Former Google Researcher's Startup Hopes to Teach AI How to Smell (cointelegraph.com) 42
"AI is already able to mimic sight and hearing," writes CNBC. And now a startup named Osmo "wants to use the technology to digitize another: smell."
Co-founded by a former Google research scientist, the company built an AI that's "superhuman in its ability to predict what things smelled like," the company's co-founder says. And he believes this might actually prove useful. "We've known that smell contains information we can use to detect disease. But computers can't speak that language and can't interpret that data yet... We will eventually be able to detect disease with scent and we're on our way to building that technology. It's not going to happen this year or anytime soon, but we're on our way."
CoinTelegraph describes how the company invented a training dataset from scratch — a kind of "smell map" with labelled examples of molecular bond associations to teach the AI to identify specific patterns. The team also hopes to develop a method to recreate smells using molecular synthesis. This would, for example, allow a computer in one place to "smell" something and then send that information to another computer for resynthesis — essentially teleporting odor over the internet. This also means scent could join sight and sound as part of the marketing and branding world.
Co-founded by a former Google research scientist, the company built an AI that's "superhuman in its ability to predict what things smelled like," the company's co-founder says. And he believes this might actually prove useful. "We've known that smell contains information we can use to detect disease. But computers can't speak that language and can't interpret that data yet... We will eventually be able to detect disease with scent and we're on our way to building that technology. It's not going to happen this year or anytime soon, but we're on our way."
CoinTelegraph describes how the company invented a training dataset from scratch — a kind of "smell map" with labelled examples of molecular bond associations to teach the AI to identify specific patterns. The team also hopes to develop a method to recreate smells using molecular synthesis. This would, for example, allow a computer in one place to "smell" something and then send that information to another computer for resynthesis — essentially teleporting odor over the internet. This also means scent could join sight and sound as part of the marketing and branding world.
Why!? (Score:2)
Re: (Score:2)
Here's your stinky AI joke (Score:2)
Person: How does AI smell?
AI: The content was filtered due to Responsible AI restrictions.
Person: Must be pretty bad if you need a filter.
Re: (Score:2)
Person: How does AI smell?
AI: The content was filtered due to Responsible AI restrictions.
Person: Must be pretty bad if you need a filter.
More like:
Person 1: How does AI smell?
Person 2: It stinks.
Re: (Score:2)
Re: (Score:2)
Cripes, they could probably finance the entire effort just by approaching natural gas companies and offering to equip drones that would patrol pipelines for leaks. Over 4% of natural gas escapes during pipeline transport, which adds up to 1,454,120 cubic feet per year just in the US. Some old pipelines lose 10% or more.
Re: (Score:2)
Calling it smell is a little bait-y, yeah, but smell is just chemical analysis, and there are plenty of useful applications for it.
Right, like replacing canines for sniffing out drugs or explosives or escaped convicts or lost children. There is even research for detecting cancer via smell.
The use cases are not the problem. The problem is getting the tech to work practically.
Re: (Score:2)
My AI's go no nose (Score:5, Funny)
Re: My AI's go no nose (Score:2)
Could it replace dogs? (Score:2)
That's the holy grail. Dogs can smell cancer, drugs, food pests, cash, etc.
Re: (Score:2)
No, it can't. Dogs are beings with real intelligence and empathy and are generally very nice to people. Much nicer than people deserve.
Re: (Score:2)
AI indeed does smell.
Re: Further proof... (Score:3)
Re: (Score:2)
Well, judging from their investment decisions they probably don't qualify on the Sapiens part.
This has enormous potential for abuse (Score:3)
On the minus side, I propose that a good odor detector could sniff out specific individuals, even in a crowd. Perhaps with augmentation by, or in addition to, facial recognition. No amount of cologne would mask to a detector the distinctive combination of chemicals given off by an individual. The only drawback would be that the AI would need the odor signature of the target.
Re:This has enormous potential for abuse (Score:4, Interesting)
>Probably not as fast or good as a dog though.
I disagree. Even if you don't have any knowledge of biology (and though I have some I'm not claiming to have much here), it stands to reason that a biological nose doesn't detect an infinite number of compounds. Noses evolve to detect molecules that affect the survival of the organism to which they are attached. Anything more than that lacks an evolutionary feedback mechanism to conserve such a trait.
Now, imagine each molecule a nose can detect is a letter, and different combinations of letters make words that represent complex substances of significance. The brain attached to a nose will have an instinctive 'dictionary' for the critical stuff that could affect survival - smelling poison, toxic gas, rotting meat... or fresh food, mates, etc.
If you can build a sensor that can detect all the letters, and plug it into an AI, you'll train it to know all the words you want it to know. And it'll be faster than biological noses.
I would actually see the issue being hardware, not software. Biological sensors can renew themselves. Artificial ones might wear out rather quickly for some types of molecules.
Re: (Score:1)
Here's a thing [cbsnews.com] A more recent thing [nih.gov] that suggests we (humans) can distinguish more than 1 trillion scents.
I wouldn't count biological noses out just yet. They're pretty damn amazing. [nih.gov]
As far as electronic noses go, it seems to me that the hardware would be the more interesting part, not the software. As far as AI goes, it's just a classification task. It's hardly ground-breaking. Hell, we've been doing it for decades. [sciencedirect.com] "ANN + Electronic nose to detect thing" is practically a science fair project these days
Re: (Score:2, Troll)
A trillion is no bigger than a game of 40 questions.
Stop being impressed by "large number", like those simpleton "journalists" you are parroting.
Re: (Score:1)
Re: (Score:2)
I'm going to go out on a limb here and say that AI isn't the missing piece of the puzzle. I'll bet dollars to donuts that we don't yet have an electronic nose that can compete with biology. If this startup is focusing on the software side and not the hardware, it's doomed to failure.
Probably true. Smell isn't like taste, where you have five basic types of receptors. The human nose appears to have about 400 different types of olfactory receptors, and presumably they all have different chemical signals they respond to. (Actually, it might be more than 400-- what I read is that there are 400 functional *genes* for olfactory receptors, and you could perhaps use those to create more than 400 different types of receptor, e.g. by mixing and matching subunits).
So to match the performance of
Re: (Score:2)
I would actually see the issue being hardware, not software. Biological sensors can renew themselves. Artificial ones might wear out rather quickly for some types of molecules.
Yeah, you'd probably need replaceable nose cartridges.
Re: (Score:3)
>Probably not as fast or good as a dog though.
I disagree. Even if you don't have any knowledge of biology (and though I have some I'm not claiming to have much here), it stands to reason that a biological nose doesn't detect an infinite number of compounds. Noses evolve to detect molecules that affect the survival of the organism to which they are attached. Anything more than that lacks an evolutionary feedback mechanism to conserve such a trait.
Now, imagine each molecule a nose can detect is a letter, and different combinations of letters make words that represent complex substances of significance. The brain attached to a nose will have an instinctive 'dictionary' for the critical stuff that could affect survival - smelling poison, toxic gas, rotting meat... or fresh food, mates, etc.
If you can build a sensor that can detect all the letters, and plug it into an AI, you'll train it to know all the words you want it to know. And it'll be faster than biological noses.
I would actually see the issue being hardware, not software. Biological sensors can renew themselves. Artificial ones might wear out rather quickly for some types of molecules.
The thing about the dog isn't the nose (which is a marvel to be sure) but the fact the dog is easily trainable and adaptable. The dog can think for itself. A dog can be trained to look for certain compounds (or people in the case of SAR) but can act on it's own and sometimes pro actively. The flipside is the dog has trouble communicating with humans and can be distracted.
We can already use technology to detect compounds, CO2 emissions, explosive chemicals, et al. the problem is a computer has no idea wea
Invest your VC millions now before it is too late! (Score:2)
And with someone from Google you can't go wrong because only about a million engineers worked for Google once.
Literally Smellovision (Score:2)
Re: (Score:2)
Congratulations. You just reinvented Smell-o-Vision, that thing literally no one wanted. OTOH, I'm sure someone will decide this is a great way to sell fart porn or something.
That's where we fire a Cruz Missile and follow up with a Trump Dump.
where the meat is still better than the machine (Score:2)
I think it's pretty surprising that in this day and age we're still using animals for pretty much anything scent-related.
- tracking lost people
- bomb detection
- foraging for mushrooms
- drug sniffing dogs
And there are apparently animals trained to smell diseases. Though it's hard to tell WHAT they're smelling. Animals can be a bit like AI.... it can WORK but you can't just lift the hood and look in and see HOW or WHY it's working. But AI training works much the same way as biological learning processes.
How does AI smell without a nose ? (Score:2)
The AI smells (Score:2)
Congratulations (Score:2)
A robot version of ... (Score:2)
Nothing New (Score:2)
As other comments confirm, this is just "something that already exists (to a certain extent) plus AI". Detecting smells would be useful, sure. What *I* would be more interested in would be the device at the "other end" that can produce a smell from a database. How would that be achieved? Are there a certain number of "base" smells that can be combined in various parts to produce the smell in question? I can't see that being the case, so how can one "replicate" a smell? Not a limited set of smells (Smellovis
That's easy (Score:2)
Just teach ChatGPT not to wash for a few weeks and the smell comes all by itself.
I applaud this! (Score:2)
police tool? (Score:2)
"Now that I've pulled you over, the wearable AI I have on smells weed."
As bad as that sounds, I'd trust AI smelling alcohol or weed 1000 times more that a rando cop that just pulled me over at 1AM. That said, all you need to beat it is an air freshener that comes up "weed smelling positive" to say while the AI is accurate, it removes any probable cause.
Plugins (Score:1)
The team also hopes to develop a method to recreate smells using molecular synthesis
Does this require cartridges? or does it just "work"? because if it just works this will revolutionize the home scent business. Imagine being able to set your home's smell with home automation?
Not to mention being able to try on cologne before you buy it, with it smelling you, calculating the difference, and letting you smell what it smells like on you.