PSA: GPT-4

[SeisK]

---

Members don't see this ad.

Will this "AI scare" lead to radiologists breaking more into the clinical rhelm?

It didn’t lead them to become more clinical during the job market glut of ‘08-‘16. No reason to think a very unlikely threat of a glut would make them do so now

 

[OnlyScans420]

Good points but how about if we assume that AI is much much faster than BI at the interpretation aspect?

Doesn't matter if the radiologist still has to review the images themselves...

 

[stay_foolish]

Visual skills have evolved in animals much longer than logic or language skills. So for human beings learning to drive could be achieved in few days compared to say becoming a Doctor or Lawyer. But for machines., visual data is order of magnitude heavier (3 hour 4K movie) vs 3 hour reading book content. In future, the amount of contribution that Radiologist could bring to his/her job will be a Lot less compared to what they have been doing now. It will be even low for specialist like oncologist, neurology or general physician. For human brain, training to be a Physician takes longer and needs hard work than trained to be a Gardener or Driver. That is why Physicians are paid a lot of money now. Currently or in near future, for Machines to be trained to be a Physician or an attorney, it will be more cost effective. This paradox of value of human learning vs machine learning is hard to digest. AI will get better and better. Just need to re-think this transition and accept the fact that Human Doctors will be less useful in the future than they are now. Fighting this and trying to stay important is dis-service to patients. Think of as an individual when your time has come to retire. You will pass the baton to next generation and just do some advisory role with low contribution. Similarly the whole profession has to think about transition from its current contributing state... Of-course everyone deserves a good living and the Onus is on the society to change current economic doctrines on how to facilitate this transition.

 

[Steve_Zissou]

Visual skills have evolved in animals much longer than logic or language skills. So for human beings learning to drive could be achieved in few days compared to say becoming a Doctor or Lawyer. But for machines., visual data is order of magnitude heavier (3 hour 4K movie) vs 3 hour reading book content. In future, the amount of contribution that Radiologist could bring to his/her job will be a Lot less compared to what they have been doing now. It will be even low for specialist like oncologist, neurology or general physician. For human brain, training to be a Physician takes longer and needs hard work than trained to be a Gardener or Driver. That is why Physicians are paid a lot of money now. Currently or in near future, for Machines to be trained to be a Physician or an attorney, it will be more cost effective. This paradox of value of human learning vs machine learning is hard to digest. AI will get better and better. Just need to re-think this transition and accept the fact that Human Doctors will be less useful in the future than they are now. Fighting this and trying to stay important is dis-service to patients. Think of as an individual when your time has come to retire. You will pass the baton to next generation and just do some advisory role with low contribution. Similarly the whole profession has to think about transition from its current contributing state... Of-course everyone deserves a good living and the Onus is on the society to change current economic doctrines on how to facilitate this transition.

This remains sci-fi for the time being.

 

[OnlyScans420]

Visual skills have evolved in animals much longer than logic or language skills. So for human beings learning to drive could be achieved in few days compared to say becoming a Doctor or Lawyer. But for machines., visual data is order of magnitude heavier (3 hour 4K movie) vs 3 hour reading book content. In future, the amount of contribution that Radiologist could bring to his/her job will be a Lot less compared to what they have been doing now. It will be even low for specialist like oncologist, neurology or general physician. For human brain, training to be a Physician takes longer and needs hard work than trained to be a Gardener or Driver. That is why Physicians are paid a lot of money now. Currently or in near future, for Machines to be trained to be a Physician or an attorney, it will be more cost effective. This paradox of value of human learning vs machine learning is hard to digest. AI will get better and better. Just need to re-think this transition and accept the fact that Human Doctors will be less useful in the future than they are now. Fighting this and trying to stay important is dis-service to patients. Think of as an individual when your time has come to retire. You will pass the baton to next generation and just do some advisory role with low contribution. Similarly the whole profession has to think about transition from its current contributing state... Of-course everyone deserves a good living and the Onus is on the society to change current economic doctrines on how to facilitate this transition.

It's clear you have a profound misunderstanding of how you actually learn and practice medicine/radiology as do most AI doomers. And I will be the first to give in when AI can compare lung nodules. For the love of God do me this one favor ChatGPT. That's all I want for Christmas

 

[stay_foolish]

It's clear you have a profound misunderstanding of how you actually learn and practice medicine/radiology as do most AI doomers. And I will be the first to give in when AI can compare lung nodules. For the love of God do me this one favor ChatGPT. That's all I want for Christmas

It is not about AI doomers. View AI as just a tool to help our lives better.

Please note AI is improving with Technology scaling and algorithm.

This is already an improvement recently when we compare AI 15years ago -> Diagnostic efficiency of artificial intelligence for pulmonary nodules based on CT scans

And also please watch recent Ted talk by Eric Topol ->

It is just about patients getting benefit and any work is just a service to other folks. If AI could compare lung nodules better, it will be Christmas gift for those who are affected. Infact just CNN's could be tuned .. and dont need to rely on ChatGPT for imaging. Work is underway where General Prupose AI like GPT4 could use other narrow specialized AI tools or software tools.

Humility is all we need to make US health care better. If Hospitals and Government regulation only allow human beings for final say on results and still Doctors paid the same, its Ok. At the end all that matters is patients getting better treatment with lower cost.

 

[VivaElBirdos]

It is not about AI doomers. View AI as just a tool to help our lives better.

Please note AI is improving with Technology scaling and algorithm.

This is already an improvement recently when we compare AI 15years ago -> Diagnostic efficiency of artificial intelligence for pulmonary nodules based on CT scans

And also please watch recent Ted talk by Eric Topol ->

It is just about patients getting benefit and any work is just a service to other folks. If AI could compare lung nodules better, it will be Christmas gift for those who are affected. Infact just CNN's could be tuned .. and dont need to rely on ChatGPT for imaging. Work is underway where General Prupose AI like GPT4 could use other narrow specialized AI tools or software tools.

Humility is all we need to make US health care better. If Hospitals and Government regulation only allow human beings for final say on results and still Doctors paid the same, its Ok. At the end all that matters is patients getting better treatment with lower cost.

Your comments read like you have very little familiarity with healthcare in general, let alone radiology

 

[OnlyScans420]

It is not about AI doomers. View AI as just a tool to help our lives better.

Please note AI is improving with Technology scaling and algorithm.

This is already an improvement recently when we compare AI 15years ago -> Diagnostic efficiency of artificial intelligence for pulmonary nodules based on CT scans

And also please watch recent Ted talk by Eric Topol ->

It is just about patients getting benefit and any work is just a service to other folks. If AI could compare lung nodules better, it will be Christmas gift for those who are affected. Infact just CNN's could be tuned .. and dont need to rely on ChatGPT for imaging. Work is underway where General Prupose AI like GPT4 could use other narrow specialized AI tools or software tools.

Humility is all we need to make US health care better. If Hospitals and Government regulation only allow human beings for final say on results and still Doctors paid the same, its Ok. At the end all that matters is patients getting better treatment with lower cos

First stop saying we, it's clear your not a radiologist. Second, from the article: "AI had a significantly higher misdiagnosis rate and a markedly lower true negative rate". Clearly u just googled AI lung nodules and hyperlinked the first article. And the article had nothing to do with COMPARING lung nodules from previous studies nor to my knowledge has there been any significant evidence in actually being able to compare studies using AI, which btw is a HUGE hurtle for useful implementation which people seem to forget about. Finding lung nodules is not the hard part. Comparing them is. Third, TED talks are just BS pop science for normies to feel like they're learning something when they are really just wasting time on their phones. Using one as evidence for anything reinforces the fact you're not a radiologist, probably not even a doctor. Fourth, this isn't about humility this is about an over hyped toy with no real utility that's being pushed by tech doomers who know nothing about radiology but want to act like AI can solve every problem because they cant get venture capitalists to invest in their 15th uber knockoff anymore.

AI is for cheating on college essays and eliminating the need for customer service reps by making it so difficult to get someone to remove the double charge from my credit card bill that I just give up and pay it anyways

 

[stay_foolish]

over hyped toy with no real utility that's being pushed by tech doomers

AI is for cheating on college essays and eliminating the need for customer service reps by making it so difficult to get someone to remove the double charge from my credit card bill that I just give up and pay it anyways

The above two phrases indicates that it will be hard for you to be open in finding optimal solutions. I do agree that I went overboard on saying that future human doctor will be less useful than they are today without knowing what type of problems future could unfold. I apologies for that. There are 1000's of Doctors, Medical students and Premeds really care and focused on the wellbeing of patients and health of the society. I apologies to all of you. I am just frustrated on Per Capita health spending of US compared to most nations in the world. It is not just the cost of Doctors, probably the whole health care system in US. GPT4, Med-Palm 2, Alphafold are no hype. I agree problems needs to be solved in AI, when they solve it will get only better than what it is today. Computing power have increased in PC's like 10,000 times over 20 years and Internet access throughput even in higher order of magnitude. Commercial LLM's are just less than 2 years old. Convolutional Neural network, GAN have started to show promises only in last 5 years. We are in AI bubble like dotcom bubble in 2000's. What happened 10 years after dotcom bubble burst in 2000-2001 ? Billions of subscribers to social media, Millions using online shopping, streaming video.. If I have to go back and live in 1990's as a kid without all these.. it is hard for me to imagine. I think there is no point talking about all these changes. We'll live through this and change ourself accordingly whatever is needed for us to be relevant. The people who could see this early might be in a better position.. thats all.

 

[Neopolymath]

The above two phrases indicates that it will be hard for you to be open in finding optimal solutions. I do agree that I went overboard on saying that future human doctor will be less useful than they are today without knowing what type of problems future could unfold. I apologies for that. There are 1000's of Doctors, Medical students and Premeds really care and focused on the wellbeing of patients and health of the society. I apologies to all of you. I am just frustrated on Per Capita health spending of US compared to most nations in the world. It is not just the cost of Doctors, probably the whole health care system in US. GPT4, Med-Palm 2, Alphafold are no hype. I agree problems needs to be solved in AI, when they solve it will get only better than what it is today. Computing power have increased in PC's like 10,000 times over 20 years and Internet access throughput even in higher order of magnitude. Commercial LLM's are just less than 2 years old. Convolutional Neural network, GAN have started to show promises only in last 5 years. We are in AI bubble like dotcom bubble in 2000's. What happened 10 years after dotcom bubble burst in 2000-2001 ? Billions of subscribers to social media, Millions using online shopping, streaming video.. If I have to go back and live in 1990's as a kid without all these.. it is hard for me to imagine. I think there is no point talking about all these changes. We'll live through this and change ourself accordingly whatever is needed for us to be relevant. The people who could see this early might be in a better position.. thats all.

You have a fundamental lack of understanding of healthcare economics in America. We spend a bunch of money because a bunch of parasitic losers in administration cost the systems greatly. Physicians cost the system something like 10%.

Healthcare is a straight up undercover jobs program in America. You are cheering for AI finding some application in rads when it should be making a bunch of losers unemployed if you really want to change things like you supposedly do.

You are getting this reaction because you have no idea what you are talking about about and then coming into a forum for radiologists to tell them they are too fat and happy. It's shockingly tone deaf.

 

[PhysMatMan]

Why waste time arguing with an AI chatbot troll?

 

[mheat3]

Physicians cost the system something like 10%.

Yes I think I saw 8% for 2023. Physician compensation is not the driver of cost and if you believe otherwise you certainly have no idea at the macroeconomics at play. Completely agree that a lot of healthcare is a jobs program and in many locations healthcare is the #1 employer. Going to be a tough cycle to break and likely won't until there is a catastrophic collapse because you know... stuff 🙂

 

[radrad5]

massive day for AI for those following. sora and gemini 1.5. wow. pretty insane. if this rate of progress keeps up, a lot of white collar work (and absolutely radiology included) could be threatened very, very soon. i stand by my 5 year prediction. buckle up.

 

[SeisK]

massive day for AI for those following. sora and gemini 1.5. wow. pretty insane. if this rate of progress keeps up, a lot of white collar work (and absolutely radiology included) could be threatened very, very soon. i stand by my 5 year prediction. buckle up.

Ok 👍

 

[yungspleen]

massive day for AI for those following. sora and gemini 1.5. wow. pretty insane. if this rate of progress keeps up, a lot of white collar work (and absolutely radiology included) could be threatened very, very soon. i stand by my 5 year prediction. buckle up.

Lol okay...nothing will be different in 5 years. But sure, just throw out random numbers

 

[Steve_Zissou]

Lol okay...nothing will be different in 5 years. But sure, just throw out random numbers

I predict that in 5 years, he will still be saying job loss will happen within 5 years.

 

[SeisK]

I predict that in 5 years, he will still be saying job loss will happen within 5 years.

Twenty years ago this guy would be one of the homeless LA panhandlers with a marker and cardboard sign saying “the end is near.”

 

[LittleFoot]

I predict that in 5 years, he will still be saying job loss will happen within 5 years.

And then after that, it will DEFINITELY be within 5 years

 

[radrad5]

i'm in the same boat as you all. i hope/ i pray i am wildly wrong. but i can also appreciate how mind boggling this technology is... look where video generation was 2 years ago. if this same rate of progress applies to other areas, not a lot of jobs are safe.

it's natural to be defensive, especially when your livelihood is at stake.

 

[SeisK]

i'm in the same boat as you all. i hope/ i pray i am wildly wrong. but i can also appreciate how mind boggling this technology is... look where video generation was 2 years ago. if this same rate of progress applies to other areas, not a lot of jobs are safe.

it's natural to be defensive, especially when your livelihood is at stake.

It’s been a common tactic among AI pushers to demonstrate a new gimmick, such as an ability to generate videos, but without overcoming the limitations of the old gimmicks. The architectural principles behind these CNNs are not new, they’re just applying it to new arenas they hadn’t before, which makes everyone go “Oooo” and garners a lot of press. But the thing that made them suck three years ago make them still suck today.

Edit: I’m predicting it now. In the next 5years we’re going to have music-making AI which is going to create some bomb jams at first glance, but upon closer analysis are going to be slightly off and not quite where we want them to be. Good enough for tiktoks, not good enough to frankly replace musicians. This AI craze is going to be a very expensive demonstration of “real close just isn’t good enough.”

 

[stay_foolish]

i'm in the same boat as you all. i hope/ i pray i am wildly wrong. but i can also appreciate how mind boggling this technology is... look where video generation was 2 years ago. if this same rate of progress applies to other areas, not a lot of jobs are safe.

it's natural to be defensive, especially when your livelihood is at stake.

This problem will not be unique to Physicians. It will impact Large numbers of Software Programmers, Attorneys, Accountants as well. The society has to change by adapting to new socio-economic doctrines. Only thing that worries me is we have categories of people who (1) Exaggerates AI will kill us. (2) Underestimate AI that it is a phony trick and in the delusion that Human Intellect has some special place in the Universe. We need more people who is altruistic and think through that If AI scales up from the current rate, how we are going to transition our society and derive answers from basic question on what is purpose of life, happiness and suffering.

 

[Steve_Zissou]

This problem will not be unique to Physicians. It will impact Large numbers of Software Programmers, Attorneys, Accountants as well. The society has to change by adapting to new socio-economic doctrines. Only thing that worries me is we have categories of people who (1) Exaggerates AI will kill us. (2) Underestimate AI that it is a phony trick and in the delusion that Human Intellect has some special place in the Universe. We need more people who is altruistic and think through that If AI scales up from the current rate, how we are going to transition our society and derive answers from basic question on what is purpose of life, happiness and suffering.

You don’t believe that human intellect is, as far as we know, unique in the universe?

 

[stay_foolish]

You don’t believe that human intellect is, as far as we know, unique in the universe?

Yes, as far as we know. Please note "As far as we know" has inherent limitation, isn't it ? Also couple of points why I have reasons to believe that human intellect could be surpassed. (1) Even though we have amazing intellectual outcomes from folks like Einstein or Spinoza, they are still limited within their field. With too much information, in modern days we can train human brain to only have narrow specialties such as Telecom Engineers or Oncologist. (2) Before homo "Something", probably the intellect of Australopithecus or even before that the intellect of Chimpanzees were probably unique in Planet Earth. Even with 100 times less connectivity than human brain, GPT4 can tackle wide amount of problem from USMLE to AP Calculus. In current form GPT4 is far less productive than Many Human beings, but both biological evolution and the technological evolution gives us all indication, the glorification of human intelligence will be just transient. But what is really important is our Joy, suffering and curiosity to find meaning for life and still plenty of mysteries to unravel in biology and physical sciences.

 

[Steve_Zissou]

Yes, as far as we know. Please note "As far as we know" has inherent limitation, isn't it ? Also couple of points why I have reasons to believe that human intellect could be surpassed. (1) Even though we have amazing intellectual outcomes from folks like Einstein or Spinoza, they are still limited within their field. With too much information, in modern days we can train human brain to only have narrow specialties such as Telecom Engineers or Oncologist. (2) Before homo "Something", probably the intellect of Australopithecus or even before that the intellect of Chimpanzees were probably unique in Planet Earth. Even with 100 times less connectivity than human brain, GPT4 can tackle wide amount of problem from USMLE to AP Calculus. In current form GPT4 is far less productive than Many Human beings, but both biological evolution and the technological evolution gives us all indication, the glorification of human intelligence will be just transient. But what is really important is our Joy, suffering and curiosity to find meaning for life and still plenty of mysteries to unravel in biology and physical sciences.

So you agree that human intellect has a special place in the universe?

You're living in a science fiction world friend. Is it possible? Sure. Is it around the corner? Hell no.

 

[stay_foolish]

. Is it possible? Sure. Is it around the corner?

Technology growth is exponential. First Electronic Calculator ( with Vacuum Tubes ) which weighs around 33 pounds came in 1963. Deep blue beat World Chess Champion in 1997, AlphaGo beat World Go champion in 2016, GPT4 released in 2023. If you look at the gap between each significant development, it is 32, 19, 7... It is possible next significant development will happen in 3 to 4 years. Also this chatGPT hype has prompted investors to jump into this. So lots of funding going on like dot com bubble. There will be a bubble burst, but this bubble also drives innovation faster. Its hard to predict. But since the trajectory, results shows promise that enables several large corporation pouring Billions is something. Its not a bad thing to be prepared and having some contingency plan in place. I feel following results of Med-Palm 2/GPT4, watching talk by Dr Eric Topol MD, Reading the book by Dr Isaac Kohane MD/PhD could help. Instead of putting our head into sand, it is better to observe the development and see how we can adapt.

 

[Steve_Zissou]

Technology growth is exponential. First Electronic Calculator ( with Vacuum Tubes ) which weighs around 33 pounds came in 1963. Deep blue beat World Chess Champion in 1997, AlphaGo beat World Go champion in 2016, GPT4 released in 2023. If you look at the gap between each significant development, it is 32, 19, 7... It is possible next significant development will happen in 3 to 4 years. Also this chatGPT hype has prompted investors to jump into this. So lots of funding going on like dot com bubble. There will be a bubble burst, but this bubble also drives innovation faster. Its hard to predict. But since the trajectory, results shows promise that enables several large corporation pouring Billions is something. Its not a bad thing to be prepared and having some contingency plan in place. I feel following results of Med-Palm 2/GPT4, watching talk by Dr Eric Topol MD, Reading the book by Dr Isaac Kohane MD/PhD could help. Instead of putting our head into sand, it is better to observe the development and see how we can adapt.

I agree that we should look forward and try to anticipate/drive how this changes our field.

As for the rest, I’ll believe it when I see it.

 

[SeisK]

Technology growth is exponential

Until it isn’t. A predictive model for growth is only trusted as far as we expect the premise underlying the model to apply… and we don’t have a premise for the model of exponential growth in this circumstance.

We actually have a counter-premise. Semiconductor chips are hitting a wall with increased computational power per unit volume because our transistors and circuits are atoms large. They can’t get smaller. Theres an upper limit to computational power density we can get to with classical machines, and we’re reaching that limit. Which means our CNNs only become more powerful by dumping more and more expensive resources into them. Moreover, quantum computing, if it were to ever arrive, solves the CNN problem less efficiently than classical machines.

Finally: AI still can’t do math. AI still can’t get the fingers right. Doesn’t matter how we train it. Doesn’t matter how we design the network. And we can’t make the networks smaller, faster, better anymore. We’re just finding new gimmicks to apply them to.

It’s not just AI doomsaying. It’s incorrect AI doomsaying.

 

[stay_foolish]

Until it isn’t. A predictive model for growth is only trusted as far as we expect the premise underlying the model to apply… and we don’t have a premise for the model of exponential growth in this circumstance.

We actually have a counter-premise. Semiconductor chips are hitting a wall with increased computational power per unit volume because our transistors and circuits are atoms large. They can’t get smaller. Theres an upper limit to computational power density we can get to with classical machines, and we’re reaching that limit. Which means our CNNs only become more powerful by dumping more and more expensive resources into them. Moreover, quantum computing, if it were to ever arrive, solves the CNN problem less efficiently than classical machines.

Finally: AI still can’t do math. AI still can’t get the fingers right. Doesn’t matter how we train it. Doesn’t matter how we design the network. And we can’t make the networks smaller, faster, better anymore. We’re just finding new gimmicks to apply them to.

It’s not just AI doomsaying. It’s incorrect AI doomsaying.

I also don't believe Quantum computing will have any practical usage in AI for many years to come. You are right on semiconductor technology node size. But, currently Nvidia's H100 is 4 nanometer. Some pilot production is already on 2 nanometer and 1 nm is in R&D. Also with advanced packaging they could stack one die one over other. Unlike human brain which has to be housed inside a small enclosure, these computing resources could reside in large datacenter rooms. We can increase the connection nodes through daisy chaining. [ Please confirm all I told with your ECE professor as well ]. So with all these additional improvements there is room for scaling up the hardware and connectivity. Yes., GPT4 is limited with its Math and reasoning and it is at its infancy. Recently Google published Alpha-Geometry which solved International Olympiad level math Proofs. But still a Long long way to go for AI to be a Genius in Math Proofs and to solve problem with higher level of abstraction. But I believe Coding, Diagnosis/Prognosis of Physicians, Aligning trial situation to a law code or previous ruling etc., could be achieved on par or better than human level in few years.

 

[SeisK]

Yeah man according to your exponential model, in 50 years we’ll all be sentient gasclouds in space.

Still gonna be reading plainfilms tho 🙄.

The point was always efficiency growth is not exponential. While there is some efficiency to still be squeezed out, there’s growing rhetoric among computational scientists, which I’m sure is reflected in all our anecdotal experiences, that AI really hasn’t been getting much better the past several years.

But I believe Coding, Diagnosis/Prognosis of Physicians, Aligning trial situation to a law code or previous ruling etc., could be achieved on par or better than human level in few years.

I don’t. It’s akin to saying we’ll get faster than light travel soon because look at how fast we went from the first flight to landing on the moon.

Maybe I should reframe: regardless of the feasibility of AI longterm overcoming human abilities in all things without monstrous infrastructure / cost investment; the implementation will be so slow it won’t make a functional difference to career physicians anyway. If only because the tedious process of verifying the safety of these things—assuming they work, which they don’t currently, and there is good reason to believe they won’t soon—takes decades.

Put another way, supercomputers have been able to solve geometry problems for a while. The problem isn’t whether they can solve it, the problem is whether it’s cheaper. I could build a multimillion dollar supercomputer, or I could just, you know, hire a mathematician salaried at $90k/yr. I don’t question whether particular problems can be solved with enough resource investment, I question the financial feasibility of the approach. And despite planned or RD’d supercomputer builds, the best datapoint for future expense estimates are recent expenses, barring an upheaval in how components are designed, and we don’t have such an upheaval on the horizon save QC, which I can personally guarantee you will not aid with the CNN problem.

 

[DoctwoB]

Anytime you project indefinite continuation of exponential growth you are likely to be wrong. There are plenty of investors who have generated a 50% or greater return in a year. Project that for 20 years with any reasonable amount of starting capital and you own the entire US economy. Most fields develop exponential growth until they run into a wall due to technical limitations

Take things that most people care more about the AI, like housing or food or gas or energy. Despite all our technology, building a house costs more now (even adjusted for inflation) then it did 20 years ago. Food and fuel costs are similar. Despite big breakthroughs in making solar/wind cheap and efficient, we still pay more per KWH. At some point you run into the fact that your limiting factors are resources and labor. Compute/AI may run into a similar resource wall, where greater things can be achieved, but only at significant cost due to compute resources required.

 

[stay_foolish]

Yeah man according to your exponential model, in 50 years we’ll all be sentient gasclouds in space.

Still gonna be reading plainfilms tho 🙄.

The point was always efficiency growth is not exponential. While there is some efficiency to still be squeezed out, there’s growing rhetoric among computational scientists, which I’m sure is reflected in all our anecdotal experiences, that AI really hasn’t been getting much better the past several years.



I don’t. It’s akin to saying we’ll get faster than light travel soon because look at how fast we went from the first flight to landing on the moon.

Maybe I should reframe: regardless of the feasibility of AI longterm overcoming human abilities in all things without monstrous infrastructure / cost investment; the implementation will be so slow it won’t make a functional difference to career physicians anyway. If only because the tedious process of verifying the safety of these things—assuming they work, which they don’t currently, and there is good reason to believe they won’t soon—takes decades.

Put another way, supercomputers have been able to solve geometry problems for a while. The problem isn’t whether they can solve it, the problem is whether it’s cheaper. I could build a multimillion dollar supercomputer, or I could just, you know, hire a mathematician salaried at $90k/yr. I don’t question whether particular problems can be solved with enough resource investment, I question the financial feasibility of the approach. And despite planned or RD’d supercomputer builds, the best datapoint for future expense estimates are recent expenses, barring an upheaval in how components are designed, and we don’t have such an upheaval on the horizon save QC, which I can personally guarantee you will not aid with the CNN problem.

The current state of AI ( Med palm-2 and GPT4) already made significant headlines with passing USMLE, Better clinical outcomes etc., The projections based on current roadmap with technology scaling, die stacking, daisy chaining, other architecture and algorithm improvement etc, will get 50-100X improvement in 5-7 years which is pretty substantial improvement with respect to current state. Beyond that Nobody knows what will happen... Could improve or stagnate.

 

[stay_foolish]

Anytime you project indefinite continuation of exponential growth you are likely to be wrong. There are plenty of investors who have generated a 50% or greater return in a year. Project that for 20 years with any reasonable amount of starting capital and you own the entire US economy. Most fields develop exponential growth until they run into a wall due to technical limitations

Take things that most people care more about the AI, like housing or food or gas or energy. Despite all our technology, building a house costs more now (even adjusted for inflation) then it did 20 years ago. Food and fuel costs are similar. Despite big breakthroughs in making solar/wind cheap and efficient, we still pay more per KWH. At some point you run into the fact that your limiting factors are resources and labor. Compute/AI may run into a similar resource wall, where greater things can be achieved, but only at significant cost due to compute resources required.

Agreed. My statements are based on current roadmap which is laid out for next 5 to 7 years. Nobody knows what will happen after 5 years. Based on the scaling that happened in 10 years, the next 5 years projection of even 50-100X is extremely conservative.

 

[yungspleen]

If you're so certain AI is going to destroy radiology why don't you go pick up some extra moonlighting shifts to make some money before the collapse of our field instead of spending that time fear mongering about things that are not going to happen.

 

[SeisK]

If you're so certain AI is going to destroy radiology why don't you go pick up some extra moonlighting shifts to make some money before the collapse of our field instead of spending that time fear mongering about things that are not going to happen.

He’s not in radiology.

 

[stay_foolish]

If you're so certain AI is going to destroy radiology why don't you go pick up some extra moonlighting shifts to make some money before the collapse of our field instead of spending that time fear mongering about things that are not going to happen.

I never mentioned AI is going to destroy radiology. I just mentioned facts on AI growth, its current roadmap and impact of AI on various career path including Physician, Programmers, Lawyers etc., Every career path re-aligns according to the latest tools available and 'supply and demand' for that work at some point. The better constructive comments could be how certain task can be augmented by current AI and how these transition to better AI could be Managed so it is helpful for both patient and Current Physicians to not have oversupply and getting trained/re-trained on what is most appropriate. There are some good comments and questions by @DoctwoB @Cognovi and few more. Its just a debate and in-depth valid points backed with facts could help us all.

 

[SeisK]

I never mentioned AI is going to destroy radiology. I just mentioned facts on AI growth, its current roadmap and impact of AI on various career path including Physician, Programmers, Lawyers etc., Every career path re-aligns according to the latest tools available and 'supply and demand' for that work at some point. The better constructive comments could be how certain task can be augmented by current AI and how these transition to better AI could be Managed so it is helpful for both patient and Current Physicians to not have oversupply and getting trained/re-trained on what is most appropriate. There are some good comments and questions by @DoctwoB @Cognovi and few more. Its just a debate and in-depth valid points backed with facts could help us all.

I think people are probably frustrated with you because you’re coming on a board of healthcare professionals and professional trainees as what I presume is an undergrad STEM major, clearly know very little about the complexities of new technology incorporation into healthcare, know nothing about the process by which new healthcare technologies are vetted prior to entering the clinical space, and know nothing about the IRL implementation of technologies and how they differ from the published idealized use cases (a symptom of the reproducibility crisis).

You’re smart and enthusiastic, and I can appreciate that, but are a bit naive and stubbornly persist when people who definitely know more about the innumerable intangible complexities of a healthcare environment, and are therefore much more qualified than you to determine expected clinical relevance, tell you “slow your roll.” Your response repeatedly has just been “look how fast things are going though in this NEW area” with clear disregard and indifference to the legitimate criticisms others have already levied against your points.

It makes it unpleasant to discuss on this board with you.

 

[OnlyScans420]

I never mentioned AI is going to destroy radiology. I just mentioned facts on AI growth, its current roadmap and impact of AI on various career path including Physician, Programmers, Lawyers etc., Every career path re-aligns according to the latest tools available and 'supply and demand' for that work at some point. The better constructive comments could be how certain task can be augmented by current AI and how these transition to better AI could be Managed so it is helpful for both patient and Current Physicians to not have oversupply and getting trained/re-trained on what is most appropriate. There are some good comments and questions by @DoctwoB @Cognovi and few more. Its just a debate and in-depth valid points backed with facts could help us ald

Dude. You come into a radiology forum with absolutely no background in radiology to talk to people that have done at least 8 years plus of higher education and even more years of specialized training/practice. And then you site Ted Talks and bull**** marketing stunts as your evidence for the AI stuff. Not only do you know nothing about radiology, but you clearly don't know about anything else. That's why people are annoyed with you. I could have a more constructive conversation with my my dead grandma, at least should wouldn't site ****ing Ted talks. You've gotta be trolling at this point

 

[hopefulraddaddypls]

I never mentioned AI is going to destroy radiology. I just mentioned facts on AI growth, its current roadmap and impact of AI on various career path including Physician, Programmers, Lawyers etc., Every career path re-aligns according to the latest tools available and 'supply and demand' for that work at some point. The better constructive comments could be how certain task can be augmented by current AI and how these transition to better AI could be Managed so it is helpful for both patient and Current Physicians to not have oversupply and getting trained/re-trained on what is most appropriate. There are some good comments and questions by @DoctwoB @Cognovi and few more. Its just a debate and in-depth valid points backed with facts could help us all.

Would you recommend a pivot out of radiology, then? I'm an m4 applying this cycle

 

[SeisK]

Would you recommend a pivot out of radiology, then? I'm an m4 applying this cycle

This is why I get mad at those kind of posts. It affects peoples lives. Case in point.

 

[hopefulraddaddypls]

This is why I get mad at those kind of posts. It affects peoples lives. Case in point.

Ive been told by various clinicians they think AI will negatively impact me, doesn't help reading stuff like this. I could always switch to IM during prelim and do cards/GI. I like radiology and would love to be a radiologist RIGHT NOW, but dang the future has me worried

 

[SeisK]

Ive been told by various clinicians they think AI will negatively impact me, doesn't help reading stuff like this. I could always switch to IM during prelim and do cards/GI. I like radiology and would love to be a radiologist RIGHT NOW, but dang the future has me worried

The guy you’re asking for advice is a baby undergrad who doesn’t know what a Fourier Transform is. You are more qualified to answer your question than the guy you’re asking.

 

[Radology]

Ah it’s fun to check on the latest and greatest doom and gloom in this field.

I remember starting residency in 2012. Back then it was outsourcing our work to India that was going to be the end of the field, scaring med students away.

Then that never happened.

Then it was all the sub specialists taking our work for themselves and leaving us with nothing but ICU plain films, scaring med students away

Then that never happened.

Now we’re onto skynet. I actually remember OP, he has been talking about this for the better part of a decade. What have I seen as I approach the midpoint of my career when I was supposed to be replaced? *yawn*. The advances in AI just look like normal advances in technology that have occurred throughout the history of medicine, and amount to “really good CAD but also still kinda crappy CAD” more than anything else. Meanwhile, we are all drowning in work. We are all begging for the supposed “AI can make you do the work of 5 radiologists” that we’ve heard about for a decade—please cybernet, “deep learn” this technology yesterday.

Freak out all you guys want “because this time it’s GENERATIVE”

 

[stay_foolish]

The guy you’re asking for advice is a baby undergrad who doesn’t know what a Fourier Transform is. You are more qualified to answer your question than the guy you’re asking.

I don't need to prove that I understand Fourier Transform ( or how I applied it ) to some anonymous person. But I am really appalled that there is no moderation in this forums for personal insults. If the moderator read most of my messages, it is about facts that I even asked to cross check with experts, apologetic when I was wrong and acknowledged when there was valid point in counter argument. I think we need Doctors who care for society and patients than being insecure and fight hard to stay relevant. People could always retrain if they have open mind.

 

[stay_foolish]

Ah it’s fun to check on the latest and greatest doom and gloom in this field.

I remember starting residency in 2012. Back then it was outsourcing our work to India that was going to be the end of the field, scaring med students away.

Then that never happened.

Then it was all the sub specialists taking our work for themselves and leaving us with nothing but ICU plain films, scaring med students away

Then that never happened.

Now we’re onto skynet. I actually remember OP, he has been talking about this for the better part of a decade. What have I seen as I approach the midpoint of my career when I was supposed to be replaced? *yawn*. The advances in AI just look like normal advances in technology that have occurred throughout the history of medicine, and amount to “really good CAD but also still kinda crappy CAD” more than anything else. Meanwhile, we are all drowning in work. We are all begging for the supposed “AI can make you do the work of 5 radiologists” that we’ve heard about for a decade—please cybernet, “deep learn” this technology yesterday.

Freak out all you guys want “because this time it’s GENERATIVE”

I also sincerely hope this AI scare is yet another "luddite fallacy".

 

[stay_foolish]

Would you recommend a pivot out of radiology, then? I'm an m4 applying this cycle

Imaging is one of the most important aspects of medical field in this technological era. It is very expensive to do CT or MRI for whatever reason ( machine cost, maintenance, training etc., ). Experts could weigh in here. When it is very important and when it is expensive, there is always huge room for improvement and hence to contribute. Only thing I was saying is the type of contribution that future Imaging experts could be different than today's radiologist and Passion triumphs Job market. The best person to make career choice is yourself and not some anonymous person in SDN like me.

 

[radrad5]

I don't need to prove that I understand Fourier Transform ( or how I applied it ) to some anonymous person. But I am really appalled that there is no moderation in this forums for personal insults. If the moderator read most of my messages, it is about facts that I even asked to cross check with experts, apologetic when I was wrong and acknowledged when there was valid point in counter argument. I think we need Doctors who care for society and patients than being insecure and fight hard to stay relevant. People could always retrain if they have open mind.

agreed. don't understand the bashing. your posts are well written, factual, and insightful.

we need that perspective.

the reality - which we all know - is that nobody knows what the next 3-10 years will look like

what we do know is that AI is progressing very, very fast. and any reasonable individual can extrapolate current trends to try to gauge where things may be in the coming years.

folks calling this tech a gimmick is laughable and in for a huge wake up call.

 

[radrad5]

Would you recommend a pivot out of radiology, then? I'm an m4 applying this cycle

reach out to as many experts as you can. reach out on twitter, linkedin etc.

try to get a well rounded, informed opinion from those with a real pulse on current state of affairs and where things are likely headed.

i can only answer for myself -- no, i would not choose radiology if i was a med student. that is how i feel. others obviously disagree. to each their own. but imo spending 6 years training for something when already in heaps of debt is just too big of a risk for me to stomach.

 

[lernaeann]

Ive been told by various clinicians they think AI will negatively impact me, doesn't help reading stuff like this. I could always switch to IM during prelim and do cards/GI. I like radiology and would love to be a radiologist RIGHT NOW, but dang the future has me worried

Eh my department is pretty pro-AI. We have had quite a few grand rounds from big places that are leading the AI research and from using the tools daily, I don’t see anything happening soon. I definitely would not listen to the pre med who keeps posting stuff out of their knowledge base. I see a lot of areas where AI can help us with work list integration, helping us do the awful cancer follow ups, and characterizing things that our eyes are not the best at seeing. Look at the fda cleared algorithms on ACR, I don’t think any of the current known algorithms are trying to replace radiologists.

Can it lower our reimbursements and job market in the future? Sure. No one knows. Keep in mind, midlevels are a bigger threat to physicians than AI at the moment. For many fields, AI+ midlevels will likely be a huge threat, especially for your example of general cardiology. As for GI, as a prelim I was allowed to do colonoscopies and endoscopies almost solo and surgeons can do them too so… I’m sure people can be trained to do them.

If you truly believe the AI hype and are scared, do something surgical. Running to IM and fellowship won’t do that much for you.

 

[hopefulraddaddypls]

reach out to as many experts as you can. reach out on twitter, linkedin etc.

try to get a well rounded, informed opinion from those with a real pulse on current state of affairs and where things are likely headed.

i can only answer for myself -- no, i would not choose radiology if i was a med student. that is how i feel. others obviously disagree. to each their own. but imo spending 6 years training for something when already in heaps of debt is just too big of a risk for me to stomach.

What would you recommend then?

 

[Steve_Zissou]

What would you recommend then?

Become a plumber, AI can’t do that. And no, I don’t mean a vascular surgeon.