Hi Gary, nice flashback to those two movies, and sadly funny collection of fails. Thank you for cataloging all these, they will serve as a useful collection for posterity!
What if we reconsider them ALL as fails - because that's what they are. Some (even many, most) could be useful fails, the rest, useless. They 'fail' to know ANYTHING about the world, the "know" not a single thing about any word's meaning, or any other symbol's meaning [math, code, music, chemical formulae, art..]. They only "know" token ordering (even that, not really, they are made to calculate it using mini nuclear power plants, literally).
GenAI, no matter how much it "reasons" (the most absurd misuse of a term ever) via "tokens" (results of dot products), has a snowball's chance in hell (including that very metaphor!!) of having ANY understanding of ANYTHING, let alone the physical world!!!! WE do - how? By living in it, including being able to live an entire life without needing to compute a SINGLE dot product. GenAI needs "modular" fission reactors, the brain needs... a Pop Tart. How much more delusional would we need to be, to claim they are equivalent?
100% of all existing AI is fake (no direct physical experience), and LLM/RAG/AgenticAI/GenAI isn't a magical exception... Neuromorphic chips are a tad bit better - first floor as opposed to the ground floor, on a journey to the moon (another metaphor that so called AI has zero clue about!). No clue about 'zero', 'clue' either, lol.
PS: 'World model' - one of the biggest oxy'moron's ever; so is 'GenAI' [which simply computes numbers].
Saty Your contempt for Generative AI is soundly based. Your comment "100% of all existing AI is fake" is not. Some people think that words have meanings, and the meanings are represented in the AI they build - Semantic AI. I would also quibble with "no physical experience" - we have no physical experience of a thousand page piece of legislation (OK, we read it and we can put it in a drawer), but our useful experience of it is abstract. Humans have a severe input limit to their conscious mind - that is the thing we are hunting, not a missing seat on a unicycle.
Jim, good points. Word do have meanings. But they are not inherent. 'Running' might be defined in the dictionary, but a 2 y/o doesn't need the definition to run, or even to be able to say 'run' - 'un' is perfectly ok as long as the kid and nearby grownups agree on its shared meaning :)
The way I see it - the 'fakeness' in 70 years to AI [to this exact date!] comes from doing just (only) symbol processing [not just in GOFAI, but traditional (classifying) ML, RL, robotics, cog archs... too]. The problem is that the knowledge repr under the symbols in symbolic AI, the data that trained the ML, the goals set up reinforcement in RL, the algorithms (eg IK, SLAM...) coded for robotics, the various modules created for brain-like architectures... are ALL human-derived. So, all AI does is derivative calculations at best, which is why they are fake.
Humans are able to process symbols for sure [that too, a lot slower and smaller than machines], but we have 'grounded' meaning (literally) (ie experience) that is symbol-free. The 1000 page law briefing is not a pile of cards that build a castle in the air, ie not words defined in terms of other words - they are about our world, our society, which we can understand even without the 1000 page briefing. IOW, in real life, symbols are optional - for AI, symbols are all there is. That is the unbridgeable gap.
A different way to differentiate the two is to ask, what is reversible? The real world isn't (rusting, dying, earthquake damage...), where all AI (including backprop epochs) can be reversed by storing all the calculations and restoring them.
OI, for a change, is brain-like. So everything I said above doesn't apply to it :)
TL;DR: I see 'faking' as originating from just doing calculations - true for AI, not true for biological brains.
Indeed. And, I didn't want any preventing, lol. What I'm pointing out are very fundamental reasons why, AI's shortcomings exist. Eg. an SDC that calculates the world is fundamentally different from a human driver. That distinction might be irrelevant with more millions of miles worth of training data being used, but it will never erase the distinction.
Grounding comes knowing what a ball is, by holding it, throwing it, sniffing it, turning it around, breaking it apart, burning it... where no calculations ever happen. Understanding of physics in this case isn't a matter of doing RL on hitting the ball. Obviously, for the very narrow case of "playing ping pong", RL can (and does) beat humans - what else would we expect, competing against a machine? On the other hand, my wanting to pick up three balls and bounce-juggle them isn't in the ping-pong playing "brain", until of course we let it do RL on 100M cycles to blindly "learn" that - by which time I move on to painting smileys on them balls :) IOW, intelligence is much more than learning one task after another, playing catch-up at best.
Words do have meanings. But they are not inherent. 'Running' might be defined in the dictionary, but a 2 y/o doesn't need the definition to run, or even to be able to say - 'un' is perfectly ok as long as the kid and nearby grownups agree on its shared meaning :) The way I see it - the 'fakeness' in 70 years to AI [to this exact date!] comes from doing just (only) symbol processing [not just in GOFAI, but traditional (classifying) ML, RL, robotics, cog archs... too]. The problem is that the knowledge repr under the symbols in symbolic AI, the data that trained the ML, the goals set up reinforcement in RL, the algorithms (eg IK, SLAM...) coded for robotics, the various modules created for brain-like architectures... are ALL human-derived. So, all AI does is derivative calculations at best, which is why they are fake. Humans are able to process symbols for sure [that too, a lot slower and smaller than machines], but we have 'grounded' meaning (literally) (ie experience) that is symbol-free. The 1000 page law briefing is not a pile of cards that build a castle in the air, ie not words defined in terms of other words - they are about our world, our society, which we can understand even without the 1000 page briefing. IOW, in real life, symbols are optional - for AI, symbols are all there is. That is the unbridgeable gap. A different way to differentiate the two is to ask, what is reversible? The real world isn't (rusting, dying, earthquake damage...), where all AI (including backprop epochs) can be reversed by storing all the calculations and restoring them. OI, for a change, is brain-like. So everything I said above doesn't apply to it :) TL;DR: I see 'faking' as originating from just doing calculations - true for AI, not true for biological brains.
“a 2 y/o doesn't need the definition to run” No, but it needed a million years of evolution to be able to run – the ilium had to grow in a different direction, which messed up the birth canal, which… humans are good if changes are small. Hopeless if changes exceed a quite severe limit – the Four Pieces Limit. Dictionaries can be quite terrible for meanings – they need to be carefully curated for Semantic AI. The 1000-page piece of legislation – not a briefing, it is a law you cannot break – it has changed our world in an abstract way, which also changed the world in a practical way – no USAID, Education gutted, etc.. Some other examples – a 60 page specification – wasted $800 million – people did not understand what it was saying – things only need to be a few pages apart to lose connection
Robodebt in Australia – people on the dole were receiving bills of $18,000 which they couldn’t pay – the government had changed income averaging without changing the legislation - $1.7 billion in reparations, 2 suicides. Solution – have the machine bring the legislation “alive” by using the exact words – no programming in a computer language.
Horizon in UK – system creating debts out of nothing – 13 suicides, unlawful incarceration, 1.7 billion pounds reparations.
Boeing 737 MAX MCAS – 346 deaths. Moved engines forward, had to add angle of attack sensor – no redundancy, no change to flight manual, ignored regulations (more symbols that have a vital purpose), lied to FAA (proud of it). Catching the extent of lying would have been hard. FAA perennially understaffed, appoints company employee as FAA inspector – could have had machine reporting to instigator. Symbols become very important at the big end of town, not the toy problems that AI people play with.
“I see 'faking' as originating from just doing calculations - true for AI, not true for biological brains.” Semantic AI has free resources – it can modify its own structure, observe the result, select the best approach. No, it doesn’t use ANNs – the person who called them that should have been tarred and feathered, and run out of town on a rail – how many careers blighted?
Jim, wow, cool examples, thank you :) Never doubted or questioned the validity of symbols (ALL of STEM is based on them :) I'm a huge fan of Stephen Wolfram, and his summer workshops offer a visual feast of what symbols can do: https://www.wolframcloud.com/obj/microsites/summerschool/projects.html). My entire line of commenting is different: at a base level, there are no symbols in brains (eg no numerical representations in any base to any bit precision that we run calculations over). Eg. sqrt(9) is "easy", sqrt(9.045677), not so much.
Not knocking symbols, not knocking AI - simply saying, there is more, and that 'more' might be that brains don't need to symbol-process, to know what the world means, in contrast to any AI (including semantic). One last example - Gibson's 'affordance' - a cat knows to snooze by a sunlit window w/o calculating light level or temperature, a tired kid sits on a flat rock when tired, without measuring area or tilt. The world 'affords' meaning directly, creatures use them w/o thinking etc. Till there is AI that does likewise, there is no parity.
Hi Jim, the years of evolution you mentioned (to be able to run) - exactly. Those years have also not introduced a von Neumann stored program architecture in any brain, because that cannot evolve by itself autopoietically. If we architect AI without long evolution (eg use brain cell organoids), they might behave similar to real brains. Then there is the mind, Self, fractured identity, mental illness, suicidal thoughts etc etc [in real brains] we might be *very* far from replicating all this :) Conversely, casting all these aside as irrelevant takes us right back to narrow AI (Arthur Samuels' checker player), which we had back in 1952 which is pre-Dartmouth AI workshop (omg).
It is interesting that we seem to be able to build a statistical model of AI-generated images in our brains after seeing quite a few. As with AI-generated text, we get a "feel" for the kind of mistakes it makes. One develops an intuitive feel for the things it gets right and wrong. So will we learn to like this AI "style"? Let's hope not.
First instance of using genAI to avoid working with a potential fraudster. Asked to look into an outside company that wanted to do business with my company. One of the founders picture looked off, and the accompanying biography had no verifiable details. Reverse image search. Totally different person, in the exact same outfit. So it was the fishy company’s obvious use of an image generator (not me using it) that tipped us off.
And all that fun and games is destroying the planet. Worse than netflix. Maybe just take out some coloring pens and be an ecologically responsible person? Being an adult here.
I am not a bass player, but I do play guitar, and I have some familiarity with how the bass is played. Tuners fading in and out of reality or not, that is not how one plays it, at all.
I think it is both. A lot of the masses are in the first category. And I count a lot of my peers in that group.
But my bosses, and executives at my (and I suspect a LOT of companies) are in the latter group. They view GenAI as a tool to allow them to at least not hire any new people, and at worst, to lay off a lot of people and let the remainders use these tools to (badly) fill in on what they used to do.
The mindless acceptance of the stuporficiallity by many people -- even by some scientists, engineers and others who are normally very skeptical and demanding of evidence is actually surprising to me.
It's as if these people simply switch off their critical faculties when they use this stuff.
If (when?) the LLMs are used for safety or security criticsl systems, the lack of human oversight is a recipe for disaster
These images would make for great visual puzzles. "Name everything that you see that is wrong with this picture. Answers on the back."
If an LLM is trained on videos, does it make fewer mistakes when asked to draw a picture than one trained on text and images? If it doesn't, this would appear to [at least partially] falsify the claim that embodying AI will improve its world models.
This seems to be a discussion about visual accuracy, yes? We’re not discussing art yet, right? I cannot, despite trying many models and many many prompts I’ve been unable to get an LLM to generate any single thing that looks unique and truly interesting. All slop. Makes me wonder how things might be different if visual artists had been involved in the training? I’m sure some of you remember when big box stores and global brands started taking over neighborhoods and city streets, homogenizing human experience. Is this any different?
On the whole, as a direct result of how it works, generative AI is a race to the middle , to mediocrity and blandness.
You are right that it is effectively a continuation of the Wallmartization of humanity, albeit in a way that affects every aspect of human society and on a much larger scale.
One thing I'm noticing from your posts is that there are a surprising number of people completely content with not looking any closer. Yet still claiming some kind of correctness about the model's generative ability. Which is ironic. The models confidently output physically, and metaphysically, inaccurate outputs. Then humans confidently ascribe these outputs the weight of a PhD. Obviously some people have neither experienced what a PhD actually does on a day to day basis nor can they imagine it. And that's what these models are being trained on. It's quite the cycle, honestly.
Just tossing this one out there since you expressed uncertainty. I'm no lighting expert, but I know a little geometry and messed around with ray tracers long ago. The position of a reflection is a function only of the positions of the viewer, reflective surface, and objects being reflected. Lighting position can change the intensity or visibility of a reflection, but not the position of it. This one actually looks not too bad.
You are right (and a reflection can certainly be at right angles to the shadow.) I agree that the reflection in that image looks believable.
On the other hand, the shadow in the first (spokeless wheel) image looks like a Komodo dragon trailing a piece of string.
But , as an amateur artist myself, i can commiserate with the chatbots. Shadows and reflections are difficult to get just right, although i must admit i never included a komodo dragon shadow in any of my paintings.
Would you describe our ability to verify the accuracy of some physical system as essentially an abstract cross modal verification system? By sense of touch, sight, sound, personal consequence or consequence of others, etc. we've built up the ability to make verification significantly less lossy. A model trained only on text and images within deterministic unimodal internal state seems incapable of developing these abstract cross modal verification systems.
Hi Gary, nice flashback to those two movies, and sadly funny collection of fails. Thank you for cataloging all these, they will serve as a useful collection for posterity!
What if we reconsider them ALL as fails - because that's what they are. Some (even many, most) could be useful fails, the rest, useless. They 'fail' to know ANYTHING about the world, the "know" not a single thing about any word's meaning, or any other symbol's meaning [math, code, music, chemical formulae, art..]. They only "know" token ordering (even that, not really, they are made to calculate it using mini nuclear power plants, literally).
GenAI, no matter how much it "reasons" (the most absurd misuse of a term ever) via "tokens" (results of dot products), has a snowball's chance in hell (including that very metaphor!!) of having ANY understanding of ANYTHING, let alone the physical world!!!! WE do - how? By living in it, including being able to live an entire life without needing to compute a SINGLE dot product. GenAI needs "modular" fission reactors, the brain needs... a Pop Tart. How much more delusional would we need to be, to claim they are equivalent?
100% of all existing AI is fake (no direct physical experience), and LLM/RAG/AgenticAI/GenAI isn't a magical exception... Neuromorphic chips are a tad bit better - first floor as opposed to the ground floor, on a journey to the moon (another metaphor that so called AI has zero clue about!). No clue about 'zero', 'clue' either, lol.
PS: 'World model' - one of the biggest oxy'moron's ever; so is 'GenAI' [which simply computes numbers].
Saty Your contempt for Generative AI is soundly based. Your comment "100% of all existing AI is fake" is not. Some people think that words have meanings, and the meanings are represented in the AI they build - Semantic AI. I would also quibble with "no physical experience" - we have no physical experience of a thousand page piece of legislation (OK, we read it and we can put it in a drawer), but our useful experience of it is abstract. Humans have a severe input limit to their conscious mind - that is the thing we are hunting, not a missing seat on a unicycle.
Jim, good points. Word do have meanings. But they are not inherent. 'Running' might be defined in the dictionary, but a 2 y/o doesn't need the definition to run, or even to be able to say 'run' - 'un' is perfectly ok as long as the kid and nearby grownups agree on its shared meaning :)
The way I see it - the 'fakeness' in 70 years to AI [to this exact date!] comes from doing just (only) symbol processing [not just in GOFAI, but traditional (classifying) ML, RL, robotics, cog archs... too]. The problem is that the knowledge repr under the symbols in symbolic AI, the data that trained the ML, the goals set up reinforcement in RL, the algorithms (eg IK, SLAM...) coded for robotics, the various modules created for brain-like architectures... are ALL human-derived. So, all AI does is derivative calculations at best, which is why they are fake.
Humans are able to process symbols for sure [that too, a lot slower and smaller than machines], but we have 'grounded' meaning (literally) (ie experience) that is symbol-free. The 1000 page law briefing is not a pile of cards that build a castle in the air, ie not words defined in terms of other words - they are about our world, our society, which we can understand even without the 1000 page briefing. IOW, in real life, symbols are optional - for AI, symbols are all there is. That is the unbridgeable gap.
A different way to differentiate the two is to ask, what is reversible? The real world isn't (rusting, dying, earthquake damage...), where all AI (including backprop epochs) can be reversed by storing all the calculations and restoring them.
OI, for a change, is brain-like. So everything I said above doesn't apply to it :)
TL;DR: I see 'faking' as originating from just doing calculations - true for AI, not true for biological brains.
Indeed. And, I didn't want any preventing, lol. What I'm pointing out are very fundamental reasons why, AI's shortcomings exist. Eg. an SDC that calculates the world is fundamentally different from a human driver. That distinction might be irrelevant with more millions of miles worth of training data being used, but it will never erase the distinction.
Grounding comes knowing what a ball is, by holding it, throwing it, sniffing it, turning it around, breaking it apart, burning it... where no calculations ever happen. Understanding of physics in this case isn't a matter of doing RL on hitting the ball. Obviously, for the very narrow case of "playing ping pong", RL can (and does) beat humans - what else would we expect, competing against a machine? On the other hand, my wanting to pick up three balls and bounce-juggle them isn't in the ping-pong playing "brain", until of course we let it do RL on 100M cycles to blindly "learn" that - by which time I move on to painting smileys on them balls :) IOW, intelligence is much more than learning one task after another, playing catch-up at best.
Words do have meanings. But they are not inherent. 'Running' might be defined in the dictionary, but a 2 y/o doesn't need the definition to run, or even to be able to say - 'un' is perfectly ok as long as the kid and nearby grownups agree on its shared meaning :) The way I see it - the 'fakeness' in 70 years to AI [to this exact date!] comes from doing just (only) symbol processing [not just in GOFAI, but traditional (classifying) ML, RL, robotics, cog archs... too]. The problem is that the knowledge repr under the symbols in symbolic AI, the data that trained the ML, the goals set up reinforcement in RL, the algorithms (eg IK, SLAM...) coded for robotics, the various modules created for brain-like architectures... are ALL human-derived. So, all AI does is derivative calculations at best, which is why they are fake. Humans are able to process symbols for sure [that too, a lot slower and smaller than machines], but we have 'grounded' meaning (literally) (ie experience) that is symbol-free. The 1000 page law briefing is not a pile of cards that build a castle in the air, ie not words defined in terms of other words - they are about our world, our society, which we can understand even without the 1000 page briefing. IOW, in real life, symbols are optional - for AI, symbols are all there is. That is the unbridgeable gap. A different way to differentiate the two is to ask, what is reversible? The real world isn't (rusting, dying, earthquake damage...), where all AI (including backprop epochs) can be reversed by storing all the calculations and restoring them. OI, for a change, is brain-like. So everything I said above doesn't apply to it :) TL;DR: I see 'faking' as originating from just doing calculations - true for AI, not true for biological brains.
“a 2 y/o doesn't need the definition to run” No, but it needed a million years of evolution to be able to run – the ilium had to grow in a different direction, which messed up the birth canal, which… humans are good if changes are small. Hopeless if changes exceed a quite severe limit – the Four Pieces Limit. Dictionaries can be quite terrible for meanings – they need to be carefully curated for Semantic AI. The 1000-page piece of legislation – not a briefing, it is a law you cannot break – it has changed our world in an abstract way, which also changed the world in a practical way – no USAID, Education gutted, etc.. Some other examples – a 60 page specification – wasted $800 million – people did not understand what it was saying – things only need to be a few pages apart to lose connection
Robodebt in Australia – people on the dole were receiving bills of $18,000 which they couldn’t pay – the government had changed income averaging without changing the legislation - $1.7 billion in reparations, 2 suicides. Solution – have the machine bring the legislation “alive” by using the exact words – no programming in a computer language.
Horizon in UK – system creating debts out of nothing – 13 suicides, unlawful incarceration, 1.7 billion pounds reparations.
Boeing 737 MAX MCAS – 346 deaths. Moved engines forward, had to add angle of attack sensor – no redundancy, no change to flight manual, ignored regulations (more symbols that have a vital purpose), lied to FAA (proud of it). Catching the extent of lying would have been hard. FAA perennially understaffed, appoints company employee as FAA inspector – could have had machine reporting to instigator. Symbols become very important at the big end of town, not the toy problems that AI people play with.
“I see 'faking' as originating from just doing calculations - true for AI, not true for biological brains.” Semantic AI has free resources – it can modify its own structure, observe the result, select the best approach. No, it doesn’t use ANNs – the person who called them that should have been tarred and feathered, and run out of town on a rail – how many careers blighted?
Jim, wow, cool examples, thank you :) Never doubted or questioned the validity of symbols (ALL of STEM is based on them :) I'm a huge fan of Stephen Wolfram, and his summer workshops offer a visual feast of what symbols can do: https://www.wolframcloud.com/obj/microsites/summerschool/projects.html). My entire line of commenting is different: at a base level, there are no symbols in brains (eg no numerical representations in any base to any bit precision that we run calculations over). Eg. sqrt(9) is "easy", sqrt(9.045677), not so much.
Not knocking symbols, not knocking AI - simply saying, there is more, and that 'more' might be that brains don't need to symbol-process, to know what the world means, in contrast to any AI (including semantic). One last example - Gibson's 'affordance' - a cat knows to snooze by a sunlit window w/o calculating light level or temperature, a tired kid sits on a flat rock when tired, without measuring area or tilt. The world 'affords' meaning directly, creatures use them w/o thinking etc. Till there is AI that does likewise, there is no parity.
Hi Jim, the years of evolution you mentioned (to be able to run) - exactly. Those years have also not introduced a von Neumann stored program architecture in any brain, because that cannot evolve by itself autopoietically. If we architect AI without long evolution (eg use brain cell organoids), they might behave similar to real brains. Then there is the mind, Self, fractured identity, mental illness, suicidal thoughts etc etc [in real brains] we might be *very* far from replicating all this :) Conversely, casting all these aside as irrelevant takes us right back to narrow AI (Arthur Samuels' checker player), which we had back in 1952 which is pre-Dartmouth AI workshop (omg).
It is interesting that we seem to be able to build a statistical model of AI-generated images in our brains after seeing quite a few. As with AI-generated text, we get a "feel" for the kind of mistakes it makes. One develops an intuitive feel for the things it gets right and wrong. So will we learn to like this AI "style"? Let's hope not.
First instance of using genAI to avoid working with a potential fraudster. Asked to look into an outside company that wanted to do business with my company. One of the founders picture looked off, and the accompanying biography had no verifiable details. Reverse image search. Totally different person, in the exact same outfit. So it was the fishy company’s obvious use of an image generator (not me using it) that tipped us off.
And all that fun and games is destroying the planet. Worse than netflix. Maybe just take out some coloring pens and be an ecologically responsible person? Being an adult here.
have you seen the amount of forever chemicals in coloring pens... yikes, you don't want to know....🤣
"You can't really dust for vomit." -- Nigel Tufnel
I am not a bass player, but I do play guitar, and I have some familiarity with how the bass is played. Tuners fading in and out of reality or not, that is not how one plays it, at all.
Alas, AI slop is taking over the world.
All of this stuff is just stuporficial.
Its mediocrity personified.
The people using it are either too dumb to notice or too indifferent to care.
I think it is both. A lot of the masses are in the first category. And I count a lot of my peers in that group.
But my bosses, and executives at my (and I suspect a LOT of companies) are in the latter group. They view GenAI as a tool to allow them to at least not hire any new people, and at worst, to lay off a lot of people and let the remainders use these tools to (badly) fill in on what they used to do.
Lather, rinse, repeat
The mindless acceptance of the stuporficiallity by many people -- even by some scientists, engineers and others who are normally very skeptical and demanding of evidence is actually surprising to me.
It's as if these people simply switch off their critical faculties when they use this stuff.
If (when?) the LLMs are used for safety or security criticsl systems, the lack of human oversight is a recipe for disaster
These images would make for great visual puzzles. "Name everything that you see that is wrong with this picture. Answers on the back."
If an LLM is trained on videos, does it make fewer mistakes when asked to draw a picture than one trained on text and images? If it doesn't, this would appear to [at least partially] falsify the claim that embodying AI will improve its world models.
It can be even more fun.... Ask the model to fix the mistakes..... Over and over and over and....
Now you have opened that door, is Open AI the new Spinal Tap of Silicon Valley? Only time will tell…
This seems to be a discussion about visual accuracy, yes? We’re not discussing art yet, right? I cannot, despite trying many models and many many prompts I’ve been unable to get an LLM to generate any single thing that looks unique and truly interesting. All slop. Makes me wonder how things might be different if visual artists had been involved in the training? I’m sure some of you remember when big box stores and global brands started taking over neighborhoods and city streets, homogenizing human experience. Is this any different?
On the whole, as a direct result of how it works, generative AI is a race to the middle , to mediocrity and blandness.
You are right that it is effectively a continuation of the Wallmartization of humanity, albeit in a way that affects every aspect of human society and on a much larger scale.
The "solution" to human evolution is dilution with AI slop.
One thing I'm noticing from your posts is that there are a surprising number of people completely content with not looking any closer. Yet still claiming some kind of correctness about the model's generative ability. Which is ironic. The models confidently output physically, and metaphysically, inaccurate outputs. Then humans confidently ascribe these outputs the weight of a PhD. Obviously some people have neither experienced what a PhD actually does on a day to day basis nor can they imagine it. And that's what these models are being trained on. It's quite the cycle, honestly.
"Not looking any closer" is the M.O. of the vast majority of people doing the vast majority of tasks. :D
Just tossing this one out there since you expressed uncertainty. I'm no lighting expert, but I know a little geometry and messed around with ray tracers long ago. The position of a reflection is a function only of the positions of the viewer, reflective surface, and objects being reflected. Lighting position can change the intensity or visibility of a reflection, but not the position of it. This one actually looks not too bad.
Fun read as always!
You are right (and a reflection can certainly be at right angles to the shadow.) I agree that the reflection in that image looks believable.
On the other hand, the shadow in the first (spokeless wheel) image looks like a Komodo dragon trailing a piece of string.
But , as an amateur artist myself, i can commiserate with the chatbots. Shadows and reflections are difficult to get just right, although i must admit i never included a komodo dragon shadow in any of my paintings.
Then again, i never included a bike wheel without spokes either.
But maybe im just way behind the curve when it comes to art.
After all, Jacson Polkack never worried about shadows in his paint splatter works. But he had lots of spokes!
It is a testament to Sam Altman's foresight that none of his AGI predictions have come true yet.
Gary has a secret decoder diagram--AI-produced--used to name guitar parts.
The headstock holds the tuners, and 4 strings pass over the nut. At the nether end, 5 strings pass over the bridge on their journey to the tailstock.
Though I think Jaco's axe was something like this.
Expect to see it when he performs next😎
It's not easy to ride a unicycle at the beach. Wet packed sand is difficult. But dry sand is almost impossible.
i have done it with really fat tires :)
Would you describe our ability to verify the accuracy of some physical system as essentially an abstract cross modal verification system? By sense of touch, sight, sound, personal consequence or consequence of others, etc. we've built up the ability to make verification significantly less lossy. A model trained only on text and images within deterministic unimodal internal state seems incapable of developing these abstract cross modal verification systems.