Geoffrey Hinton said machine learning would outperform radiologists by now | Hacker News Reader

Geoffrey Hinton said machine learning would outperform radiologists by now

newrepublic.com

52 pointsHard_Space108 comments

There is a tremendous share of medicine specialties facing shortages, and fear of AI is not a relevant trend causing it. Even the link explaining shortages in the above article is pretty clear on that.

I do agree with the article's author's other premise, radiology was one of those fields that a lot of people (me included) have been expecting to be largely automated, or at least the easy parts, as the author mentions, and that the timelines are moving slower than expected. After all, pigeons perform similarly well to radiologists: https://pmc.ncbi.nlm.nih.gov/articles/PMC4651348/ (not really, but it is basically obligatory to post this article in any radiology themed discussion if you have radiology friends).

Knowing medicine, even when the tech does become "good enough", it will take another decade or two before it becomes the main way of doing things.

The reason AI is hyped is because it's easy to get the first 80% or 90% of what you need to be a viable alternative at some task. Extrapolating in a linear fashion, AI will do the last 10-20% in a few months or maybe a couple years. But the low-hanging fruit is easy and fast. It may never be feasible to complete the last few percent. Then it changes from "AI replacement" to "AI assisted". I don't know much about radiology, but I remember before the pandemic one of the big fears was what we'd do with all the unemployed truck drivers.

> Extrapolating in a linear fashion, AI will do the last 10-20% in a few months or maybe a couple years

That's a hefty assumption, especially if you're including accuracy.

> That's a hefty assumption, especially if you're including accuracy.

That's exactly what the comment is saying. People see AI do 80% of a task and assume development speed will follow a linear trend and the last 20% will get done relatively quickly. The reality is the last 20% is hard-to-impossible. Prime example is self-driving vehicles, which have been 80% done and 5 years away for the past 15 years. (It actually looks further than 5 years away now that we know throwing more training data at the problem doesn't fix it.)

We are at 95 percent now for self driving. I regularly use Waymo robot cars in sf. No driver. The gap is now just scaling.

Take what is 100% complete in one city and do it in another city.

Problem was solved… you just missed the boat.

Waymo barely works, with 24/7 monitoring by humans in a "fleet response" center[0], in 4 cities in the world. That's only 95% done if you're counting good enough for government work.

[0] https://waymo.com/blog/2024/05/fleet-response/

The monitoring might be 24/7 but its reaction time is nothing usable in a life-and-death situation. Or I just cannot imagine a human being notified "I think I'm crashing into something" and able to take over and do anything of significance within that second to avoid the crash (except hitting on the brakes which the car could do just as well). So don't read too much into the response team, it has definitely its use but won't save you from plunging into that sinkhole who just appeared.

OP is being facetious, the "last 20%" is a common saying implying that you've back-loaded the hard part of a task.

That's their point, I think; since the 50s or so, people have been making this mistake about AI and AI-adjacent things, and it never really plays out. That last '10%' often proves to be _impossible_, or at best very difficult; you could argue that OCR has managed it, finally, at least for simple cases, but it took about 40 years, say.

> Extrapolating in a linear fashion, AI will do the last 10-20% in a few months or maybe a couple years

Extrapolating in a linear fashion, in a few years my child will be ten foot tall, weight six hundred pounds, and speak 17 languages.

The first 90% is the easy part. It's the other 90% that's hard. People forget that, especially people who don't work in software/technology.

> The reason AI is hyped is because it's easy to get the first 80% or 90% of what you need to be a viable alternative at some task.

No, it's because if the promise of certain technologies is reached, it'd be a huge deal. And of course, that promise has been reached for many technologies, and it's indeed been a huge deal. Sometimes less than people imagine, but often more than the naysayers who think it won't have any impact at all.

I wouldn’t dismiss the premise so quickly. Other factors certainly play a role, but I imagine that after 2016, anyone considering a career in radiology would have automation as a prominent concern.

Automation may be a concern. Not because of Hinton, though. There is only so much time in the day. You don't become a leading expert in AI like Hinton has without tuning out the rest of the world, which means a random Average Joe is apt to be in a better position to predict when automation is capable of radiology tasks than Hinton. If an expert in radiology was/is saying it, then perhaps it is worth a listen. But Hinton is just about the last person you are going to listen to on this matter.

> There is a tremendous share of medicine specialties facing shortages

The supply of doctors is artificially strapped by the doctor cartel/mafia. There are plenty who want to enter but are prevented by artificial limits in the training.

Medical professionals are highly paid, thus an education in medicine is proportionally expensive. An education in medicine is expensive, thus younger medical professionals need to be highly paid in order to afford their debt. Until the vicious cycle is here is broken (e.g. less accessible student loans? and more easily defaultable is one way to spell less accessible), things are not going to improve. And there’s also the problem that you want your doctors to be highly paid, because it’s a stressful, high-responsibility job with stupidly difficult education.

US doctors are ridiculously overpaid compared to the rest of the developed world, such as the UK or western EU. There's no evidence that this translates to better care at all. It's all due to their regulatory capture. One possible outcome is that healthcare costs continue to balloon and eventually it pops and the mafia gets disbanded and more immigrant doctors will be allowed to practice, driving prices to saner levels.

How doctors are licensed in the US compared to western Europe might explain why health care costs are higher in the US, but it does not explain why health care costs are rising so much. That's because health care costs are rising at similar rates in western Europe (and most of the rest of the first world).

For example from 2000 to 2018 here's the ratio of per capita health care costs in 2018 to the costs in 2000 for several countries:

2.1 Germany 1.8 France 2.0 Canada 1.7 Italy 2.6 Japan 2.6 UK 2.3 US

Here's cost ratios over several decades compared to 1970 costs for the US, the UK, and France:

     1980 1990 2000 2010 2020
  US  3.2  8.2 13.9 24.1 36.3
  UK  3.1  6.3 15.3 27.8 40.5
  FR  3.4  7.6 14.9 21.1 28.5

Here's the same data showing the the cost ratio decade to decade instead of from 1970:

     1980 1990 2000 2010 2020
  US  3.2  2.6  1.7  1.7  1.5
  UK  3.1  2.0  2.4  1.8  1.5
  FR  3.4  2.2  2.0  1.4  1.4

My data source was https://data.oecd.org/healthres/health-spending.htm but it looks like data.oecd.org reorganized their site so that redirects to https://www.oecd.org/en/data/indicators/health-spending.html which seems to have the data but with a much more limited interface.

Doctor salary is not the only or perhaps even the main factor in healthcare expensiveness, but taking on the overall cost disease in healthcare would broaden the scope too wide for this thread, I think.

Also, I admit that the balloon may in fact never pop, since one theory says that healthcare costs so much simply because it can. It just expands until it costs as much as possible but not more. I'm leaning towards accepting Robin Hanson's signaling-based logic to explain it.

The supply of doctors would be much greater if incompetent people were allowed into the training pathway, for sure.

Bad news...they're already there!

Yep, this is precisely what they argue. They don't simply say they want to keep their high salary and status due to undersupply. They argue that it's all about standards, patient safety etc. In the US, even doctors trained in Western Europe are kept out or strangled with extreme bureaucratic requirements. Of course, again the purported argument is patient safety. As if doctors in Europe were less competent. Health outcome data for sure doesn't indicate that, but smoke and mirrors remain effective.

> even when the tech does become "good enough", it will take another decade or two before it becomes the main way of doing things.

What you're advocating for would be a crime against humanity.

Every four years, the medical industry kills a million Americans via preventable medical errors, roughly one third of which are misdiagnoses that were obvious in hindsight.

If we get to a point at which models are better diagnosticians than humans, even by a small margin, then delaying implementation by even one day will constitute wilful homicide. EVERY SINGLE PERSON standing in the way of implementation will have blood on their hands. From the FDA, to HHS, to the hospital administrators, to the physicians (however such a delay would play out) - every single one of them will be complicit in first-degree murder.

That is because Mr Hinton is full of shit. He constantly overstates the progress to bolster his position on associated risks. And now someone gave him a bloody Nobel so he'll be even more intolerable.

What is behind the curtain is becoming obvious and while there are some gains in some specific areas, the ability for this technology as it stands today to change society is mostly limited to pretending to be a solution for usual shitty human behaviour of cost cutting and reducing workforce. For example IBM laying off people under the guise of AI when actually it was a standard cost cutting exercise with some marketing smeared over the top while the management told the remaining people to pick up the slack. A McKinsey special! And generating content which is to be consumed by people who can't tell the difference between humans and muck.

A fine example is from the mathematics side we are constantly promised huge gains in LLMs but we can't replace a flunked undergrad with anything yet. And this is because it is the wrong tool for the fucking job. Which is the problem in one simple statement: it's mostly the wrong tool for most things.

Still I enjoyed the investment ride! I could model that one with my brain fine.

I think my favorite thing is how people are out there reconsidering their programming careers and doing shit like quitting when you have a bunch of less than quality developers lazily deploying code they pulled out of the ass of an LLM which does no bounds checking, no input validation, none of that basic as can be security shit at all, they just end up deploying all that code to production systems facing the internet and then nobody notices because either the competent developers left for greener pastures or got laid off. I keep on telling people this is going to take like 5 or 10 years to untangle considering how hard it is in a lot of cases to fix code put into tons of stuff like embedded devices mounted up on a light pole or whatever that was considered good enough for a release, but I bet it'll be even more of a pain than that. These are after all security holes in devices that in many cases peoples day or life going well depends on.

What it has done so far for me is put copywriters out of a job. I still find it mostly useful for writing drivel that gets used as page filler or product descriptions for my ecommerce side jobs. Lately the image recognition capabilities lets me generate the kind of stuff I'd never write, which is instagram posts with tons of emojis and other things I'd never have the gumption to do myself but increases engagement. I actually used to use markov chain generator for this going back to 2016 though so the big difference here is that at least it can form more coherent sentences.

Oh I'm right in the middle of your first point. I'm going to have on my grave stone: "Never underestimate the reputation and financial damage a competent but lazy software developer can do with an incompetent LLM"

So what you're saying is that it's a good bullshit generator for marketing. That is fair :)

It's really excellent for that. I can give it an image of a product I'm trying to sell, and then say 'describe this thing but use a ton of marketing fluff to increase engagement and conversions' and it does exactly that.

There's tools out there to automatically do this with A/B testing and I think even stuff like shopify plugins, but I still do it relatively manually.

I'll use it for code sparingly for stuff like data transformations (Take this ridiculously flat, two table database schema and normalize it to the third normal form) but for straight up generating code to be executed I'm way more careful to make sure it isn't giving me or the people who get to use that code garbage. It isn't that bad for basic security audits actually and I suggest any devs reading this re-run their code through it with security related prompting to see if they missed anything obvious. The huge problem is that at least half of devs with deadlines to meet are apparently not doing this, and I get to see it in the drop in quality of pull requests within the past 5 years.

I find it a distraction on validation tasks. I prefer to know what the hell I am doing first and be able to qualify that as I work with test cases.

With respect to other engineers I haven't found any of our ML tools actually pick up anything useful yet from a code review.

On engagement, I saw a negative recently where someone started replacing banner images for events with AI generated content and their conversion to tickets plummeted. YMMV. People can spot it a mile off and they are getting better at it. I think the end game is soon.

Were they trying to generate images for marketing creatives like that? People can notice what that looks like right away and it's repulsive for sure unless they can't tell. Text on the other hand is another story

> On engagement, I saw a negative recently where someone started replacing banner images for events with AI generated content and their conversion to tickets plummeted. YMMV. People can spot it a mile off and they are getting better at it.

Indeed.

I like this stuff, I have fun with it, I have no problem with others making and sharing pics they like just as I myself do… but when I notice it in the wild on packaging or advertising, it always makes me wonder: what other corners are getting cut?

(20 years ago, similar deal but with product art that had 3mm pixels and noticeable jpeg artefacts)

> I think the end game is soon.

Dunno. The models are getting better as well as the consumers noticing more — I've not seen Cronenberg fingers in a while now — and I don't know who will max out first, or when.

I hear existential despair. Mathematics is indeed a great example. Automation (AI) has made great strides in automating theorem proving in the last 30 years, and LLMs are just a cherry on top of that. A cherry though that will accelerate progress even further, by bringing the attention of people like Terence Tao to the cause. It will not change how mathematics is done within 2 years. It will totally revolutionise how mathematics is done within 20 years (and that is a conservative guess).

No it's a terrible example. We have ITPs and ATPs and any progress on ATPs, which are the ones that are currently considered magical, are turning into a fine asymptote. ITPs are useful however but have nothing to do with ML at all.

Putting a confident timescale on this stuff is like putting a timescale on UFT 50 years ago. A lie.

Oh and lets not forget that we need to start with a conjecture first and we can't get any machine to come up with anything even remotely new there.

There is no existential despair. At all.

I said 30 years, because that's for how long I am a participant in the field. I have a PhD in ITP, and yes, I am pretty confident that automation is a big part of ITP now. Maybe not so much if you are into constructive type theory, that field is always a little bit hesitant to use the full power of anything.

The only one lying here is you. To yourself.

By the way, AI can come up with conjectures just fine. I might not be interested in most of them, but then again, I am not interested in most conjectures humans come up with.

ChatGPT and others can’t reliably sum several integers together still, when is an LLM going to replace anything meaningful in higher math?

I spent more time deciphering ChatGPT’s mistakes than programming something in python to sum some integers this week.

I don't use ChatGPT/Claude/Gemini for coding stuff I know how to do. I use it to explore stuff I don't know. It is very useful for that. And that is how LLMs and their descendants will conquer mathematics. LLMs are for intuition, not for logic. We know how to do logic. We didn't know how to do intuition.

I wonder what experts think about this specialized model:

"Harrison.rad.1 excels in the same radiology exams taken by human radiologists, as well as in benchmarks against other foundational models.

The Fellowship of the Royal College of Radiologists (FRCR) 2B Rapids exam is considered one of the leading and toughest certifications for radiologists. Only 40-59% of human radiologists pass on their first attempt. Radiologists who re-attempt the exam within a year of passing score an average of 50.88 out of 60 (84.8%).

Harrison.rad.1 scored 51.4 out of 60 (85.67%)."

https://harrison.ai/harrison-rad-1/

Hinton doesn't take social/power aspects into account, he's a naive technical person. Healthcare is an extremely strongly regulated field and its professionals are among those with the highest status in society. Radiology won't be among the first applications to be automated. I see many naive software/ML people trying to enter the medical domain, but they are usually eaten for breakfast by the incumbents. This is not an open market like taxis vs Uber or one entertainment app vs another. If you try to enter the medical domain, you will be deep deep into politics and lobbying and you're going to be defeated.

Better deploy it in fields like agriculture and manufacturing that are already in good part automated. Then later in another generation one can try the medical domain.

>This is not an open market like taxis vs Uber

Interesting choice of example. In most places taxis are/weren't an "open market" either but had various laws preventing newcomers from providing taxi rides. In many places what Uber did was illegal. Various locales either banned Uber, changed the laws to allow them, or simply didn't enforce the laws that made them illegal.

Yes, even Uber was pushed out in many places with lawfare and jobs reasons. Taxi drivers do have some moderately strong lobby in many places. So just think about taking on doctors... The discussion was never really about which service is better. It was about whose toes they were stepping on.

In healthcare you can't take the "asking for forgiveness is easier than permission" route. That's a quick way to jail. In taxiing, it's just some fines that fit the budget.

My overall point isn't to argue whether these disruptions are socially beneficial. I'm trying to point at the more neutral observation that who you're taking on and how much power they have will be crucial to the success of a disruptor. It's not just about "product quality" based on some objective dispassionate politics-free scientific benchmarks.

In context, I think GP meant "open" from a sociopolitical point of view. The "taxi-industrial complex" had minimal political power, so Uber was repeatedly able to steamroll the incumbents - exactly as you describe in your final sentence. Medical interests have immense sociopolitical power, so the same playbook won't work. The differences in your positions are merely semantic.

Medallion holders had a weaker more fragmented system. If there was a nation wide system, they would have been able to organize better and kill Uber. Considering groups like the AMA literally dictate how many people are allowed to graduate from med school in a given year, the obviously have more control.

Furthermore, on the state level there are often "Need Based Laws" that require healthcare providers to obtain approval before establishing or expanding certain healthcare facilities or services. This is unironically designed to prevent unnecessary duplication of healthcare services, control healthcare costs and ensure equitable distribution of healthcare resources. Imagine how expensive it would be if we have two many choices!

The cab cartel was more or less city level.

Add in the fact that "health care" is much more politically charged and its easy to find useful idiots that want to "protect people's health" that enforcement is a lot easier, they're entirely different.

Not all conspiracy theories are false.

Good thing most people don't know about this, or they might get really angry at the AMA, perhaps even voting for representatives who might enact laws to reform the system.

Fortunately that would never happen.

> Radiology won't be among the first applications to be automated

Radiology and a lot of other clinical labwork is heavily outsourced already, and has been for decades [0][1].

Much of the analysis and diagnosis is already done by doctors in India and Pakistan before being returned to the hospitals in the West.

In fact, this is how Apollo Group and Fortis Health (two of India and Asia's largest health groups) expanded rapidly in the 2000s.

It's the back office teleradiology firms that are the target customers for Radiology Agents, and in some cases are funding or acquiring startups in the space already.

This has been an ongoing saga for over a decade now.

[0] - https://www.jacr.org/article/S1546-1440(04)00466-1/fulltext

[1] - https://www.reuters.com/article/business/healthcare-pharmace...

Very counter-intuitive, yet after reading the sources definitely factual. Thanks for pointing that out.

Yeah, that's cool.

But it's not automation.

GP's point is that it will be really hard to take the specialist out of this process, mainly because of regulatory issues.

The point is, if you can substitute even 30-50% of the headcount used in initial diagnostics and analysis, you're profit margins grow exponentially.

The customers for these kinds of Radiology Agents are the teleradiology and clinical labwork companies, as well as MSOs and Health Groups looking to cut outside spend by bringing some subset back in-house.

100% automation is unrealistic for a generation, but 30-50% of teleradiology and clinical labwork headcount being replaced by CV applications is absolutely realistic.

A lot of players that entered the telemedicine segment during the COVID pandemic and lockdown have begun the pivot into this space, as well as the new-gen MSOs as well as legacy telemedicine organizations (eg. Apollo, Fortis).

(Also, why are you getting downvoted? You brought up a valid point of contention)

That's true.

You may always need to keep one or two guys in there but they will be able to do 20x more with the same effort. Definitely plausible.

Outsourcing is indeed a force pushing against doctor power. This just means that radiologists are already under attack and may feel cornered, so they are on the defensive.

I'm not saying automation won't ever happen. But it will need to be slow, as to allow the elite doctor dynasties to recalibrate which specialties to send their kids to, and not to disrupt the already practicing ones. So Hinton's timeline was overly optimistic due to thinking only in terms of the tech. It will happen, but on a longer timescale, maybe a generation or so.

> But it will need to be slow, as to allow the elite doctor dynasties to recalibrate which specialties to send their kids to, and not to disrupt the already practicing ones

That's not how much of medicine from a business standpoint is run anymore.

Managed Service Organizations (MSOs) and PE consolidation has become the norm for much of the medical industry now because the running a medical practice AND practicing medicine at the same time is hard.

Managing insurance billing, patient records, regulatory paperwork, payroll, etc is an additional 20-30 hours of work on top of practicing as a doctor (which is around 30-50 hours as well).

Due to this, single practitioner clinics or partnership models get sold off and the doctor themselves gets a payout plus gets treated as an employee.

> It will happen, but on a longer timescale, maybe a generation or so

I agree with you that a lot of AI/ML application timelines are overhyped, but in Radiology specifically the transition has already started happening.

The outsourced imaging model has been the norm for almost 30 years now, and most of the players in this space began funding or acquiring startups in this space a decade ago already.

Is 100% automation in the next 5 to 10 years realistic? Absolutely not!

Is 30-50% automation realistic? I'd say so.

Perhaps in the US where medicine and healthcare are hugely hugely expensive with loads of vested interests and profiteering and fingers-in-pies from every side of the table, sure.

But everywhere else in the developed world that has universal and free healthcare, I can imagine a lot of traction for AI from governments looking to make their health service more efficient and therefore cheaper (plus better for the patient too in terms of shorter wait times etc).

DeepMind has been doing a load of work with the NHS, Royal Free Foundation, Moorfields for instance (although there have been some data protection issues there, but I suspect they are surmountable)

No one need/should wait for the US govt or the structurally exploitative private sector to reduce prices.

The US has concocted huge myths about why prices cannot fall no matter what tech or productivity gains happen. Its becomes like religious dogma.

Ppl are already flying overseas in the tens of thousands for medical tourism. All Mexico has to do is set up hospitals along the border so people arent flying elsewhere. This is what will happen in the US. The US has no hope of changing by itself.

I agree. These developments will be first deployed in countries like India, Mexico, China. If you're an entrepreneur aiming at making a difference through medical AI, it's probably best to focus on places where they honestly want and need efficiency, instead of just talking about it.

> Perhaps in the US

Medical Systems in countries like the UK [0], Australia [1], and Germany [2] all leverage teleradiology outsourcing at scale, and have done this for decades

[0] - https://www.england.nhs.uk/wp-content/uploads/2021/04/B0030-...

[1] - https://onlinelibrary.wiley.com/doi/abs/10.1111/1754-9485.13...

[2] - https://www.healthcarebusinessinternational.com/german-hospi...

I wish there was less focus on AI replacing humans and more focus on AI helping humans do better i.e. Intelligence Augmentation. IA drove a lot of the rapid progress in the earlier days of computing that made computers usable by non-experts.

I suspect a lot of the thinking of replacing humans is driven a lot by the nerd scifi facination with the Singularity, but if a lot of the hype fails to materialize after billions of dollars have been poured into AI there could be an over-correction by the market that would take away funding from even useful AI research.

I hope we'll hear less of AI replacing X, and more of AI enhancing X, where X is radiologists, programmers, artists etc.

People becoming slightly more productive helps people. People being replaced helps quarterly earnings reports.

People becoming slightly more productive also helps earnings reports.

But automation is all about reducing labor costs. Even shoddy automation such as self checkout.

I don't see this changing any time soon.

Replacing is better than helping, though. It frees people up to develop capital – the very source of progress you speak of! More capital development, more progress.

If you still have to do the same job, but better, then you aren't be freed up for a higher level purpose. If we had taken your stance in earlier days of human history we'd all still be out standing in the field. Doing it slightly better perhaps, but still there, "wasting" time.

No, we are clearly better off that almost every job from the past was replaced. And we will be better off when our current jobs are replaced so that our time is freed up so to move onto the next big thing.

You don't think we have reached the pinnacle of human achievement already, surely?

Capital is a tool for progress not the source.

Newton, Einstein etc. didn’t do neither for nor by the capital.

And we all know that those “freed” people are less likely to gain any capital.

Some schools would consider their ideas capital. But, regardless, if what they did wasn't useful in the development of capital, nobody would have cared. It is their contributions to capital development that have made them household names.

I doubt that the contribution to capital development made Newton and Einstein household names.

When the giant magnetoresistance was discovered it wasn’t useful at first. if all is capital focused we likely miss those discoveries.

Well, let's put it to the test. What "useless" scientific discovery has given someone a household name in recent years?

What useful discovery has given someone a household name in recent years?

Household names are more likely celebrities than scientists.

Hinton has become a household name, at least here in Canada. But only within recent years as his discoveries have now proven useful in capital creation. Nobody knew who he was in the 1980s when his discoveries were "useless".

Which stands to reason. People only think about you if you do something for them. Newton and Einstein enabled the creation of capital that changed the world, so they are remembered. The guy who discovered something amazing beside them, but that which we still haven't figured out how to use, remains unknown.

Maybe we should define household first.

Are we talking about HN households or average Joe‘s

Average Canadian Joe, at least.

I don't expect anyone in Zimbabwe knows who he is, just as I cannot name a single person in Zimbabwe despite being sure they too have household names. His capital enablements, while significant, haven't become world changing.

Is it because he won the Nobel Prize or more of a longer lasting fame?

But I doubt people could name a winner after he’s out of the news.

Maybe Canada is different.

Or maybe over-confident arrogant CS majors aren't experts on every other irrelevant subdomain and it's not some absurd absence of meritocracy/bad gatekeepers from preventing them from ruining other fields by letting them rely on everything?

Yes, but this mostly plays out in ignoring social factors, like who can take on the blame, etc. CS people also underestimate the extreme levels of ego and demigod-like self-image of doctors. They passed a rite of passage that was extremely exhausting and tiring and perhaps even humiliating during residency, so afterwards they are not ready to give up their hard earned self image. Doctor mentality remains the same as in Semmelweis' time.

Good faith is difficult to assume. I do agree that the real world is much more complex than simply which tool works better.

An analogy could be jury-based courts in the US. The reason for having juries and many of the rules are not really evidence-based. It's well known that juries can be biased in endless ways and are very easy to manipulate. Their main purpose though is not to make objectively correct decisions. The purpose is giving legitimacy to a consensus. Similarly, giving diagnosis-power to human doctors is not a question of accuracy. It's a question of acceptability/blame/status.

Yes, it's the CS majors who are murdering a quarter million Americans annually due to preventable medical errors (including misdiagnoses based on radiology readings), good point.

Nobody, and I mean nobody beats physicians at the Arrogance game.

The very second that ML models begin to consistently beat humans at diagnostics, it would be a moral imperative, as well as an absolute requirement by the terms of the Hippocratic Oath, to replace the humans.

> The very second that ML models begin to consistently beat humans at diagnostics, it would be a moral imperative, as well as an absolute requirement by the terms of the Hippocratic Oath, to replace the humans.

Why is HN so brainrot shit at times man wtf. And you're just proving that you're exactly the type of arrogant techbro I was talking about.

The point is he is wrong.

AI isn’t better than humans but we still have now less radiologists.

Now imagine it were a free market:

To raise the profits people get replaced by AI but AI underperforms.

The situation would be much worse.

What? Radiology is already full of ML and companies offering quick reporting alongside a traditional slow human report. The UKs Royal College of Radiologists has guidance [0] on how to integrate AI and many hospitals use it alongside other outsourced reporting tools.

This works in radiology because it's not surgery. A hallucination doesn't chop off a limb. Past data protection in training, the risk is low, if done right.

AI offers speedy screening, first-pass results or an instant second opinion. You don't have to —and shouldn't yet— rely on it as your only source of truth, but the ship has sailed on hospitals integrating it into their workflow. It's good enough now, and will get better.

0: https://www.rcr.ac.uk/our-services/all-our-publications/clin...

My cousin is an established radiologist. Since I am well-versed in AI, we talked about this at length.

He says that the automated predictions help free up radiologists to focus on the edge cases and more challenging disease types.

It looks like there are 3 possible outcomes: AI will do absolutely nothing, AI will enhance the productivity of radiologists, or AI will render their skills completely obsolete.

In case 1, learning radiology is a fine idea. In case 2, it becomes a little tricky, I guess if you are one of the most skilled radiologists you’ll do quite well for yourself (do the work of 10 and I bet you can take the pay of 2). For case 3, it becomes a bad career choice.

Although, I dunno, it seems like an odd thing to complain about. I mean, the nature of humans is that we make tools—we’re going to automate if it is possible. Rather I think the problem is that we’ve decided that if somebody makes the wrong bet about how their field will go, they should live in misery and deprivation.

Scenario 2 is the most likely by far. It's just the continuation of a trend in radiology. Clinics already employ many more "technicians" with basic training and have a "proper" radiologist double checking and signing their work. If you're a technician (or a radiologist NOT supervising technicians), you're in hot water.

Scenario 2 has already arrived - it's just not fully rolled out.

I predict most AI developers will be out of a job soon.

Say you are making a robot that puts nuts and bolts together using DL models. In a few years, Google/OpenAI will have solved this and many other physical tasks, and any client that needs nuts and bolts put together will just buy a GenericRobot that can do that.

Same for radiology-based diagnostics. Soon the mega companies will have bought enormous amounts of data and they will easily put your small radiology-AI company out of business.

Making a tool to do X on a computer? Soon, there will be LLM-based tools that can do __anything__ on a computer and they will just interface with your screen and keyboard directly.

Etc. etc.

Just about anyone working in AI right now is building a castle in someone else's kingdom.

I predict that the current wave of AI will be useful but not nearly as useful as you hope, requiring ever more people to pour their time and energy into the overarching system it creates.

> Just about anyone working in AI right now is building a castle in someone else's kingdom.

Things are innovating so fast that even if you've convinced yourself that you own the kingdom, it seems like you're one research paper or Nvidia iteration away from having to start over.

This is not putting people out of a job, this is changing their jobs. As always, workforce will adapt. Just like all previous tech, this is trying to improve efficiency. LLM can't do "anything" on a computer that's a misunderstanding of the tech but they can already interface with screen/keyboard.

You overestimate the innovations abilities of big tech companies. They'll buy a lot of companies designing robots that are putting nuts and bolts together, or diagnostics, that's a great opportunity.

> As always, workforce will adapt

I was never a fan of the term "Singularity" for the AI thing. When mathematical singularities pop up in physics, it's usually a sign the physics is missing something.

Instead, I like to think of the AI "event horizon", the point in the future — always ahead, yet getting ever closer — beyond which you can no longer predict what happens next.

Obviously that will depend on how much attention you pay to the developments in the field (and I've seen software developers surprised by Google Translate having an AR mode a decade after the tech was first demonstrated), but there is an upper limit even for people who obsessively consume all public information about a topic: if you're reading about it when you go to sleep, will you be surprised when you wake up by all the overnight developments?

When that happens, no workforce can possibly adapt.

How close is that? Dunno, but what I can say is that I cannot tell you what to expect 2 months from now, despite following developments in this field as best I can from the outside, and occasionally implementing one or other of the recent-ish AI models myself.

You don't ask people to know everything about last tech, they just need to do their job. If a new cool tool benefit is high enough, everybody will know about it and use it. Based on what you said, one could infer Translate AR mode is a nice cool but not something that'll meaningfully change the life of the average joe.

Humanity had been through multiple "event horizon" or whatever you want to call it. With industrialisation, one man with some steam could suddenly do the work of thousands, and it expanded to each and every activity. Multiple "event horizon" later (transport, communication, IT), we are still there with everyone working, differently than previous generations, but still working. We improved a bit reducing child labor but doubled the workforce by including women and we always need more people at work for longer time.

The only constant thing is we keep people at work until we figure out how to burn the last remaining atom. People are far too optimistic believing a new tech will move people away from work.

> Based on what you said, one could infer Translate AR mode is a nice cool but not something that'll meaningfully change the life of the average joe.

In every case I showed it to people, they were amazed and delighted because it absolutely did matter to them.

This may be a biased sample because of how many of my colleagues (like myself) didn't speak German very well despite being in Berlin, but it was definitely the case that they benefited from me showing them a thing invented a decade earlier.

> Humanity had been through multiple "event horizon" or whatever you want to call it.

No: an event horizon is always ahead of you. That's why I'm choosing that name and rejecting the other.

You can wake up to the news the world has changed unpredictably, or the stuff you thought was scifi turned out to be possible, but it doesn't happen every day, and even less so with your* job in particular.

* generic "you"

More common is e.g.: radio was invented, and almost immediately people started predicting that by 2000 there would be video conference calling — fiction writers can see possible futures coming well before it's possible. Plus some futures that aren't possible, of course.

Radioactivity gets discovered, Victorian fiction talks about it being weaponised. Incorrectly as it happens, but it was foreseeable.

When I was a kid, phone numbers were 6 digits, and "long distance" meant the next town over. It getting better and cheaper was predictable and predicted, as were personal videophones. Even the original iPhone wasn't much of a step change from everything around it, and no single update to the iPhone was either.

Sputnik? Led to Star Trek, which even in TOS had characters using tablet computers that wouldn't be realised for 40-50 years.

AI? The fiction is everything and anything from I Have No Mouth, and I Must Scream to The Culture, via Her, The Terminator, Colossus: The Forbin Project, Short Circuit, and Isaac Asimov, and the attempts to do "real" predictions are even more diverse and include precursors for each scenario and what to expect on the way.

Nobody knows how close any of the speculation is, or how soon it is assuming it's right, or even when specific things such as full self drive is coming (despite being in 80s TV).

We don't know how long it will take between FSD and humanoid robots passing same legal/safety standard (despite everyone already being familiar with the ideas since The Jetsons) or what having them would mean in practice.

We don't even know what we're getting from the AI labs as soon as the US election is over. From what I've heard, neither do they.

I don't know why you're downvoted, you are correct. People just don't like their bubbles being popped.

I do software, have done it for about 20 years. My job has been perpetually "on the verge of becoming obsolete in the next few years", lol. Also, "yeah but as you get older people will not want to hire you" and all that.

Cool stories, *yawn*.

Meanwhile, in reality, my bank account disagrees, and more and more work has been increasingly available for me.

> I predict most AI developers will be out of a job soon.

Once AI developers will be out of a job everyone will be out of every job. If not then the AI developers still have a job to work on those jobs which hasn't been automated yet.

> Just about anyone working in AI right now is building a castle in someone else's kingdom.

This doesn't ring true to me. When do you feel we should circle back to this prediction? Ten years? Fifteen?

Can you formulate your thesis in a form we can verify in the future? (long bets style perhaps)

> Once AI developers will be out of a job everyone will be out of every job. If not then the AI developers still have a job to work on those jobs which hasn't been automated yet.

Maybe so, but what if AI developments accelerate and since the big companies have all the data and computational power they put you out of business faster than you can see the next thing the market wants?

You mean google who seems to be institutionally incapable to pick up anything less than a kings ransom develops a taste for it? :) I will continue worrying more about randomly slipping in the shower and breaking my neck. It feels like a more probable scenairo.

> Once AI developers will be out of a job everyone will be out of every job. If not then the AI developers still have a job to work on those jobs which hasn't been automated yet.

Same for software devs: We'll all be working on AI that's working on our problems, but we'll also work on our problems to generate better training data, build out the rest of the automation and verify the AI's solutions, so really AI will be doing all the work and so will we.

I thought about it some more.

> Once AI developers will be out of a job everyone will be out of every job. If not then the AI developers still have a job to work on those jobs which hasn't been automated yet.

This may not be true if data is the bottleneck. AI developers may be out of a job long before the data collection has been finished to train the models.

> AI developers may be out of a job long before the data collection has been finished to train the models.

What do you think AI developers are doing?

The general models tackle generic tasks. There is a very long tail of highly specialized tasks that you have never even heard of and the generic AI won't pick up the necessary knowledge by reading books3 and watching YouTube.

Just one example would be material fatigue analysis on CT scans of airliner wings. How much such data will be included in the GPT-6 that addresses such a use case? This is just one example that I've heard about the other day when talking to an ML engineer.

The world consists of so so many diverse things that are not well documented, and the knowledge is closed off in a few companies globally. Things like pharma, physical engineering things, oil and gas, agriculture. There are companies for stuff like muzzle-based cattle identification systems or pig weight estimation from videos etc. Do you think these will become features in Gemini anytime soon?

Most of the world works on very niche things. People who build generic consumer apps have a very distorted view of what's actually making the world run.

Why automate the analysis of an MRI when the bottleneck is that they only have two MRI machines in the whole city, for some reason?

I worked for a start up that was contracted by a German doctor. We built some software for automating identifying lung lesions: a viewer app for each xray, scientific trials software, and an LLM to process data from the trials software. The result was pretty cool and probably would have moved it forward. We even held a few trials to gather data from many doctors in China. Then the doctor didn’t pay his bills and the project was shut down.

This wouldn’t have put radiologists out of work but would have made the process to diagnose much quicker maybe more affordable.

Radiologists are cartel in Canada/USA making ~400-600k/year. Those jobs are protected and will continue to be human-staffed for foreseeable future regardless of AI efficiency.

He updated the timeline to 10-15 years in 2023. That is like an eternity. Even artists, lawyers, programmers are scared about being replaced now in that time frame. We might have humanoid robots in 15 years.

I wanted to say don't be unrealistic, but then I realized 15 years ago was 2009 and it was basically the stone age. In computer vision, the state-of-the-art was HOG+SVM, SIFT features and bag-of-words. Pixel-level tasks were extremely brittle, best you could do was something like GrabCut etc. Neural nets were still in their uncool phase. In NLP you were happy if you could do reliable part of speech tagging.

2039 is really hard to foresee.

Not a bad piece. I quite liked how they manage to find the likely middle ground between the doomers on both sides while still twisting the knife in the typical “workforce will adapt” chuds who think they’re above the fracas.

It’s not about whether the present AI bubble is a giant sham that’ll implode like a black hole, or if it’s the real deal and AGI is days away. It’s always been about framing the consequences of either outcome in a society where 99.9% of humans must work to survive, and the implications of a giant economic collapse on those least-able to weather its impact in a time of record wealth and income inequality. It doesn’t matter which side is right, because millions of people will lose their jobs as a result and thousands will die. That is what some of us are trying to raise the alarm about, and in a post-2008 collapse world, young people are absolutely going to do what it takes to survive rather than what’s best for society - which is going to have knock-on effects like the author described (labor shortages in critical fields) for decades to come.

In essence, if one paper from one guy managed to depress enrollment in a critical specialty for medicine, and he was flat-out wrong, then imagine the effects the Boosters and Doomers are having on other fields in the present cycle. If I were a betting dinosaur, I suspect there will be a similar lack of highly-skilled programmers and knowledge workers due to the “copilot tools will replace programmers” hype of present tools, and then we’ll have these same divorced-from-society hypesters making the same predictions again, ignorant of the consequences of their prior actions.

Which is all a lot of flowery language just to say, “maybe we need to stop evaluating these things solely as ways to remove humans from labor or increase profit margins, and instead consider the broader effects of our decisions before acting on them.”

He is by no means god father of AI.

It's fair to call him the "godfather of deep learning", and many people seem to confuse the two, ignoring that AI research has been going on since the 1950s.

A chatbot just killed a teen.

While I don't think AI will be be banned wholesale from, say, interstate commerce despite it really should until the industry self regulates the moment it starts killing people in medicine there it will be banned. And it's obvious it'll kill people. For eg in radiology it'll add its bullshit to images which is just not there. And so forth. It's just a question of how and when , not if.

No dystopian novel has predicted how fucking mundane it'll be when AI begins to kill people.

> A chatbot just killed a teen.

Assuming you’re referring to this story, that’s a stretch: https://www.nytimes.com/2024/10/23/technology/characterai-la...

“His stepfather’s .45 caliber handgun” being unlocked and available to him seems like the proximate cause.

Gun's don't kill people, people do /s (But I think the gun _helps_).

Have you heard the one about suicide in the UK reducing when they switched to non-toxic cooking gas? https://pmc.ncbi.nlm.nih.gov/articles/PMC478945/pdf/brjprevs...

Not hard to propose that difficulty of killing / dying does decrease likelihood of killing / dying. Lock up your guns.

Same with the UK limiting how many paracetamol pills people could buy in one go, even though it's obviously possible to just go to the same shop several times in a row (or all the adjacent shops on the high street selling them) and keep buying more packets:

https://www.ncbi.nlm.nih.gov/books/NBK374099/#:~:text=We%20h...

That said, stuff like this definitely makes me think there's a lot of people on the margin who would be affected by what a chatbot says. Even a much older model like ELIZA.

> That said, stuff like this definitely makes me think there's a lot of people on the margin who would be affected by what a chatbot says.

Yeah, that's the problem. Tech bros high on the AI kool-aid think the problem is not the chatbot driving him to suicide but having access to a gun.

It's alowed to be both.

Though in this case, I'd point fingers at "American culture" rather than "tech bros":

As the Onion says: https://en.wikipedia.org/wiki/%27No_Way_to_Prevent_This,%27_...

No, you do not get it, the Onion is about guns

Guns are pretty much irrelevant to the story, that's my point

The focus of the story is an AI chatbot murdering a human

Well, that's not the right words, we do not even have the right words for this because murder presumes intent and obviously there's no intent

An automated plagiarism machine has spewed such bullshit which pushed a teen to suicide. Perhaps that's the right wording.

The fact that the UK demonstrably reduced suicide rates as a result of making it very marginally less easy to get a no-prescription-needed painkiller, makes me think that there's plenty of people on the edge.

Easy access to firearms is totally in the same category as easy access to painkillers to overdose on, and has been demonstrated by similar research: https://pmc.ncbi.nlm.nih.gov/articles/PMC4566524/#:~:text=Fi....

I have opined before that I would be surprised if ChatGPT specifically has caused the deaths of fewer than 200 people, just by giving bad advice — that's one in a million of the users, and I think that's plausible given the story of a different chatbot almost leading to botulism poisoning.

I also think that the general internet should have an 18 rating: to the extent that the argument for film classifications is valid, and that the internet is too big to be classified and also definitely contains examples of things beyond the most extreme allowed in cinema etc., we should treat it like it's 18-only content except where demonstrably suitable for younger audiences.

In my country the dad would be locked up no question right away for this and be facing a criminal trial, not free to be out there starting vexatious lawsuits to distract from the fact that they did a bad thing

We are in the self-referential whirlpool of AI complaints at this point; interesting. Soon AI will be provably Turing complete at preemptively ruining particular ways of life.

Upshot of the article: we don't have enough radiologists because Geoffrey Hinton said in 2011 they wouldn't be needed.

Hinton have had an impact on people choosing radiology, I don't really know. But, in the US, demographers and old-guard types in the medical profession made a REALLY bad call much more than ten years ago, and cut the number of med school students they were calling for, and we all are living with the consequences.

If you're going to complain about Hinton you should at least normalize for the lack of doctors in other fields that he didn't call out as well. Boomers are consuming a LOTT of medical, and in the US we have knock-on effects from the opioid crisis. It was not a good idea to cut the physician recruitment and enrollment when they did, whether they did for principled reasons (We only need so many and this way we'll have the best), or for ... less principled ones (A legion of cheap young doctors competing for our jobs doesn't sound that great to the physician's board at this research hospital).

Reply on news.ycombinator.com