I'm sure I've plugged this before, but the following essay by Cosma Shalizi expresses Karpathy's view in more generality: the singularity already happened, it was called the industrial revolution.
Thanks for the coherent article. It is within my 70 years of life to see technical innovation lifecycle and maturation cycles are generational. Think about how long it too us to put wheels on suitcases, or move from mainframe computers to networked desktops. No biz plan stands contact with the consumer and the rate of change in technology. It is way too early to know the true value of Ai. It is evolving and it is a genie now out of the bottle. Remember the REDHAT business that grew from Shareware. There will be candy and blood before the answer is known.
This is a great summary of the limitations of the different “sides” of the AI debate. What you’re suggesting sounds very similar to the Ai as Normal Technology folk.
Mostly that AI will follow a similar trajectory as our other groundbreaking technologies (electricity, automobiles, etc) and take decades to become fully integrated into our economy.
That does seem at odds with the current build out of data centres. How can their cost possibly be covered if we’re not seeing much economic benefit from LLMs?
The US economy is really big, and AI investments today are preparation for AI demand many years into the future. So it's possible that the data centers will be valuable enough to justify the investment, but not so valuable that they move the needle very much in terms of economy-wide GDP growth.
Someone observed that the huge sums Google spends on data center buildouts are less than the even higher amount of free cash flow their business generates. If true, perhaps the damage done by the bubble will be minimal. Google has to do something with that free cash flow, which might instead go into buy backs or dividends.
Here's another analogy for where we are with machine learning/AI: When electronic computers were first developed in the 40s/50s, people were astounded by their calculating speed. The computers could crunch numbers in minutes or hours that took people months to complete. The power awed folks such that The Forbin Project and other computers-take-over depictions soon followed.
Now we have machine learning and its progeny LLMs which are astoundingly good at pattern matching. It can find things faster and more accurately than people ever could. These matching/prediction tools create the gigantic databases of patterns used by Chatbots and enable the lookups. And of course we are astounded by this capability and many are predicting computers-take-over scenarios.
The machine learning/LLM tools will find their uses, but like David and Karpathy say, it will take decades. The current business models are very questionable -- I don't see how they add up. But the folks trying to sell us social-media-like engaging Chatbots have very very deep pockets, and may be able to write off the first attempts to use the technology.
I re-read this article and I think I was bit too harsh. I should mention that pretty much on the Gary Marcus side of the spectrum of views on AI and I have very little tolerance for AI boosters. I still see little evidence that LLMs will deliver significant economic value.
The observation on productivity occurring when the technology is adapted into workflows is the critical point. It needs to be built in a scalable solution. That means embedding in applications. Microsoft is doing a decent job of this with Copilot being embedded in Excel and writing your emails. Google as well with Search.
It’s not some separate technology. It just gets embedded in applications that makes people more productive.
This means much more rapid adoption (through embedding) than electrification because factories don’t need to be redesigned…electric poles don’t need to be dug into the ground etc.
I think it remains to be seen how AI will be adapted into workflows. Adding chatbots to existing applications like Excel or Google search is certainly one possibility. But there are also likely to be other applications that may be more significant in the long run. Think about self-driving cars, for example. At scale I expect this to have a big economic impact, but it's probably going to take another 10-15 years before a majority of cars on the road are autonomous.
I appreciate your point on cars, but do we need AI to enable self driving cars? We already have self-driving cars that use LIDAR, visual systems etc. Perhaps we should split the thinking for the short term and then longer term. Let me touch on the near future.
For now, embedding AI into the applications gets the return. Microsoft charges $57/month for its enterprise suite of applications. Copilot (Microsoft AI) is another $30/month. So, a 50% jump in recurring revenue. Google is similar, their plans are just a bit cheaper ($35 enterprise/$20 AI). I think this is consistent with your comments on the McKinsey study.
The question that you could help with is whether we need LLMs to support these applications or can it be a more limited model. For example, Copilot embedded with Excel. Wouldn't the model underlying this be simpler as its basically a data analysis engine? If this is so, is it cheaper to run (less tokens/electricity) and cheaper to improve/train. Same goes for summarizing meetings (Teams), writing emails etc.
It'd be interesting to get your take on narrow models for particular applications/functions vs. these generalized models that are very difficult to build and expensive to run.
If we assume that is the case (and I truly mean 'If' because I don't know if it is true), then we think about the implications on the need for energy, data farms etc. We're already hearing about cheaper models from China. Is this how they're thinking? It's the Rolex vs Casio strategy. They both tell time...
If you have the contacts, I'd like to know how companies that use AI for economic gains do it. For example, I suspect that Meta and Amazon are both leaders in using AI to improve productivity or, in some cases, to redesign their operations to be more effective based on AI. Can you find out more about where the benefits they harvest come from?
I also understand that Intuit has put a lot of effort into integrating AI into its products in such a way that users feel like the product is more powerful. Do you know what they do? Apparently, the useful AI features are very specific to the product.
I'm sure I've plugged this before, but the following essay by Cosma Shalizi expresses Karpathy's view in more generality: the singularity already happened, it was called the industrial revolution.
http://bactra.org/weblog/699.html
Thanks for the coherent article. It is within my 70 years of life to see technical innovation lifecycle and maturation cycles are generational. Think about how long it too us to put wheels on suitcases, or move from mainframe computers to networked desktops. No biz plan stands contact with the consumer and the rate of change in technology. It is way too early to know the true value of Ai. It is evolving and it is a genie now out of the bottle. Remember the REDHAT business that grew from Shareware. There will be candy and blood before the answer is known.
This is a great summary of the limitations of the different “sides” of the AI debate. What you’re suggesting sounds very similar to the Ai as Normal Technology folk.
Mostly that AI will follow a similar trajectory as our other groundbreaking technologies (electricity, automobiles, etc) and take decades to become fully integrated into our economy.
That does seem at odds with the current build out of data centres. How can their cost possibly be covered if we’re not seeing much economic benefit from LLMs?
The US economy is really big, and AI investments today are preparation for AI demand many years into the future. So it's possible that the data centers will be valuable enough to justify the investment, but not so valuable that they move the needle very much in terms of economy-wide GDP growth.
Someone observed that the huge sums Google spends on data center buildouts are less than the even higher amount of free cash flow their business generates. If true, perhaps the damage done by the bubble will be minimal. Google has to do something with that free cash flow, which might instead go into buy backs or dividends.
Here's another analogy for where we are with machine learning/AI: When electronic computers were first developed in the 40s/50s, people were astounded by their calculating speed. The computers could crunch numbers in minutes or hours that took people months to complete. The power awed folks such that The Forbin Project and other computers-take-over depictions soon followed.
Now we have machine learning and its progeny LLMs which are astoundingly good at pattern matching. It can find things faster and more accurately than people ever could. These matching/prediction tools create the gigantic databases of patterns used by Chatbots and enable the lookups. And of course we are astounded by this capability and many are predicting computers-take-over scenarios.
The machine learning/LLM tools will find their uses, but like David and Karpathy say, it will take decades. The current business models are very questionable -- I don't see how they add up. But the folks trying to sell us social-media-like engaging Chatbots have very very deep pockets, and may be able to write off the first attempts to use the technology.
I regret that I signed up for a paid subscription to this newsletter and will not be renewing my subscription
haha what? I did too off of this article but I don't understand the hate.
I’m sorry you are disappointed! Would love to hear more about why.
I re-read this article and I think I was bit too harsh. I should mention that pretty much on the Gary Marcus side of the spectrum of views on AI and I have very little tolerance for AI boosters. I still see little evidence that LLMs will deliver significant economic value.
The observation on productivity occurring when the technology is adapted into workflows is the critical point. It needs to be built in a scalable solution. That means embedding in applications. Microsoft is doing a decent job of this with Copilot being embedded in Excel and writing your emails. Google as well with Search.
It’s not some separate technology. It just gets embedded in applications that makes people more productive.
This means much more rapid adoption (through embedding) than electrification because factories don’t need to be redesigned…electric poles don’t need to be dug into the ground etc.
I think it remains to be seen how AI will be adapted into workflows. Adding chatbots to existing applications like Excel or Google search is certainly one possibility. But there are also likely to be other applications that may be more significant in the long run. Think about self-driving cars, for example. At scale I expect this to have a big economic impact, but it's probably going to take another 10-15 years before a majority of cars on the road are autonomous.
I appreciate your point on cars, but do we need AI to enable self driving cars? We already have self-driving cars that use LIDAR, visual systems etc. Perhaps we should split the thinking for the short term and then longer term. Let me touch on the near future.
For now, embedding AI into the applications gets the return. Microsoft charges $57/month for its enterprise suite of applications. Copilot (Microsoft AI) is another $30/month. So, a 50% jump in recurring revenue. Google is similar, their plans are just a bit cheaper ($35 enterprise/$20 AI). I think this is consistent with your comments on the McKinsey study.
The question that you could help with is whether we need LLMs to support these applications or can it be a more limited model. For example, Copilot embedded with Excel. Wouldn't the model underlying this be simpler as its basically a data analysis engine? If this is so, is it cheaper to run (less tokens/electricity) and cheaper to improve/train. Same goes for summarizing meetings (Teams), writing emails etc.
It'd be interesting to get your take on narrow models for particular applications/functions vs. these generalized models that are very difficult to build and expensive to run.
If we assume that is the case (and I truly mean 'If' because I don't know if it is true), then we think about the implications on the need for energy, data farms etc. We're already hearing about cheaper models from China. Is this how they're thinking? It's the Rolex vs Casio strategy. They both tell time...
If you have the contacts, I'd like to know how companies that use AI for economic gains do it. For example, I suspect that Meta and Amazon are both leaders in using AI to improve productivity or, in some cases, to redesign their operations to be more effective based on AI. Can you find out more about where the benefits they harvest come from?
I also understand that Intuit has put a lot of effort into integrating AI into its products in such a way that users feel like the product is more powerful. Do you know what they do? Apparently, the useful AI features are very specific to the product.
Thanks.
Excellent analysis, thank you!
There's also the challenge that Generative AI has to, simply put, stop being Degenerative AI to vulnerable individuals.