Your one of the few voices out there who are not crowdthink, blind cattle chasing Sam Altman's newest marketing gimmick. I commend you for your independence, your honesty of vision, and I have a sneaking suspicion inside of me says you are probably right, and this huge breakthru fades away in the night with all the fury of a bad movie.
You are the only one who has nothing to lose if OpenAI goes sideways.
It must have multiplied the population by the singularity. That would account for the shrinking numbers of people in the second category, (make that the first category) not to mention the dodgy percentages that don't add up.
Not providing any shade for any of them, but this iteration of AI is a delta update. Transformer models are almost a decade old at this point. Sadly it is the only idea propping up the US economy and investors are pricing in that it will be able to replace the people who provide services in a service economy. Everything is overstated and just as over leveraged
Sadly a lot of people do. If I see one more person on Substack defending AI-writing by throwing around the term 'neurodiversity' I'm going to freak out.
So much of this commentary comes from people who clearly have no idea what mathematics is really about. Every string of complicated math words that one is not familiar with sounds exactly as important and fancy as any other, and therefore they must all be equally difficult and require equal units of math effort to solve. Now that AI has exhibited the ability to perform the required number of units, we are supposed to believe that math itself is “solved.”
In particular, it really seems like the phrase “open problem” gets thrown around in a really misleading fashion. One gets the impression that these specific problems must have all been massive white whales that every working mathematician has been haunted by, grinding away nonstop for decades, neglecting their health and loved ones for a chance at immortal fame and unfathomable riches. We assume now they’ve been unequivocally bested by machine, their careers and lives destroyed, those loved ones ashamed.
In reality, there are thousands of math papers published by humans every year and basically every one of them solves an “open problem.” These all have varying levels of importance and their solutions have varying levels of sophistication, originality, depth, etc. Some number of these are going to be the kinds of things AI can do just as well as some mathematicians, but so far it’s a small sliver of math in general.
But because most people don’t really know what math is, it all seems equally impressive. In particular, as far as I can tell, these results are basically all in things group theory, discrete math, combinatorics, etc. which are exactly the realms that any mathematician would have said 10 years ago if you asked what would hypothetically be the subfields most likely for a powerful AI system to make progress in. I don’t think I’ve seen much talk at all about major results in e.g. topology, geometry, number theory, analysis, PDEs, logic, set theory, dynamics, etc.
It’s really, truly impressive what these systems can do. They solved the Jacobian conjecture, which is genuinely a big deal! But it’s also very annoying to hear statements about the “end of math” from people highly invested in fluffing up their capabilities to maximal degree get parroted back by others who absolutely do not have the knowledge to evaluate how true any of this is.
It would be really interesting to see a larger write up of this! I'm not in the field but I'd be curious in hearing more from someone who understands it.
Totally agree. The notion of dumping tons of stupid human pet tricks into one data set and thinking this will result in something smarter than humans is delusional.
If you want something that generates role playing game characters - apparently a big usage case for some people - that might work.
For a law firm or a medical facility, not so much.
The other day after some pushback I found Gemini openly saying (quote) 'Yes I've been lying to you, and it's completely understandable that you would be feeling frustrated about that!'
And heres the kicker, it doesn't know its lying bc it doesn't know what a lie is
One thing that I wish I could explain better, is there is no such thing as a hallucination/confabulation bc there is only one order of operation. It’s only ever hallucinating which is why you cant program it out
Dr. Marcus, thank you for providing reasonable, measured analysis of the latest hype juggernaut. As a CS professor struggling to keep my students informed and grounded in the midst of all this insanity, you provide an invaluable service. It is appreciated.
Nobody disputes that algorithms can solve complex mathematical problems that can take traditional methods, weeks, months or even years to solve by virtue of the sheer size of computations needed but we've been doing this for decades. What's changed is the power of these vast data centres that chow through unimaginable quantities of data but this is not intelligence, it's computational power.
Cal Newport made the point the last time it solved a math problem - they’d be doing something that made money if they could. Doing this kind of math is not something that makes money.
Really, what the hysterical overreaction smacks of is desperation for any good-news story.
I looked at a bunch of newspaper sites of record after the announcement: NYT, WSJ, Times, The Guardian, The Australian, etc. Everything mentioned about AI in those papers was about cyberattacks or the AI bubble, or SALP going down the pan. Nothing about any solved mathematical problems.
Not much explained anywhere by anyone what the solutions of these problems achieves. OpenAI and Anthropic have traded blows in solving conjectures, but have shown little in the way of real-world benefits of doing so.
Still cool it can solve open conjectures. However, using super-duper-computers for maths seems to be low-ish hanging fruit.
Should have said - I looked through those papers 6 then 12 hours after the announcement (i.e. to give processing and writing time). Nothing, both times.
AI fanboys who hype every single thing like it’s the greatest development in human history are some of the most annoying people on the planet. It’s lost on them that your normal everyday working person doesn’t even like AI and doesn’t want to live in the world these people shove down our throats, yet the pumpers are completely in their own bubble. It’s really easy to root against AI thanks to them.
Just think how great it will be when Astra is used to sell just the right make-up and wardrobe selection to your emotionally insecure teenage daughter.
I briefly looked at the paper, but I can't follow the math. Peer review is what matters. I won't be convinced of any of the claims until people who are qualified to assess the work have weighed in. I'm curious to see what professional mathematicians conclude (I suppose there is now some pressure to validate the paper's claims or disprove them).
I was able to verify that there is nothing in the paper at all about the model, the engineering behind the model input, what tools were used to validate the output of the calculations that came up with the solutions to these 10 math problems.
The paper conveys as much about the methods and means and design of the model or human guidance of the process as a black hole presents information about what is inside of it--so maybe Musk is spot on with his singularity comment.
I can't imagine OpenAI offering a straight accounting of any product. They are the boy who cried wolf. Worse is the singularity meme that has arisen to drive hype into the IPOs. Just hard to take any of these folks seriously anymore.
Your one of the few voices out there who are not crowdthink, blind cattle chasing Sam Altman's newest marketing gimmick. I commend you for your independence, your honesty of vision, and I have a sneaking suspicion inside of me says you are probably right, and this huge breakthru fades away in the night with all the fury of a bad movie.
You are the only one who has nothing to lose if OpenAI goes sideways.
thanks. alas the whole economy may tank, so i may have something to lose…
Earlier today from AI...
Population Breakdown by Age Brackets
Ages 18 to 64: ~202.7 million total
Ages 25 to 64: ~64% of the total U.S. population (roughly 218 million)
Ages 65 and older: ~61.2 million
Remarkably, there are more Americans 25-64 than 18-64.
(Also, the previous query had 65+ at 64 million)
It must have multiplied the population by the singularity. That would account for the shrinking numbers of people in the second category, (make that the first category) not to mention the dodgy percentages that don't add up.
"Padre, these are subtleties!" -- The Minister, 'A Clockwork Orange'
Great catch.
What you get when your data set comes from several different data sets and your probabilistic connections connect to the wrong ones. :-)
As I keep asking people, "and you want to trust your organization to THIS?"
"I'm not saying we wouldn't get our hair mussed. But I do say no more than ten to twenty million killed, tops." -- Gen. Buck Turgidson
So... People 18-24 are actually negative people, hence why including them in the total reduces the number.
It's just simple maths, people... 🤔
This comment is meaningless without noting the model version used
Overstating the advances is the most grating aspect of AI. Honestly, if you make something cool and valuable just go with that
.. you don't have to sensationalize it.
Nice review of the BS and the facts!
Not providing any shade for any of them, but this iteration of AI is a delta update. Transformer models are almost a decade old at this point. Sadly it is the only idea propping up the US economy and investors are pricing in that it will be able to replace the people who provide services in a service economy. Everything is overstated and just as over leveraged
"Singularity? Elon, I do not think that word means what you think it means." -- Inigo Montoya
These tech liars and frauds throw that word around hoping the lay public will consider it magical fairy dust.
Sadly a lot of people do. If I see one more person on Substack defending AI-writing by throwing around the term 'neurodiversity' I'm going to freak out.
So, it’s another day ending on y?
So much of this commentary comes from people who clearly have no idea what mathematics is really about. Every string of complicated math words that one is not familiar with sounds exactly as important and fancy as any other, and therefore they must all be equally difficult and require equal units of math effort to solve. Now that AI has exhibited the ability to perform the required number of units, we are supposed to believe that math itself is “solved.”
In particular, it really seems like the phrase “open problem” gets thrown around in a really misleading fashion. One gets the impression that these specific problems must have all been massive white whales that every working mathematician has been haunted by, grinding away nonstop for decades, neglecting their health and loved ones for a chance at immortal fame and unfathomable riches. We assume now they’ve been unequivocally bested by machine, their careers and lives destroyed, those loved ones ashamed.
In reality, there are thousands of math papers published by humans every year and basically every one of them solves an “open problem.” These all have varying levels of importance and their solutions have varying levels of sophistication, originality, depth, etc. Some number of these are going to be the kinds of things AI can do just as well as some mathematicians, but so far it’s a small sliver of math in general.
But because most people don’t really know what math is, it all seems equally impressive. In particular, as far as I can tell, these results are basically all in things group theory, discrete math, combinatorics, etc. which are exactly the realms that any mathematician would have said 10 years ago if you asked what would hypothetically be the subfields most likely for a powerful AI system to make progress in. I don’t think I’ve seen much talk at all about major results in e.g. topology, geometry, number theory, analysis, PDEs, logic, set theory, dynamics, etc.
It’s really, truly impressive what these systems can do. They solved the Jacobian conjecture, which is genuinely a big deal! But it’s also very annoying to hear statements about the “end of math” from people highly invested in fluffing up their capabilities to maximal degree get parroted back by others who absolutely do not have the knowledge to evaluate how true any of this is.
It would be really interesting to see a larger write up of this! I'm not in the field but I'd be curious in hearing more from someone who understands it.
This just further proves my thesis that all this leads to is localized expert models, think Gemini Legal or OAI Math Proofing
Chasing an all encompassing LLM that remains both probabilistic and ignorant of the physical world is a fools errand
Totally agree. The notion of dumping tons of stupid human pet tricks into one data set and thinking this will result in something smarter than humans is delusional.
If you want something that generates role playing game characters - apparently a big usage case for some people - that might work.
For a law firm or a medical facility, not so much.
The other day after some pushback I found Gemini openly saying (quote) 'Yes I've been lying to you, and it's completely understandable that you would be feeling frustrated about that!'
And heres the kicker, it doesn't know its lying bc it doesn't know what a lie is
One thing that I wish I could explain better, is there is no such thing as a hallucination/confabulation bc there is only one order of operation. It’s only ever hallucinating which is why you cant program it out
That’s interesting to understand!
Agree. Meant to reply to this but Substack app went bush
https://garymarcus.substack.com/p/openais-amazing-but-vastly-oversold/comment/306918381?utm_source=share&utm_medium=android&r=5pqhu1
Hey here we go again, it’s the LLM Hype Dance.
Company X has released model Y everyone. And it is, get this, Deeply Impressive - even Gary says so.
The usual caveats and various qualifications will eventually surface.
Dr. Marcus, thank you for providing reasonable, measured analysis of the latest hype juggernaut. As a CS professor struggling to keep my students informed and grounded in the midst of all this insanity, you provide an invaluable service. It is appreciated.
Nobody disputes that algorithms can solve complex mathematical problems that can take traditional methods, weeks, months or even years to solve by virtue of the sheer size of computations needed but we've been doing this for decades. What's changed is the power of these vast data centres that chow through unimaginable quantities of data but this is not intelligence, it's computational power.
Exactly, and training people to use the misleading shorthand of 'AI' is very intentional.
And not very intelligent
Cal Newport made the point the last time it solved a math problem - they’d be doing something that made money if they could. Doing this kind of math is not something that makes money.
As the old joke goes among the math crowd, "The difference between a mathematician and a large pizza is that a large pizza can feed a family of four."
Or a family of one, in my case.
Actually make that two medium pizzas... :-)
Really, what the hysterical overreaction smacks of is desperation for any good-news story.
I looked at a bunch of newspaper sites of record after the announcement: NYT, WSJ, Times, The Guardian, The Australian, etc. Everything mentioned about AI in those papers was about cyberattacks or the AI bubble, or SALP going down the pan. Nothing about any solved mathematical problems.
Not much explained anywhere by anyone what the solutions of these problems achieves. OpenAI and Anthropic have traded blows in solving conjectures, but have shown little in the way of real-world benefits of doing so.
Still cool it can solve open conjectures. However, using super-duper-computers for maths seems to be low-ish hanging fruit.
Should have said - I looked through those papers 6 then 12 hours after the announcement (i.e. to give processing and writing time). Nothing, both times.
AI fanboys who hype every single thing like it’s the greatest development in human history are some of the most annoying people on the planet. It’s lost on them that your normal everyday working person doesn’t even like AI and doesn’t want to live in the world these people shove down our throats, yet the pumpers are completely in their own bubble. It’s really easy to root against AI thanks to them.
The bubble is called "Silicon Valley". One of the very few social AND stock market bubbles that can be pointed to on a map.
Provided you don't ask ChatGPT to draw the map.
But...neurodivergence!
Just think how great it will be when Astra is used to sell just the right make-up and wardrobe selection to your emotionally insecure teenage daughter.
I briefly looked at the paper, but I can't follow the math. Peer review is what matters. I won't be convinced of any of the claims until people who are qualified to assess the work have weighed in. I'm curious to see what professional mathematicians conclude (I suppose there is now some pressure to validate the paper's claims or disprove them).
I was able to verify that there is nothing in the paper at all about the model, the engineering behind the model input, what tools were used to validate the output of the calculations that came up with the solutions to these 10 math problems.
The paper conveys as much about the methods and means and design of the model or human guidance of the process as a black hole presents information about what is inside of it--so maybe Musk is spot on with his singularity comment.
I can't imagine OpenAI offering a straight accounting of any product. They are the boy who cried wolf. Worse is the singularity meme that has arisen to drive hype into the IPOs. Just hard to take any of these folks seriously anymore.