Even analytical AI needs to be questioned and validated before use.
I wouldn’t trust a AI to ID mushrooms for consumption.
I forget the details, but there was a group training a diagnostic model (this was before “AI” became the popular term), and it was giving a lot of false positives. They eventually teased out that it was flagging low quality images because most of the unhealthy examples it was trained on came from poorer countries with less robust healthcare systems; hence the higher rates of the disease and lower quality images from older technology.
I've seen a similar thing where the machine learning model started associating rulers with cancer because the images it was fed with known cancer almost always also had a ruler to provide scale to measure the size of the tumor
It's like these geoguesser not guessing the country by the plants and streets or houses, but by the camera angle and some imperfections only occuring in pictures taken in that country.
"When a measure becomes a target, it ceases to be a good measure"
People are saying you shouldn't use AI to identify edible mushrooms, which is absolutely correct, but remember that people forage fruits and greens too. Plants are deadly poisonous at a higher rate than mushrooms, so plant ID AI has the potential to be more deadly too.
And then there's the issue that these ID models are very America and/or Europe centric, and will fail miserably most of the time outside of those contexts. And if they do successfully ID a plant, they won't provide information about it being a noxious invasive in the habitat of the user.
Like essentially all AI, even when it works it's barely useful at the most surface level only. When it doesn't work, which is often, it's actively detrimental.
I actually think AI for mushroom identification is okay, but as a step in the process. Sometimes you see a mushroom and your like "what is that?". Do a little scan to see what it is. Okay now you have an idea of what it is, but then comes the next part! https://mushroomexpert.com/ there you can go through the list and see if you get a positive ID.
Like if you're not 100% positive you know what you're foraging why would you take the risk.
Just dealt with an AI bot this morning when I called a law office. They try so hard to mimic humans. They even added background people talking sounds. But was 100% easily given away when they repeated the same response to my asking to speak with a human. "I will gladly pass on your message to (insert weird pause) "Bill""
There was an interesting story on NPR last week about someone experimenting with AI agent clones of himself. Even his best attempt sounded pretty obvious thanks to stuff like that.
probably because Google lens is made to be all-purpose, if you had a model that has been specifically trained to recognize plants, it wouldn't make such obvious mistakes
Try out Flora Incognita, it usually works pretty well, at least in the UK. Though I'm kinda bemused that while often I give it one blurry photo of a leaf and it immediately spits out the 97% confidence answer (and yes, the result's example photos do match the real plant's other parts too), but occasionally it has me take painstaking photos of like 5 different aspects of the plant, then thinks for more than a minute and goes "eeh, 37% that it's this [completely different looking] plant, otherwise no clue".
Flora incognita is a nice one I've tried. It's a spinout from a university. The best use is to take its guess and then click through, and you can see example images of the plant it is guessing so you can judge for yourself
If it looks like a carrot, don't touch it, to dig it up, and especially don't eat it. There are tons of plants in the same familuly of plants that look nearly identical or extremely similar that will give you an extremely bad day, month, year, or death.
"AI"'s best uses are the Machine Learning aspects we were already using before we started calling things AI
It's so painful watching this tech be forced down our throats by marketing departments despite us already discovering all of its best use cases long before the marketing teams got ahold of the tech.
Exactly this. GenAI is pretty much dogshit all around. Machine learning as a concept is nothing new at all, and shoving it under the same marketing umbrella does so much more harm than good.
One of the issues with LLM is that it attracted all attention. Classifiers are generally cool, cheap and saved us from multiple issues (ok face recognition aside 🙂)
When the AI bubble will burst (because of LLM being expensive and not good enough to replace a person even if they are good in pretending to be a person) all AI will slow down… including classifiers, NLP, etc
All this because the AI community was obsessed by the Turing test/imitation game 🙄
Turing was a genius but heck if I am upset with him for coming with this BS 🤣
It made sense in the context it was devised in. Back then we thought the way to build an AI was to build something that was capable of reasoning about the world.
The notion that there'd be this massive amount of text generated by a significant percentage of the world's population all typing their thoughts into networked computers for a few decades, coupled with the digitisation of every book written, that could be stitched together in a 1,000,000,000,000-byte model that just spat out the word with the highest chance of being next based on what everyone else in the past had written, producing the illusion of intelligence, would have been very difficult for him to predict.
Remember, Moore's Law wasn't coined for another 15 years, and personal computers didn't even exist as a sci-fi concept until later still.
If you ever are an executive and you need to explain a product idea you made up and don't want to bother with actual proof of concept, then AI has got you.
If you want custom porn, AI has got you.
These are its two competing functions. And judging by my last foray into what AI is about, the latter is winning hard.
But there is endless technology that grifters label as "AI". Ofc some of this technology will be useful. But under capitalism all technology is developed by and for the benefit of the disgustingly privileged via violent control.
I've found so many fucking errors in AI summaries that I don't trust shit from AI searches when a direct link to a source or wiki could give me better summarized info.
I guess it's an evolution, but I'm really hoping these mutations prove inferior and it dies off already. But capitalism won't have that with their sunk cost fallacy driven insistence that I just use the inferior product.
That’s why I said you have to ask for the source. It’s summary isn’t good but that you can describe something to it in human language instead of focusing on keywords to start your search and that the sources it gives aren’t just ads yet is useful.
The difference is that plant identification is no longer an interesting area for AI research. It was ”AI” 10 years ago, but now it’s more or less a solved problem.
Honestly I like when it writes for me too. I tend to be very blunt and concise in my messaging and AI just puts that corporate shine and bubbliness on my messages that everyone seems to feel is important.
59 Comments
Takeshidude@lemmy.world · 151 pts · 332d
analytical AI is great
generative AI is cancer
MyTurtleSwimsUpsideDown@fedia.io · 73 pts · 332d
Even analytical AI needs to be questioned and validated before use.
I wouldn’t trust a AI to ID mushrooms for consumption.
I forget the details, but there was a group training a diagnostic model (this was before “AI” became the popular term), and it was giving a lot of false positives. They eventually teased out that it was flagging low quality images because most of the unhealthy examples it was trained on came from poorer countries with less robust healthcare systems; hence the higher rates of the disease and lower quality images from older technology.
Ageroth@reddthat.com · 35 pts · 331d
I've seen a similar thing where the machine learning model started associating rulers with cancer because the images it was fed with known cancer almost always also had a ruler to provide scale to measure the size of the tumor
shrugs@lemmy.world · 15 pts · 331d
It's like these geoguesser not guessing the country by the plants and streets or houses, but by the camera angle and some imperfections only occuring in pictures taken in that country.
"When a measure becomes a target, it ceases to be a good measure"
monogram@feddit.nl · 19 pts · 332d
ceenote@lemmy.world · 33 pts · 332d
Mass surveillance is bad regardless of whether or not AI is part of it.
technocrit@lemmy.dbzer0.com · 5 pts · 332d
Well there's no "AI" so we don't have to worry about it. Just a bunch of disgusting rich bros spying on us or worse.
technocrit@lemmy.dbzer0.com · 7 pts · 332d
This kind of wacky absolutism is part of the problem. Esp since "AI" doesn't even exist.
https://www.downtoearth.org.in/governance/lavender-wheres-daddy-and-the-ethics-of-ai-driven-war
s@piefed.world · 54 pts · 331d
Definitely do not use AI or AI-written guidebooks to differentiate edible mushrooms from poisonous mushrooms
Zoomboingding@lemmy.world · 29 pts · 331d
It clearly says "plant identification"~
SuperIce@lemmy.world · 8 pts · 331d
Could still have difficulty accurately identifying the difference between a delectable tea and a deadly poison though
greedytacothief@lemmy.dbzer0.com · 3 pts · 331d
Or you found laborador tea and it's both delectable and poisonous!
ColeSloth@discuss.tchncs.de · 1 pts · 331d
Touche
Aeao@lemmy.world · 8 pts · 331d
Honestly don’t use any guide book or advice. If you aren’t 100 percent sure on your own maybe just walk away.
My self personally… even if I was 99.999 percent sure it still wouldn’t be worth the risk. I’ll just buy some mushrooms.
punkfungus@sh.itjust.works · 36 pts · 331d
People are saying you shouldn't use AI to identify edible mushrooms, which is absolutely correct, but remember that people forage fruits and greens too. Plants are deadly poisonous at a higher rate than mushrooms, so plant ID AI has the potential to be more deadly too.
And then there's the issue that these ID models are very America and/or Europe centric, and will fail miserably most of the time outside of those contexts. And if they do successfully ID a plant, they won't provide information about it being a noxious invasive in the habitat of the user.
Like essentially all AI, even when it works it's barely useful at the most surface level only. When it doesn't work, which is often, it's actively detrimental.
greedytacothief@lemmy.dbzer0.com · 5 pts · 331d
I actually think AI for mushroom identification is okay, but as a step in the process. Sometimes you see a mushroom and your like "what is that?". Do a little scan to see what it is. Okay now you have an idea of what it is, but then comes the next part! https://mushroomexpert.com/ there you can go through the list and see if you get a positive ID.
Like if you're not 100% positive you know what you're foraging why would you take the risk.
crazycraw@crazypeople.online · 36 pts · 332d
rage against the machine learning
Gerudo@lemmy.zip · 2 pts · 331d
Good nerdcore band name
ThePantser@sh.itjust.works · 29 pts · 332d
Just dealt with an AI bot this morning when I called a law office. They try so hard to mimic humans. They even added background people talking sounds. But was 100% easily given away when they repeated the same response to my asking to speak with a human. "I will gladly pass on your message to (insert weird pause) "Bill""
A_Union_of_Kobolds@lemmy.world · 10 pts · 332d
There was an interesting story on NPR last week about someone experimenting with AI agent clones of himself. Even his best attempt sounded pretty obvious thanks to stuff like that.
DeathByBigSad@sh.itjust.works · 3 pts · 331d
They even add fake keyboard typing sounds now 💀
Blackfeathr@lemmy.world · 22 pts · 332d
Idk, when I Google Lensed a Sunflower plant, the AI told me it was a Peruvian ground apple....
It also has a lot of trouble identifying lambs quarters and other common wild weeds.
renzhexiangjiao@piefed.blahaj.zone · 19 pts · 332d
probably because Google lens is made to be all-purpose, if you had a model that has been specifically trained to recognize plants, it wouldn't make such obvious mistakes
Fiivemacs@lemmy.ca · -3 pts · 332d
doubt
kazerniel@lemmy.world · 2 pts · 331d
Try out Flora Incognita, it usually works pretty well, at least in the UK. Though I'm kinda bemused that while often I give it one blurry photo of a leaf and it immediately spits out the 97% confidence answer (and yes, the result's example photos do match the real plant's other parts too), but occasionally it has me take painstaking photos of like 5 different aspects of the plant, then thinks for more than a minute and goes "eeh, 37% that it's this [completely different looking] plant, otherwise no clue".
nocturne@slrpnk.net · 15 pts · 332d
Try seek by inaturalist
skavj@lemmy.zip · 3 pts · 331d
Flora incognita is a nice one I've tried. It's a spinout from a university. The best use is to take its guess and then click through, and you can see example images of the plant it is guessing so you can judge for yourself
nocturne@slrpnk.net · 1 pts · 329d
Downloaded it, will give it a try.
Cevilia@lemmy.blahaj.zone · 15 pts · 331d
odelik@lemmy.today · 4 pts · 331d
Word to the wise.
If it looks like a carrot, don't touch it, to dig it up, and especially don't eat it. There are tons of plants in the same familuly of plants that look nearly identical or extremely similar that will give you an extremely bad day, month, year, or death.
SoftestSapphic@lemmy.world · 15 pts · 331d
"AI"'s best uses are the Machine Learning aspects we were already using before we started calling things AI
It's so painful watching this tech be forced down our throats by marketing departments despite us already discovering all of its best use cases long before the marketing teams got ahold of the tech.
nfreak@lemmy.ml · 1 pts · 331d
Exactly this. GenAI is pretty much dogshit all around. Machine learning as a concept is nothing new at all, and shoving it under the same marketing umbrella does so much more harm than good.
Jankatarch@lemmy.world · 13 pts · 331d
Honestly good rule about Machine Learning is just "predicting = good, generating = bad." Rest are case by case but usually bad.
Predict inflation in 3 years - cool
Predict chance of cancer - cool.
Generate image or mail or summary or tech article - fuck you.
Generating speech from text/image is also cool but it's kind of a special case there.
Aeao@lemmy.world · 13 pts · 331d
It didn’t mention the star finder apps!
Save yourself the money and time. It’s Venus. That cool star you’re looking at? Yeah that’s Venus. Just trust me.
Lifter@discuss.tchncs.de · 5 pts · 331d
You don't need AI toshoew a star map. This is the one and only use for Augmented Reality though.
Aeao@lemmy.world · 3 pts · 331d
Yeah I actually paid for the full version of mine… even though it’s always Venus
lavander@lemmy.dbzer0.com · 12 pts · 331d
One of the issues with LLM is that it attracted all attention. Classifiers are generally cool, cheap and saved us from multiple issues (ok face recognition aside 🙂)
When the AI bubble will burst (because of LLM being expensive and not good enough to replace a person even if they are good in pretending to be a person) all AI will slow down… including classifiers, NLP, etc
All this because the AI community was obsessed by the Turing test/imitation game 🙄
Turing was a genius but heck if I am upset with him for coming with this BS 🤣
skisnow@lemmy.ca · 4 pts · 331d
It made sense in the context it was devised in. Back then we thought the way to build an AI was to build something that was capable of reasoning about the world.
The notion that there'd be this massive amount of text generated by a significant percentage of the world's population all typing their thoughts into networked computers for a few decades, coupled with the digitisation of every book written, that could be stitched together in a 1,000,000,000,000-byte model that just spat out the word with the highest chance of being next based on what everyone else in the past had written, producing the illusion of intelligence, would have been very difficult for him to predict.
Remember, Moore's Law wasn't coined for another 15 years, and personal computers didn't even exist as a sci-fi concept until later still.
brucethemoose@lemmy.world · 1 pts · 331d
I dunno about that. We got a pile of architecture research out of it just waiting for some more tests/implementations.
And think of how cheap renting compute will be! It’s already basically subsidized, but imagine when all those A100s/H100s are dumped.
CompactFlax@discuss.tchncs.de · 8 pts · 332d
Friendly reminder that automatic transmissions were sometimes considered to be artificial intelligence.
“Fuck business idiots waxing poetic about the inestimable value of LLMs” isn’t a good community name though.
pelespirit@sh.itjust.works · 3 pts · 332d
Automatic transmissions aren't trying to take your creative job.
ZombiFrancis@sh.itjust.works · 8 pts · 331d
If you ever are an executive and you need to explain a product idea you made up and don't want to bother with actual proof of concept, then AI has got you.
If you want custom porn, AI has got you.
These are its two competing functions. And judging by my last foray into what AI is about, the latter is winning hard.
Tattorack@lemmy.world · 2 pts · 331d
"Winning hard". Nice choice of words.
chosensilence@pawb.social · 7 pts · 332d
genAI is the enemy. other kinds are useful.
popcap200@lemmy.ml · 6 pts · 332d
To be fair, some of the memes are bangers as well.
Prunebutt@slrpnk.net · 5 pts · 332d
I've yet to see a good one.
pelespirit@sh.itjust.works · 3 pts · 332d
I think memes are okay too. Memes are usually recycled jokes anyway.
technocrit@lemmy.dbzer0.com · 4 pts · 332d
There is no "AI".
But there is endless technology that grifters label as "AI". Ofc some of this technology will be useful. But under capitalism all technology is developed by and for the benefit of the disgustingly privileged via violent control.
brbposting@sh.itjust.works · 4 pts · 330d
Plant ID is soooo disappointing - works sometimes though.
Always gotta run the ID, web search for images of the recommendation, compare images to plant.
Semantic search can be helpful:
Guess OP image could be about e.g. Perplexity repeatedly HAMMERING (no caching?) the beautiful open web and slopping out poor syntheses.
Soapbox@lemmy.zip · 4 pts · 331d
AI noise reduction and spot removal tools for photo editing get a pass too.
ILikeBoobies@lemmy.ca · 3 pts · 331d
It’s a decent evolution of the search engine but you have to ask it for sources and it’s way too expensive for it’s use case.
AeonFelis@lemmy.world · 3 pts · 331d
You've misspelled "descent"
odelik@lemmy.today · 1 pts · 331d
Is it?
I've found so many fucking errors in AI summaries that I don't trust shit from AI searches when a direct link to a source or wiki could give me better summarized info.
I guess it's an evolution, but I'm really hoping these mutations prove inferior and it dies off already. But capitalism won't have that with their sunk cost fallacy driven insistence that I just use the inferior product.
ILikeBoobies@lemmy.ca · 1 pts · 330d
That’s why I said you have to ask for the source. It’s summary isn’t good but that you can describe something to it in human language instead of focusing on keywords to start your search and that the sources it gives aren’t just ads yet is useful.
jlow@discuss.tchncs.de · 3 pts · 331d
+1 for WhoBird
30p87@feddit.org · 0 pts · 332d
That's known as image identification with ML though, not "AI". The difference? Capitalism.
magic_lobster_party@fedia.io · 9 pts · 332d
The difference is that plant identification is no longer an interesting area for AI research. It was ”AI” 10 years ago, but now it’s more or less a solved problem.
30p87@feddit.org · 8 pts · 332d
Primarily, it's not interesting financially and therefore for marketing.
Grimy@lemmy.world · 1 pts · 331d
https://en.wikipedia.org/wiki/Artificial_intelligence
technocrit@lemmy.dbzer0.com · 0 pts · 332d
They're all known as "apps" because that's all they are. Like Angry Birds if we told people the piggies were "AI".
artyom@piefed.social · -3 pts · 332d
Honestly I like when it writes for me too. I tend to be very blunt and concise in my messaging and AI just puts that corporate shine and bubbliness on my messages that everyone seems to feel is important.