A bunch of small but important items in recently-breaking AI news: Tesla settled their lawsuit with the Huang family over Walter Huang’s death, for an undisclosed amount of money. “Although Huang’s family acknowledges he was distracted while the car was driving, they argued Tesla is at fault because it falsely marketed Autopilot as self-driving software. They alleged Tesla knew that Autopilot was not ready for prime time and had flaws that could make its use unsafe.” The settlement is striking in part because Musk had previously said this
We are constructing a Real Time/Real World experiment to discover just how destructive stochastic parrots spewing word salad joined with the Argumentum ad Populum Fallacy is.
Re the "scarlet letter": This has become the new normal in science fiction magazines over the last year. The use of AI in SF art or storytelling is toxic. The highest-reputation, best-known, best-paying magazines won't accept AI submissions, with language like this quote from Asimov's: "Statement on the Use of “AI” writing tools such as ChatGPT: We will not consider any submissions written, developed, or assisted by these tools. Attempting to submit these works may result in being banned from submitting works in the future." And if they accidentally run an AI-generated cover, SF presses have been known to _withdraw it and apologize._
I’m very pro AI progress, but I really appreciate the counter balance to the hype that you, @michael Spencer and @alberto Romero are offering recently. Very important to keep a sane, sober perspective on everything.
Another cracker of a post! I'm helping our academics to deal with AI and to see it from various angles. Posts like these are necessary counterweights to the hyperbolic gushing of the AI companies themselves.
I'm sure lawyers, judges, investors, and bankers will find screwed up document chronology (say during discovery process of a huge lawsuit) and 85% factual accuracy at best totally okay for summarizing long documents and it totally won't generate liability. NOT!
I'm not a mathematician nor a stastician but it seems to me if you try to scale up to 450 or 4500 incident free minutes - are you just asking to increases the chances of a screw up? Incident free minutes of any duration may only mean you just got lucky. Correct?
And they are incapable of solving the simplest of cryptograms, an art that can be taught to a 10-year-old. The "make a guess - evaluate the guess - refine the guess" cycle is impossible for a token muncher to achieve.
As far as full self driving goes, it is going to crash and burn the same place the Darpa Autonomous Vehicle did: machine vision. The universe is analog, neural nets are digital and "see" only a sample of what's out there. They are therefore irretrievably vulnerable to spoofing and jamming.
I think it was well done, but I am not particularly interested in tests that are conducted with multiple prompts, because I can't separate the effect of the human devising the prompt with the effect of the prompt being processed. I assess test results with two criteria: minimal prompts, and an accompanying argument that the result could not be achieved by simple pattern recognition and therefore an emergent property may have been demonstrated. But that's just me, YMMV.
In a benign environment, maybe. In a hostile one, absolutely not. Both visual and lidar spoofing have been demonstrated. I can send a pattern of dots to a collision avoidance system that fools it into thinking it's looking at an encroaching vehicle and the same pattern would be rejected by a human.
This stuff has been studied by defense establishments for decades, to a depth which makes the work of commercial machine vision researchers look like child's play. We already have devices for sale that will unlock an electronically locked car. Next step is a device that you can point at an FSD vehicle and cause it to jump into the next lane or slam on its brakes. Or cause it to not see objects in its path.
There's a reason that all aircraft collision avoidance systems are based on communication between entities rather than one-sided deduction of the situation.
"This is an AI Free Zone! Text created by Large Language Models is spreading rapidly across the Internet. It's well-written, artificial, frequently inaccurate. If you find a mistake on Spaceweather.com, rest assured it was made by a real human being."
Hypothesis regarding genAI summarization of scientific literature: summarization tools such as ScopusAI will be biased to select literature that reflects the biases of non expert internet users if the model was trained on poor quality internet data.
This seems to be true when I put things like “is aluminum in deodorant safe” into the tool (it is almost certainly fine for nearly everyone, but people on the internet think it’s not). Someone with more time than me could write a paper.
Meanwhile, apparently it’s exciting that these tools, which were almost certainly trained on the NYTimes crossword puzzle, can solve the NYTimes crossword puzzle. Please hold your applause.
LLMs do exactly one thing: output strings of text in response to inputted strings of text. Every claim of a "new ability" is based on a human being imagining what the LLM must *really* be doing under the hood. We ask ourselves "how could a computer algorithm accomplish this?" and a lot of people really want to answer "it is a nascent but rapidly improving human-like intelligence".
But the default answer should always be "it's doing what it was designed to do", which is output one token at a time using a probability distribution across available tokens, whose values are determined by weights established during its long training and reinforcement process. Those who tout "emergent" new abilities are saying we should ascribe properties of human intelligence to this process: "reasoning" and "learning" and "guessing" and such. As though it can't just be really good at outputting strings of text similar to the text it was trained on. As though it hasn't seen a gazillion crossword puzzles, and gazillion programming problems, a gazillion psychology experiments, etc.
I am already getting numb to all the ridiculous BS currently happening in the AI field, and it's going to get a lot worse once they really hit the wall.
Tesla has been misleading people about the abilities of their cars for decades. They should get what they deserve.
As to self-driving cars in general, nobody, including people who do not like current methods, have an easy path forward.
The best approach, as for any very large and complex project, is to divide it into manageable pieces, do as much honest physics modeling as you can, get best sensors, use best methods, practice caution, and scale up gradually. It is a problem well-worth solving.
Wish he could just go back to his boring business; get autonomous self-driving over there, and roll out AI with some subterranean community to find out what is going to be the societal tradeoff in gain vs damage.
Excellent, looking forward. By the way, really appreciate this substack. I read The Algebraic Mind during my Masters over two decades ago, when it was paired with Churchland's The Engine of Reason for a course on AI/connectionism - quite an inspired combo, looking back - and it's fascinating that so much of what you talked about then sort of fizzled out. Now it's suddenly all come back and even though many of the same arguments are made, the AI/engineer community does not like history (or reasoning), it just has to get in there with its newer and better hammer, to somehow get a problem sorted. But now I'm there as a cognitive psychologist amongst AI folks, and I feel the pain: every time they are willing to entertain reasoning for as long as I'm in the room; but once out, philosophy is off the table, it's just about fixing it.
Hi Jack, pray tell, who, exactly, is shrugging off neural network statistical outlier error-caused deaths as the “the price of doing business”?
Who are these callous, brutal beasts that can look in the face of the deceased families - the children who cannot ever see their fathers again, the parents who never thought they had to bury their children, the soul mates, the beloved siblings - and tell them with a straight face, “but for the good of humanity”?
What guarantees do we have that the unregulated self driving industry won’t repeat these horrific deaths?
Have these pondering people tried to imagine the final seconds the drivers had, to feel the horror and fear, to hear the screams before being torn to shreds/decapitated (yes, true dat)/or burnt to death?
What were the dreams, joys and excited feelings that these same dead drivers had when, earlier in time, they got into their newly purchased vehicles for the first time?
Apples vs oranges… so it’s hard to answer your question.
People choose to be angry/drunk/irresponsible behind the wheel.
But in autonomous driving mode, it’s natural for drivers to lose concentration when the computer takes over. People _actively_ have to fight the attention drift. Their lives depend on it, but few appreciate that.
Folks could argue that a rideshare/taxi and a robotaxi are equally as dangerous - driver could suffer a stroke; the computer could glitch and output an error (“our team is actively working on it”) in the middle of a freeway.
But a robotaxi though might drive erroneously into a dangerous neighbourhood. There’s no driver to yell at, and the AI maps engine may repeatedly arrive at the deduction that it has done its job.
Think it’s too early to comment on driverless buses.
We are constructing a Real Time/Real World experiment to discover just how destructive stochastic parrots spewing word salad joined with the Argumentum ad Populum Fallacy is.
Re the "scarlet letter": This has become the new normal in science fiction magazines over the last year. The use of AI in SF art or storytelling is toxic. The highest-reputation, best-known, best-paying magazines won't accept AI submissions, with language like this quote from Asimov's: "Statement on the Use of “AI” writing tools such as ChatGPT: We will not consider any submissions written, developed, or assisted by these tools. Attempting to submit these works may result in being banned from submitting works in the future." And if they accidentally run an AI-generated cover, SF presses have been known to _withdraw it and apologize._
wow!
I’m very pro AI progress, but I really appreciate the counter balance to the hype that you, @michael Spencer and @alberto Romero are offering recently. Very important to keep a sane, sober perspective on everything.
Another cracker of a post! I'm helping our academics to deal with AI and to see it from various angles. Posts like these are necessary counterweights to the hyperbolic gushing of the AI companies themselves.
I'm sure lawyers, judges, investors, and bankers will find screwed up document chronology (say during discovery process of a huge lawsuit) and 85% factual accuracy at best totally okay for summarizing long documents and it totally won't generate liability. NOT!
I'm not a mathematician nor a stastician but it seems to me if you try to scale up to 450 or 4500 incident free minutes - are you just asking to increases the chances of a screw up? Incident free minutes of any duration may only mean you just got lucky. Correct?
And they are incapable of solving the simplest of cryptograms, an art that can be taught to a 10-year-old. The "make a guess - evaluate the guess - refine the guess" cycle is impossible for a token muncher to achieve.
As far as full self driving goes, it is going to crash and burn the same place the Darpa Autonomous Vehicle did: machine vision. The universe is analog, neural nets are digital and "see" only a sample of what's out there. They are therefore irretrievably vulnerable to spoofing and jamming.
What are your thoughts on this then - https://www.oneusefulthing.org/i/143372526/we-still-dont-know-the-full-capabilities-of-current-frontier-models
I think it was well done, but I am not particularly interested in tests that are conducted with multiple prompts, because I can't separate the effect of the human devising the prompt with the effect of the prompt being processed. I assess test results with two criteria: minimal prompts, and an accompanying argument that the result could not be achieved by simple pattern recognition and therefore an emergent property may have been demonstrated. But that's just me, YMMV.
When you go digital, it isn't always (ever?) possible to know if the baby got thrown out with the (digitized) bathwater.
In a benign environment, maybe. In a hostile one, absolutely not. Both visual and lidar spoofing have been demonstrated. I can send a pattern of dots to a collision avoidance system that fools it into thinking it's looking at an encroaching vehicle and the same pattern would be rejected by a human.
This stuff has been studied by defense establishments for decades, to a depth which makes the work of commercial machine vision researchers look like child's play. We already have devices for sale that will unlock an electronically locked car. Next step is a device that you can point at an FSD vehicle and cause it to jump into the next lane or slam on its brakes. Or cause it to not see objects in its path.
There's a reason that all aircraft collision avoidance systems are based on communication between entities rather than one-sided deduction of the situation.
I want to give a callout to Spaceweather.com which was one of the first with a no-llm message, and I have always cherished them since.
https://spaceweather.com/
"This is an AI Free Zone! Text created by Large Language Models is spreading rapidly across the Internet. It's well-written, artificial, frequently inaccurate. If you find a mistake on Spaceweather.com, rest assured it was made by a real human being."
Hypothesis regarding genAI summarization of scientific literature: summarization tools such as ScopusAI will be biased to select literature that reflects the biases of non expert internet users if the model was trained on poor quality internet data.
This seems to be true when I put things like “is aluminum in deodorant safe” into the tool (it is almost certainly fine for nearly everyone, but people on the internet think it’s not). Someone with more time than me could write a paper.
Meanwhile, apparently it’s exciting that these tools, which were almost certainly trained on the NYTimes crossword puzzle, can solve the NYTimes crossword puzzle. Please hold your applause.
Here is a link to the Steven Overly podcast: https://www.politico.com/podcasts/tech.
The image in this post is just an image, and not a link to the podcast.
@Gary what do you think about the claims made by Ethan here? https://www.oneusefulthing.org/p/what-just-happened-what-is-happening/comments
LLMs do exactly one thing: output strings of text in response to inputted strings of text. Every claim of a "new ability" is based on a human being imagining what the LLM must *really* be doing under the hood. We ask ourselves "how could a computer algorithm accomplish this?" and a lot of people really want to answer "it is a nascent but rapidly improving human-like intelligence".
But the default answer should always be "it's doing what it was designed to do", which is output one token at a time using a probability distribution across available tokens, whose values are determined by weights established during its long training and reinforcement process. Those who tout "emergent" new abilities are saying we should ascribe properties of human intelligence to this process: "reasoning" and "learning" and "guessing" and such. As though it can't just be really good at outputting strings of text similar to the text it was trained on. As though it hasn't seen a gazillion crossword puzzles, and gazillion programming problems, a gazillion psychology experiments, etc.
I am already getting numb to all the ridiculous BS currently happening in the AI field, and it's going to get a lot worse once they really hit the wall.
Tesla has been misleading people about the abilities of their cars for decades. They should get what they deserve.
As to self-driving cars in general, nobody, including people who do not like current methods, have an easy path forward.
The best approach, as for any very large and complex project, is to divide it into manageable pieces, do as much honest physics modeling as you can, get best sensors, use best methods, practice caution, and scale up gradually. It is a problem well-worth solving.
Meanwhile, even though self-driving cars are definitely ready by 2017, 2019, 2021, or maybe never, we'll see AGI by next year, according to Musk: https://www.theguardian.com/technology/2024/apr/09/elon-musk-predicts-superhuman-ai-will-be-smarter-than-people-next-year
Wish he could just go back to his boring business; get autonomous self-driving over there, and roll out AI with some subterranean community to find out what is going to be the societal tradeoff in gain vs damage.
i am saving that one for a separate piece :)
Excellent, looking forward. By the way, really appreciate this substack. I read The Algebraic Mind during my Masters over two decades ago, when it was paired with Churchland's The Engine of Reason for a course on AI/connectionism - quite an inspired combo, looking back - and it's fascinating that so much of what you talked about then sort of fizzled out. Now it's suddenly all come back and even though many of the same arguments are made, the AI/engineer community does not like history (or reasoning), it just has to get in there with its newer and better hammer, to somehow get a problem sorted. But now I'm there as a cognitive psychologist amongst AI folks, and I feel the pain: every time they are willing to entertain reasoning for as long as I'm in the room; but once out, philosophy is off the table, it's just about fixing it.
Hi Jack, pray tell, who, exactly, is shrugging off neural network statistical outlier error-caused deaths as the “the price of doing business”?
Who are these callous, brutal beasts that can look in the face of the deceased families - the children who cannot ever see their fathers again, the parents who never thought they had to bury their children, the soul mates, the beloved siblings - and tell them with a straight face, “but for the good of humanity”?
What guarantees do we have that the unregulated self driving industry won’t repeat these horrific deaths?
Have these pondering people tried to imagine the final seconds the drivers had, to feel the horror and fear, to hear the screams before being torn to shreds/decapitated (yes, true dat)/or burnt to death?
What were the dreams, joys and excited feelings that these same dead drivers had when, earlier in time, they got into their newly purchased vehicles for the first time?
Let us ponder!
Apples vs oranges… so it’s hard to answer your question.
People choose to be angry/drunk/irresponsible behind the wheel.
But in autonomous driving mode, it’s natural for drivers to lose concentration when the computer takes over. People _actively_ have to fight the attention drift. Their lives depend on it, but few appreciate that.
Folks could argue that a rideshare/taxi and a robotaxi are equally as dangerous - driver could suffer a stroke; the computer could glitch and output an error (“our team is actively working on it”) in the middle of a freeway.
But a robotaxi though might drive erroneously into a dangerous neighbourhood. There’s no driver to yell at, and the AI maps engine may repeatedly arrive at the deduction that it has done its job.
Think it’s too early to comment on driverless buses.