What this covers

Ege Erdil and Tamay Besiroglu argue against the near-term artificial superintelligence scenario by reframing what constrains economic growth. Rather than raw intelligence, they identify complementarity as the binding bottleneck: automating real work requires not just better reasoning but vast complementary inputs—compute, capital, data, supply chains, and labor reallocation. The conversation with Dwarkesh Patel spans the economic theory of growth, the trajectory of AI capabilities, and the political implications of different technological futures, grounded throughout in skepticism toward pure-cognition models of progress.

The discussion moves across several contested zones. On compute, Besiroglu argues that remaining scaling headroom is limited to roughly 3–4 orders of magnitude, and that returns to research effort in software show no acceleration pattern supporting a software-only singularity. On economics, both speakers challenge the intelligence explosion framing itself—comparing it unfavorably to a "horsepower explosion" during the Industrial Revolution—and argue that growth will arrive along the extensive margin of new products rather than as unbounded acceleration. They invoke learning curves and Moravec's Paradox to explain why capabilities that impress humans narrow our view. On governance, they apply bargaining theory of war to suggest concentrated AI takeover is economically irrational, and they return repeatedly to limits of reasoning and predictable humility—the observation that detailed advance planning fails in practice—to argue against high-confidence policy blueprints. The overall case rests on treating AI as one complementary input among many, not as the singular driver of an explosive discontinuity.

Sharpest takeaway

Erdil and Besiroglu argue that AGI is roughly three decades away because automating real work requires many complementary capabilities and massive complementary inputs (compute, capital, data, supply chains), not just smarter reasoning — and that this same bottleneck logic rules out a software-only intelligence explosion while still permitting broad, diffuse explosive economic growth.

  • Intelligence/reasoning is not the binding bottleneck on growth; complementary inputs and capital accumulation are
  • AI progress historically tracks compute scaling, and remaining scaling headroom is limited to 3-4 orders of magnitude
  • Because the transition is diffuse and economically driven, value lock-in and concentrated takeover are unlikely, favoring near-term focus and classical liberalism under uncertainty

The claims · ranked53 claims · weighted by value

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.

0.81

Detailed advance planning for novel transitions is unreliable: before WWII the British government expected hundreds of thousands of aerial-bombing casualties in the first weeks, but actual casualties over the entire six-year war were fewer than that, due to boring practical realities (daytime bombing got planes shot down, nighttime bombing was imprecise, firefighters limited damage, aircraft were costly) — so plans made today about the AI transition will likely become obsolete as we learn.

factualhigh valueestablishednovelty 3/4durability 3/4· Ege Erdil

“in all of Britain there were fewer casualties from air bombing in the entire six years of the Second World War than the British government expected in the first few weeks of the war”

0.79

Contrary to its usual pessimistic invocation, the O-ring model (output as a product of many inputs) is actually optimistic about bottlenecks because it implies constant marginal product allowing unbounded scaling; if one input fails to scale you become less efficient but can still grow a lot — so Tyler Cowen should take his own bottleneck view more seriously, yet the correct conclusion is still explosive growth once AI can flexibly substitute across tasks.

factualhigh valuecontestednovelty 4/4durability 3/4· Tamay Besiroglu

“the O-ring model is taking the product of many, many inputs... But actually this is pretty optimistic from the point of view of having fewer bottlenecks.”

0.77

The huge discontinuity between humans and other animals is not mainly raw intelligence but a vastly more efficient training-data modality: culture and language let humans learn from explicit communication and imitation rather than only natural-world experience, creating a feedback loop of selection pressure — implying the human-to-superintelligence jump need not replicate this animal-to-human discontinuity since AIs are not artificially data-starved.

causalhigh valuecontestednovelty 3/4durability 3/4· Ege Erdil

“what seems to have changed with humans compared to other animals is that humans became able to have culture and they have language, which enables them to have a much more efficient training data modality compared to animals”

0.77

The 'intelligence explosion' framing is not a useful concept; it's like calling the Industrial Revolution a 'horsepower explosion' — raw intelligence (like raw physical power) was only one of many complementary changes across sectors that drove the acceleration, so focusing on it misrepresents how the AI transition will actually unfold.

normativehigh valuecontestednovelty 3/4durability 3/4· Tamay Besiroglu

“It's kind of like calling the Industrial Revolution a horsepower explosion. Sure, during the Industrial Revolution, we saw this drastic acceleration in raw physical power, but there are many other things that were maybe equally important”

0.75

The strength of bottlenecks depends on substitutability: materials like aluminum are fairly substitutable (use a worse metal and just pay more in fuel), but input pairs like labor-and-capital, skilled-and-unskilled work, or leadership-and-army-size are far less substitutable — excellent leadership with only 100 soldiers gets you nowhere.

factualhigh valueestablishednovelty 2/4durability 3/4· Ege Erdil

“you have excellent leadership, but your army only has 100 people. You're not going to get very far.”

0.75

The Industrial Revolution's defining change was not producing more of the same pre-industrial goods (more crops, more farmhouses) but the emergence of entirely different products across nearly every sector — transportation, food, healthcare — so the coming AI-driven expansion of consumption will likewise come mainly along the extensive margin of new product variety rather than just more of today's goods.

factualhigh valueestablishednovelty 2/4durability 3/4· Tamay Besiroglu

“along pretty much every main sector of the economy, we just had many different products that are totally different from what was being consumed prior to that”

0.73

Superhuman intelligence is a coherent idea — like superhuman Go ELO, AI capability curves cross the human level without it being privileged — but ASI and AGI are not particularly useful concepts because AI capability profiles are jagged, making averages meaningless, and you can have drastic acceleration without an AI better than humans at everything, or an ASI with no acceleration if it's too slow or expensive; better to reason directly about effects on the world.

definitionhigh valuecontestednovelty 3/4durability 3/4· Tamay Besiroglu

“AI systems have these very jagged profiles of capability. So you have to somehow take some notion of average capabilities and what exactly does that mean, it just feels really unclear.”

0.73

Empirical estimates of the returns to research effort in software (across linear/SAT solvers, computer vision, RL, language modeling) are ambiguous about whether scaling cognitive effort produces accelerating growth or merely exponential growth; combined with the fact that software progress tracks hardware progress (~30%/year matching Moore's law, with the deep-learning acceleration coinciding with compute scaling), this argues against a hyperbolic faster-than-exponential software-only singularity.

factualhigh valuecontestednovelty 3/4durability 3/4· Tamay Besiroglu

“the estimates are a bit ambiguous about whether this results in this acceleration or whether it results in just merely exponential growth”

0.73

The naive model that value misalignment causes conflict is wrong; the conflict literature shows war also requires additional factors like poor information about relative strengths or strong indivisible/sacred commitments — so if AI is powerful and the value at stake is large, the efficient outcome is a negotiated settlement rather than war, making violent takeover less likely though peaceful disempowerment remains possible.

causalhigh valuecontestednovelty 3/4durability 3/4· Tamay Besiroglu

“that's actually not what the literature on what causes conflict says creates conflict. It's not just misalignment, it's also other issues like having a bad understanding of the relative strengths”

0.73

People overrate how smart individual AI copies will be and underrate AIs' collective advantages that biology denies humans: AIs can be copied with all their tacit knowledge (copy an Ilya or Elon to staff every role), avoiding the dilution, turnover, and aging that erode effective human teams; firms currently have selection and variation but lack high-fidelity replication, so adding that final evolutionary ingredient enables far faster, more intense firm evolution.

factualhigh valuecontestednovelty 3/4durability 3/4· Dwarkesh Patel

“Firms right now have two of the three relevant criteria for evolution; they have selection, and they have variation, but they don't have high fidelity replication.”

0.73

There is no fundamental reason economic growth cannot be far faster than today; it is currently slow largely because humans are the key bottleneck — they supply labor, play crucial roles in discovery, and complement capital in ways machines cannot substitute — so growth and productivity end up bottlenecked by the growth of the human population, a constraint AI removes by scaling the (AI) labor force.

causalhigh valuecontestednovelty 3/4durability 3/4· Ege Erdil

“the growth of the economy and growth productivity just ends up being bottlenecked by the growth of human population”

0.73

Capabilities like reasoning look cheap and easy in retrospect (only ~1% of compute on RL) only because we stand on a huge stack of technology built over years; in 2015 building a reasoning model would have looked insurmountable, requiring internet-scale data, tens of thousands of GPUs, and distillation innovations — so agency may similarly look simple in hindsight while taking years of complementary innovation to actually achieve.

causalhigh valuecontestednovelty 3/4durability 3/4· Tamay Besiroglu

“it looks easy from our point of view, standing on this huge stack of technology that we've built up over the past five years or so, but at the time, it would have been very hard”

0.73

Automating AI research is both harder than people credit — current models excel at compact, closed benchmark tasks (like METR's) but lack the many other competences needed to be a researcher in messy real contexts — and even if research were automated, much software progress has historically been driven by compute scaling enabling more and larger experiments rather than by cognitive effort alone.

causalhigh valuecontestednovelty 3/4durability 3/4· Ege Erdil

“automating research is, first of all, more difficult than people give it credit for. I think you need more skills to do it and definitely more than models are displaying right now.”

0.73

Moravec's paradox — that tasks hard for humans (chess, advanced math, 100-digit multiplication) are easy for AI while tasks easy for humans are hard for AI — is explained by evolutionary recency: skills that emerged recently (language ~100,000 years ago, chess) had little time and weak fitness pressure to be optimized, so AI makes fast progress on exactly the things that impress us, which means we should not over-update on those impressive narrow capabilities about general competence.

causalhigh valuecontestednovelty 3/4durability 3/4· Tamay Besiroglu

“the tasks that humans seem to struggle on and AI systems seem to make much faster progress on are things that have emerged fairly recently in evolutionary time”

0.73

Because compute and cognitive effort are complementary, either can be the bottleneck — so having the most compute doesn't guarantee winning (a dysfunctional culture can waste it), but the key question for a software-only singularity is how many orders of magnitude software efficiency can improve before a fixed compute stock becomes the binding constraint, an experiment that has essentially never been run so intuitions about it are weak.

causalhigh valuecontestednovelty 3/4durability 3/4· Ege Erdil

“just as A can bottleneck you, B can also bottleneck you”

0.73

There is no clean temporal separation where invention precedes deployment; technological progress is an interplay between having ideas and building, learning, and producing (learning curves) — nobody 20 years ago had the blueprint for 2025 solar panels, and aircraft depended on prior smelting improvements, so capital accumulation and complementary inputs are intrinsic to discovery itself, not just downstream of it.

causalhigh valuecontestednovelty 3/4durability 3/4· Tamay Besiroglu

“There isn't this temporal difference where it's first you do the invention... often there's this interplay between the actual capital buildup and the innovation.”

0.73

Much of the fear about AI is really a fear of change and especially fast change; biological humans recognizably making all important decisions a million years from now would be surprising even without AI, so the distinctive feature of AI is that it compresses inevitable change into a narrower period rather than introducing a uniquely alien vector of change.

normativehigh valuecontestednovelty 3/4durability 3/4· Ege Erdil

“a bunch of the things that people fear about AI are just things they fear about change and fast change. So the thing that's different is that AI has a prospect of accelerating much of this change”

0.73

Even if Baumol-type bottlenecks are maximally powerful, explosive growth still follows because workers displaced from automated tasks reallocate to non-automated ones (e.g. software engineers becoming physical workers at much higher wages), and a quantitative objection must specify the bottlenecked sector's output share and the elasticity of substitution — qualitative Baumol arguments alone don't permit a growth-rate prediction.

causalhigh valuecontestednovelty 3/4durability 3/4· Ege Erdil

“even if bottlenecks are maximally powerful, even if you just look at all the tasks in the economy and literally take the worst one for productivity growth, you would still get a lot of increase in output because of this reallocation”

0.73

Economists who object to explosive growth by asking how much more valuable hairdressing or restaurants can become are making a levels argument against a rate-of-change claim; this is the wrong kind of objection and amounts to special pleading, since economists would not accept the same levels argument against a slower growth rate continuing over a longer period.

normativehigh valuecontestednovelty 3/4durability 3/4· Ege Erdil

“We're talking about the rate of change, and you're objecting to it by making an argument about the absolute level of productivity.”

0.73

Because AIs can be aligned and fine-tuned to have the preferences you want — unlike human workers whom you can only select — AI firms could dissolve foundational frictions of human firms such as the principal-agent problem (misaligned incentives between worker, manager, and shareholders), and combined with higher bandwidth (copies of a 'mega-founder' reviewing every pull request and press release) could maintain coherent vision at scale.

causalhigh valuecontestednovelty 3/4durability 3/4· Ege Erdil

“AIs can be aligned. You get to control the preferences of your AI systems in a way that you don't really get to control the preference of your workers.”

0.72

Even though AIs will eventually drive most of the economy and could coordinate to take over if they wanted, they likely won't — analogous to how the US could invade Guatemala but doesn't because the value is low — because AIs grow up within human economic norms where undermining coordination has costs, and if they already capture nearly all income, the marginal benefit of expropriating humans is small.

forecasthigh valuecontestednovelty 3/4durability 2/4· Ege Erdil

“if you already are getting almost all of the income in the economy, will be fairly small”

0.71

A core underlying disagreement is the power of reason: the speakers hold that abstract, deductive, and even Bayesian reasoning is far less powerful than others think because the real world has enormous richness and detail you cannot reason about but must observe — which explains both why detailed AI policy plans are futile and why AI's economic impact requires real-world experiments and deployment rather than pure cognition.

normativehigh valuecontestednovelty 3/4durability 4/4· Tamay Besiroglu

“this underlying latent thing is, what is the power of reason? How much can we reason about what might happen?”

0.69

Over the past decade AI gained major capability 'unlocks' (game play, language, abstract reasoning/coding/math) at a rate of roughly one every three years or one per three orders of magnitude of compute scaling; since AGI needs several more such unlocks (long-horizon coherence, agency, full multimodality) and only ~3-4 orders of magnitude of compute scaling remain before data centers consume a non-trivial fraction of world output, full automation should take decades, not years.

causalhigh valuecontestednovelty 3/4durability 2/4· Tamay Besiroglu

“that happened on the order of once every three years or so, or maybe one every three orders of magnitude of compute scaling”

0.69

There won't be a separate 'AI economy' running on an island; it is far more efficient for AI to operate within the existing human economy, benefiting from established supply chains and markets, rather than rebuilding everything from scratch — the notion of a separate AI economy is downstream of the (rejected) software-only singularity view.

forecasthigh valuecontestednovelty 3/4durability 2/4· Tamay Besiroglu

“it's just much more efficient for AI to operate in our economy and benefit from the existing supply chains and existing markets rather than set up shop on some island somewhere and do its own thing”

0.69

Delaying broad automation by one year is very costly: using concave utility functions, a year of delay is worth on the order of tens of trillions of dollars in deferred consumption, and in mortality terms each year of delay could mean roughly 100-200 million people alive today dying — so unless you hold a strong longtermist view with high confidence in your leverage over the far future, accelerating is justified.

normativehigh valuecontestednovelty 3/4durability 2/4· Tamay Besiroglu

“that's worth on the order of tens of trillions of dollars per year in consumption. That is roughly the amount consumers might be willing to defer in order to bring forward the date of automation one year.”

0.69

Pausing AI to do alignment research for 20 years at fixed compute would yield little, because compute scaling is essential to making alignment progress — trying to do alignment in 2016 with 2016 compute budgets would have discovered none of the useful things known today, so a pause leaves you in a similar position years later.

causalhigh valuecontestednovelty 3/4durability 2/4· Ege Erdil

“Imagine you were trying to make progress on alignment in 2016 with the compute budgets of 2016. Well, you would have gotten nowhere, basically.”

0.69

A quantitative growth prediction: the human brain does roughly 1E15 flop/second, an H100 does computation on that order, an H100 costs ~$30K, and if it ran human-brain software earning human wages (~$50-100K/year), it would pay itself back on the order of a year — implying an economy doubling time of about a year, a specificity that qualitative Baumol objections fail to match.

forecasthigh valuecontestednovelty 3/4durability 2/4· Tamay Besiroglu

“it costs on the order of 30 K per H100. And so you get a doubling time of maybe on the order of a year.”

0.68

AIs have a structural advantage over humans because every human must learn from scratch over a lifetime — creating massive duplication — whereas an AI can learn once in a single large training run and then be deployed everywhere, and doubling compute yields both more copies and smarter copies, an additional source of economies of scale.

factualhigh valuecontestednovelty 2/4durability 3/4· Ege Erdil

“for humans, every human has to learn things from scratch... So in human learning, there's a ton of duplication, while for an AI system, it could just learn once.”

0.68

R&D is genuinely important and necessary but is less important in relative terms than other inputs (capital, labor), none of which alone is sufficient for growth; just as you can't reach a Dyson sphere by scaling only labor and capital, you also can't reach it by scaling only TFP — you must scale everything at once, and credit attribution should follow factor elasticities (labor ~0.6, capital ~0.3).

causalhigh valuecontestednovelty 2/4durability 3/4· Ege Erdil

“It is less important in relative terms than some other things, none of which are by themselves sufficient to enable this growth.”

0.68

People underestimate how responsive important people are: if what you send is interesting and high quality, important people are very likely to respond, so being more aggressive about reaching out and less ashamed of looking dumb is high-value career advice, especially because productivity depends on being plugged into a community given the limits of solitary reasoning.

normativehigh valuecontestednovelty 2/4durability 3/4· Ege Erdil

“if what you send to them is interesting and high quality, then it's very, very likely that they will respond”

0.68

Tyler Cowen's point that most of Sub-Saharan Africa lacks reliable clean water despite the required intelligence not being scarce supports the view that intelligence is not the bottleneck on technological progress or growth — many other factors (capital, human capital, scaling the whole economy) bind — which is an objection to the 'geniuses in a data center' model that the speakers accept.

causalhigh valuecontestednovelty 2/4durability 3/4· Tamay Besiroglu

“intelligence isn't the bottleneck that's holding back technological progress or economic growth. It's like many other things.”

0.68

Bullish forecasters overestimate automation speed because they think jobs are simpler than they are; what looks like a job (e.g. 'a travel agent books flights') is actually a small fraction of what the person does, so automating that one task neither automates the job nor has much economic impact — full automation requires many more competences than people credit.

causalhigh valuecontestednovelty 2/4durability 3/4· Ege Erdil

“if you look at the fraction of their time on the job that they spend on doing that, it's a very small fraction of what they actually do”

0.68

Because we should be more epistemically uncertain about AI's trajectory than about almost anything in history, the right societal response is freedom, decentralization of both knowledge and decision-making, and avoiding high-volatility centralized moves (like nationalizing AI to ensure an aligned software-singularity) that are brittle and contingent on a single worldview.

normativehigh valuecontestednovelty 2/4durability 3/4· Dwarkesh Patel

“how we as a society deal with topics on which we are this uncertain is just to have freedom, decentralization, both decentralized knowledge and decentralized decision making take the reins and not to do super high volatility centralized moves”

0.68

One should discount the long-term future not because it is morally worthless but because it is extremely hard to anticipate the consequences of present actions on it; the right strategy is to focus on near-term beneficial actions (aligning current AI to value happiness and dislike suffering, building institutional capacity to intervene later) while maintaining flexibility rather than committing to detailed long-horizon plans.

normativehigh valuecontestednovelty 2/4durability 3/4· Ege Erdil

“you should just discount the future. Not for a moral reason, not because the future is worthless or something, but because it's just very hard to anticipate the effects of your actions”

0.66

The world is not bottlenecked by a lack of good reasoning, so simply having abundant superhuman reasoners will not by itself substantially accelerate technological change and growth; acceleration requires complementary innovations across industries, supply-chain upgrades, demand, and a growing economy as a whole.

causalhigh valuecontestednovelty 3/4durability 3/4· Tamay Besiroglu

“It isn't the case that the world today is totally bottlenecked by not having enough good reasoning, that's not really what's bottlenecking the world's ability to grow much more substantially.”

0.66

Central planning could become relatively more viable with AI because information gathering and processing can be disaggregated (cheap sensors feed centralized processing that benefits from economies of scale and bigger 'planner' models), and the incentive rationale for markets weakens with aligned AI — but it likely still won't be optimal, because as the economy grows more complex the need for information processing rises alongside the improved ability to do it (Apple could plan ancient Uruk's economy but not today's world economy).

forecasthigh valuecontestednovelty 3/4durability 3/4· Tamay Besiroglu

“communication bandwidth being potentially much, much greater than it is today. In the current world, the information gathering and the information processing are co-located”

0.66

The idea that one actor will use a software-singularity AI to 'lock in' a stable utility function for millions of years looks unlike anything in history; even digital information suffers rapid link rot and digitization has been associated with faster, not slower, cultural change — so the argument that we can preserve utility functions forever is weak.

factualhigh valuecontestednovelty 3/4durability 3/4· Ege Erdil

“that just looks very unlike anything that has happened in the past. So I'm kind of confused why people think it's very plausible.”

0.66

The end of slavery was driven primarily by deep economic forces of the Industrial Revolution that made coerced labor inefficient (e.g. Russia abolished serfdom in the 1860s without British pressure; harsh colonial labor became less prevalent), rather than by abolitionist persuasion — so future-shaping values tend to be determined by which values are functional in the technological-economic environment, not by contingent actions of individuals.

causalhigh valuecontestednovelty 3/4durability 3/4· Ege Erdil

“the change in values associated with the Industrial Revolution made it so that slavery just became an inefficient thing to sustain in a bunch of ways”

0.65

Large reasoning models solve impressive problems chiefly by leveraging their vast knowledge (they know more than any human), approaching problems very differently from creativity-driven humans; despite enormous knowledge they show no examples of producing even a small genuinely interesting novel mathematical concept, revealing they recombine rather than innovate.

factualhigh valuecontestednovelty 3/4durability 2/4· Ege Erdil

“has a reasoning model ever come up with a math concept that even seems slightly interesting to a human mathematician? And I've never seen that.”

0.64

In AI systems the correlation between narrow impressive skills and general competence is weaker than in humans: the model best at competitive programming (o3-mini high) is not the best at actually helping you write code or capturing enterprise coding revenue (which goes to Claude), showing benchmark prowess and real-world value diverge.

factualhigh valuecontestednovelty 3/4durability 1/4· Tamay Besiroglu

“the strongest systems on competitive programming are not even the ones that are best at actually helping you code”

0.64

Claude Plays Pokemon illustrates a key gap: the model has vast explicit knowledge (it can tell you what to do if stuck in Mount Moon) yet still gets stuck there for 48 hours while playing, showing that having knowledge and being able to leverage it through long-horizon action are very different and the latter is not easy.

factualhigh valuecontestednovelty 3/4durability 1/4· Ege Erdil

“it has explicit knowledge, but then when it's actually playing the game, it doesn't behave in a way which reflects that it has that knowledge”

0.64

A drop-in remote worker capable of doing literally everything a human can do remotely is roughly 20 years away (around 2045), in contrast to the 2027-2030 timelines common in San Francisco AI circles.

forecasthigh valuecontestednovelty 3/4durability 1/4· Ege Erdil

“Maybe for me, that would be around 2045.”

0.63

Regulation is the strongest argument against explosive growth: the world has a surprising ability to coordinate against pursuing certain technologies (e.g. human cloning), so there's perhaps a 10-20% chance of effective global regulatory coordination — enforced via sanctions on defectors — that slows things enough to prevent explosive growth, though AI is less likely than cloning to be blocked because it's more valuable and nationally strategic.

forecasthigh valuecontestednovelty 2/4durability 2/4· Ege Erdil

“the world does have a surprising ability to coordinate on just not pursuing certain technologies”

0.63

Gwern's argument that the AI period will be more geopolitically unstable than the Industrial Revolution because a century of progress is compressed into a decade is weakened by the integrated view of technology: progress isn't an orthogonal vector where one country gets ballistic missiles and another railroads, but a churning further along an interconnected tech tree, so leads may be less decisive than feared (a one-year lead may not suffice to risk war against a nuclear-armed rival).

forecasthigh valuecontestednovelty 2/4durability 2/4· Dwarkesh Patel

“if you have this more integrated perspective about what it takes to get to ballistic missiles and to railroads, then you might think 'no, basically this isn't some orthogonal vector. You're just churning on the tech tree further and further'”

0.61

China's growth differs from the AI scenario in a key way: China scaled capital, technology adoption, and human capital but did not scale up its labor force, whereas AI scales the (AI) labor force simultaneously with capital — and getting both to scale together yields much faster growth than China's example; still, China in the 1990s-2000s is a useful intuition pump for what 30% growth feels like, albeit slower than forecast.

factualhigh valuecontestednovelty 2/4durability 3/4· Ege Erdil

“in China, you see a massive amount of capital accumulation... But you're not seeing a huge scale up in the labor force. While for AI, you should expect to see a scale up in the labor force as well”

0.52

Given limited time (unlike an AI model), one should aggressively prioritize reading the key pieces of empirical literature that people who think well about a topic recommend — and deeply engage them (e.g. spend a day on a transformer-circuits paper rather than skimming) — rather than generally reading widely, while also valuing high-information-density sources like Twitter over arXiv for some purposes.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Tamay Besiroglu

“you have to aggressively prioritize what you're going to spend your time reading”

0.45

OpenAI reaching $100 billion in revenue (its contractual AGI marker) would not be a big update because people pay enormous sums for ordinary things like oil; only revenue on the order of $500 billion would constitute strong evidence of transformative capability.

normativehigh valuespeaker onlynovelty 2/4durability 1/4· Tamay Besiroglu

“the update would come if it was, in fact, $500 billion in revenue or something like that. But then I would certainly update quite a lot. But a hundred billion, that seems pretty kind of likely to me.”

0.40

ChatGPT exemplifies cheap 'unhobbling': roughly 1% of additional compute spent on post-training a pretrained internet corpus was enough to make a model highly competent at chatbot conversation, and reasoning similarly emerged from a small fraction of compute spent on RL.

factualcontestednovelty 2/4durability 2/4· Dwarkesh Patel

“ChatGPT is an example of this unhobbling, where 1% of additional compute spent on getting it to talk in a chatbot-like fashion with post training is enough to make it competent”

0.32

The original GPT-4 became roughly 100 times cheaper to run for the same capability by the time of GPT-4o, illustrating the algorithmic efficiency gains that a software-driven feedback loop could compound by increasing the population of AI workers and the number of parallel experiments runnable per unit compute.

factualestablishednovelty 2/4durability 1/4· Tamay Besiroglu

“It's like, maybe a hundred times for the same capability or something.”

0.31

There is a tradeoff between how accessible considerations are to individual humans and how technologically advanced the world is; in a very sophisticated world many things will be beyond individual human comprehension, just as much is already not understood today, and that's an acceptable trade — better to live in a world with advanced technology and products you can enjoy even if you don't understand how they work.

normativespeaker onlynovelty 2/4durability 2/4· Tamay Besiroglu

“you have this trade-off between accessibility and maybe how advanced the world is”

0.26

The key measure of national capacity for the AI transition is the scale of total economic output in dollars, not the raw number of people or firms; by that measure the US (and its allies, who control relevant AI inputs China lacks) leads China across many dimensions, though China leads in some, so one shouldn't over-update toward China being a much bigger deal than already appreciated.

factualcontestednovelty 1/4durability 1/4· Tamay Besiroglu

“the scale of the overall economy, which is just measured in how much is being produced in terms of dollars. There, the U.S. is ahead.”

0.23

Epoch was founded because academia could not publish interesting work on grounded AI forecasting and industry was not focused on producing useful insights, leaving a gap that a dedicated org of curious 'misfits' could fill — timed just before ChatGPT to enable more grounded discussion of AI's future.

factualspeaker onlynovelty 1/4durability 1/4· Tamay Besiroglu

“academia and industry is kind of dropping the ball on a bunch of important questions that academia is unable to publish interesting papers on. Industry is not really focused on producing useful insights.”

0.14

Introductory framing of the episode's guests and topic.

factualestablishednovelty 0/4durability 0/4· Dwarkesh Patel

“Today, I'm chatting with Tamay Besiroglu and Ege Erdil.”