The 2045 Superintelligence Timeline:...

1. Key Themes

The AI Bubble Question: Follow the Inference Revenue, Not the Hype

The central indicator for whether AI is in a bubble isn't speculation or future promises—it's current inference spending and profitability. Companies are already profitable on existing models when excluding future R&D costs, which fundamentally differentiates this from typical bubble dynamics.

Speaker 2 substantiated this: "Right now, the amount of money companies are actually earning in profit, not including the cost to develop the models initially, seems to be like very positive, such that if they stop developing bigger and bigger models and just stick with the ones they've had, they'd have earned a profit pretty quickly at the current margins" [00:03:30]. Speaker 3 added: "People are spending a lot on these models...They're presumably doing this because they're getting value from them...That's a pretty solid sign" [00:05:18].

Data Center Infrastructure as the Real Bottleneck (Or Lack Thereof)

The narrative that energy or infrastructure will constrain AI progress is overblown. The research reveals that massive data centers are being built at unprecedented speed, with companies willing to pay premium costs for solutions that bypass traditional grid timelines.

Speaker 2 explained: "My actual take away from a lot of this has been oh we're not having too much trouble scaling up...these plans are going really quite fast and it's not obvious that people would actually have the finances and desire to do them faster" [00:53:34]. He added: "There are expensive technologies that exist right now you could pay for solar power plus batteries this is fairly small lead times it might cost twice as much as normal power but that's still way less than your GPUs" [00:54:06].

The Exponential Political Response Threshold

When AI causes rapid, visible labor market disruption (hypothetically 5% unemployment increase in six months), political response will be swift, decisive, and unprecedented—similar to COVID stimulus but with unpredictable direction.

Speaker 2 articulated: "The public's reaction to this will determine a lot. There will be very, very strong feelings about AI once this happens...I don't know what that will look like for AI. But I think it's like everything else in AI. It's like exponential, which means it will pass the point of people sort of care about it to people really care about it quite fast...wherever we end up, there will be this certain thing which we would have considered unimaginable a year ago" [00:55:42].

2. Contrarian Perspectives

Math is Surprisingly Easy for AI, Not the Pinnacle of Intelligence

Contrary to popular belief that mathematical reasoning represents the deepest form of intelligence, it appears to be one of the easier domains for AI to master—challenging our assumptions about what capabilities indicate true intelligence.

Speaker 2 stated: "Math team is unusually easy for AI. I'm good to be honest. People often think claims about it being like this, you know, intuitive, deep thing that it would mean that AI has achieved something, some huge level of intelligence for it to solve. I think in practice, this is just like, you know, making a piece of art. It turns out to be farther down the capability stream than people might have guessed" [00:50:52]. Speaker 3 compared: "We sort of had this with chess decades ago, right? Like, computers solved chess very well and everyone was thinking of this as the pinnacle of reasoning and everyone as there is what kind of concluded by, oh well, of course, computers can do chess" [00:41:35].

Researchers Need Experimental Compute More Than We Think

The belief in a "software-only singularity" where AI can recursively improve itself through pure research without massive experimental compute is likely wrong. The evidence shows companies spending similar or greater amounts on experimental compute than final training runs.

Speaker 2 explained: "One thing that needs to happen in order for a software only singularity to occur is you need to be in this world where scaling up the amount of researcher R&D time...allows you to improve AI enough that it makes up for the lack of being able to scale experimental compute...what we found is that compute...the training runs that are being used for robotics are like a hundred times smaller than the training runs that are being used for frontier models...I don't think the people who claim otherwise are like crazy...but I don't think we have good evidence that researchers and just researchers would be able to speed things up without doing more experiments" [00:08:32].

The 90% Code Written by AI Metric is Meaningless

While Dario Amodei predicted 90% of code would be written by AI, this metric conflates lines of code with actual programmer productivity. The distinction between "lines written" versus "cognitive work done" reveals why this benchmark doesn't capture what matters.

Speaker 3 noted: "If you were to measure how many lines of code are being written quote unquote by like tab completion then it's probably quite high. But you don't necessarily expect that that's taking on that much of the programmers really hard work" [00:16:00]. Speaker 2 added: "I know I'm not like the average coder at all. But it's definitely I don't think it's such a wild prediction at this point that 90% of code is going to be written by AI...It's by a lot of like people's intuitive sense in terms of like I was 90% of the job of a programmer being done by AI's definitely not" [00:15:11].

3. Companies Identified

Anthropic

AI research company building large language models, notably Claude.

Why mentioned: Most likely to operate the first gigawatt-scale data center, ahead of other major labs—a surprising lead in infrastructure buildout.

Quote: "We learned that the most likely candidate to have the first gigawatt skill data center is Anthropic which would not have been my pick um but Anthropic Amazon's new Carlisle a project reindeer development seems on track to come online in January" [00:51:08] - Speaker 2

Microsoft (with OpenAI)

Technology company building massive AI infrastructure for OpenAI's use.

Why mentioned: Has the largest concrete data center plans underway (not just marketing), with the Fairwater facility using more than half of New York City's power consumption.

Quote: "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 um in Mount Pleasant um they're calling it Microsoft Fairwater um and that one's going to be uh use a size use not quite as much power as New York City but I think more than half" [00:51:38] - Speaker 2

xAI

AI company founded by Elon Musk, operating the Colossus data center.

Why mentioned: Notable for aggressive timeline with Colossus 2 coming online shortly after Anthropic's facility; also exemplifies the pattern of starting data centers before grid connection.

Quote: "Followed shortly thereafter by Colossus 2...Abelene was an example xai classus one is a prominent example of just finding ways around this that are expensive" [00:51:20, 00:54:39] - Speaker 2

NVIDIA

GPU manufacturer whose sales serve as primary indicator of AI compute spending.

Why mentioned: Sales figures provide the most reliable metric for assessing whether AI investment represents a bubble.

Quote: "How much are they spending? You can look at Nvidia and how much they're selling each year and you can see whether it keeps on growing and you can see whether this fuss is kind of looking good to continue" [00:00:40] - Speaker 3

4. People Identified

Dario Amodei

CEO of Anthropic.

Why mentioned: His bullish predictions (90% of code written by AI within six months, country of geniuses in a data center by 2026-2027) represent a more aggressive timeline view, likely driven by belief in rapid R&D automation feedback loops.

Quote: "My model at least, which I don't know of, it's right, but what it is is that they think a bit more like the people who believe in you automate R&D and that gives you very quick takeoff. So they see it as like, yep, we're working on these AI's that are great for kind of research engineering type coding and at some point they're going to be useful and that's going to rapidly accelerate us to develop the next ones and then it's going to be quick progress" [00:13:09] - Speaker 3

5. Operating Insights

Use Computer Use Agents for Multi-Step Research Tasks

AI computer use agents have reached the threshold of practical utility for complex research workflows involving navigation of multiple non-standardized databases—a capability that emerged within the last year.

Speaker 2 shared: "This was the first year I found computer use actually useful. We use chat GPT 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...chat GPT does...it's able to navigate this such that I can just ask it to find me permits on a data center in a particular city and it will come back with air pollution permits and like tax abatement documents and all of this stuff that let me learn a huge amount" [00:28:02]

Monitor Experimental Compute Spending as Leading Indicator

Companies are spending "many times more money" on experimental compute than on final training runs. This reveals where true research constraints lie and predicts future capabilities better than announced models.

Speaker 2 explained: "What we found is that like compute...experimental compute...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 actually being released" [00:09:03]

Track Revenue Growth for Ground Truth on AI Value

The most reliable signal for AI's actual economic impact is revenue growth from inference, not training costs or capability benchmarks. Current profitable inference revenue validates the technology's real-world value.

Speaker 3 noted: "The big indicator being how much people are spending on stuff like compute...you can look at Nvidia and how much they're selling each year and you can see whether it keeps on growing" [00:02:39]

6. Overlooked Insights

Vision Capabilities as Hidden Bottleneck for Computer Use

While much attention focuses on reasoning and context length, vision capabilities may be artificially limiting computer use applications. Models struggle with GUI manipulation in ways that suggest the visual processing component is a constraint.

Speaker 3 observed: "I do think that there is a sense in which models are a little bit artificially hobbled by their vision capabilities...a common pattern you see when you try to get models to do stuff with a gooey if they kind of get a bit confused about manipulating it...you're not able to just easily look back on the thing and see kind of how I was wrong you instead go down like some awful bed end of just I'm just going to click this again and again and again" [00:26:44]

The Synthetic Data Feedback Loop is Already Working

Pre-training and post-training are synergistic in ways that create valuable data cycles: better models generate usage data that improves next-generation pre-training. This quiet feedback loop may be more important than dramatic breakthroughs.

Speaker 3 explained: "It seems as if plausibly like even a load of this stuff is quite synergistic. You develop a better model, you use post-training stuff to make it better. You get a load of data of the model actually being used successfully or not. A load of that can probably go into pre-training next time" [00:06:41]