will

[will @willessd]

@YounisJoseph There is a much larger and more rigorous ongoing study of AMIE with Included Health, to my knowledge.

[will @willessd]

@pmddomingos You are right. But that also doesn’t mean it won’t have incredible utility and ability to replace all jobs. I don’t understand why @ylecun and others keep emphasizing this. Yes, it matters if you are strictly talking about AGI in principle. But practically, no effect.

[will @willessd]

@HaQ_mAn @Arturo256395169 I’m an attending in Canada/US. Laws are manmade and change all the time. It’s an artificial constraint, not a physical one. AI and a midlevel can do a lot. Physicians tend to think AI is well suited to clean and rigid tasks. That couldn’t be further from the truth.

[will @willessd]

@HaQ_mAn @Arturo256395169 Great argument. Nonprocedural outpatient practices are in for a reckoning. It’s baffling how much of a misunderstanding fellow clinicians have of modern ML. You can close your eyes and block your ears all you want. Doesn’t change reality.

[will @willessd]

@HaQ_mAn @Arturo256395169 I don’t disclose anymore than that online. My specialty is also irrelevant to my argument. I was in ML research previously, so I have a unique understanding of both.

[will @willessd]

@HaQ_mAn @Arturo256395169 I am a doctor. This is laughable incorrect.

[will @willessd]

@docmilanfar That is a very broad brush stroke way if describing clinical medicine. A very significant portion of it is algorithmic, and the rest is just surface level pattern matching.

The “messiness” can be solved with a human in the loop, need not be a physician.

[will @willessd]

@Berci Funny thing is all that stuff wouldn’t bother AI nor would it be anywhere near as time consuming

[will @willessd]

@YounisJoseph Also, what “obvious latent risks in these models and their implementation”?

You got inside info on what model they used, what guardrails implemented, what aspects are deterministic, how the system is designed overall?

2/2

[will @willessd]

@YounisJoseph This is a pilot staged rollout. That’s how you gather RWE. You can’t have it both ways by crying wolf because there’s no RWE and at the same time preventing RWE from being gathered.
1/

[will @willessd]

@neiltyson I too can say things using vague terms to create unfalsifiable statements.

Concretely define “understanding”.

[will @willessd]

@realLongWangMD @BBillDavis And yet you have still to provide any substance to your argument other than playing semantics.

Again, God help your patients.

[will @willessd]

@realLongWangMD @BBillDavis You have yet to provide any substance instead of doing mental gymnastics.

Your patients must love your communication skills. God help them.

[will @willessd]

@realLongWangMD @BBillDavis Wow. That’s a fantasy. I’m a physician but know there are fields incredibly more intellectually difficult than clinical medicine.

Also, your point is favourable for AI, ironically. Moravec’s Paradox.

Law is a human construct, not a physical constraint. Changes all the time.

[will @willessd]

@DrSiyabMD Why

[will @willessd]

@realLongWangMD @BBillDavis Based on what? You think clinical medicine is intellectually too complex for AI? Because that’s laughable

[will @willessd]

@mcuban Fundamental misunderstanding of economics.

Tldr:

If AI can do your entire job, it will fully replace. If it can do some, it will make more efficient (jevons) and depending on elasticity will lead to more jobs or less.

We are currently at the “some” stage.

[will @willessd]

@ylecun @KenRoth You could argue everyone buys votes. You can pay even if it’s not directly cash.

[will @willessd]

@SeanLangenfeld The goal of AI is to create intelligence - what allows you to perform surgery. Not your domain expertise that has been acquired *due* to intelligence. It may not be there yet (and data is a barrier in surgery) but by no means does that means there is some fundamental barrier.
4/4

[will @willessd]

@SeanLangenfeld You are treating this as something that is deterministic and hard coded, which *would* require significant domain knowledge. Although still required (but nowhere near the same degree), this is a completely different paradigm. These are learning algorithms. 3/