"average person eats 3 spiders a year" factoid actualy just statistical error. average person eats 0 spiders per year. Spiders Georg, who lives in cave & eats over 10,000 each day, is an outlier adn should not have been counted
This reminds me of quicksand. In the 1980s, kids in my circle were scared of it. I recall a middle school teacher describing how we could safely escape it. But I grew up in Michigan and to this day have never seen quicksand. I’m also pretty sure now that quicksand fatalities must be practically nonexistent.
Re. Gwern, I have to say that the scaling "prediction" is another example of a very low bar. Almost everybody working in AI believed it (many still do), largely on the basis of earlier generations of AI, where many researchers were adamant that their approach would have worked out if only they had been funded for a supercomputer. It's also the assertion of the famous (infamous?) "Attention is all you need" paper that kicked off the LLM space race.
Now for a slightly higher bar: Did Gwern also predict that the improvement would tail off, and if so, when? (Answer: roughly, GPT4) or even according to some observers, plateau entirely or worse, decline by some measures? This is only a slightly higher bar because to any objective observer, as it was clear that once you had plagiarized all of the quality content on the internet (again, roughly GPT-4), you were reduced to scraping ever lower quality stuff if you wanted to increase the quantity. Of course, there are things you can do to improve the models themselves, but as long as you are stuck working with LLMs, you can make results faster and cheaper, but there's not much you can do to make them better other than hand tuning to knock down bad results.*
However, it's even worse than that. GPT4 is also the point at which the web began to be populated with AI slop, which could be generated far faster than quality, human-made content. So now AI models are eating their own poop, so to speak, and unlike rabbits, that's not good for them. (The technical term is model collapse.)
Anyway, the key thing is that the "singularity" that AI stans and self-identified futurists like to bang on about is not going to come from an improvement curve that curves down rather than up, and it's not going to come from LLMs.
----
*The one _possible_ exception is coding assistants like Claude Code, which work in a domain that it's possible LLMs might get quite good at. But based on studies so far as well as uptake, the jury is still out on that. On the plus side, at least as long as uptake is low, Claude Code is learning from code generated by other code generators...
I really love the way you write - thank you! I was wondering about the experiment about estimating the speed of the cars in the video because isn't guessing a speed different to constructing memory? Doesn't memory at least start with experiencing the thing, yet in the case of the video they were not in the car and were never aware of the actual speed?
...not sure how I randomly stumbled upon this substack. But this is good dang stuff. Keep going. Tiny pressures on the universe of words could eventually lead humanity somewhere good.
I was a work colleague of your Mom’s when she was at EHOVE Career Center. I have a daughter about your age and we would often share stories about our amazing children. I enjoy your Experimental History articles and enjoy even more your reading of them. Tell your mom that Mary Jean said hello.
(10) the thing about PC Insider, PC World and PC Magazine which all existed at roughly the same time (I know because I worked in a number of retail PC stores back in the day), is that the Editor's Choice Award which each mag ran a version of all consistently went to the company produce which bought the most advertising the previous period. Insane you say? Perhaps. I AM NOT SAYING THAT SPIDERS PURCHASE ADVERTISING. Well, maybe I am. IDK. Good work, as always.
(4) Choice overload/underload sounds exactly like the totality of each human’s existence. What if in every moment there was just one choice and it was “Am I trusting Love?”
(12) I’m sure when they were designing the universe they were having lots of fun and probably dancing and maybe drank a bit too much wine.
Adam, with regard to your theory that moral decline is an illusion: I have two counter-examples. First off, there's good evidence that drivers are increasingly more likely to run red lights (link below). Second, there's anecdotal evidence that people are less likely to RSVP to invitations. I recently had to plan a birthday celebration, and I encountered that problem (also, two people finally RSVPed that they were coming, but they never showed up). I don't have solid statistics, though, so the RSVP thing could be false.
For those unaware of the classsic:
"average person eats 3 spiders a year" factoid actualy just statistical error. average person eats 0 spiders per year. Spiders Georg, who lives in cave & eats over 10,000 each day, is an outlier adn should not have been counted
Thanks again for bringing joy to the internet.
So glad you're on this planet.
This reminds me of quicksand. In the 1980s, kids in my circle were scared of it. I recall a middle school teacher describing how we could safely escape it. But I grew up in Michigan and to this day have never seen quicksand. I’m also pretty sure now that quicksand fatalities must be practically nonexistent.
"Bigger than a breadbox" was, in the 20th century, a feasible way to convey size data conversationally.
>> Officer, how fast was I going when ...
:D :D :D
Re. Gwern, I have to say that the scaling "prediction" is another example of a very low bar. Almost everybody working in AI believed it (many still do), largely on the basis of earlier generations of AI, where many researchers were adamant that their approach would have worked out if only they had been funded for a supercomputer. It's also the assertion of the famous (infamous?) "Attention is all you need" paper that kicked off the LLM space race.
Now for a slightly higher bar: Did Gwern also predict that the improvement would tail off, and if so, when? (Answer: roughly, GPT4) or even according to some observers, plateau entirely or worse, decline by some measures? This is only a slightly higher bar because to any objective observer, as it was clear that once you had plagiarized all of the quality content on the internet (again, roughly GPT-4), you were reduced to scraping ever lower quality stuff if you wanted to increase the quantity. Of course, there are things you can do to improve the models themselves, but as long as you are stuck working with LLMs, you can make results faster and cheaper, but there's not much you can do to make them better other than hand tuning to knock down bad results.*
However, it's even worse than that. GPT4 is also the point at which the web began to be populated with AI slop, which could be generated far faster than quality, human-made content. So now AI models are eating their own poop, so to speak, and unlike rabbits, that's not good for them. (The technical term is model collapse.)
Anyway, the key thing is that the "singularity" that AI stans and self-identified futurists like to bang on about is not going to come from an improvement curve that curves down rather than up, and it's not going to come from LLMs.
----
*The one _possible_ exception is coding assistants like Claude Code, which work in a domain that it's possible LLMs might get quite good at. But based on studies so far as well as uptake, the jury is still out on that. On the plus side, at least as long as uptake is low, Claude Code is learning from code generated by other code generators...
I really love the way you write - thank you! I was wondering about the experiment about estimating the speed of the cars in the video because isn't guessing a speed different to constructing memory? Doesn't memory at least start with experiencing the thing, yet in the case of the video they were not in the car and were never aware of the actual speed?
...not sure how I randomly stumbled upon this substack. But this is good dang stuff. Keep going. Tiny pressures on the universe of words could eventually lead humanity somewhere good.
I was a work colleague of your Mom’s when she was at EHOVE Career Center. I have a daughter about your age and we would often share stories about our amazing children. I enjoy your Experimental History articles and enjoy even more your reading of them. Tell your mom that Mary Jean said hello.
(10) the thing about PC Insider, PC World and PC Magazine which all existed at roughly the same time (I know because I worked in a number of retail PC stores back in the day), is that the Editor's Choice Award which each mag ran a version of all consistently went to the company produce which bought the most advertising the previous period. Insane you say? Perhaps. I AM NOT SAYING THAT SPIDERS PURCHASE ADVERTISING. Well, maybe I am. IDK. Good work, as always.
(4) Choice overload/underload sounds exactly like the totality of each human’s existence. What if in every moment there was just one choice and it was “Am I trusting Love?”
(12) I’m sure when they were designing the universe they were having lots of fun and probably dancing and maybe drank a bit too much wine.
I knew it was the bear all along.
Thanks Adam, great work as always. Love you.
Adam, with regard to your theory that moral decline is an illusion: I have two counter-examples. First off, there's good evidence that drivers are increasingly more likely to run red lights (link below). Second, there's anecdotal evidence that people are less likely to RSVP to invitations. I recently had to plan a birthday celebration, and I encountered that problem (also, two people finally RSVPed that they were coming, but they never showed up). I don't have solid statistics, though, so the RSVP thing could be false.
https://www.timesleader.com/news/1663465/over-a-thousand-lives-lost-every-year-as-red-light-running-rages-on