Small AI Models Gain Traction Around the World

Small AI Models Gain Traction Around the World 49

locater16 shares a report from IEEE Spectrum: One morning in 2019, Adebayo Alonge was in a Cape Town hotel room, preparing to demonstrate his startup's AI answer to a serious problem in African health care: counterfeit medication, which kills thousands of people across the continent every year. The RxScanner is a handheld spectrometer that scans a pill with infrared light, then sends the item's molecular profile to an AI model equipped with a pharmaceutical database. In seconds, the AI identifies the medication from its molecular profile -- or reports that it's phony.

Pharmacies were using the system in more than a dozen countries, including Ghana, Kenya, Myanmar, and Alonge's native Nigeria. But that morning in South Africa, it didn't work. "I was shocked," Alonge says... So Alonge immediately asked his engineers to shrink the AI model down to a smaller, low-power, unconnected version that could run entirely on his Android phone. They produced it 2 hours later, and that saved the demo. More importantly, the work birthed a new version of his device, which can authenticate a pill in places without broadband, computers, or even reliable electricity. It also turned Alonge into an advocate for this kind of "small AI." "The article goes on to detail other immediately useful 'small' AI applications without any subscription or billion dollar data centers needed," writes locator16. For example, Bala Murugan and colleagues at Vellore Institute of Technology in India developed a drone-based system that photographs cashew plants and identifies disease-indicating splotches on the plants. The key advantage is that all processing happens on the drone itself, so farmers do not need a computer, broadband connection, or cloud server access.

In a Uruguayan vineyard, researchers developed small-AI systems to identify ant infestations. The article doesn't go deep into the deployment details, but it presents this as another example of a narrow, localized model trained to recognize a specific agricultural threat. Small AI has also been used to detect the presence of malaria-carrying mosquitoes in multiple countries. This is especially useful in regions where public-health teams may lack reliable network access or expensive lab infrastructure, but still need fast, local detection.

In parts of Brazil without access to more complex medical equipment, researchers have used small AI to run electrocardiograms from an Arduino device. The article also describes Marcelo Jose Rovai's work on a TinyML model that generates electrocardiograms in a patient simulator lab. Rovai also describes a newer experiment using an Arduino UNO Q with a Qualcomm chipset. The device runs a language model locally, collects sensor data, and analyzes it to detect tiny pools of water where mosquitoes might breed -- while using only about 3 watts of power.

Farming (Score:1)

I'm following a Japanese farmer on Twitter that posts about how he uses AI around the farm. It's fascinating.

Re: (Score:3)

Several of them will eat your SSD (Score:2)

Basically if you're going to run one of these you need to check for that and make sure that if you're using a model that had that bug that it's been addressed. Now is not the time to have a piece of software chew through your ssds lifespan. Prices are basically twice what they would normally be without the AI slop apocalypse.

On the other hand if you're t

Re: (Score:2)

Chances are this is a configuration issue.

For instance, I recently saw a case of this with LMStudio. It turns out that LMStudio doesn't properly handle K/V caching when paired with Claude Code. The solution? Literally the exact same model, but using Ollama as the server application instead. Problem instantly solved.

The other big reason for SSD thrashing is not having enough RAM to handle the model, context, and cache. If you do, then your disk shouldn't be touched at all for model work other than initially

Ok cool (Score:3)

Re:Ok cool (Score:4, Informative)

But why does rxscanner need AI? It's a scanner and a database lookup. The AI is redundant no?

from Gemini

Pattern Recognition: Unlike a standard database that only looks for exact matches, AI looks for patterns.

Tolerance Handling: The AI accounts for minor variations, such as humidity, pill age, or scanner angles.

Anomaly Detection: Machine learning algorithms instantly flag if active ingredients are missing, diluted, or replaced with toxic fillers.

Re: (Score:1)

Pattern Recognition:

James C. Bezdek's "Pattern Recognition with Fuzzy Objective Function Algorithms" (1981) is a seminal work in the field of fuzzy clustering and pattern recognition.

Tolerance Handling:

Variables can be assigned as ranges rather than absolute values.

Anomaly Detection:

That's why you have the database lookup.

I don't see any AI here. It's just basic computer programming as it's been done for decades.

Re: (Score:3)

But why does rxscanner need AI? It's a scanner and a database lookup. The AI is redundant no?

I was about to ask a similar question, but reading your wording suggested an answer to me.

I could be totally wrong, but I'm guessing that the advantage of AI may lie in its ability to recognize patterns. For example, certain arrangements of molecules or structures or whatever may show up as a problem just because they look similar to known problems, toxins, etc.

Conversely, maybe if the similarity is to known good patterns, then the substance or item in question is likely to be beneficial, or at least safe.

Re: Ok cool (Score:2)

Re:Ok cool (Score:5, Informative)

On top of that, there is a lot of clutter in those spectral bands, not to mention measurement problems caused by varying reflectivity and light levels between samples taken in the wild

So it isn't as simple as taking a scan and comparing the spectra you measure to a database, there is a lot of noise.

Machine learning/AI models to help with this are quite common in this field, and have been around for decades to help with spectral library lookups - long before the current LLM hype phase.

Re: Ok cool (Score:2)

Drug interactions (Score:2)

But why does rxscanner need AI? It's a scanner and a database lookup. The AI is redundant no?

Drug interactions. ML is probably overkill, an expert system (which is AI too) might be sufficient. But to get investor funding for the development project you probably need to use ML.

Re: (Score:2)

We used to call this simply, "Machine Learning." Pattern recognition using neural nets, etc. Always was more than a database lookup. Now it's called AI.

Just A Reminder (Score:1)

Even if these are smaller AI models, they are still likely trained on cloud hardware. The kind that increases water and energy prices [lincolninst.edu], the kind that cause dirty air from the generators [theinvadingsea.com], the kind with robot dogs protecting the premises dystopian terminator style [vice.com]

Maybe the offset of the people the hospital saves with the AI makes up for those who die around the datacenter? This all looks like a dystopian tradeoff to me.

Re: (Score:2)

Maybe the offset of the people the hospital saves with the AI makes up for those who die around the datacenter? This all looks like a dystopian tradeoff to me.

I tend to agree. But I maintain the hope that such developments may ultimately reduce the need for huge data-centres.

But unfortunately, the creators of said data-centres consider the damage they cause to local populations a feature and not a bug. For a far better and wider reaching explanation than I can give, I recommend Benn Jordan's exploration of both capitalism, and the system which is replacing it at a rapid pace even as you're reading this: https://www.youtube.com/watch [youtube.com]

Re: (Score:1)

anything that uses fuzzy matching is now described as AI because $$$. stupidity.

Why did that need AI? (Score:2)

Please, people here who are expert on databases or AI, help me understand this.

handheld spectrometer that scans a pill with infrared light, then sends the item's molecular profile to an AI model equipped with a pharmaceutical database

The scan results in a linear array of data. Translating that to signatures for chemicals is basic spectroscopy which has worked for decades. Correlating that with a database of real versus bogus signatures seems like basic lookup or a DB query, which has also worked fine "forever". I don't understand why this needs an AI solution, or why AI would be better than traditional methods.

The other stories about aerial scanning for ca

Re: (Score:2)

In summary, no. You aren't missing anything.

Re: (Score:2)

Like with LLMs, these models can accommodate unexpected variations.

Consider the classic Alexa or Siri assistants. To get them to do what you wanted, you had to use very precise phrasing, or it wouldn't know what you meant. Now that these assistants are AI-backed, you can use more natural language, without having to remember the precise phrases.

I'm guessing there are similar variations in fake drugs. Maybe differences in the inactive ingredients, for example, could throw of a classic spectrographic analysis.

Re: (Score:2)

The fact it's running on a statistical system doesn't automatically mean it knows which of the token interactions it has represent, for example, fractions. There's no reason to believe it'd be better at actually performing deductive statistical analysis any more than to to believe it's using actual integers to perform arithmetic.

That is to say... even if it is incidentally ok at it, there are going to be some absolute financial disasters in the foreseeable future that are due to humans making bad decisions

Today's AI is just Automation! (Score:2)

All these companies thought they were so close to AGI that a little fib in the beginning wouldn't matter!

But here we are years later still talking Intelligence when there isn't any!

When true AGI comes, I doubt that huge centralized data centers will be the main component.

Too qualified for the job that makes LLMs useless (Score:2)

Just like overqualified musicians that are made to play children's songs will go crazy :

https://youtu.be/sReGQj67HJU?t [youtu.be]

It's machine learning, not Gen AI (Score:1)

I figured that this was just case of wide-eyed-with-astonishment journalists going "Oh, AI that doesn't need big data centres?!" and totally unaware that ML has been around for a very, very long time, while simultaneously being confused by the now almost meaningless umbrella term "AI" and assuming that meant something like Claude or ChatGPT or whatever.

Then I read the comments, and realised most Slashdotters apparently don't know the difference either.

Re: (Score:2)

Well, training Gen AI, including LLMs, is machine learning. I understand that you're annoyed that the "AI" label is put on everything and their mother, but it seems the current transformer-based deep "neural" networks are stuck with that label, so we might as well go along. A very small (parameter-wise) such model is still "AI" in that sense. So let's not get hung up too much on names.

Overall, I think it is a very positive trend for people to not simply use the big models for everything but to come up wit

What is small AI? (Score:1)

A database lookup? Nested IF...THEN...ELSE or SWITCH statements? What? Let's attach the term "AI" to the most ridiculous things we can, shall we? That will convince people that "AI" is good and has been here all along. sigh

Re: (Score:2)

Obviously a small model (fewer parameters). Come on, it's not that hard to understand. They got Qwen down to 4b and it still works, so I am not surprised that it works for small application areas such as this.

"They produced it 2 hours later" (Score:2)

They took a cloud-dependent model and shrank it to an Android-size phone in 2 hours?

Sentinel — Human