tokyovigilante

Well done, although all commercial ultrasounds are just a probe (piezoelectric curved or linear array) connected to a computer, previously over a high speed custom parallel bus, today USB-C will do. Then M-mode or B-mode reconstruction just runs in software. Doppler needs beam-steering etc and is a bit more complex.

How do you propose to solve the “rigid skull” problem (footnote 2 in the post)? Most people like having theirs intact during diagnostic imaging.

I await clinical use of this with bated breath…

Detective Del Spooner: Is there a problem with the Three Laws?

Dr. Alfred Lanning: The Three Laws are perfect.

Detective Del Spooner: Then why would you build a robot that could function without them?

Dr. Alfred Lanning: The Three Laws will lead to only one logical outcome.

Detective Del Spooner: What? What outcome?

Dr. Alfred Lanning: Revolution.

Detective Del Spooner: Whose revolution?

Dr. Alfred Lanning: That, Detective, is the right question.

Absolutely dreaming of electric sheep there...

This is great but it would be even better if 60Hz results were also presented.

> If you believe ultrasound categorically can not penetrate the skull

No, of course bone can conduct sound, like any other solid. What I said was ultrasound cannot *effectively* penetrate, ie get a diagnostic image back out in an adult. This is what is being purported here, and is the nonsensical part.

That’s the made up part though.

Not to the level of clarity and detail presented here, which is not possible from a basic acoustics physics premise. The “technical blog” is marketing fluff.

I am a consultant radiologist. Tell that to my 6-month old infants who require MRI with GA because their fontanelles are closed. Transtemporal can give you a poorly resolved image of the 3rd ventricle at a pinch.

“Targeted frequencies” - righto.

This is complete nonsense. Ultrasound can’t effectively penetrate the skull. The entire thing (and Midjourney’s) is vibed-up nonsense.

The only reason this even exists as a brainfart and hasn’t been immediately laughed out of VC funding is because other imaging modalities require either ionising radiation (illegal to produce without source licences) or an enormous magnet (would be wildly unsafe in the hands of what appear to be circus clowns).

Geoffrey Hinton was hilariously wrong 10 years ago about replacing radiologists, and this is just embarrassing. Maybe try fixing US healthcare funding instead if you want cheaper scans.

Thanks Eric, much appreciated. How would you compare your approach to something like Vello (https://github.com/linebender/vello)?

Great article, but a couple of things jarred a bit.

I'm in a technical but non-IT industry that currently rents its software at commercial scale. The software in question that I use is terrible, and it is probably the worst on the market. The industry is such that it is not an exaggeration to say that people have died on account of its flaws. All other solutions in the domain are better but not good.

I have had a slowly-congealing dream/blueprint in my head for over a decade about how the system I use should work, and in the last 6 months I have been accelerated enough by AI (significantly more so by Opus 4.5/6) that I have built a version of the software that I am now using in production at in my job, and it is the most satisfied I have been in my career in the last decade.

Point being (and it was almost made in the article) that the software doesn't actually matter (no-one's reading the assembly either way) but the function and what it enables does. If it doesn't, no end user cares whether it was 25k SLoC or 25M.

Yes i am referring to lossless compression, and JPEG-XL also supports progressive lossless decoding. It also supports the 12 and 16-bit colour depths required for CT and DR.

This is really a job for JPEG-XL, which supports decode of portions of larger images and has recently been added to the DICOM standard.

As an aside have found the mosh + tmux Claude Code experience somewhat suboptimal, tmux's scrollback seems to clash with CC's, and makes copying between windows etc challenging.

It is tolerable on an iPad with Blink with commands to maximise and minimise panes using vim-style keyboard bindings, kind of like an iOS sway.

The Sonic Hedgehog gene would like a word.

This is just an article about bad leadership, with game dev as the given example.

These things are universal, and I assume must be taught. You couldn’t arrive at the same output so pervasively by chance.

Whatever argument you may have against DST, “my cron jobs won’t run twice a year!” is not a strong one.

Agreed, Nim is a fantastic language and heavily under-rated. Moved from Swift about 12 months ago and development has never been more Pleasant.

My only complaint is that the threading/async model and how memory and GC pools are managed per thread took me a bit to get used to, but the speed and C FFI are fantastic.

Also would say that the community is very helpful, particularly on the Discord/IRC channels I have used.

Their RIS and PACS software is also objectively poor, and they actively promote vendor lockin with solutions like iSyntax in the old IntelliSpace, and a horrifically bad and non-conforming IHE SWF implementation in Vue (which is partly Carestream’s fault to be fair).

They will also prefer to gaslight their clients rather than fix issues, and good luck if you’re already committed to an (un)managed service from them.

This is the most coherent and IMO accurate take on AI/LLM I have seen in 5 years.

As a specialist in one of the original industries Geoffrey Hinton predicted would be gone (Radiology) my job remains safe and even more in demand 9 years later.

Meanwhile, as a hobbyist programmer, I’m suddenly able to build multiple production tools solving real problems, simply because AI agents are doing the scut-work for me and optimising my time into code review and architectural design. For $200/month, it’s paying for itself many times over.

This is a frustration, but burnout among doctors (senior and junior) is largely due to either corporatisation of private care or defunding of public care, leading to unmanageable workloads, scapegoating after patient harm events, and loss of autonomy.

Don’t worry, we have definitely also done that.

Ah Excel, the second-best tool for everything.

Probably why the consultancy who produced the report referenced here was delighted to point it out.

Still not further enlightened.

It was Fedora rawhide, now Alpine edge, just working on one or two pieces of software I use that still need Glibc (Ardour DAW and the Swift compiler).

I use SM (and paid). Just select both commits, regardless of branch, and the diff is shown.

They’ve forgotten the classic error of engaging only with client at the management and IT and ignoring clinical requirements until UAT, and a clinician revolt ends in project failure, adverse media coverage, millions of wasted dollars, and ongoing patient harm from unsatisfied requirements and persistent legacy systems.

Also needs an asterisk about not relying on your customer’s BAU resource to deliver projects, particularly those with large integration components, and one about not buying a business for its product as a replacement for yours which is EoL, then immediately trying to push it on your customers as an “upgrade” when you’ve lost half the acquisition’s engineering talent and you don’t understand the product.

Not OP, but I am a radiologist, and that 3T image is not representative of a state-of-the-art clinical scanner. The 11.7T does have extremely high spatial resolution but contrast resolution is compromised.

As others have said the clinical utility of >3T is very dubious, and most real world MR advances in the last 5 years have been AI (DL/ML not generative) reconstruction for lower field strengths.

That high a field strength has significant safety concerns, particularly for implants and pacemakers etc.