aware of the "Fee For Services" payment model.
The array of vectors of force acting on medical interactions between patients and health care providers is bewildering. To use a classic example in the US, there are a ton of factors that go into the high rate of cesarean sections that occur here. To attribute huge, systemic problems to t
Re: (Score:2)
In the government-funded healthcare model, which I presume you live with, there are a different set of problems. Namely, the government wants to control costs, so it is more likely to deny options or treatments that should be considered.
Yes, the American system has problems, but every system has problems.
Nope. (Score:1)
Re: (Score:2)
We have just the thing you need. (Score:2)
Endeavors like this all too often start with "the company" or the technology dudes having a ghee whiz idea divorced from the reality of the primary matter, in this case doctoring and health care. So, of course they are going to talk it up like the greatest thing ever. It's not.
The old adage - If you think you know the answer, you don't understand the problem.
Some things that struck me as funny, or too stupid or obvious to be taken seriously :
Unlike the base model, Med-PaLM-2 has been trained on questions and answer from medical licensing exams, along with a curated set of medical expert demonstrations. That gives it expertise in answering health-related questions
These AI bots are supposed to be trained on large language model
good for idiots (Score:2)
Chatbots are good for answering questions from idiots. Questions like "I have a sore throat, does that mean I have cancer in my foot."
Nontrivial questions will fail to be answered either correctly or at all, but your access to a human who can actually answer it will be severely restricted.