
The AI Awakening: Implications for Business and the Economy – Business & Generative AI
What this covers
This is a single-speaker presentation by Erik Brynjolfsson, given at Wharton's Business & Generative AI Workshop in September 2024. It begins with why AI counts as a general-purpose technology. He sets out the three changes that made machine learning powerful: vastly more compute, abundant digital data, and better algorithms. He then places generative AI as a third wave, in which machines create content rather than only execute instructions or predict outputs. Scaling laws and the AI Index serve as evidence of rapid progress and as the rationale for hyperscaler spending.
The empirical center is a call-center study of an LLM-based assistant from Cresta. It found a 14% gain in resolutions per hour, with the largest gains, roughly 30-35%, among the lowest-skill quintile and near zero for the most skilled. He reads this as the reverse of the usual skill-biased pattern. Agents also performed better than those who had never had the tool when it suffered outages, which he takes as a sign of learning rather than dependence. He adds that other studies are finding similar patterns.
The later sections are more contested. He argues that The Turing Trap is a problem both of ambition and of income distribution, and that labor income is a decentralized mechanism that could fail if machines make workers unnecessary. He also floats the worry that large productivity gains could hurt the macroeconomy if demand lags, using 1930s farming as the analogy. On complementarity, Garry Kasparov's centaur chess is the example, followed by the point that AlphaZero has outgrown human advice. Other topics include the Stages of Automation (Levels 0-5)) applied to autonomous vehicles, and the difficulty of interpreting neural networks. On interpretability he cites Eric Schmidt and Henry Kissinger on an coming "oracle era", and compares it to human confabulation about our own motives. He closes with the incentives, tax rules and framing that currently push toward automation.
Brynjolfsson argues that AI is the most general-purpose technology ever created, with transformative economic potential, but current development incentivizes full automation (the 'Turing trap') which could concentrate wealth and reduce labor income; instead, human-AI augmentation should be the goal, and policymakers must actively reshape incentives to ensure shared prosperity.
- AI as GPT is fundamentally different from prior technologies because it can improve continuously and spawn complementary innovations across all domains
- Generative AI shows inverse skill bias—benefiting less-skilled workers most—by codifying tacit knowledge, opposite to prior technological disruption patterns
- Full automation to human parity creates infinite productivity but zero labor income, making shared prosperity impossible without active policy intervention
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
When call center agents were stratified by skill level (quintiles), the least skilled workers benefited most from AI assistance with productivity gains near 30-35% for the lowest skill quintile, whereas the most skilled workers experienced near-zero or slightly negative effects, reversing the historical pattern of skill-biased technological change.
“when we broke up the agents into uh five different uh quintiles um what we found was actually was the least skilled workers who benefited the most not the most skilled workers this was not skill bias technical change uh this was was helping the less skilled workers um if you remember earlier um the overall benefit was about 14% but for the least skill workers it was closer to 30% on this metric some of the measures it was up to 35% and the most skill workers you know um it's like not statistically significant than zero if if you really want to say actually the points it's even a little bit negative”
The 'Turing trap' is the danger that pursuing the goal of making machines indistinguishable from humans—following Alan Turing's imitation game test—will paradoxically reduce prosperity and increase inequality because: (1) it's not ambitious enough (merely replicating current human jobs), and (2) it leads to zero labor income if machines can do all human tasks, making labor income distribution impossible and destabilizing.
“I am confident about AI boosting productivity I'm much more worried despite what I just said about it consistently reducing income inequality even though it may help out on the skills side it could shift the value between capital and labor and capital is much more concentrated than labor so if it shifts value over to Capital it may lead to a concentration of wealth and therefore a concentration of power and a particularly concerning aspect of that is um is making machines that closely imitate humans... I don't think it's a good test at all I also think it's even worse as a goal... the goal of what he called strong AI was that they would do the exact jobs that humans do and I think it's tempting to go that way but um it's really not what I think we should be doing”
Machine learning fundamentally differs from traditional programming because instead of humans writing explicit instructions (inputs → program → outputs), machine learning learns the function from training data (inputs and outputs → learned function → outputs), enabling systems to capture tacit knowledge and patterns without explicit codification.
“most coding that you guys have all done um we all do um you take set of inputs and you write down a program a set of instructions and then the computer figures out how to turn that into outputs you have to know exactly what the instructions are... now the good news is that still covers a lot of stuff there's billions not trillions of dollars worth of value in this Paradigm that has been created... the new slide software 2.0 is different um instead of inputs being applied to a function you give it inputs and outputs from training data and then it figures out what the function is so this is a new way of writing software you don't have to know what the function is in advance it can take tacet knowledge it can take patterns in the world and learn what's going on that's why it's called machine learning”
AI is the most important technology of our era, possibly of all time, because it is a general-purpose technology (GPT) that is pervasive, improvable over time, and able to spawn complementary innovations, which are the characteristics that drive economic growth and have historically included steam engines and electricity.
“I think AI is the most important technology of our era you know maybe of all time and that's because AI is a GPT and and in this context GPT means general purpose technology not generative pre-chain Transformer sort of sucks that the AI folks have stolen that acronym from Economist but um and it takes all of those boxes that uh uh uh Tim bresnahan and Manuel shenberg defined of being pervasive able to be improved over time and able to Spa complimentary Innovations and the reason that's important is because gpts basically is what drive the economy they drive economic growth steam engine uh electricity they're the most important Technologies and AI is the GPT of of our era of our generation”
Replicating existing human jobs is too low a ceiling for AI because living standards historically increased not from zero-labor production of existing goods but from invention of entirely new goods and services never previously available
“suppose he had actually succeeded just hypothetically and suppose he had followed Allen Turing or or nilson Nelson's advice let's make machines that do everything that humans do everything the ancient Greeks were doing sounds pretty good what would that mean well it would mean that if you needed to make clay pottery completely automated weaving tunics done H drawn carts machines will do it all you're sick no problem we're going to have the machines burn the incense just like the humans used to well the good news is M humans wouldn't have to work anymore machines would do everything but would our living standards really be that much higher you know we have piles of Greek ears in the corner and like the hor drawn carts would always work but you can see fundamentally that was not what lifted living standards it wasn't sucking the labor out it was doing new things inventing new things things that we never did before”
Alan Turing's 1950 'imitation game' test defined AI as a machine indistinguishable from humans in conversation, making human intelligence the standard by which to measure artificial intelligence, but Brynjolfsson argues this was 'a pretty good idea' at the time and is now 'a terrible idea' both as a test and especially as a goal.
“Alan Turing um wrote the iconic uh paper the definition that's come to become a definition of artificial intelligence the Turing test in his 1950 paper and it basically was based on as it says there in the title The subtitle the imitation game um and he thought that the the way you should Define ai was just like a human human was the standard for intelligence and an AI was indistinguishable from a human ergo it was intelligent and I to think that was a pretty good idea now I think it's a terrible idea”
Three major changes enabled machine learning to become powerful: compute power increased by millions of times compared to decades ago, digital data became abundant (before the 1990s most data was analog), and algorithmic improvements were made, all improving by many orders of magnitude.
“three things have changed one we have much more computer power like millions of times times more computer power than people did a few decades ago and it takes a lot of computer power to do these inferences we have much more data um we didn't have a lot of Digital Data back before the 1990s uh most data was analog and you really need huge amounts of data and finally uh we have some improvements in the algorithms and those are also important but all three of those things have been proved by many orders of magnitude and it's taken something that was an idea and made it something that's incredibly powerful”
Brynjolfsson hypothesizes that massive productivity gains from AI (which he expects) may paradoxically harm macroeconomy if demand doesn't keep up with supply increases: the 1930s farm example shows that even 2-3x productivity gains left farmers worse off because prices fell and total revenue declined—AI could trigger similar demand-supply mismatch.
“although I'm not super into business Cycles I do have a hypothesis that um when you have these tremendous productivity gains which I'm a productivity Optimist I made you know a bet that it's going to go a lot faster um that that always doesn't always turn out well for the macroeconomy paradoxically you get a lot more production and prices fall but if the fed and others don't manage it correctly um demand um may not keep up with the increased Supply and then what happens well in the 1930s what we had was a lot of farmers Having excess production they had tractors and other technologies that allowed them to you know increase productivity by double digit maybe triple digit percentag in some cases so you had all this additional output but we didn't eat like a lot more you know apples and bananas and oranges so um prices fell and they actually made less money than they did when they were producing less”
Interpreting neural networks remains fundamentally challenging: researchers like Jeff Hinton (called 'the Godfather' of neural networks) 'don't totally understand' how LLMs reach conclusions; multiple papers and researchers (Schmidt, Kissinger, Hoetler) argue we may be entering a new era where we must 'trust the Oracle' without full causal understanding.
“IO sber you know he has a nice little video in 60 minutes you can show you how to build a an llm or model you write like two or three PES a code but then you go and you have data and it trains for hours days weeks months and then it's writing all those connections in the neural Nets literally billions in some cases trillions and he doesn't know what all those connections do Jeff Hinton says he doesn't understand what's inside of He's The Godfather um they don't totally understand how they how these systems come to their conclusions”
Garry Kasparov created 'centaur chess' (freestyle chess) after losing to Deep Blue in 1997, demonstrating that human-machine teams could outperform either alone; however, AlphaZero is now so advanced that human advice is as irrelevant as a 'gnat buzzing around' giving suggestions—showing the complementarity model breaks down at sufficiently high AI capability.
“Gary caspero he lost a deep blue in 1997 98 I think it was 97 um and afterwards he's like okay machine can beat me but a human and machine together can beat any machine um so he created centur chests or freestyle chests um and this was very successful for for a while and it fits with what I was talking about the Turing trap having humans and machines work together um but it turns out that's not true anymore Alpha zero is so good that if a human tries to give it advice it's kind of like a a knat buzzing around you saying oh do this do that it's so far ahead of what we humans can do that that we really have little or nothing to add anymore”
Generative AI represents a third wave distinct from traditional programming and machine learning, where machines create novel content rather than just executing instructions or predicting outputs, and this capability surprised even computer scientists.
“but now we're in a third wave which this conference is about which is uh the third wave of the second Machine Age so it's not the just we have to have instructions or learning now it's actually machines that create stuff generative Ai and it's it's happening this one uh is catching a lot of people by surprise including I I walk around the halls of the gates building in at Stanford where all the computer scientists are that I move my office there from the economics Department um because I want to be close to them and I asked them and they are genuinely surprised”
Multiple independent studies beyond the Cresta research are finding similar patterns: quick productivity gains, human-machine outperforming either alone, and less-skilled workers benefiting more than more-skilled workers, suggesting these are not study-specific findings but emerging robust patterns worth documenting more systematically.
“there are is some evidence accumulating I think we still have to see whether there's some you know this is a universal I'm sure it's not a universal but how common this pattern is... the only point I want to make is that while this was one study um I think the patterns that I just described they seem to be showing up a lot that you know number one would getting pretty quick productivity gains we're often having humans and machines outperforming machines alone and we're often having in all these studies having the less skilled workers benefit more than the more skilled workers”
Scaling laws (defined by Darrell Modi and others) show surprisingly predictable performance improvements: as compute, data, and parameters increase by many orders of magnitude, AI model performance improves in a consistent pattern that has held historically and is motivating hyperscaler investments of billions to hundreds of billions of dollars.
“we think it may be progressing is because of scaling laws... as you add more compute more data more parameters there's a surprisingly predictable uh Improvement in performance at least it has been historically and all the hyperscalers you know around a few blocks around from here are putting in billions tens of billions some say up to hundreds of billions of dollars to go down this will we run out of data first or run out of compute first”
Labor income is inherently decentralized as a distribution mechanism: if you don't pay a factory worker, farmer, or professor, you don't get their output; but if machines eliminate the need for workers, there is no economic bargaining power to ensure they continue to receive income, making the distribution system unstable unless maintained through charity, law, or policy.
“one of the things about Labor income is it's inherently decentralized if you don't pay your factory worker if you don't pay your farmer you don't pay your professor whatever then you don't get the output of that person um but if you don't need them anymore then there's no economic bargaining power you may decide to do it out of Charity or kindness or maybe there's a law that's passed and you have to do it but you can see that this may not be a stable equilibrium where these people don't have power or we don't have power depending on if you're a laborer or Capital owner and so that's that's something I worry a little bit about is uh is we have to think a little bit about how are we going to have not just Prosperity but shared Prosperity”
Autonomous vehicles are progressing through layers of automation (0-5, from no automation to full automation), with Waymo currently performing full autonomy (level 5) in limited geographic areas like San Francisco and Phoenix but still facing cost barriers preventing widespread deployment, while Tesla uses a different technical approach (leaner, more data) to achieve partial autonomy and cost competitiveness.
“we're getting there but took a while and we're working our way through these different uh layers of automation hopefully you guys have all heard of you know the six layers of automation healthfully labeled from zero to five um and weo is pretty much over here now you know you don't really need a human it's just the costs aren't quite there uh Tesla is somewhere um I have a Tesla um it's a little scary sometimes when I try to put on autopilot I'm have my fingers crossed but we had a couple near misses um but you know it's they're tring these two different approaches uh weo has a lot of Technology lar in particular um and a lot more compute power Tesla is doing it leaner trying to get more data and we'll see which way converges to a cost effective”
The AI Index, curated by Brynjolfsson and team, aggregates 520+ pages of progress metrics across technical, economic, social, and academic domains, providing quantitative evidence of rapid AI advancement across multiple dimensions rather than anecdotal observations.
“we started this thing called AI index because there's no one place to see all the progress and all the all the evidence on that and so here's one of the charts from our latest report um you can go to the the website there um and and download about 520 pages of charts like this um that try to assemble we try to do some some curation and a validation of the data to see how things are progressing on technical economic social academic lots of different metrics”
Stephen Wolfram argues that LLMs contain layers of linguistic understanding and abstraction that no human has consciously noticed before and may be beyond human comprehension in their full detail, suggesting that AI systems may access or create knowledge structures inaccessible to human minds.
“stepen wolf says that llms have layers of understanding of um of even language our language that no human ever noticed before or and can understand because it's so it's on a whole another level”
Machine learning systems perform better on common, frequent questions (where training data is abundant) than on rare or novel questions (where data is sparse), while humans excel at exception handling and can muddle through with limited data, creating a complementary division where machine-human teams outperform either alone on moderately common tasks.
“if you look at the kinds of question there's some questions that come up very frequently like how do I change my password I'm locked out of the system you know um where do I update and there are some questions that only appear once or twice in the data or hardly at all um and uh you know like some complicated tax questions happen to be Financial tax software um the machine Learning System can't really learn about it doesn't know how to answer this because it's never seen it before has lots of data on the common questions... the machine Learning System does much better on the more want they have more data for we humans on the other hand can muddle our way through um even when we don't have a lot of data on it we're good at exception handling... the humans are better over here the machines are better over here actually in the middle is where the machines turn out to have the biggest value”
Humans are often less understood about their own decision-making than AI systems are understood: when humans introspect about motivations, they often give 'confabulations' (post-hoc stories) rather than accessing their actual neural mechanisms, similar to how we can't fully understand the billions of neurons in our brains—so the philosophical problem of AI interpretability mirrors an unsolved problem in human self-knowledge.
“In fairness humans do this themselves like when when I tell you you know why I did something what my motivations were I'm not sure I'm really understanding the whatever 70 billion neurons in there and how they're all connecting I'm giving you kind of sometimes a what's the word a confabulation kind of a a story that hangs together that makes maybe sense to me and sense to you but it may not get it the the real you know it's it's there's a lay of complexity we can get to but it's a it's a interesting New World we're entering with some philosophical and cognitive implications”
In a study of call center agents using a large language model-based AI system (Cresta, founded at Stanford), when the AI system provided suggestions, call center agents experienced a 14% improvement in resolutions per hour, with this productivity gain being statistically significant and robust across multiple metrics and econometric methodologies.
“on a whole bunch of different metrics um we immediately found that this was before and this was after you just see there was a noticeable shift and uh that's statistically significant uh on uh the number of resolutions you can do an hour or how fast you can do them the number of PA you can do um How likely you are to resolve them um if you want to do a little bit more econometrics you can do difference and differences and you know it's pretty clear jump there... no matter how you slice it whatever methodology you do you get a a clear result there so that was that was was pretty clean”
Data may become the bottleneck limiting AI improvement before compute becomes limiting, but multiple alternatives exist: synthetic data generation (especially for coding tasks, where AI can generate training examples that train better models recursively), identifying efficiency (finding that ~1% of parameters in LLMs are truly useful), and metacognition (planning approaches that improve performance orders of magnitude without requiring large new datasets).
“data is probably going to be the bottleneck and we'll have to figure out either ways of creating synthetic data other ways around it... imagine if we could find a more efficient way to find that that 1% that's one example or imagine we do planning um had a conversation with Bill Gates and he's very excited about what he calls uh metacognition which is you know you guys know that an llm predicts the next word and so if it writes a you know a 5,000 word essay it starts with you know the dog you know it just predicts each word and somehow it's like weirdly like coherent... but most of the time it's not how I write papers um I I start with what do I do you guys you know I write make an outline and then I gather evidence for the different parts and then I adjust the outline... if you ask an LM what are the five steps you should do in writing a paper and then you say do step one make the outline do step two do step three it does a lot better than if it just starts with predicting each word”
There are currently 'excess incentives' pushing technologists, policymakers, and managers toward excessive automation (cost-cutting focus) rather than augmentation, creating misalignment with social welfare; this imbalance should be reoriented through conversation-shifting, tax structure changes, and regulatory rebalancing.
“there are excess incentives right now there should be a balance between full automation some places like dull dirty unpleasant jobs we should do more of those um Dangerous Ones um but other places you know it's not as as effective and I argue that technologists uh policy makers and managers all have excess incentives to push too much to one side than the other and we should we should rebalance those”
Technologists can be persuaded to aim at different AI development targets (augmentation vs. automation): when Brynjolfsson presented the augmentation frame at an AI conference, technologists responded positively saying 'no problem we can make AI that augments people' and 'why didn't you tell us'—showing that the goal direction is not determined by technological constraints but by framing.
“I spoke at at I clear that one of the big AI conferences Yan invite me to give a kind of a keynote there um I gave this other Paradigm and all these people came up to me these technologist and they said oh I didn't realize that you want you want us to aim at this target instead of that Target no problem we can do this target we can make AI that augments people um we can make AI that extends capabil I mean why didn't you tell us would we be happy to do that and so I think there's actually much more um agency in where the technology goes”
Eric Schmidt and Henry Kissinger (with Dan Hutleer) wrote in the Wall Street Journal that technological progress may force humans into a new epistemic relationship with technology: moving from a 'superstition era' (supernatural explanations) to a 'science era' (mechanistic understanding) to a new 'oracle era' (trusting results without understanding mechanisms).
“Eric Schmidt and uh the late Henry Kissinger wrote and Dan Hut leer wrote a kind of a long a little bit sprawling piece in the Wall Street Journal like less than a year ago where they said that there may be a new kind of relationship we have to our technology you know a thousand years ago we were all kind of supers humans were all kind of superstitious you know the rain came and we had like the rain God must have done it or whatever then we had science and we understood okay here's here's our our model and we created Technologies like you were saying okay this is the how to power the steam engine... they say that now we may be going back to an Era where we just have to trust the Oracle we don't know what's actually why it's giving the answer it's giving it's almost kind of mystical and spooky”
There is a natural progression through six layers of automation (levels 0-5) for any task in the economy; different tasks progress at different speeds, and Brynjolfsson hypothesizes there are systematic patterns predicting which tasks advance faster or slower, though this remains an open empirical question.
“imagine not just chess not just cars but every task in the economy going through these stages and my hypothesis is that there's a a natural progression over time some of them go faster some of them go slower if you're interested in thinking about this with me at some point Come and Talk to me afterwards to send me an email because I think I'd like to try to create a taxonomy of when are you going when are you going to go through these faster when you going to go slower will we eventually be able to get here for everything or or there's some things that you'll never get there for”
When the AI system went down (random outages), agents who had been using it continued to perform better than agents who never had access, suggesting they internalized the system's suggestions rather than becoming dependent on it as a 'crutch,' indicating the system functioned as a learning tool rather than causing skill atrophy.
“wonderful for us researchers not for the company every once in a while the system went down it just crashed um and that created a natural experiment as far as we know this was random although maybe the Russians were doing it or something I don't know um and um when the system went down um we could see what happened to these poor people who had been relying on the system were they like lost now well it's kind of noisy because it we don't have nearly as much data on te points when they're down um but if it looks to me just squint a little bit even statistically significant that they contined to do better so people who had been using the system who had access to the system they internalized some of that so rather than it becoming a crutch it actually was more of a learning tool”
Brynjolfsson distinguishes 'AGI' (ambiguously defined as matching human intelligence, which he rejects) from 'transformative AI' (the meaningful economic concept: technology that transforms almost every job, task, process, and company)—the latter is what economists should focus on regardless of whether it's 'truly intelligent.'
“there's another category I'll call transformative AI which is relevant to economists like me which is I'm not going to debate whether it's truly intelligent or whatever I want what I care about is it changing the economy it transforming almost every job almost every task almost every process almost every company if the entire economy is being transformed transformative Ai and there isn't a real good metric of that but but going back to AGI how many of you guys are familiar with metaculus”
John Maynard Keynes (via quotes often attributed to him but actually from a different economist) said that 'most businesspeople are slaves to some defunct Economist without even knowing it,' suggesting business ideology is shaped by economic frameworks and can be changed through alternative frames.
“I think it was John Mayer KES who said that most business peoples are slaves to some defunct Economist without even knowing it that's you guys that's us so um I think that if we can you change the conversation and come up with a philosophy of what are the benefits and costs and come up with these principles um we may be able to to get people to think a little bit differently”
More broadly, the question of what technology does is not 'what will technology do to us' but 'what do we want to do with technology,' and the only way to answer this is through careful empirical and theoretical research by engaged stakeholders—emphasizing human agency over technological determinism.
“the question is not what technology is going to do or what technolog is going to do to us it's what we want to do with technology and the only way we're going to be able to answer that is by doing the kind of careful research that you're all doing in this room so thanks very much for giving me a chance to talk about”
Image classification performance using neural networks has improved dramatically, reaching a point where machines now outperform humans on many image recognition tasks, as shown by image classification crossing the 'human parity' threshold on benchmarks like ImageNet.
“you can see the uh the rapid progress on lots of different metrics I mean like even the first one their image classification that's actually quite a dramatic change um with image net it used to be that you could barely use it as recognizing images and now it's it's in many applications better than humans I put that dotted line there um because it has a lot of economics significance when a machine can do something better than a human”
Large language models tested on standardized exams (SAT, PSAT, GRE) now score at or above the 50th to 90th percentile or better compared to human test-takers, as documented in Microsoft's 'Sparks of AGI' paper with 100+ pages of benchmark charts.
“Eric horid and BBC and the others at uh Microsoft had this you know nice paper uh Sparks of AGI 100 pages of of charts like this one um that I'm sure you've seen about Pro you know that that these can now do the SATs or the PSATs or the Gres or whatever test you took to get into grad school as well as uh you know 50% or 90% or more more of uh of humans”
Within five years, humans will reconceptualize human intelligence as a very narrow, specific kind of intelligence rather than 'general' intelligence, because AI will outperform humans on many dimensions (matrix multiplication, understanding ultrasound, etc.), fundamentally reframing the meaning of AGI and general intelligence.
“my prediction you can write this down within five years people are going to think of human intelligence as a very narrow kind of intelligence they won't think of it as general intelligence B oh humans that's like a very specific narrow kind but they can't multiply matrices they can't we see understand Ultrasonics or you know whatever um but”
Project Apollo is Brynjolfsson's initiative to develop the economic theories, benchmarks, and institutions necessary to govern and understand an economy where AI has much broader transformative capabilities (possibly by 2032 or sooner), requiring both empirical and theoretical work because policymakers are 'flying blind' without such frameworks.
“I'm especially excited about right now is a project to understand the economics of transformative AI we're calling it kind of ambitiously project Apollo which will Define what are the economic theories benchmarks institutions that will govern and understand an economy where AI has much broader capabilities maybe as soon as 2032 maybe sooner maybe later um I'm not sure what the right year is but if there's any reasonable probability of it being in that range I need to get to work on that question... until we have an understanding of those things I think policy makers will be Flying Blind and so there's some empirical work to be done there's some theoretical work to be done there may be some simulations and other things you can do using the these Technologies”
Allen Greenspan told Brynjolfsson (or vice versa: Brynjolfsson asked Greenspan) that digital technologies speed up both information availability and corporate reaction time equally—a 'Red Queen' phenomenon where Fed and companies see faster but improve at the same rate, so overall economic control doesn't improve despite faster processing.
“long time ago Allen Greenspan asked me a similar question about um a digital Technologies just speeding everything up actually I take a back I asked him the question I said digital Technologies are speeding everything up and on the FED can learn what's happening in the economy much faster and are you're able to respond more quickly and he said it was actually like a red queen phenomena that they were able to see things faster but all the companies were able to react faster so they weren't any better”
AGI (artificial general intelligence) predictions on Metaculus have moved dramatically closer: in 2022 the median prediction was 2057 (35+ years away), in 2023 it was 2040 (17 years away), and by 2024 (checking that morning) it was 2032 (8 years away), with detailed multi-page specifications for what counts as AGI including specific benchmark tests and physical task requirements.
“in 2022 the prediction was we would have General AI in 2057 so whatever it is like 30 plus years away 34 years from then um last year the prediction was would be in40 this year the prediction is 2032 I just checked it this morning so uh at the end of 2032 now they the nice thing is they have a very detailed multi-page very detailed description of what they mean by General intelligent has to pass certain tests um the ler prize it has to do certain you know uh physical tasks as well um so it's it's pretty well defined”
Bill Gates is 'very excited' about metacognition as an approach where LLMs explicitly plan steps (e.g., outline → gather evidence → adjust outline → draft → edit) rather than just predicting the next word, and this planning-based approach performs 'a lot better' than token-by-token prediction without requiring large new datasets.
“had a conversation with Bill Gates and he's very excited about what he calls uh metacognition which is you know you guys know that an llm predicts the next word and so if it writes a you know a 5,000 word essay it starts with you know the dog you know it just predicts each word and somehow it's like weirdly like coherent you know it's pretty good I sometimes I write that way but but most of the time it's not how I write papers um I I start with what do I do you guys you know I write make an outline and then I gather evidence for the different parts and then I adjust the outline and then I I maybe make a first draft and then I see how it's coherent and so if you ask an LM what are the five steps you should do in writing a paper and then you say do step one make the outline do step two do step three it does a lot better than if it just starts with predicting each word”
Large language models operate with only about 1% of parameters being truly useful while 99% are 'junk', suggesting enormous efficiency gains possible if sparse useful subsets can be isolated
“a large language model has all these parameters billions of parameters it turns out apparently that a small fraction of only 1% are really useful for almost all the answers the other ones are just kind of junk so imagine if we could find a more efficient way to find that that 1% that's one example”
Earlier technologies (steam engine, electricity) required humans to develop understanding before deployment, but historically sometimes the technology came before the understanding, and this pragmatic precedent suggests AI may follow a similar path where we deploy before fully understanding.
“real historians will tell you that sometimes the technology actually came before the understanding but that's okay we kind of like you know figured it out”
The iRobot movie from the 2030s (set in Chicago 2035) included a scene where a human character challenged a robot by asking 'can a robot write a symphony, can a robot take a blank canvas and turn it into a masterpiece,' expecting the answer to be 'no,' but by 2024, generative AI can do both of these tasks, indicating we are ahead of science fiction timelines.
“I think most of you have probably seen iRobot with Will Smith and there's this iconic scene there where the you know he's trying to give the robot a hard time he says well can a robot write a symphony can a robot take a blank canvas and turn into a masterpiece and during in the movie it's obvious you know robot can't do that no way I mean that's that's ridiculous this was actually supposed to be set in Chicago 2035 and so of course the robots well can't do that so haha humans are better well if they were to do a realistic version of the movie not in 2035 but in 2024 the robot would say yes yeah of course it can do those things um and a lot more so it we are ahead of schedule in terms of what even science fiction writers were imagining”
Customer satisfaction (sentiment) improved measurably when agents had access to AI system, while agent sentiment was already near-ceiling and showed minimal variation or improvement
“you could measure customer sentiment...customers you know there's a whole range of different uh sentiment in the customers I was hoping to be more variation in the agents but it turns out all the agents are very very nice so um almost all of them um so there's not much variation on the agent side um so what happened when you introduced the AI well customers got happier um they were nicer and they were more likely to be happy and satisfied at least according to the way that they talked so that that was that was a good thing agents H it was hard to tell”
Historical precedent: AI research in 1988 focused on hand-coded expert systems and rule-based reasoning, fundamentally different from modern machine learning's data-driven approach
“I'm old enough that when I first started teaching courses on AI back in 1988 um I talk courses about um expert systems rule-based systems we would hand code a lot of different rules and then have the machine carry them out”
Academic tests (SAT, GRE, PSAT) are known to be imperfect measures of practical intelligence and real-world value, so Brynjolfsson acknowledges this limitation but still uses them as quantitative benchmarks for AI capability progress, with the caveat that the benchmarks must be supplemented by productivity studies.
“being AC we all know that that academic uh tests are not a good measure of real practical value right so um we want to look at some of the the productivity effects”
Gabriel Auerbach and Brynjolfsson have written recently about questions raised by scaling laws, suggesting there may be finite limits to certain approaches or resource constraints becoming binding
“it raises a bunch of questions about productivity inequality industrial concentration that Gabriel anger wrote and I wrote about uh less than a year ago”