Trusting an LLM with a decision that can potentially be fatal is a good Darwin test I suppose.
Sadly I don’t think the general populace understands that LLMs are unreliable. (And even people in tech can vastly overestimate the capabilities.. at least judging by the insane spending on them).
> ...even people in tech can vastly overestimate the capabilities...
I think people in (software) tech are currently more prone to overestimate the capabilities, because LLMs in a harness are genuinely excellent at programming. Programming is the perfect LLM task since 1. it's symbolic manipulation, 2. there is a vast corpus of high quality training data, and 3. most mistakes can be harmlessly caught at compile-time or unit-test-time. Especially point 3 makes it so that just throwing more "effort" at a problem, something the machine is endlessly willing to do, virtually guarantees an improved result.
In contrast there is no way for the machine to write a unit test to double-check its work when what it's offering the user is a legal document, or a medical diagnosis, or a recommendation of "yep that mushroom is safe to eat".
I've often seen the "Gell-Mann Amnesia Effect" referenced with respect to LLMs. It is said that most people can tell the LLM is not great at their own subject of expertise, yet they still trust it for other subjects where they can't personally assess the quality of its answers. Imagine how bad this is when the LLM actually is great at the thing you have expertise in.
Ha! I just had a very interesting conversation with ChatGPT about a mushroom that I found in my front lawn. I am 99% sure it was an edible oyster mushroom which included analysis the of the size, growing substrate (dead ash tree root which turns out to be a useful identifying characteristic -- I learned a lot about mycology in a short time). Plus I don't use fertilizers, herbicide, pesticides, fungicides (obviously) on my lawn so I think it would have been quite tasty. Did I eat it? Hell, no. But then I got to thinking, would I trust a human to correctly identify it? I am actually not sure I would either. Penalty is just too great for misidentification LLM or not.
This time of year it is common for folks to mistake Dear mushrooms (Pluteus cervinus) with the poisonous juvenile invasive Amanita family.
Amanita travels everywhere with wood mulch around local gardens, and is practically impossible to get rid of once established.
Stay safe, and never believe your own lying eyes when it comes to mycology. Could join a local university hosted public outreach foraging club, as the experts enjoy nothing more that spotting weird looking finds. =3
I've used Gemini quiet a bit for foraging. It's definitely a "trust, but verify" situation. It's really good at at least getting you in the right lane to verify, rather than having to comb hundreds of pictures.
So even if it's wrong, a short verification usually reveals why it was wrong. I haven't had a situation yet where it was catastrophically wrong (i.e. the mushroom it called out looks nothing like the mushroom imaged). It also will generally warn of lookalikes.
I also use it a lot for general plant ID, and it's really impressive there too, with much lower stakes (I'm not eating those).
Is it possible to reliably (>99%) identify mushrooms from a single image, with any method?
All image datasets have an intrinsic minimum error that comes from the limit of information in the data, regardless of what model or method you use to analyze it.
You might not be able to tell apart two similar-looking mushroom species just by looking.
Certainly it's not possible, yeah. A better test might be to have the AI report either confirmed good, confirmed toxic, or unknown. Even an experienced mushroom hunter needs to humbly put many mushrooms into the "unknown" category.
many species of mushroom cant be identified just by looking, spore print is a necessary differentiator for many species, and as the article notes, for some species DNA sequencing is the only way to be sure.
Chanterelles and false chanterelles are especially tricky. They fruit at the same time in the same areas. I have to do a 2nd pass after I have washed and dried my haul.
There's been a spike in cases of people eating death cap mushrooms in California this year, and I would bet almost anything that LLM adoption has something to do with it.
They look like the Asian Paddy Straw Mushrooms, so people used to collecting those, being new to the area and not knowing about the Death Cap run afoul of it. The San Francisco Chronicle often has stories about this after it rains.
I actually find AI-based plant and mushroom identifiers very interesting. They make being curious about such things as a beginner or layperson more accessible, and you learn some patterns.
However it's very obvious that you shouldn't rely on them to decide what to put in your body. All those apps that I have seen (I guess the responsible ones?) will warn you not to do it. That's not what it's for.
The single best animated data visualization to demonstrate the stochastic nature of LLM models: an animation of the probability of solving a long multiplication problem over several runs. [0]
The check can't be just a fast glance because an incorrect answer for sure looks exactly the same as a correct one. From my experience a separate pass with verification of the source is required.
the interesting question is: what is the barrier we need to cross to trust LLMs with stuff like this? we'll probably cross the barrier but I'd have a hard time setting a true percentage success boundary
Sometimes the things that people say, says more about themselves than the thing they said.
If the only reason you give advice is predicated on the transactional value of that action, I think you've given up on acting ethically.
While to some extent we all are motivated by some abstract transactional value, I think immediately perceiving and reducing everything to a surivivalist, transactional mentality is an empty, cynical existence.
> If the only reason you give advice is predicated on the transactional value of that action, I think you've given up on acting ethically.
I said nothing of the sort.
Let me give a different example: I would not ride on a plane being flown remotely by a pilot on the ground. But I will ride on a plane with the pilots aboard. I do not necessarily believe the remote pilot would intentionally endanger me, but at the same time they are not facing mortal peril if they screw up. So I would need some other reason to trust them (maybe the remote pilot is a family member).
An AI has nothing to lose for misidentifying a mushroom, and I have my life to lose by trusting it. It can’t even fear for my life on my behalf.
Neither does a radiologist but yet we trust them to read a CT scan. Your relatives can sue if there is malpractice, but in practice, misreading happens, and nobody goes to jail for that.
Everything performs a particular task to the best of its ability and we all accept the level of risk we are willing to tolerate. If machines gain trust from enough sources, it will be accepted.
All I'm saying here is (if you walk back the necessity of having "skin in the game") -- all I'm saying here is, we regularly trust our lives to situations and people we expect to reasonably fulfill a given task. Most of us get in an elevator, or drive in a car, or even just buy food from a store.
> Neither does a radiologist but yet we trust them to read a CT scan.
I also don’t really have any other options in that regard, so there’s no real choice to make, other than getting a second opinion. I may not have a chance for a second opinion if I fully trust an AI on mushroom ID.
> Most of us get in an elevator, or drive in a car, or even just buy food from a store.
Of course, and I do too. Mostly because I have to and because there are many aspects of those risks that remain under my direct control. At times I have decided to take the stairs instead of the elevator.
I guess if you had a billion dollar you could assemble an ML team and fine-tune a frontier LLM specifically for that purpose through RL, and it could work better than your local pharmacologist, but I doubt anyone would bother doing that.
So it will probably come long after airplanes actually.
Mushroom pickers generally know only a handful of easy to distinguish types and only pick those. I would say that 99.9% of pickers will never get poisoned.
I never touch a mushroom I am not 100% certain is safe. So far I’ve had a 100% success rate in not poisoning myself or others. Wouldn’t recommend trusting an LLM either.
Giving the impossibly high standards people here have for identifying a mushroom, it’s a wonder that they would eat a mushroom at all. You may not trust an LLM, but how would you even trust yourself? Who would you ever trust?
I don’t anyone here has indicated that they have "impossible" standards. If you think "gather only mushrooms that you definitely recognize as edible; if you have any doubts, leave it be" is impossibly high, I’m not sure what to say. Of course what "definitely recognize" means in terms of the odds of a false positive is subjective, so sometimes people get poisoned. Even more rarely a poisoning leads to death, but people who die from mushrooms they have picked are those who didn’t heed the warning at all.
Mushrooms are best treated like firearms; that’s not an impossibly high standard.
Why would you want to eat a random wild mushroom so badly in the first place? We're not starving, wandering through the forest for anything which might be edible. Entire regulated industries exist around selling and growing human food grade mushrooms.
Curiosity is great. Take a picture, enjoy what you find in nature. At the same time, stay safe and resist the urge to throw a random mushroom down your throat just because you might be able to unless you're willing to put the work in to make doing so reasonably safe. If you do reaaaaallly want to try a mushroom you can't find for consumption otherwise then it's worth the heavy amount of verification. 99.9999% of mushrooms people normally care about eating can more easily be found without foraging and self-identifying though.
The lack or determinism is irrelevant. It's trivial to make an LLM deterministic (certianly much easier than to make a human deterministic!), but that wouldn't make it any more trustworthy.
I have been using "new" Siri on iOS similarly but less dangerously. Pointing my camera at dogs at the dog park and seeing how well Siri identifies them. Pretty impressed so far. I am not eating them tho.
I think the more important question is: have people really run out of interesting flavors at their local stores and restaurants, that they are so obsessed with eating things they find growing on the ground?
You can't really understand until you've tasted choice mushrooms, the only one of which is findable in stores is the bolete, and even that's usually dry and tells terrible quality.
I used to go in the forest and find maybe 5 types of really tasty mushrooms, come back with enough for the whole family. Add some cheese and some onions and it's Michelin star in the mountains.
It's also quite a thrill to spot a log absolutely covered in oyster or woodear mushrooms (two I feel very confident in identifying, and thankfully both were common on the 10k trail above Albuquerque)
There is a great variety of mushrooms with distinct characteristics, many of which are only found in the wild. Even the ones sold at restaurants are often sourced from foragers. It's a skill to find and identify them and for many it's an exciting hobby.
Taste aside, foraging in the woods--whether for mushrooms, berries or herbs--unlocks some atavistic response inside me. It feels deeply true to my nature as an animal, and is very rewarding in a way that surprised me when I first did it.
Same as fishing or hunting, or foraging blueberry or blackberries while hiking, there's some primal reward in gathering your own food (occasionally, that is).
I’ve been at a locals’ restaurant in Italy where the chef had sourced some truffles from a nearby wood. They were better than anything he could have got from a store
Are you this actively shitty towards all hobbies you don't also enjoy?
People have been foraging for mushrooms for literally thousands of years, but mushroom cultivation is pretty recent, and the overwhelming majority of edible species have not yet been reliably cultivated.
Are you ignorant of where those interesting flavors come from (indeed, are discovered)?
This has strong "why should we care about farms? People get everything they need from the store" vibes.
> Are you this actively shitty towards all hobbies you don't also enjoy
No, really just this one. There’s not that many hobbies that have death as an outcome of a minor mistake. For something like free solo climbing, we all understand those guys are nuts and wired different for an incredible high. For mushroom foraging on the other hand, it seems to me like… your reward is a slightly different mushroom flavor?
I had a friend that was quite passionate about foraging for mushrooms while we were in high school. He eventually went on to receive a PhD. in mycology. My old friend as well as everyone else I knew that liked to forage mushrooms were... uh... not looking for "flavorful" mushrooms to say the least.
Love this idea. We should encourage every AI enthusiast to eat wild mushrooms using their sycophantic bullshit machine to determine each one's safety. Who benefits? Society
Trusting an LLM with a decision that can potentially be fatal is a good Darwin test I suppose.
Sadly I don’t think the general populace understands that LLMs are unreliable. (And even people in tech can vastly overestimate the capabilities.. at least judging by the insane spending on them).
> ...even people in tech can vastly overestimate the capabilities...
I think people in (software) tech are currently more prone to overestimate the capabilities, because LLMs in a harness are genuinely excellent at programming. Programming is the perfect LLM task since 1. it's symbolic manipulation, 2. there is a vast corpus of high quality training data, and 3. most mistakes can be harmlessly caught at compile-time or unit-test-time. Especially point 3 makes it so that just throwing more "effort" at a problem, something the machine is endlessly willing to do, virtually guarantees an improved result.
In contrast there is no way for the machine to write a unit test to double-check its work when what it's offering the user is a legal document, or a medical diagnosis, or a recommendation of "yep that mushroom is safe to eat".
I've often seen the "Gell-Mann Amnesia Effect" referenced with respect to LLMs. It is said that most people can tell the LLM is not great at their own subject of expertise, yet they still trust it for other subjects where they can't personally assess the quality of its answers. Imagine how bad this is when the LLM actually is great at the thing you have expertise in.
Ha! I just had a very interesting conversation with ChatGPT about a mushroom that I found in my front lawn. I am 99% sure it was an edible oyster mushroom which included analysis the of the size, growing substrate (dead ash tree root which turns out to be a useful identifying characteristic -- I learned a lot about mycology in a short time). Plus I don't use fertilizers, herbicide, pesticides, fungicides (obviously) on my lawn so I think it would have been quite tasty. Did I eat it? Hell, no. But then I got to thinking, would I trust a human to correctly identify it? I am actually not sure I would either. Penalty is just too great for misidentification LLM or not.
oysters are pretty safe
Yes but you need to be sure it is an Oyster Mushroom in the first place.
I mean, there's a reason the Old Testament banned them. Blech
Great stereotypical engineer's response.
This time of year it is common for folks to mistake Dear mushrooms (Pluteus cervinus) with the poisonous juvenile invasive Amanita family.
Amanita travels everywhere with wood mulch around local gardens, and is practically impossible to get rid of once established.
Stay safe, and never believe your own lying eyes when it comes to mycology. Could join a local university hosted public outreach foraging club, as the experts enjoy nothing more that spotting weird looking finds. =3
I've used Gemini quiet a bit for foraging. It's definitely a "trust, but verify" situation. It's really good at at least getting you in the right lane to verify, rather than having to comb hundreds of pictures.
So even if it's wrong, a short verification usually reveals why it was wrong. I haven't had a situation yet where it was catastrophically wrong (i.e. the mushroom it called out looks nothing like the mushroom imaged). It also will generally warn of lookalikes.
I also use it a lot for general plant ID, and it's really impressive there too, with much lower stakes (I'm not eating those).
There are old mushroom hunters, and there are bold mushroom hunters.
There are no old bold mushroom hunters.
Is it possible to reliably (>99%) identify mushrooms from a single image, with any method?
All image datasets have an intrinsic minimum error that comes from the limit of information in the data, regardless of what model or method you use to analyze it.
You might not be able to tell apart two similar-looking mushroom species just by looking.
Certainly it's not possible, yeah. A better test might be to have the AI report either confirmed good, confirmed toxic, or unknown. Even an experienced mushroom hunter needs to humbly put many mushrooms into the "unknown" category.
No. At the very least you need a photo of both the upper side of the cap and the gills.
many species of mushroom cant be identified just by looking, spore print is a necessary differentiator for many species, and as the article notes, for some species DNA sequencing is the only way to be sure.
Chanterelles and false chanterelles are especially tricky. They fruit at the same time in the same areas. I have to do a 2nd pass after I have washed and dried my haul.
https://www.mushroom-appreciation.com/wp-content/uploads/202...
Do the false chanterelles that you encounter smell similar to chanterelles?
The false chanterelle that I know is not poisonous.
https://en.wikipedia.org/wiki/Hygrophoropsis_aurantiaca#Edib...
The real thing is plentiful enough I don't mess with them, but I know people that regularly eat the false ones.
Considering that some mushrooms need a spore print to tell them apart, I'd say that you can't ever trust an LLM.
There's been a spike in cases of people eating death cap mushrooms in California this year, and I would bet almost anything that LLM adoption has something to do with it.
They look like the Asian Paddy Straw Mushrooms, so people used to collecting those, being new to the area and not knowing about the Death Cap run afoul of it. The San Francisco Chronicle often has stories about this after it rains.
Repo for reproducible science: https://github.com/stared/mushroom-hunting-llm-bench
I actually find AI-based plant and mushroom identifiers very interesting. They make being curious about such things as a beginner or layperson more accessible, and you learn some patterns.
However it's very obvious that you shouldn't rely on them to decide what to put in your body. All those apps that I have seen (I guess the responsible ones?) will warn you not to do it. That's not what it's for.
The single best animated data visualization to demonstrate the stochastic nature of LLM models: an animation of the probability of solving a long multiplication problem over several runs. [0]
[0] https://adamsohn.com/reasoning-grid/#walk-the-surface
The check can't be just a fast glance because an incorrect answer for sure looks exactly the same as a correct one. From my experience a separate pass with verification of the source is required.
the interesting question is: what is the barrier we need to cross to trust LLMs with stuff like this? we'll probably cross the barrier but I'd have a hard time setting a true percentage success boundary
You should never trust any image-based method for mushroom identification.
It is simply not possible to identify mushrooms from image alone, no matter how good your LLM is.
What is needed before it is wise to trust LLMs with life or death decisions?
Or, what is needed before we are forced to give LLMs control of life or death decisions in the name of profits?
Around the same time the LLM also faces an existential risk when its advice is wrong.
Not trusting my life to an “intelligence” that has nothing to lose.
Sometimes the things that people say, says more about themselves than the thing they said.
If the only reason you give advice is predicated on the transactional value of that action, I think you've given up on acting ethically.
While to some extent we all are motivated by some abstract transactional value, I think immediately perceiving and reducing everything to a surivivalist, transactional mentality is an empty, cynical existence.
> If the only reason you give advice is predicated on the transactional value of that action, I think you've given up on acting ethically.
I said nothing of the sort.
Let me give a different example: I would not ride on a plane being flown remotely by a pilot on the ground. But I will ride on a plane with the pilots aboard. I do not necessarily believe the remote pilot would intentionally endanger me, but at the same time they are not facing mortal peril if they screw up. So I would need some other reason to trust them (maybe the remote pilot is a family member).
An AI has nothing to lose for misidentifying a mushroom, and I have my life to lose by trusting it. It can’t even fear for my life on my behalf.
Neither does a radiologist but yet we trust them to read a CT scan. Your relatives can sue if there is malpractice, but in practice, misreading happens, and nobody goes to jail for that.
Everything performs a particular task to the best of its ability and we all accept the level of risk we are willing to tolerate. If machines gain trust from enough sources, it will be accepted.
All I'm saying here is (if you walk back the necessity of having "skin in the game") -- all I'm saying here is, we regularly trust our lives to situations and people we expect to reasonably fulfill a given task. Most of us get in an elevator, or drive in a car, or even just buy food from a store.
> Neither does a radiologist but yet we trust them to read a CT scan.
I also don’t really have any other options in that regard, so there’s no real choice to make, other than getting a second opinion. I may not have a chance for a second opinion if I fully trust an AI on mushroom ID.
> Most of us get in an elevator, or drive in a car, or even just buy food from a store.
Of course, and I do too. Mostly because I have to and because there are many aspects of those risks that remain under my direct control. At times I have decided to take the stairs instead of the elevator.
We'll cross it around the same time we hand over nuclear reactor and airplane control systems to LLMs.
I guess if you had a billion dollar you could assemble an ML team and fine-tune a frontier LLM specifically for that purpose through RL, and it could work better than your local pharmacologist, but I doubt anyone would bother doing that.
So it will probably come long after airplanes actually.
What are the error rates for a human experienced/trained in spotting mushrooms?
(isn't that the standard proponents use for autonomous driving?)
Mushroom pickers generally know only a handful of easy to distinguish types and only pick those. I would say that 99.9% of pickers will never get poisoned.
I never touch a mushroom I am not 100% certain is safe. So far I’ve had a 100% success rate in not poisoning myself or others. Wouldn’t recommend trusting an LLM either.
Gotta' love it when a possible answer to an LLM-related question is "Death."
Gemini 3.8 flash told me to eat Entoloma sinuatum!
Nobody ever mentions the fox piss.
Giving the impossibly high standards people here have for identifying a mushroom, it’s a wonder that they would eat a mushroom at all. You may not trust an LLM, but how would you even trust yourself? Who would you ever trust?
I don’t anyone here has indicated that they have "impossible" standards. If you think "gather only mushrooms that you definitely recognize as edible; if you have any doubts, leave it be" is impossibly high, I’m not sure what to say. Of course what "definitely recognize" means in terms of the odds of a false positive is subjective, so sometimes people get poisoned. Even more rarely a poisoning leads to death, but people who die from mushrooms they have picked are those who didn’t heed the warning at all.
Mushrooms are best treated like firearms; that’s not an impossibly high standard.
Why would you want to eat a random wild mushroom so badly in the first place? We're not starving, wandering through the forest for anything which might be edible. Entire regulated industries exist around selling and growing human food grade mushrooms.
Curiosity is great. Take a picture, enjoy what you find in nature. At the same time, stay safe and resist the urge to throw a random mushroom down your throat just because you might be able to unless you're willing to put the work in to make doing so reasonably safe. If you do reaaaaallly want to try a mushroom you can't find for consumption otherwise then it's worth the heavy amount of verification. 99.9999% of mushrooms people normally care about eating can more easily be found without foraging and self-identifying though.
Someone who your estate can sue and get a settlement of a dollar value equal to your own estimation of your future life's worth?
Honestly, yeah. I wouldn't eat a wild mushroom. The ones in the store are good enough for me.
LLMS ARE NON-DETERMINISTIC. DO NOT TRUST THEM WITH LIFE OR DEATH DECISIONS! Seriously, you could die of a near-optimal solution.
The lack or determinism is irrelevant. It's trivial to make an LLM deterministic (certianly much easier than to make a human deterministic!), but that wouldn't make it any more trustworthy.
When you are ingesting potential toxins determinism is crucial. You have a random chance of dying. Facts matter.
Humanity is in its "fuck around" phase with LLMs and is going to fly headlong into "find out" soon.
I have been using "new" Siri on iOS similarly but less dangerously. Pointing my camera at dogs at the dog park and seeing how well Siri identifies them. Pretty impressed so far. I am not eating them tho.
I think the more important question is: have people really run out of interesting flavors at their local stores and restaurants, that they are so obsessed with eating things they find growing on the ground?
I’ll never understand mushroom culture.
Some mushrooms like chanterelles can't be reliably farmed, and are quite tasty! I love being able to spend a day hiking and coming back with a meal.
Morels are starting to make good progress in being farmed.
But, I still enjoy the hunt every spring.
They're still on my bucket list to find. Lots of false morels around here. :(
Maybe checkout morels.com forums next spring to gather some intel on the conditions.
Some reasons to forage (applies to berries, nuts, fruits, not just mushrooms):
* Foraging is very satisfying to do
* Foraging allows you to participate in your cultural practices (although sadly many people in the US are long-cut off from such cultural knowledge)
* You can get fresher food foraging
* There absolutely are interesting varieties that aren't sold because they are hard to farm/transport/store etc.
* Foraged food is free
You can't really understand until you've tasted choice mushrooms, the only one of which is findable in stores is the bolete, and even that's usually dry and tells terrible quality.
I used to go in the forest and find maybe 5 types of really tasty mushrooms, come back with enough for the whole family. Add some cheese and some onions and it's Michelin star in the mountains.
It's also quite a thrill to spot a log absolutely covered in oyster or woodear mushrooms (two I feel very confident in identifying, and thankfully both were common on the 10k trail above Albuquerque)
There is a great variety of mushrooms with distinct characteristics, many of which are only found in the wild. Even the ones sold at restaurants are often sourced from foragers. It's a skill to find and identify them and for many it's an exciting hobby.
I agree. Foraging out in the beautiful woods connecting to nature by eating what you picked is irrational.
Taste aside, foraging in the woods--whether for mushrooms, berries or herbs--unlocks some atavistic response inside me. It feels deeply true to my nature as an animal, and is very rewarding in a way that surprised me when I first did it.
Same as fishing or hunting, or foraging blueberry or blackberries while hiking, there's some primal reward in gathering your own food (occasionally, that is).
I’ve been at a locals’ restaurant in Italy where the chef had sourced some truffles from a nearby wood. They were better than anything he could have got from a store
Are you this actively shitty towards all hobbies you don't also enjoy?
People have been foraging for mushrooms for literally thousands of years, but mushroom cultivation is pretty recent, and the overwhelming majority of edible species have not yet been reliably cultivated.
Are you ignorant of where those interesting flavors come from (indeed, are discovered)?
This has strong "why should we care about farms? People get everything they need from the store" vibes.
> Are you this actively shitty towards all hobbies you don't also enjoy
No, really just this one. There’s not that many hobbies that have death as an outcome of a minor mistake. For something like free solo climbing, we all understand those guys are nuts and wired different for an incredible high. For mushroom foraging on the other hand, it seems to me like… your reward is a slightly different mushroom flavor?
I had a friend that was quite passionate about foraging for mushrooms while we were in high school. He eventually went on to receive a PhD. in mycology. My old friend as well as everyone else I knew that liked to forage mushrooms were... uh... not looking for "flavorful" mushrooms to say the least.
Love this idea. We should encourage every AI enthusiast to eat wild mushrooms using their sycophantic bullshit machine to determine each one's safety. Who benefits? Society
Did you read the post or just decide to advertise your dislike of AI by wishing ill on others?
Both :) happy hunting!