Nov 24, 2025 · 58m · a16z
The 2045 Superintelligence Timeline: Epoch AI’s Data-Driven Forecast
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The a16z Show, host Erik Torenberg and co-host Marco Mascorro interview Epoch AI researchers David Owen and Yafah Edelman to evaluate empirical trends, economic impacts, hardware bottlenecks, and realistic timelines surrounding artificial general intelligence and superintelligence.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 3.5 Guest teaching 3.7 Guest disagreement 2.5 The host pushing back 2.3
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 52:09 Rejection of cluster scaling bottleneck premise
The guest directly dismisses the host's premise about cluster scaling constraints, stating that people are 'approximately wrong that there's something stopping us' and characterizing energy complaints as minor compared to GPU capital expenditure.
Hardest push from the host ▶ 38:06 Host insists on strict criteria for math benchmarkWhen the guest attempts to cross-examine and dilute the question's premise ('does it have to solve it on its own?'), the host firmly holds ground and enforces clear, unassisted criteria.
Biggest teaching moment ▶ 47:39 Empirical reveal on robotics compute disparityThe guest educates the host using Epoch AI's dataset, revealing that robotics training compute is 100x smaller than frontier LLM models and reframing robotics as an economic hardware issue rather than a software AI breakthrough issue.
The host holds their own ▶ 12:36 Host confronts guest with specific industry quotesThe host demonstrates deep preparation by citing Dario Amodei's exact predictions regarding 90% code generation and data centers filled with a country of geniuses to challenge Epoch AI's forecasting models.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Evaluating the Macro AI Spending and Bubble Debate | 1 | 3 | 1 | 0 | The host opens with a broad question on whether AI spending is a bubble. The guest provides a detailed breakdown of hardware sales and inference revenue, explaining why current unit economics do not indicate an immediate bubble. | |
| Pre-Training Scaling Trends and Post-Training Synergies | 3 | 4 | 2 | 2 | The host presses on why Epoch AI does not forecast a self-improving software singularity. The guest gently reframes the question, highlighting that experimental compute bottlenecks limit purely algorithmic R&D takeoff. | |
| Algorithmic Bottlenecks, Human Learning, and Capability Progression | 2 | 5 | 3 | 1 | When the co-host compares machine learning backpropagation to how human children learn, the guest pushes back against comparing AI to human cognition, arguing that we understand AI optimization much better than child development. | |
| Evaluating Bullish AI Timelines and Code Generation Metrics | 4 | 4 | 2 | 3 | The host directly confronts the guest with Dario Amodei's aggressive coding automation predictions. The guest analyzes the premise, distinguishing between simple tab-completion lines of code and full job task automation. | |
| Labor Market Disruption and Job Automation Projections | 4 | 3 | 1 | 2 | The host cites industry terminology like 'middle-to-middle' automation to query labor market projections. The guests agree and expand with a task-based labor economics framework. | |
| Higher Education and Career Guidance in an AI Era | 2 | 2 | 1 | 1 | A lighthearted discussion on higher education choices in an AI world where the hosts and guests humorously dismiss the concept of prompt engineering as a long-term career degree. | |
| Computer Use Agents: Benchmarks, Challenges, and Applications | 4 | 3 | 2 | 2 | The co-host demonstrates domain familiarity by referencing OSWorld and WebArena benchmarks. The guest explains how vision error loops and context window growth hinder computer-use agents. | |
| Macroeconomic Impact, GDP Growth Projections, and AGI Models | 4 | 4 | 4 | 2 | The host cites Tyler Cowen's GDP estimates to test macroeconomic models. The guests debate each other on extreme takeoff scenarios versus moderate GDP growth projections. | |
| Evolving Evaluation Benchmarks for Frontier AI Models | 4 | 3 | 1 | 1 | The co-host brings up SWE-bench and MMLU saturation. The guest details why future evaluation benchmarks will require significantly larger financial and computational budgets. | |
| Predicting AI Timelines for Unsolved Mathematical Discoveries | 4 | 4 | 5 | 5 | When the host asks for a timeline on AI solving a major math problem, the guest cross-examines the host on exact parameters. The host holds ground and strictly defines unassisted, major mathematical proofs. | |
| Scientific Discovery Timelines: Biology and Physical World Constraints | 3 | 4 | 2 | 2 | The host shifts focus to biology and medicine timelines. The guest explains why physical world experimentation introduces friction absent in formal mathematics. | |
| Forecasting the Timeline to Artificial Superintelligence | 3 | 3 | 2 | 2 | The host pushes for a timeline on artificial superintelligence. The guest outlines their 2045 modal timeline while explaining why long-range forecasting models breakdown near AGI. | |
| Hardware Economics, Data Constraints, and Physical Robotics | 4 | 5 | 3 | 2 | The co-host raises physical robotics and world models. The guest educates the host with empirical findings showing robotics training runs are 100x smaller than LLMs, reframing robotics as a hardware unit economics issue. | |
| Infrastructure Expansion: Data Center Projects and Energy Constraints | 5 | 5 | 6 | 6 | The host repeatedly questions data center bottlenecks and energy constraints. The guest explicitly rejects the host's premise, claiming people are 'approximately wrong' and that energy costs pale in comparison to GPU investments. | |
| Political Reactions, Governmental Intervention, and Policy Trends | 5 | 3 | 2 | 3 | The host cites Leopold Aschenbrenner's nationalization thesis to ask how governments will respond to fast AI progress. The guests draw parallels to rapid emergency legislation during COVID-19. |
Statements from this episode (25)
Epoch AI: 20-30% chance AI causes sudden 5% unemployment spike
“The, like, interesting scenario to think about, you know, 20% chance, 30% chance something like this will happen in the next decade is like, you know, a five percent increase in unemployment over, over a very short period of time, like six months, due to AI.”
Epoch AI: Most AI compute is spent on product inference, not training
“It does seem as if most compute gets spent on inference that companies don't so far regret like using to offer their products.”
Epoch AI: AI labs would be highly profitable if they paused scaling
“Right now the amount of money companies are actually earning in profit, not including the cost to develop the models initially, is, seems to be, like, very positive, such that if they stop developing bigger and bigger models, and just stick with the ones they'…”
David Owen: AI pre-training receives less focus due to post-training progress
“It seems as if pre-training is comparatively less of a focus than it was before, partly because, like, you have this exciting new direction of, well, new, newish direction of post-training where they've done so much about reasoning”
David Owen: Post-training usage data generates feedback loops for pre-training
“A lot of this stuff is quite synergistic. You develop a better model. You, like, use post-training stuff to make it better. You get a load of data of the model actually being used successfully or not. A lot of that can probably go into pre-training next time.”
David Owen: Experimental compute receives far more funding than final training runs
“As far as we can tell, experimental compute, which you seem to need to do research, is also, is receiving a similar amount of money, and that in fact it's receiving many times more money than the final training runs that are actually, of the models that are ac…”
David Owen: Software-only AI singularity is impossible without scaling physical compute
“I tend to lean towards, no, you actually need to do more experiments, and that means you can't get this software-only singularity”
Epoch AI: Humanity understands AI learning mechanisms better than human learning
“I think we know a lot more about how AIs learn right now than we know about how humans learn”
Epoch AI: No slowdown in AI capabilities has been observed yet
“Yeah, I don't, I definitely don't think we've seen any slowdown yet in capabilities from any of these concerns people have.”
Epoch AI: 90% of code will eventually be written by AI
“I mean, far more than 90% of the code I write is written by AI these days but I know I'm not like the average coder at all, but it's definitely, I don't think it's like a wild prediction at this point that 90% of code is going to be written by AI.”
Epoch AI: Developer subscription revenue is the top indicator of AI utility
“I think at the end of the day, the most reliable indicator here is going to be how much money these people are making from programmers. And from, you know, subscriptions in general, and it's a lot of money. I think there's definitely indications that people ar…”
Epoch AI: AI will automate at least 10% of jobs within a decade
“I think I would be surprised if there were not five percent of jobs that exist now which AI has automated away over the course of the next decade. I'd honestly, I'd be surprised if it's not 10% of the jobs that exist now, I think.”
David Owen: Vision limitations cause computer-use agents to get stuck in UI loops
“I do think that there is a sense in which models are a little bit artificially hobbled by their vision capabilities. Like, it does seem as if a common pattern you see when you try to get models to do stuff with a GUI is they kind of get a bit confused about ma…”
David Owen: Epoch AI used ChatGPT agents to search local data center permits
“We used ChatGPT agent in our data center research, because a lot of what we have to do is find permits, which are all going to be on janky county by county databases of error permits for, you know, the county that Abilene, Texas is in and I don't know what dat…”
Epoch AI: Current trends project a few percent GDP boost by 2030
“My best guess on current trends is this fairly well-defined, you know, few percent of GDP in 2030 thing, which is already pretty crazy by economic standards.”
Epoch AI: Human-level AI means 30% GDP growth or human extinction
“Assuming in the next 10 years we get AI that is capable of doing any remote job as well as any human. I think, you know, 30% GDP growth seems like a lower bound on something that's reasonable. Assuming you get, this is a big assumption that a lot of people are…”
Yafah Edelman: AI may solve the Riemann hypothesis within five years
“I would not be surprised if AI solves, like, a major unsolved math problem, like the Rayman hypothesis, or similar in the next five years.”
Epoch AI: Biological AI breakthroughs lag math due to real-world constraints
“It definitely seems plausible, but there's a lot of other concerns there where it needs to It needs to be able to, like, actually do experiments, and get data, and interact with the real world for a lot of these in a way that does not need to happen at all for…”
Yafah Edelman: The modal timeline for artificial superintelligence is 2045
“I have, I think I am on the record as saying that the median timeline I discussed or the modal timeline, sorry I think it's modal, yeah which might be on the early side compared to where my median is is, you know, twenty-forty-five was where, when I did the po…”
David Owen: AI will automate all remote work within 25 years
“Like, my sort of, you know, guesses, my, like, judgmental forecasts, to use the fancy term, for just kind of can do any remote work tasks. Probably have a median of about, like, 20, 25 years. I kind of struggle to imagine a world where that happens, and people…”
Yafah Edelman: Robotics training runs use 100x less compute than frontier LLMs
“What we found is that like compute it, the training runs that are being used for robotics are like a hundred times smaller than the training runs that are being used for Than the training runs that are being used for, like, frontier models.”
Epoch AI: Anthropic and Amazon are building the first gigawatt data center
“For instance, we learned that the most likely candidate to have the first gigawatt scale data center is Anthropic, which would not have been my pick but Anthropic Amazon's New Carlisle Project Rainier development seems on track to come online in January follow…”
Epoch AI: Microsoft's Mount Pleasant data center will consume half NYC's power
“The one we found that's actually seriously underway and has permits and is, you know, setting up the electrical infrastructure for is one by Microsoft, which is going to be used by OpenAI, at least in part in Mount Pleasant. They're calling it Microsoft Fairwa…”
Epoch AI: Energy availability will not bottleneck AI scaling
“At the end of the day, though, it's just like, does not, there seem to be enough solutions, especially if you are as willing to pay as you, as people are in AI, that I don't really expect it to be a significant bottleneck.”
Epoch AI: Government attention to AI will double or triple annually
“I think that right now we've seen this thing in revenue and finances where it's been doubling or tripling every year. And my default assumption is that attention that AI gets from policymakers and governments is going to follow a similar trend where it will do…”