The scoop: X-ray imaging of various strictures for scientific purposes produces colossal reams of data, previously hard to analyze. Meta provides machine analysis, both segmentation and classification, using unsupervised learning models.
> a fully reconstructed, semantically labeled 3D volume delivered back to the scientist physically standing at the beamline [x-ray] instrument, ready for interpretation while the experiment is still running. Total turnaround: approximately 15 minutes.
Gonna need to have a talk with LLNL. I'm sure they didn't choose the name, but seems a tad leaning in to use the name of tech from Star Trek meant to produce untold abundance that instead became an unintentional doomsday device.
[flagged]
That is — incredible
Really fascinating writeup — I particularly appreciated how they didn't hesitate to delve into the load bearing design choices — the implications are staggering.
this page hijacks your tab's back button history :\
how so...back button history is fine for me (Chrome/Windows)
Well Meta has hijacked my privacy. Now IRL Meta can track me from the people wearing their sunglasses.
ELI5.
AI model inspects hundreds of thousands of scientific images. A job that previously took an expert roughly a month can now be completed in around 15 minutes.
SAM 3 (Segment Anything Model 3) and DINOv3, projects released by Facebook, were used to do science and research.
Whenever Meta claims their models are open source, you have to double-check.
SAM License (https://github.com/facebookresearch/sam3/blob/main/LICENSE):
> iv. Your use of the SAM Materials will not involve or encourage others to reverse engineer, decompile or discover the underlying components of the SAM Materials.
> v. You are not the target of Trade Controls and your use of SAM Materials must comply with Trade Controls. You agree not to use, or permit others to use, SAM Materials for any activities subject to the International Traffic in Arms Regulations (ITAR) or end uses prohibited by Trade Controls, including those related to military or warfare purposes, nuclear industries or applications, espionage, or the development or use of guns or illegal weapons.
> b. If you institute litigation or other proceedings against Meta or any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the SAM Materials, outputs or results, or any portion of any of the foregoing, constitutes infringement of intellectual property or other rights owned or licensable by you, then any licenses granted to you under this Agreement shall terminate as of the date such litigation or claim is filed or instituted. You will indemnify and hold harmless Meta from and against any claim by any third party arising out of or related to your use or distribution of the SAM Materials.
The DINOv3 License (https://github.com/facebookresearch/dinov3/blob/main/LICENSE...) is similar but with the model names swapped.
It's always nice when a model's weights are released, but Meta's models are not open source because their weights always come with weird restrictions.
We've seen b before. React was controversially released with a similar clause and ultimately Facebook dropped it and used a standard open source license. Their lawyers really love this idea for some reason.
Was it released with it? Iirc they tried moving to such a license from bsd, which contains no patent grants whatsoever, and hence they thought it was a strict upgrade. The community did not agree so they moved back.
Do their lawyers love the idea? Hard to say. My guess is, Zuck really likes the idea, so he keeps directing them and their chief counsel to try stuff like this.
It's almost as if having a CEO of a company placed beyond the control of the corporate board is a bad thing.
This is a pretty common clause when a company open sources something that was previously locked away. The AOM AV1 codec has a similar rule for its patent pool to discourage trolling.
I'm no fan of Facebook or its social effects but I can't deny the wonderful downstream effect of their open source.
Popular microscopy models like Cellpose[0] have leaned heavily on the cornucopia of open and SOTA power. I have no doubt thousands of biologists have benefitted from the capabilities these models bring. I think it was unthinkable just 5 years ago that a single biologist with just a laptop could do mass-segmentation at this kind of fidelity.
Then there's Napari and it's plugin ecosystem[1] that wouldn't exist without the Chan Zuckerberg Initiative. Again, I'm not trying to glaze them but as someone in the biotech/microscopy space I can't understate how often I use and benefit from their open source.
[0] https://cellpose.readthedocs.io/en/latest/models.html [1] https://chanzuckerberg.com/rfa/napari-plugin-grants/
> but I can't deny the wonderful downstream effect of their open source.
ReactJS has been a very mixed-bag...
> ReactJS has been a very mixed-bag...
This is a very confused statement to be made, specially after the last decade where React established itself as the de facto way of writing any and every SPA.
Is it yet another example of "there are only two kinds of languages: the ones people complain about and the ones nobody uses."
I think the universe of SPAs that should exist is far smaller than the universe of SPAs that does exist, and the hype around react is largely to blame.
> the ones people complain about and the ones nobody uses
The funny thing is that prior to 2015 "the ones nobody uses" were the academically-designed ivory tower FP languages; perpetually relegated to being a minority curiosity because FP is immensely difficult to wrap one's head around at first, especially if all you've-ever-known was imperative C or VB and (compromised) OOP in Java and C#.
I remember when Google's overqualified design-team for Angular decided that Angular2+ is too good for Promise-based fetch() - instead everyone should be forced to learn RxJS's way-of-doing-things - cue the agony, wailing and the gnashing of teeth.
...I think being forced to live through that experience did 2 things:
1) It forced an entire generation of front-end web-devs (and VB6 refugees) to get familiar with FP-like concepts.
2) It also made those same front-end web-devs swear-off ever using a Google-maintained SPA framework in future.
...which set the stage for the mass-adoption of ReactJS:
1) It made effective use of FP concepts, which many devs were now accustomed-to after having survived Angular2.
2) It wasn't made by Google.
React is more of a plague than a wonderful downstream effect or mixed bag. It’s essentially cemented Js as the way to build a website even if you don’t need the complexity. It is the leader in brain dead Js evangelism.
Credit where credit is due, though: ReactJS became #1 on its merits.
...yes, Redux (not React) deserves to be #1, but Redux isn't a complete, all-under-one-roof framework the way React is; but regardless of that: The Redux/React approach is just fundamentally a better design than what we had before: stateful-controls/widgets and two-way data-binding.
If you'd like to relive how UI devs suffered throughout the 1990s, 2000s, and most of the 2010s I invite you to try making a native Windows 11 desktop UI using WinUI3 using the MVVM (anti-) pattern: nothing but mutable objects of indeterminable state getting caught in infinite-loops or unbound recursion due to INotifyPropertyChanged - and Microsoft is still pretending that's the "right" way to build a UI.
Sorry am ranting on about something I have very little control over; it's just frustating.
WinUI3 is horrible, but that is an implementation detail. WPF isn’t so bad (of course for business reasons no one should use either).
I still see WPF and AppKit as both being superior UI technologies to web+react, if you ignore the challenges of distribution.
I made several large Cocoa (macOS AppKit, Obj-C) apps in 2004-2018 using the classic Smalltalk-inspired MVC patterns and this was never a problem. Sometimes writing the controller boilerplate was boring, but at least there were no observer loops, it was simple to debug where things happened, and undo/redo were reliable to implement through the framework.
Maybe Microsoft just never was very good at these 1990s style GUIs?
> nothing but mutable objects of indeterminable state getting caught in infinite-loops or unbound recursion due to INotifyPropertyChanged
Rant on brother/sister. I'm a year into building a large WPF (because reasons) app for my employer and some days its very difficult.
> ...yes, Redux (not React) deserves to be #1,
This is a rather perplexing statement to make.
If anything, Redux only plays a role in adding syntactic sugar over React's support for reactive programming.
More importantly, the general guidelines are that you should only bother with Redux if your SPA is already considerably complex.
> Redux only plays a role in adding syntactic sugar over React's support for reactive programming
Redux doesn't have anything to do with any kind of syntactic-sugar for Reac.
> the general guidelines are that you should only bother with Redux if your SPA is already considerably complex
If one already has an (non-React) SPA which is not already built following Redux then you're stuck with what you've got; it isn't really possible to take an existing SPA of any complexity and deciding to "make it use Redux" because it completely up-ends practically every aspect of SPA application-design, at least as far as how "traditional" (non-React) SPAs do things.
If you'll permit me to oversimplify: I tell people that React is a basically an opinionated Redux kernel wrapped around a JSX-based DOM+component/widget system (, which then means you now need a JS build system too); or conversely: Redux is like React, except without JSX components.
...or am I misunderstanding you?
> Meta's open-source approach makes this possible
I'm sorry, was this article drafted in 2024 and never updated?
This sort of random out of date comment is a common LLM writing trope. Though it is true that the segment anything model is open source.
https://github.com/facebookresearch/sam3
They have another with calling A100s modern.
> A100 GPUs — the high-performance computing chips that power today's most advanced AI systems.
They might not have written it with AI, but the article has a lot of em dashs and colons and not this but that statements.
The models in question have source and weights available. AFAIK not fully FOSS because the weights require registration to access.
SAM3.1 was released and open sourced March 2026.
But what's the point of correcting you? People will continue to propagate their own lies.
Sure, I just find using OSS grand standing, and trying to educate the audience about the benefits of OSS a bit distasteful after Meta pulled away from OSS for their main flagship models.
As a former proposal specialist (B2B, B2G, non DoD) I looked into the Genesis Mission procurement site and process.
Unless someone can correct me, the total amount of grant monies is $280,000,000 or so.
It became obvious that it’s not worth my time to engage in the “mission” as they call it, even if I could benefit some worthwhile causes.
That’s a pittance and pretty insulting to the purported benefit of funding scientific endeavors. I’m not even attempting to be political here. $280 Million versus $XX Billion for warfighting is a seriously gross misallocation of public monies, IMHO.
Total lackluster reporting on the scale and scope of the actual numbers, but not surprising.
This is a really weird comparison. The Chan Zuckerberg foundation is nowhere involved with any war. All they’re doing is making money available for research. In which universe is $280m not enough?
When the people running it have a net worth of over $150 billion.
It's a weird world where $280M is not considered a lot of money.
So, not LLM models, right?
Also on this:
> The numbers are staggering: The DOE's light and neutron source facilities now produce tens of petabytes of data annually
Come on, petabytes are not staggering for entreprise software.
I'm not sure exactly which enterprises you have in mind, but sure: quantities which can be expressed as "a year's worth fits on my desk" should not be described as "staggering", and 10 PB of hard drives will (just about) fit on my desk.
The LHC, on the other hand, that generates a petabyte a second and has to throw most of it away for obvious reasons:
https://www.itnews.com.au/news/computing-for-the-large-hadro...
I used to work with images representing scans of brain tissue - for a full brain visualization at one horizontal slice terabytes was a common measure and the resolution of those images wasn't even particularly detailed - all the full resolution stuff was taken of tiny sub-sections of interest. This was also two decades ago - so I'm sure they've upped their game.
My company does whole-brain scans of mice on a Zeiss Z.1. Lower resolutions are typically in the low hundreds of GB. Higher-resolutions and multichannel staining can get you in the TB range. When a typical client is doing 10's of brains it definitely adds up. But even for us (and I consider us a smaller operation) we aren't output PB's. So I'd consider the above claim still pretty impressive.
Yeah - our multi TB scan from two decades ago was a dolphin brain image which is a fair bit larger than mouse - and it was a stained sample that ended up being used as our demo image frequently because the contrast dyes set well and resulted in a very pretty visual overall.
I didn't meant to say that PBs of image data is common place - we had no image that approached that size - but people outside the domain of microscope scan results might be unfamiliar with just how chonky these image files would get traditionally.
I think a big thing that the enterprise comparison misses is that this is petabytes worth of dense data that has to be put through fairly heavy processing (some of which is currently custom for the specific experiment) and studied by a human.
It isn't just a giant database of small files and metadata blindly feeding a recommender system.
Several petabytes is definitely a staggering amount of data in that context of being analyzed by human eyes to extract some scientific value.
This fits with my impression of the 'personality' of various models:
Meta: Perceptive (strong vision)
Gemini: Fastest
Claude: Smartest
OpenAI: Prettiest
I would replace Gemini with DeepSeek. I also find OpenAI smarter than Claude but Claude is better for API ergonomics and frontend
They aren't talking about vision LLMs though. SAM and DINO are vision models, no LLM involved.
Why do you think OpenAI is the prettiest?
Claude often makes better looking interfaces and designs. And I think OpenAI has solved more open math/ stats/ CS problems.
Their text output often includes more emojis and seems formatted better. Also their image models subjectively looks better than the rest.
> Claude often makes better looking interfaces and designs
How do you even qualify this? Either by "Well, when you're not specifying anything about it in the prompt" and then it almost doesn't matter at all, or by what actually goes into the prompt, then again it doesn't matter at all what model you use, more about the person driving it.
> And I think OpenAI has solved more open math/ stats/ CS problems.
Still surprises me that OpenAI seems to lead in this one weird niche, I wonder what causes GPT to be able to routinely pull this off, there was one instance where some random 18 year old broke some mathematical question without knowing more than high school math if I remember correctly, all because of GPT.
I love how consistently none of us even vaguely consider Grok an actual player
To Grok's credit I think it's fairly good as a creative writing tool because it can be very "spontaneous" and it naturally seems to use an informal style. It also lacks a lot of the words and phrasing Claude and OpenAI get hyper-fixated on.
IDK if this is emergent from being trained on an endless trough of Twitter shitposts but compared to how stiff the rest are, I consider it a feature. I wouldn't use it for anything important though, heh.
Ah creative writing - the domain where AI is failing to actually generate any economic value while absolutely devastating the ways in which people have traditionally found artistic expression.
The models might be good but the product design, user story, and marketing is so terrible that it’s difficult to see it as more than an also-ran
Even that is kind tbh. The company is so poorly run and the leader so controversial that it makes it irresponsible to build anything serious that relies on their products. Other than the rocket part of the business, everything else under the SpaceX umbrella is a nonstarter.
Starlink arguably is not the rocket part and its the most profitable piece, it held together the rocket side of SpaceX and he expanded research and development.
> The company is so poorly run
In what specific ways is the company very poorly run, say compared to a comparable outfit?
It's so bizarrely run that it's hard to decide what counts as a "comparable outfit"? It's a Space company crammed together with several deeply unprofitable and embarrassing investment companies the founder bought as a part of a transparent grift to cover his mistakes and his only argument is a vague promise of "data centers in space", after over a decade of promising similarly ambitious things that have never happened.
Compared to the other AI companies? It's got even less revenue and product/branding decisions that have it fundamentally incompatible with any of the customers that would result in it becoming a leader.
Compared to other companies of similar revenue? That would be like... uh... Carnival Cruises? lol the fact that it has a valuation where it does is only due to corruption and politics. Oh, sorry Carnival Cruises had actual revenue last year, SpaceX was -$1.9B.
Maybe I'm just in my media bubble, care to share the experts out there explaining how it's a great company whose valuation makes any sort of sense?
Qwen: Zestiest
Nemo: Straightest
DeepSeek: Craftiest
And, of course, Grok: Sir-not-appearing-in-this-listest
Drunkleist?
Muskiest
It's hilarious to me that one man has so single handedly destroyed the concept of "musk" (the smell) being attractive.