Sure. Someone trained a TensorFlow model and deployed it as a web service on a cloud provider. When the boss's demo didn't work they typed "Tensorflow deploy to android device" into Google and copied and pasted Gemini's six-liner for converting the model to Tensorflow Lite.

Translating that to signatures for chemicals is basic spectroscopy which has worked for decades.

That's AI. That type of problem, and the methods you're thinking of, are triumphs of classic AI.

The qustion you're probably asking is why would you use a large language model to do this. The answer is, you wouldn't, that would be stupid. Since they apparently got it working locally on a phone in two hours, I doubt very much that's what they're doing.

The question you might be trying to ask is why would you use machine learning instead of one of those classic engineered solutions.

There are several reasons:

1) ML models are cheaper and faster to develop
2) ML models are usually faster, smaller and more robust than engineered ones
3) That classic "engineered" solution you're thinking of might actually be ML anyway, e.g. a linear regression model (which is a special case of a neural network, BTW).
4) Engineered solutions often rely on not-quite-true approximations, like the system being linear time invariant. Frequently a great deal of money is spent on hardware to make those approximations sufficiently true, and that hardware money can sometimes be saved by using a less restrictive model.

It's the other way around. Everything the GP described is AI, including "just a databse lookup." AI is the field concerned with using machines to solve problems that people are good at and machines are usually bad at. Pattern recognition is a classic AI problem. If you solve it with a database lookup, that's AI. AI describes the problem, not the method used to solve it.

Machine learning is a class of methods that work pretty well on AI problems. The two seem synonymous because you probably wouldn't use ML to solve a non-AI problem and ML-style AI is today so overwhelmingly succesful that anybody who didn't study AI pre-2010s has never even heard of symbolic AI.

Spectroscopy doesn't give you a fingerprint. It gives you a whole bunch of fingerprints all mixed together. One approach is to very carefully measure the spectrogram for very pure samples of everything you're interested in, put it all in a database, then use an optimizer to solve a linear system of all of those possible spectra. That's computationally intensive, often not terribly robust if there's something in there you haven't got a spectrum for, and spectra often aren't really linear combinations anyway.

The alternative is to collect a bunch of data and let a machine learning model figure it out. That model might be linear regression or it might be a neural network. Today you'd probably do the latter because of the not necessarily linear thing. Either way, AI. It's usually faster and smaller than the hand engineered version.

Even fuzzy matching wouldn't need AI

Fuzzy matching IS AI.

BOTH of those are AI. They've been fighting it out for which one is the best for longer than the term "artificial intelligence" has existed.

Yes, it is "all that different". You're comparing conceptual similarity, not data similarity. Data (and usually) appear similar without being conceptually similar, or appear different while actually being conceptually similar. A dolphin, a lungfish, a coelacanth, a tuna, a shark, and a hagfish may all appear visually to be the same sort of "thing" (fish), but they're all actually radically different things. What you actually want to be comparing is the core conceptual aspects of things, not their outward (raw data) appearances.

The SpaceX conglomerate has a lot of pre-IPO investors who are not Elon Musk.

Elon Musk wants to put people on Mars. That's what SpaceX is for, and they're not shy about saying it. SpaceX needed $75 billion and hyping datacentres let them achieve that with Musk surrendering as little of his share as possible. He's already gotten "his money" even though it's in SpaceX's bank account and not his. It will be used to achieve his goals just as much as if it had been paid to him personally in shipping containers full of $100s. Better, given corporate vs. personal income tax rates.

Musk also has a bunch of pissed off investors he convinced to pay $40 billion for Twitter. They were looking at an 80% loss. First he rolled Twitter into a second rate AI company but that didn't really work very well. Pitching space datacentres justified SpaceX acquiring the remains of Twitter, and the IPO gives those investors about a 3x payout, which they are free to start collecting a month from now.

I don't think there's going to be a dump. SpaceX may well want some more money. They will probably design a satellite with big solar panels and a rack of GPUs, and launch some. They might even sell a few for Trump-style Star Wars or something. But I don't think they're going to launch a million of them or make much money doing it, and I think they know it.

Oh, and I forgot to include the "why". TL/DR: it just works better.

To form a latent, the network fundamentally has to learn correlations, in order to be able to accurately compress and then decompress the information in its input space. Due to the limited size of the pinch, you simply can't store every possible data-space representation of a given input. The encoding down to these vectors must persist despite the sort of "real-world" variations that occur in the input space.

Within the latent space (latents are vectors), you can think of individual subconcepts as being directions with in the latent space. In the "king" / "queen" case, for example, adding "man" times some scalar shifts you forward and backwards along degrees of masculinity. In the case of a spectrum, these vectors represent various fundamental molecular properties of the material being scanned. By computing the cosine distance vs. known entries in the vector db and picking the most similar entries, you are in effect minimizing the distance between the molecular properties of the samples.

By contrast, you certainly CAN minimize distances in data space! The problem is that data space lacks correlations. Two things being physically similar in data space doesn't actually mean that much if the things that are different are extremely definitionally-critical aspects. By contrast, definitionally-critical aspects inherently become a key part of vector encoding. Trying to work in text space, for example, a distance algorithm might declare that "king" and "kingdom" are textually similar, but "queen" does not appear at all similar to it. But it's fundamentally similar in vector space.

For YHVH's sake, does nobody in this thread know what a vector DB is?

Translating that to signatures for chemicals is basic spectroscopy which has worked for decades.

The emerging standard for spectroscopy is to translate the spectrum into a latent vector and match against vector keys. For example, Spec2Vec.

This is not what you may at first be thinking of when you see vector, e.g. "the spectrum itself is represented as a vector". These vectors are latents, not spectra.

The most common way to create a latent is via an autoencoder. You task a neural network with recreating its input and put a "pinch" in the middle of the network. This forces the network, to maximize its predictive accuracy, to create a dense conceptual representation of the concepts at hand at the pinch: a latent. The left hand side of the network becomes a latent encoder (detecting concepts in the input space and encoding them into the latent space), while the right hand side becomes a latent decoder (translating the latent space back to data space).

(Note that Spec2Vec is a bit different - it's based on Word2Vec and treats the peaks and neutral losses as "words" and the entire spectrum as a "document", seeking to learn the correlations between peaks - but this is getting out into the weeds)

One of the properties of latents (in most latent spaces) is that they let you interpolate concepts (for the classic example, the latent for "king", minus the latent for "man", plus the latent for "woman", resembles the latent for "queen"), and they let you measure the conceptual distance between concepts (cosine similarity)

Once you have a trained encoder and decoder, a vector database can handle vector queries. The ideal situation would be that all rows would be keyed with vectors and the cosine similiarity to your query key is calculated vs. the table keys, but since that's not practical to query every row, Approximate Nearest Neighbor algorithms are used to approximate that. Hits matching a desired minimum level of similarity are returned. Note that the decoder is no longer needed, you only care about encoding the latents.

Search engines widely use vector databases, BTW, though they usually also combine them with text matching ("hybrid search") - the latter is fast and cheap, so it doesn't hurt to add it.

It's not "fuzzy matching", it's a vector database. You're looking up with conceptual keys into conceptual-indexed tables. The keys are latents, not fuzzy string-matching rules.

Um, no. AI was founded as an academic discipline in 1956, at the Dartsmouth Workshop. The term has been in widespread use for seven decades.

Also, AI != ML. ML is an AI technique. Most AI today is ML, but it wasn't always that way. Symbolic AI, Rules-Based AI, etc used to dominate before NNs advanced enough to push them aside. Expert Systems are a classic example of non-ML AI.

Digging more into RxScanner:

As far as I can tell, 100% of the information out there about its claimed efficacy comes from the manufacturer itself, with zero independent studies. Innovations for Poverty Action (IPA) Nigeria+ USAID + Bloom Public Health initiated a RCT that was supposed to be evaluating it against lab tests with results to be published in April 2025, but as far as I can tell, the results were never published.

From a technological perspective, if it is what it says at all, one would expect that it could be pretty good at confirming something as genuine and in pristine condition but pretty bad at telling if something is counterfeit. E.g. a generic of the same drug, or a slight formulation change (such as the binder or coatings), or some meaningless degradation during storage, will give a different result. Sounds like they're trying to bypass this by using AI to tell what differences are meaningful or not. But without third party validation.... *shrug*

Also, the manufacturer initially marketed itself as a hardware and SaaS company, but has now been broadening out, making e.g. a B2B marketplace where the sell drugs to pharmacies (RxDelivered) and a financing and point-of-sale software service for those pharmacies (RxPay). So as far as supply chain auditing goes, i's a red flag when the "independent" quality tester also operates the marketplace selling the goods (something that they try to spin as a "trusted network"). Because they obviously benefit from any biases to detect their goods as legit but flag competitors as fake.

As it stands, I don't think one can say it's an outright grift, but on the other hand, I wouldn't put too much trust in this company as it stands.

This is not fuzzy string matching. This is a vector database. Please learn the difference.

Vector databases return conceptual matches, not fuzzy data matches (such as fuzzy string matches).

Also: "The results of drug sample scans are recorded and stored on the blockchain", which isn't something that affects the choice of AI models, but does make me immediately suspicious that this whole thing is a grift / scam.

What are you talking about? You generally run inference on the GPU. Sometimes in RAM. Almost never on swap. It is certainly possible to run a model without being able to load the whole thing into memory, but it's comically slow.