Fable remains in limbo, with renewed hope that we will get it back soon (45% by tomorrow, 69% by July 1, nice.) The full capabilities post is now available.
> and seems not to fully grok the principle of ‘you can just act sensibly in response to mildly concerning information’ despite explicitly stating that principle many times.
You *can* but how many people actually*do*? Will our system overall act sensibly? What's possible is not under dispute, what's likely is.
On refine.ink, I've tried it and it's useful. My view, however, is that it's expensive relative to its usefulness. I also worry about a particular provider getting entrenched. Econometrica now uses refine.ink for final-round screening, I am reliably informed. The next time I want to send a paper there, I guess I'd better find out what refine.ink thinks first. As Tyler likes to say: "solve for the equilibrium".
Fortunately, it's easy to build your own version of this system and operate it at a fraction of the cost: see e.g. https://coarse.ink. An even better approach, and the one I use, is to create agent-based skills to use with Claude Code as you work on a research paper. Unlike with refine.ink, I can give Claude Code access to full-text of any paper I build on or cite in .md format. This is a real game-changer: give it a try. Good tools for converting .pdfs to markdown that can run on a MacBook Pro are mineru and marker:
"You can just act sensibly in response to mildly concerning information" is exactly right, but only if you are not a human with normal human emotions and behaviors.
This scanner might be a net positive in the world, I don't know. What I'm reacting to is the out-of-hand dismissal that humans act like humans.
Could I hear "look at these 10 lumps we found in your body, they could be cancer but they are probably not," and not lay awake at night with worry? I'm not sure. Could my loved ones? I know they cannot. This is IMO a common, normal, and predictable response in many ordinary people. And things that can be predicted should be part of the input, not dismissed, when we make a judgement about a technology or policy.
This feels like a classic Rationalist failure mode. "The rational way to think is X, so with my Rationalism hat on, I should make decision Y that assumes people will act like X." But if Rationalism is about making better decisions that lead to better outcomes, then this is not good Rationalism. This is making decisions that would lead to better outcomes in a world that does not exist, the world of spherical cows and Homo rationalis. Scott's analysis might be wrong, but he is at least evaluating this tech in the real world. "You can just act sensibly" is not.
The first response to this will be "You can just not get tested," but I hope it's obvious how the above applies to that as well. By contrast "Yes, these predictable human behaviors exist, but taking them into account, that still should not prevent access to the technology," this is a very rational argument that can be made (even if I might or might not agree with its conclusion).
Lots of juice in this post, but the main thing I feel wild about is how various senior figures within the medical ‘establishment’ (:P) haven’t updated their baselines of disclosure in light of COVID. “Hmm, our ability to predict second order consequences of information dissemination may not be as finely calibrated as we believed!”
It is becoming increasingly ridiculous to take Ploymarket’s odds regarding Fable seriously, since they move in direct accordance with twitter shitposting.
I'm in the weights but only for opus 4.8 (barely, but there!). all the other AIs think I'm a politician (presumably they're confusing me with ayelet shaked)
I got top 8% for "Reddit user artifex0", though the models seemed to have only a very vague recollection of my account and reported details that were all slightly off.
I don't think I dismissed the idea of doing incremental scans and taking the difference "out of hand". My point is that this is priced into everyone's discussions of how useful this is. "Doing a scan twice" isn't some kind of insane technological leap that it took an AI company to think of. Existing whole body MRIs are often done yearly. The reason this doesn't help much is that dangerous things aren't particularly well-distinguished from benign things based on how new they are or how quickly they grow (or rather, the degree to which they are is already included in the calculations I gave earlier).
The amount of coverage the Midjourney scanner is getting is just bizarre. New imaging modalities are invented not infrequently, however, almost none of them make it past the lab. The current suite of medical imaging tools are actually very good, and while it's possible to beat them on one or two metrics, it's hard to do better as a whole. Not impossible, but you should treat announcements about revolutionizing medical imaging at the same level of credence as claims to have found a cure for cancer based on results in mice.
The real story here is that Midjourney failed as an AI company. They are no longer a top contender among image generation models (ranked #81 on artificial analysis, which is even worse than I thought). And they are trying to pivot as a last ditch attempt to get more funding.
As for specific reasons to think this will fail - the biggest issue is that the images on their website are completely non diagnostic. You can't claim to revolutionize imaging until you can actually produce useable images. Preferably with some clinical data to show that it can pick up pathology prospectively.
More specific reasons to think Ultrasound tomography won't work very well - sound waves scatter more than X-rays. The further away the target, the more noise gets introduced. Which is why ultrasound is traditionally a targeted modality - you place the scanner directly next to the area of concern. Ultrasound also doesn't penetrate bone very well. You can get a signal via full wave inversion, but so much of the sound waves are attenuated that you get a very noisy image that has so far not been clinically useful.
And on the point of harms from whole body screening, you're missing two things. The first is the third part payer problem - people aren't bearing the costs of screening and intervention - those costs are distributed, leading to a tragedy of the commons. If costs were totally out of pocket then sure this wouldn't be an issue. But second - lawsuits. A physician can decide not to intervene on findings that are 90% likely to be irrelevant. But in the remaining 10% of cases, there is potential for getting sued.
Re: "Grok, like the internet, is for porn."... Well, to twist two lines from Alexander James Adams's musical paean to Pan, "Creature of the Wood" https://www.youtube.com/watch?v=9-K2wLNtKT8 , SOTA LLMs are already "Not Beast, not God, and yet not Man" - and the cited application area notes their progress towards the "I shall be thy lover! " line :-)
"I see people saying ‘no this is good writing it’s just in the AI style’ and seriously, no, stop, this is a walking joke. It’s like something Garrison Keillor would have written in 5 minutes as a radio sketch to make fun of pompousness, except also it’s AI. You have to hear it in ‘gritty film noir style radio story narrator’ voice to really appreciate its horridness."
nit: I find the story unobjectionable, personally. But I read very little fiction. Perhaps it is a matter of whether one has seen cliched styles so many times that they raise one's hackles - and I haven't, so they don't raise mine?
On section: "Middle Of The Journey", you don't actually say what it is you're talking about or give a link to what you're talking about.
[Edit:] "After you step on a platform, Midjourney's scanner will submerge you in water at a rate of 2 inches per second. Your body passes through a ring made of half a million [ultrasonic tranceivers about 1/5th of a mm square each]"
"the company is developing the machine with handheld ultrasound device maker Butterfly Network. Midjourney signed a licensing agreement with Butterfly Network in November 2025, securing exclusive rights to its ultrasound-on-chip technology. The project is led by Tom Calloway PhD, Head of Medical at Midjourney."
These are 1-10MHz MEMS non-piezo capacitor actuator/sensor phased arays with 8960 elements and all the ADCs, drivers and most of the processing on-board, quite a sophisticated design. The ADCs are only 10-bit, but fast and there's 1 for every 8 array elements. The focal depth and frequency can be dynamically varied for high-res skin or lower-res deep scanning. Regular piezos would only operate well at a single frequency. The units Butterfly sells with a more developed version of this chip cost $3900 [not bad!] + about $300/yr. for the SW license [bah!]. There will be 50 to 60 such chips in a ring scanner.
I had a very similar immersion ultrasound idea in 2007, with a much more sophisticated imaging layout and processing, than their ring design (giving whole-body gross internal anatomy at video frame rates, ~1000x their speed). Initially, the object was ultrasonic neuromodulation (then new), but I quickly realized the scanning applications would be more valuable, envisioning a large population whole-body scanning project, together with a longitudinal study of child development, both with whole-genome sequencing, full-spectrum bloodwork (not lots of specific tests, but as complete a quantitative analysis of all cell types and molecules present as possible), mental testing (many subtests, of perception, reaction and reasoning abilities). With a few hundred thousand individuals, we could get a much better idea of phenotypical effects of genetic variants. The anatomic information alone would allow generative models to produce human models with any sort of age, size, proportions, features, with completely natural movement, shape changes, etc., including muscle and fat, skin structure, elastomeric and non-rigid mechanics, etc. for use in simulations, including movies and games.
I feel like you're a bit too credulous with the Midjourney thing, even though I think they might end up with a killer product anyway, first just the vibe of the website should be a clue that they're over selling, but a quick prompt to Claude basically "how real is this?" confirmed.
Some things they pointed out:
The number of scanning chips quoted on the site ("hundreds of thousands") and their size ("as tiny as grains of sand") is magnitudes off from the actual prototype (40 and normal chip sized).
The bone/air problem is real and extremely hard to get around by all known physics. Air is basically a no go (although there aren't a lot of organs entirely surrounded by air, it still makes reconstruction super difficult and it's not something they've demonstrated) and bone is technically doable but at much lower resolution than an MRI.
Getting to "the spa" depends on solving a bunch of hard problems and the timeline is just unrealistic.
A bunch of other companies are trying to do similar things and their claims are more modest.
See my comment above, the tech is already retail. Brain imaging won't be very good, probably, but there are many processing techniques and even fairly simple hardware tech that hasn't been applied yet.
Well yes, that was part of Claude's point I believe. The crowning achievement of the tech seems to be a breast scanner, can't recall if it's from butterfly, that meets or exceeds MRI in usefulness. Pretty amazing but breasts are apparently absolutely ideal for ultrasound and the rest of the body is between somewhat and way more difficult. Overall point is that Midjourney are promising something way more advanced than they have demonstrated, maybe they can pull it off but maybe not and it might take a lot more than 5 years.
Btw total aside but one thing I learned is that brain imaging is already pretty good in babies, when the skull hasn't yet fused, and it's been demonstrated in animals by making small holes in the skull and replacing with an ultrasound transparent material, seems that with the skull issue taken care of, functional ultrasound, where blood flow is measured by the doppler effect, is totally feasible!
Cutting holes in people's skulls is unpopular. But if you have arrays in direct contact with the skull over most of the cranium, with gradient-index acoustical impedance matching, it's mostly the interior reflections off the skull that attenuate the signal and the arrays on opposite sides of the skull get near-normal incidence of the beams, cutting reflection, plus opposing pairs make the attenuation power of 2 instead of 4, as it is with one array. Add in the data from the other arrays and apply the compressed-sensing algorithms that improved MRI resolution per magnetic field so much, scan longer, and I think you can make out cortical area quite well, and at least fair resolution of hippocampus, corpus callosum and such bigger deep structures, but maybe not tthe fiddly little bits in the limbic system.
Here's an alternative take on the weird Bezos piece.
Jeff does understand what transformative AGI does to work in society. He is basic AGI-pilled nerd happy about an endpoint where machines do all the work and humans have to decide what to find meaningful rather than working
But, he knows that is a wild scary future for typical folk and governments to consider. It is normal to have a job. People want to have a job. How could we possibly transition to an economy where people don't get paid to work?
His solution is to change the definition of what a job is. A post-scarcity economy where no one needs to work gradually manifests as everybody being boss of a company they alone work for.
It's quite a neat way to sell the future in the frame of today.
Of course, he somehow isn't ASI-pilled, troubled by alignment or acknowledging that likely future, because that would make him sad and scared.
> and seems not to fully grok the principle of ‘you can just act sensibly in response to mildly concerning information’ despite explicitly stating that principle many times.
You *can* but how many people actually*do*? Will our system overall act sensibly? What's possible is not under dispute, what's likely is.
On refine.ink, I've tried it and it's useful. My view, however, is that it's expensive relative to its usefulness. I also worry about a particular provider getting entrenched. Econometrica now uses refine.ink for final-round screening, I am reliably informed. The next time I want to send a paper there, I guess I'd better find out what refine.ink thinks first. As Tyler likes to say: "solve for the equilibrium".
Fortunately, it's easy to build your own version of this system and operate it at a fraction of the cost: see e.g. https://coarse.ink. An even better approach, and the one I use, is to create agent-based skills to use with Claude Code as you work on a research paper. Unlike with refine.ink, I can give Claude Code access to full-text of any paper I build on or cite in .md format. This is a real game-changer: give it a try. Good tools for converting .pdfs to markdown that can run on a MacBook Pro are mineru and marker:
https://github.com/opendatalab/mineru
https://github.com/datalab-to/marker
"You can just act sensibly in response to mildly concerning information" is exactly right, but only if you are not a human with normal human emotions and behaviors.
This scanner might be a net positive in the world, I don't know. What I'm reacting to is the out-of-hand dismissal that humans act like humans.
Could I hear "look at these 10 lumps we found in your body, they could be cancer but they are probably not," and not lay awake at night with worry? I'm not sure. Could my loved ones? I know they cannot. This is IMO a common, normal, and predictable response in many ordinary people. And things that can be predicted should be part of the input, not dismissed, when we make a judgement about a technology or policy.
This feels like a classic Rationalist failure mode. "The rational way to think is X, so with my Rationalism hat on, I should make decision Y that assumes people will act like X." But if Rationalism is about making better decisions that lead to better outcomes, then this is not good Rationalism. This is making decisions that would lead to better outcomes in a world that does not exist, the world of spherical cows and Homo rationalis. Scott's analysis might be wrong, but he is at least evaluating this tech in the real world. "You can just act sensibly" is not.
The first response to this will be "You can just not get tested," but I hope it's obvious how the above applies to that as well. By contrast "Yes, these predictable human behaviors exist, but taking them into account, that still should not prevent access to the technology," this is a very rational argument that can be made (even if I might or might not agree with its conclusion).
Lots of juice in this post, but the main thing I feel wild about is how various senior figures within the medical ‘establishment’ (:P) haven’t updated their baselines of disclosure in light of COVID. “Hmm, our ability to predict second order consequences of information dissemination may not be as finely calibrated as we believed!”
YES – for fuck's sake already
It is becoming increasingly ridiculous to take Ploymarket’s odds regarding Fable seriously, since they move in direct accordance with twitter shitposting.
I'm in the weights but only for opus 4.8 (barely, but there!). all the other AIs think I'm a politician (presumably they're confusing me with ayelet shaked)
I got top 8% for "Reddit user artifex0", though the models seemed to have only a very vague recollection of my account and reported details that were all slightly off.
I don't think I dismissed the idea of doing incremental scans and taking the difference "out of hand". My point is that this is priced into everyone's discussions of how useful this is. "Doing a scan twice" isn't some kind of insane technological leap that it took an AI company to think of. Existing whole body MRIs are often done yearly. The reason this doesn't help much is that dangerous things aren't particularly well-distinguished from benign things based on how new they are or how quickly they grow (or rather, the degree to which they are is already included in the calculations I gave earlier).
The amount of coverage the Midjourney scanner is getting is just bizarre. New imaging modalities are invented not infrequently, however, almost none of them make it past the lab. The current suite of medical imaging tools are actually very good, and while it's possible to beat them on one or two metrics, it's hard to do better as a whole. Not impossible, but you should treat announcements about revolutionizing medical imaging at the same level of credence as claims to have found a cure for cancer based on results in mice.
The real story here is that Midjourney failed as an AI company. They are no longer a top contender among image generation models (ranked #81 on artificial analysis, which is even worse than I thought). And they are trying to pivot as a last ditch attempt to get more funding.
As for specific reasons to think this will fail - the biggest issue is that the images on their website are completely non diagnostic. You can't claim to revolutionize imaging until you can actually produce useable images. Preferably with some clinical data to show that it can pick up pathology prospectively.
More specific reasons to think Ultrasound tomography won't work very well - sound waves scatter more than X-rays. The further away the target, the more noise gets introduced. Which is why ultrasound is traditionally a targeted modality - you place the scanner directly next to the area of concern. Ultrasound also doesn't penetrate bone very well. You can get a signal via full wave inversion, but so much of the sound waves are attenuated that you get a very noisy image that has so far not been clinically useful.
And on the point of harms from whole body screening, you're missing two things. The first is the third part payer problem - people aren't bearing the costs of screening and intervention - those costs are distributed, leading to a tragedy of the commons. If costs were totally out of pocket then sure this wouldn't be an issue. But second - lawsuits. A physician can decide not to intervene on findings that are 90% likely to be irrelevant. But in the remaining 10% of cases, there is potential for getting sued.
I just want to register that I find it sad that "Hansonian medicine" doesn't mean the opposite of what it means!
Whole lotta Bayes points riding on this Midjourney scanner debate. Interested to check back in circa 2028 and see how it all shakes out.
"now is the time to inform regulators on how we should build a world with safe, frontier, open intelligence."
Please don't be shy to start Nathan!
<mildSnark>
Re: "Grok, like the internet, is for porn."... Well, to twist two lines from Alexander James Adams's musical paean to Pan, "Creature of the Wood" https://www.youtube.com/watch?v=9-K2wLNtKT8 , SOTA LLMs are already "Not Beast, not God, and yet not Man" - and the cited application area notes their progress towards the "I shall be thy lover! " line :-)
</mildSnark>
"I see people saying ‘no this is good writing it’s just in the AI style’ and seriously, no, stop, this is a walking joke. It’s like something Garrison Keillor would have written in 5 minutes as a radio sketch to make fun of pompousness, except also it’s AI. You have to hear it in ‘gritty film noir style radio story narrator’ voice to really appreciate its horridness."
nit: I find the story unobjectionable, personally. But I read very little fiction. Perhaps it is a matter of whether one has seen cliched styles so many times that they raise one's hackles - and I haven't, so they don't raise mine?
On section: "Middle Of The Journey", you don't actually say what it is you're talking about or give a link to what you're talking about.
[Edit:] "After you step on a platform, Midjourney's scanner will submerge you in water at a rate of 2 inches per second. Your body passes through a ring made of half a million [ultrasonic tranceivers about 1/5th of a mm square each]"
"the company is developing the machine with handheld ultrasound device maker Butterfly Network. Midjourney signed a licensing agreement with Butterfly Network in November 2025, securing exclusive rights to its ultrasound-on-chip technology. The project is led by Tom Calloway PhD, Head of Medical at Midjourney."
[From: https://www.engadget.com/2196998/midjourney-full-body-ultrasonic-scanner/ ]
The Midjourney and Butterfly network pages are stuffed with B-Ark worthy maketing pablum. Here's a technical article on the design of the ultrasound chips: https://www.pnas.org/doi/full/10.1073/pnas.2019339118
These are 1-10MHz MEMS non-piezo capacitor actuator/sensor phased arays with 8960 elements and all the ADCs, drivers and most of the processing on-board, quite a sophisticated design. The ADCs are only 10-bit, but fast and there's 1 for every 8 array elements. The focal depth and frequency can be dynamically varied for high-res skin or lower-res deep scanning. Regular piezos would only operate well at a single frequency. The units Butterfly sells with a more developed version of this chip cost $3900 [not bad!] + about $300/yr. for the SW license [bah!]. There will be 50 to 60 such chips in a ring scanner.
I had a very similar immersion ultrasound idea in 2007, with a much more sophisticated imaging layout and processing, than their ring design (giving whole-body gross internal anatomy at video frame rates, ~1000x their speed). Initially, the object was ultrasonic neuromodulation (then new), but I quickly realized the scanning applications would be more valuable, envisioning a large population whole-body scanning project, together with a longitudinal study of child development, both with whole-genome sequencing, full-spectrum bloodwork (not lots of specific tests, but as complete a quantitative analysis of all cell types and molecules present as possible), mental testing (many subtests, of perception, reaction and reasoning abilities). With a few hundred thousand individuals, we could get a much better idea of phenotypical effects of genetic variants. The anatomic information alone would allow generative models to produce human models with any sort of age, size, proportions, features, with completely natural movement, shape changes, etc., including muscle and fat, skin structure, elastomeric and non-rigid mechanics, etc. for use in simulations, including movies and games.
I feel like you're a bit too credulous with the Midjourney thing, even though I think they might end up with a killer product anyway, first just the vibe of the website should be a clue that they're over selling, but a quick prompt to Claude basically "how real is this?" confirmed.
Some things they pointed out:
The number of scanning chips quoted on the site ("hundreds of thousands") and their size ("as tiny as grains of sand") is magnitudes off from the actual prototype (40 and normal chip sized).
The bone/air problem is real and extremely hard to get around by all known physics. Air is basically a no go (although there aren't a lot of organs entirely surrounded by air, it still makes reconstruction super difficult and it's not something they've demonstrated) and bone is technically doable but at much lower resolution than an MRI.
Getting to "the spa" depends on solving a bunch of hard problems and the timeline is just unrealistic.
A bunch of other companies are trying to do similar things and their claims are more modest.
See my comment above, the tech is already retail. Brain imaging won't be very good, probably, but there are many processing techniques and even fairly simple hardware tech that hasn't been applied yet.
Well yes, that was part of Claude's point I believe. The crowning achievement of the tech seems to be a breast scanner, can't recall if it's from butterfly, that meets or exceeds MRI in usefulness. Pretty amazing but breasts are apparently absolutely ideal for ultrasound and the rest of the body is between somewhat and way more difficult. Overall point is that Midjourney are promising something way more advanced than they have demonstrated, maybe they can pull it off but maybe not and it might take a lot more than 5 years.
Btw total aside but one thing I learned is that brain imaging is already pretty good in babies, when the skull hasn't yet fused, and it's been demonstrated in animals by making small holes in the skull and replacing with an ultrasound transparent material, seems that with the skull issue taken care of, functional ultrasound, where blood flow is measured by the doppler effect, is totally feasible!
Cutting holes in people's skulls is unpopular. But if you have arrays in direct contact with the skull over most of the cranium, with gradient-index acoustical impedance matching, it's mostly the interior reflections off the skull that attenuate the signal and the arrays on opposite sides of the skull get near-normal incidence of the beams, cutting reflection, plus opposing pairs make the attenuation power of 2 instead of 4, as it is with one array. Add in the data from the other arrays and apply the compressed-sensing algorithms that improved MRI resolution per magnetic field so much, scan longer, and I think you can make out cortical area quite well, and at least fair resolution of hippocampus, corpus callosum and such bigger deep structures, but maybe not tthe fiddly little bits in the limbic system.
Here's an alternative take on the weird Bezos piece.
Jeff does understand what transformative AGI does to work in society. He is basic AGI-pilled nerd happy about an endpoint where machines do all the work and humans have to decide what to find meaningful rather than working
But, he knows that is a wild scary future for typical folk and governments to consider. It is normal to have a job. People want to have a job. How could we possibly transition to an economy where people don't get paid to work?
His solution is to change the definition of what a job is. A post-scarcity economy where no one needs to work gradually manifests as everybody being boss of a company they alone work for.
It's quite a neat way to sell the future in the frame of today.
Of course, he somehow isn't ASI-pilled, troubled by alignment or acknowledging that likely future, because that would make him sad and scared.