Nov 23, 2023·edited Nov 23, 2023Liked by Gary Marcus
From an ML perspective, LLMs are an amazing achievement. Who would have expected such impressive (from the perspective of an external observer) performance from such a simple NN model trained with vast amounts of (relatively low-quality) data, and vast amounts of compute?
From an AGI perspective, however, if your objective was to build literally the *worst* AGI possible (relatively weak cognition (lots of data plus statistical induction plus analogical reasoning, but no deduction or abduction), extremely shallow (if any) internal model of the physical universe, appallingly misaligned with aggregate human preferences, no understanding of how it actually works inside, let alone mathematically well-founded), then you couldn't have done a better job.
This AGI perspective is an original and insightful view. I agree, but want to sharpen it on the data aspect: LLMs actually have all their shortcomings /despite/ being trained on /super-high-quality/ data. All the books, Wikipedia articles, scientific publications, etc contain the destilled essence of an immense number of though processes, often pursued over decades or even centuries and by the brightest human minds. Imho this incredibly valuable input is the key source of the powers of LLMs. Just imagine what could emerge from really understanding all of that content!
What's really wrong at Open AI is that all female board members are out, whereas all the men involved still have roles. Adam D’Angelo and Ilya Sutskever were 50% of one faction. Altman and Brockman 100% of the other faction. Net result…. Although technically Altman, Brockman and Sutskever are no longer on the Board, they are still in place. And we got two new rich white guys, Larry Summers and Bret Taylor. I’ve lost track of the score but I think we’re around -12 for Human Intelligence vs 0 for AGI.
On the other hand, one of the most absurd, boneheaded powerplays in relevant history was carried out by the women who are no longer on the board, who both are EA doomers.
I've not kept score on that front, but you seem to be fixated on the gender angle and not addressing the bizarre interim EA CEO appointment, the board trying to merge OpenAI and Anthropic as well.
It was an EA hostile takeover that absolutely failed, and mere coincidence that those "power players" on the board who were female also just-so-happened to be EA doomers.
D'Angelo and Sutskever came around when their feet were put to the fire, the same cannot be said for the cultists
Haha. Me too. My take is that billions of dollars will be wasted on LLMs and OpenAI in the quest for AGI. $100B were wasted on AVs. Investors seemed to have not learned anything from that painful lesson. You and I can save investors busloads of cash but they won't listen. I don't know about you but I'm very affordable. My advice is free.
To say that $100B were wasted on self-driving cars is the cocky statement of a pointy-head who has nothing to show for it.
Cruise's problems were with rushed engineering not "limits of neural nets".
Self-driving cars are the finest achievement of AI up-to-date. A real system in the messy world, made up of many moving parts, drawing from many computer science disciplines, all working together very reliably (Waymo).
Of course there are still problems to be solved. But what we learn will help with large scale deployment of robots and will advance the AI discipline. While academics still sneer and still have nothing.
After $100B and more than a decade of furious work, the AV industry is not any closer to solving the number one problem of self-driving cars: the corner or edge case problem. That counts as a serious waste of money to me. The problem cannot be solved with deep learning or scaling. I'm sorry that you are offended.
There is no one thing called "corner case problem". What we have, is a very complex world. Nothing will ever be foolproof.
Waymo solves problems as we know best. With a lot of engineering work and using the best ideas out there. When better ideas come, they will be adopted.
Your criticism rings hollow since your points are obvious and you have no solution.
To say "self-driving cars industry isn't any closer than 10 years ago" is an academic point with no value. You have to ask what are the concrete challenges that are being encountered on the road, what models were used to address those, and how has it been working.
Waymo has worked very hard on on the challenges they encountered. The goal is to push the accident rate to be very low.
While nothing will ever be foolproof (including humans), you are making an assumption that self driving AI will be forgiven when it makes a mess. I suspect that both the general public and legislators will have little understanding.
I would also like to see the AI trying to navigate any of the "trolley problems" and get away with whatever choice it makes. I would be keen to hear what Waymo does when the proverbial hits the fan: do you prioritize safety of the passengers or 3rd party that suddenly entered the road? Would you take calculated risk and try to save lives at the expense of crashig the car?
Sadly, there will be situations where people will be killed on the road no matter what the AI driver does. I am quite sure Waymo invested a very large number of resources in modeling and evaluating such scenarios.
Any AI car crash, especially having fatalities, will be thoroughly investigated. If, like Cruise, they are found to have done a sloppy job, Waymo will get in trouble. If Waymo cars have fewer deaths than people-driven cars, and looks like the machine did all it could, or somewhat close to it, it will likely be compelled to make improvements.
Or possibly it's a reference to “the most familiar Q is portrayed by John de Lancie. He is an extra-dimensional being of unknown origin who possesses immeasurable power over time, space, the laws of physics, and reality itself, being capable of altering it to his whim. Despite his vast knowledge and experience spanning untold eons, he is not above practical jokes for his own personal amusement, for a Machiavellian or manipulative purpose, or to prove a point. He is said to be almost completely omnipotent and he is continually evasive regarding his true motivations.”
The cynic in me suspects that this whole debacle, along with the hint of a huge breakthrough, is largely marketing (given the how much rapid commercialization was pushed at OAI dev-day). OAI is facing growing competition in the GAI space, and what better way to to crank up the hype engine and keep OAI front-and-center in the news. Reminds me of the crypto pile-on, but with potential for significantly greater negative societal impact (near-term and long-term)
"A.I. is a field that has brilliant people painting wildly diverging but also persuasive portraits of where this is going. The venture capital investor Marc Andreessen emphasizes that it is going to change the world vastly for the better. The cognitive scientist Gary Marcus depicts an equally persuasive scenario about how all this could go wrong."
Yep. The Bing example is just too funny (and too telling). The core capability of this technology is not being truthful, it is being creative (with the truth, among other things)
Great stuff as normal Gary! What I would add about so-called breakthroughs can be summed up by suggesting a read of "Crossing the Chasm," by Geoffrey Moore. He nicely explains the route from early creation and early-adopters to a real industry, and what that takes (tons of additional investment, trial and error, interaction with customers, cost management, distribution systems, building after-market systems, etc.). There is so much willingness to believe the hype (driven by greed and FOMO). I am just watching the Bloomberg long video on FTX/SBF. It would be so instructive for people to see this and be reminded of this kind of people worshiping the Sam Altman is getting. Sam is human and flawed, just like SBF. We are at the earliest stages of creating and understanding how to use AI. We also have quantum computing, robotics and gene-editing breakthroughs on the way. How some of that may integrate to create new products and services is beyond anyone's understanding. That is the long term game which will demand a lot of risk taking. Keep up the good work Gary! At some point a scenario-based learning process may provide some guidance as we learn our way forward.
Nov 23, 2023·edited Nov 23, 2023Liked by Gary Marcus
Perhaps Q* could be a homage to Star Trek's "Q" character - an immensely powerful, god-like being capable of manipulating time, space, and reality itself.
**EDIT: I see someone else already alluded to this in a previous comment
Sutskever was the primary instigator. The other board members held the positions that they were appointed to hold. Providing viewpoints of the dangers of AGI.
The people to blame here are not the people doing their jobs but the people who created the ridiculous boneheaded structure and then allowed it to lose board members to the point of not functioning as a board - and to be clear it was not ever intended to function as a fiducial board. And the people who set that up, and put it in play, are the people who are still in power.
I keep getting back to the amazement about us humans. It is so clear that 'breakthroughs' aren't that. That 'understanding' isn't that (it can't). That 'learning' is not an ability of these models, but it is a misplaced word that stands for optimising parameters in a very large formula (actually, reacting on the prompts — zero-shot, one-shot, few-shot — might be called the 'learning' it is capable of, and that is really shallow).
But while it is pretty clear none of this is anyway a step on the road to OpenAI's professed goal of AGI, the world is awash with people who are convinced it is. Why? This hype is telling us more about the limitations of human intelligence, than about the development of AGI.
In the meantime, the real issue indeed is how these technologies in the hands of 'evil humans' are going to wreak havoc in our societies made up of (all of us) intellects that have little power breaking free of convictions.
I don't think Q* is a big breakthrough. But it does not need to be. This is incremental work.
I think current methods are not a dead-end, and they have a lot to give. What matters for now is assistants that are becoming smarter and more reliable. These methods may also hint at future directions to remove their limitations.
This is basically what I said yesterday in response to the Reuter's report, but from a more philosophical and biting angle (and mentioning the Longtermism and Effective Altruism at work behind the scenes with some on the board and in the AI field in general).
you debunked something that wasn't announced, released, or demonstrated, well done... but yeah it's true there was a lot of hype and not a lot of actual info
existing chatgpt does poorly on novel math and programming problems, sort of BSes its way through stuff it has seen before in training data, scores on high school math in the GPT-4 paper were not good.
From an ML perspective, LLMs are an amazing achievement. Who would have expected such impressive (from the perspective of an external observer) performance from such a simple NN model trained with vast amounts of (relatively low-quality) data, and vast amounts of compute?
From an AGI perspective, however, if your objective was to build literally the *worst* AGI possible (relatively weak cognition (lots of data plus statistical induction plus analogical reasoning, but no deduction or abduction), extremely shallow (if any) internal model of the physical universe, appallingly misaligned with aggregate human preferences, no understanding of how it actually works inside, let alone mathematically well-founded), then you couldn't have done a better job.
I await further details of Q* with trepidation...
This AGI perspective is an original and insightful view. I agree, but want to sharpen it on the data aspect: LLMs actually have all their shortcomings /despite/ being trained on /super-high-quality/ data. All the books, Wikipedia articles, scientific publications, etc contain the destilled essence of an immense number of though processes, often pursued over decades or even centuries and by the brightest human minds. Imho this incredibly valuable input is the key source of the powers of LLMs. Just imagine what could emerge from really understanding all of that content!
What's really wrong at Open AI is that all female board members are out, whereas all the men involved still have roles. Adam D’Angelo and Ilya Sutskever were 50% of one faction. Altman and Brockman 100% of the other faction. Net result…. Although technically Altman, Brockman and Sutskever are no longer on the Board, they are still in place. And we got two new rich white guys, Larry Summers and Bret Taylor. I’ve lost track of the score but I think we’re around -12 for Human Intelligence vs 0 for AGI.
yep smh
On the other hand, one of the most absurd, boneheaded powerplays in relevant history was carried out by the women who are no longer on the board, who both are EA doomers.
I've not kept score on that front, but you seem to be fixated on the gender angle and not addressing the bizarre interim EA CEO appointment, the board trying to merge OpenAI and Anthropic as well.
It was an EA hostile takeover that absolutely failed, and mere coincidence that those "power players" on the board who were female also just-so-happened to be EA doomers.
D'Angelo and Sutskever came around when their feet were put to the fire, the same cannot be said for the cultists
"Me being me, I called bullshit ..."
Haha. Me too. My take is that billions of dollars will be wasted on LLMs and OpenAI in the quest for AGI. $100B were wasted on AVs. Investors seemed to have not learned anything from that painful lesson. You and I can save investors busloads of cash but they won't listen. I don't know about you but I'm very affordable. My advice is free.
To say that $100B were wasted on self-driving cars is the cocky statement of a pointy-head who has nothing to show for it.
Cruise's problems were with rushed engineering not "limits of neural nets".
Self-driving cars are the finest achievement of AI up-to-date. A real system in the messy world, made up of many moving parts, drawing from many computer science disciplines, all working together very reliably (Waymo).
Of course there are still problems to be solved. But what we learn will help with large scale deployment of robots and will advance the AI discipline. While academics still sneer and still have nothing.
After $100B and more than a decade of furious work, the AV industry is not any closer to solving the number one problem of self-driving cars: the corner or edge case problem. That counts as a serious waste of money to me. The problem cannot be solved with deep learning or scaling. I'm sorry that you are offended.
There is no one thing called "corner case problem". What we have, is a very complex world. Nothing will ever be foolproof.
Waymo solves problems as we know best. With a lot of engineering work and using the best ideas out there. When better ideas come, they will be adopted.
Your criticism rings hollow since your points are obvious and you have no solution.
To say "self-driving cars industry isn't any closer than 10 years ago" is an academic point with no value. You have to ask what are the concrete challenges that are being encountered on the road, what models were used to address those, and how has it been working.
Waymo has worked very hard on on the challenges they encountered. The goal is to push the accident rate to be very low.
While nothing will ever be foolproof (including humans), you are making an assumption that self driving AI will be forgiven when it makes a mess. I suspect that both the general public and legislators will have little understanding.
I would also like to see the AI trying to navigate any of the "trolley problems" and get away with whatever choice it makes. I would be keen to hear what Waymo does when the proverbial hits the fan: do you prioritize safety of the passengers or 3rd party that suddenly entered the road? Would you take calculated risk and try to save lives at the expense of crashig the car?
Sadly, there will be situations where people will be killed on the road no matter what the AI driver does. I am quite sure Waymo invested a very large number of resources in modeling and evaluating such scenarios.
Any AI car crash, especially having fatalities, will be thoroughly investigated. If, like Cruise, they are found to have done a sloppy job, Waymo will get in trouble. If Waymo cars have fewer deaths than people-driven cars, and looks like the machine did all it could, or somewhat close to it, it will likely be compelled to make improvements.
Or possibly it's a reference to “the most familiar Q is portrayed by John de Lancie. He is an extra-dimensional being of unknown origin who possesses immeasurable power over time, space, the laws of physics, and reality itself, being capable of altering it to his whim. Despite his vast knowledge and experience spanning untold eons, he is not above practical jokes for his own personal amusement, for a Machiavellian or manipulative purpose, or to prove a point. He is said to be almost completely omnipotent and he is continually evasive regarding his true motivations.”
https://en.wikipedia.org/wiki/Q_(Star_Trek)
ha I am not omnipotent but most of these blog posts are written primarily for my own personal amusement lol
Elon alludes to a Q*Anon reference... https://twitter.com/elonmusk/status/1727493310606954897
The cynic in me suspects that this whole debacle, along with the hint of a huge breakthrough, is largely marketing (given the how much rapid commercialization was pushed at OAI dev-day). OAI is facing growing competition in the GAI space, and what better way to to crank up the hype engine and keep OAI front-and-center in the news. Reminds me of the crypto pile-on, but with potential for significantly greater negative societal impact (near-term and long-term)
Gary is mentioned in this article by David Brooks:
https://www.nytimes.com/2023/11/23/opinion/sam-altman-openai.html
"A.I. is a field that has brilliant people painting wildly diverging but also persuasive portraits of where this is going. The venture capital investor Marc Andreessen emphasizes that it is going to change the world vastly for the better. The cognitive scientist Gary Marcus depicts an equally persuasive scenario about how all this could go wrong."
Yep. The Bing example is just too funny (and too telling). The core capability of this technology is not being truthful, it is being creative (with the truth, among other things)
Large Lying Models (credit to Melanie Mitchell) :-)
damn that’s good
Great stuff as normal Gary! What I would add about so-called breakthroughs can be summed up by suggesting a read of "Crossing the Chasm," by Geoffrey Moore. He nicely explains the route from early creation and early-adopters to a real industry, and what that takes (tons of additional investment, trial and error, interaction with customers, cost management, distribution systems, building after-market systems, etc.). There is so much willingness to believe the hype (driven by greed and FOMO). I am just watching the Bloomberg long video on FTX/SBF. It would be so instructive for people to see this and be reminded of this kind of people worshiping the Sam Altman is getting. Sam is human and flawed, just like SBF. We are at the earliest stages of creating and understanding how to use AI. We also have quantum computing, robotics and gene-editing breakthroughs on the way. How some of that may integrate to create new products and services is beyond anyone's understanding. That is the long term game which will demand a lot of risk taking. Keep up the good work Gary! At some point a scenario-based learning process may provide some guidance as we learn our way forward.
Perhaps Q* could be a homage to Star Trek's "Q" character - an immensely powerful, god-like being capable of manipulating time, space, and reality itself.
**EDIT: I see someone else already alluded to this in a previous comment
Sutskever was the primary instigator. The other board members held the positions that they were appointed to hold. Providing viewpoints of the dangers of AGI.
The people to blame here are not the people doing their jobs but the people who created the ridiculous boneheaded structure and then allowed it to lose board members to the point of not functioning as a board - and to be clear it was not ever intended to function as a fiducial board. And the people who set that up, and put it in play, are the people who are still in power.
I keep getting back to the amazement about us humans. It is so clear that 'breakthroughs' aren't that. That 'understanding' isn't that (it can't). That 'learning' is not an ability of these models, but it is a misplaced word that stands for optimising parameters in a very large formula (actually, reacting on the prompts — zero-shot, one-shot, few-shot — might be called the 'learning' it is capable of, and that is really shallow).
But while it is pretty clear none of this is anyway a step on the road to OpenAI's professed goal of AGI, the world is awash with people who are convinced it is. Why? This hype is telling us more about the limitations of human intelligence, than about the development of AGI.
In the meantime, the real issue indeed is how these technologies in the hands of 'evil humans' are going to wreak havoc in our societies made up of (all of us) intellects that have little power breaking free of convictions.
Hi Gerben, we humans have this most useful thing - our bodies :) LLMs, not so much...
Human stupidity is far more dangerous than AGI.
I don't think Q* is a big breakthrough. But it does not need to be. This is incremental work.
I think current methods are not a dead-end, and they have a lot to give. What matters for now is assistants that are becoming smarter and more reliable. These methods may also hint at future directions to remove their limitations.
This is basically what I said yesterday in response to the Reuter's report, but from a more philosophical and biting angle (and mentioning the Longtermism and Effective Altruism at work behind the scenes with some on the board and in the AI field in general).
https://twitter.com/itsgottabenew/status/1727517681849626904
https://gist.github.com/B-R-P/89db51ca89a5170a88b107bce15c76f9
Small program making use of q-star equation/algorithm
thanks!
you debunked something that wasn't announced, released, or demonstrated, well done... but yeah it's true there was a lot of hype and not a lot of actual info
given that it solved math word problems, might be related to this work https://openai.com/research/improving-mathematical-reasoning-with-process-supervision
existing chatgpt does poorly on novel math and programming problems, sort of BSes its way through stuff it has seen before in training data, scores on high school math in the GPT-4 paper were not good.