[amelius]

The main invention is in the electronic nose. The kid just did the plumbing of connecting it to some ML library.

Of course, the electronic nose itself is a work of plumbing too, where some existing gas sensors are put on a pcb.

In short, nothing seems really new here, but the application is interesting. I guess it's always interesting when people start looking for correlations in data and get some positive results, so from that point of view it is noteworthy.

[hwillis]

The sensor (the four gas sensors on the board) was created by a third person.

The artificial nose is the TinyML model which trains on the sensor data (CO, NO2, ethanol, VOCs) to detect arbitrary scents by their signatures in those four categories.

The fungal pneumonia detector wires up a whole API with azure etc. and trains the model specifically to recognize pneumonia, based on an actual science experiment which grew and measured the fungus in artificial lungs.

As far as I'm concerned, both Caleb and Benjamin had brilliant ideas, executed them fantastically, and created something that may be truly useful. A $40 sensor that can detect disease just by breathing on it is more of a tangible contribution to humanity than many software engineers make in their life and almost certainly more than 99% of us did before the age of 14.

[wesleywt]

It is great that the kid is involved and interested in these technologies. Whether it works is another issue. You are going to need metrics such a LODs, LOQs, sensitivity and specificity to determine if this beats the gold standard tests.

[amelius]

The reason why no engineer has made this yet, is because the medical data was not available to them.

The innovation is in the data, not in the ML.

[hwillis]

You are wrong. The kid GREW FUNGUS IN ARTIFICIAL LUNGS using a sterile field made out of a plastic bin with dish gloves cut into it. He didn't need medical data. Anyone could have done this- anyone with the intelligence and creativity that this kid has.

[dekhn]

is it possible you're overstating what the kid did a bit here? The history of sterile fields shows that even great scientists take decades to debug contamination that causes false positives and negatives. ASn experiment like this can be easily thrown off by any number of variables that weren't carefully controlled for.

[amelius]

You need medical data if you want to validate your results for a real medical disorder.

Fabrication of data is not very useful if you have to gather data for validation anyway.

[pmarreck]

> The innovation is in the data, not in the ML.

can't you kind of say this about all ML? That the main driver in ML is 99% the quality of the training data and 1% the specific details of the neural networks used?

[amelius]

For most real-world applications of ML, yes. Of course, what happens in ML-research is different (where e.g. new networks for new modalities are invented).

But back to the topic, I bet the kid didn't even invent their own neural network topology, but just pulled a predefined network from a library, perhaps without even knowing it. Which is ok, because that is how most people use ML.

[dkersten]

I think the news is less "13 year old revolutionized medicine" and more "13 year old used ingenuity to create cool thing". Or at least, that's how I think it should be read. Focusing too much on the end result is likely not the best outcome because, while its a working prototype, many interesting prototypes never make it to fully end user capable system and there are a lot of hurdles to overcome to get it there.

But that doesn't take away from the fact that its a cool project and the kid did a great job in coming up with it and executing on it! Its definitely far beyond what most people achieve, nevermind 13 year olds.

[SubiculumCode]

This is a story about inspiration and achievement. Objective facts are less important than the message, IMO, especially after decades of "try-hard" being ridiculed in the U.S.

[deckar01]

No, the author designed the ML training project also. The last section of the Make article is how to send the data to Edge Impulse and configure the ML training [0].

Mentioning Microsoft Azure IoT Central in the article and the video is odd, because you definitely don't need that to complete this project. It seems to be a feature that the Microsoft employee added to the GitHub project their self [1].

Caleb mentioned in the video that a co-worker of his aunt authored the research paper about detecting bacterial pneumonia from VOC levels. Everything else feels like a Microsoft marketing hype train that went off the rails.

[0]: https://makezine.com/projects/second-sense-build-an-ai-smart...

[1]: https://github.com/kartben/artificial-nose/blob/master/firmw...

[kartben_]

With this reasoning, aren't most things a work of plumbing, and nothing ever really new? And isn't it how innovation happens, at the end of the day?

[tinc293]

Well, first of all, you can see those glove holes are too small, that's not the hands of a 13-year-old who mines his own silica. Disgraceful.

[dekhn]

The vast majority of work is plumbing, but there are still new things.

What matters isn't making new things, it's making new things that work, and bringing them all the way to completion.

The vast majority of innovation isn't exciting discovery, it's coming up with tests to convince yourself you actually did what you thought you did and sharing those results with regulators. That's the difference between Theranos and GRAIL.

I just built a tardigrade detector neural network; nobody has done that exact thing before, and now i'm talking to the world's leading tardigrade researcher because my tool might be helpful in answering an unanswered important tardigrade question.

But thousands if not millions of individuals worked over hundreds to thousands of years to bring us that tardigrade detector; all I did was take advantage of that to label 100 images! I've tried being an ML researcher; my conclusion is the best models are distilled from postdoc tears, and we merely retrain those models without pain.