Anthropic's R&D acceleration confirmed: the driver is not compute, but algorithm and data quality
Anthropic admits in IPO filing that R&D may have undergone ‘dramatic acceleration’; METR's independent assessment shows: this acceleration comes not from more compute, but from algorithm and data quality itself being driven forward by existing intelligence.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
IPO introduces market discipline when internal risk awareness escalates
Anthropic's IPO filing adds new language: following the current model roadmap, the company ‘may very soon’ hit the threshold in its own Responsible Scaling Policy for ‘AI-driven dramatic acceleration,’ and admits it cannot yet trust Claude to independently design the next-frontier model unsupervised. Prakash sees the company internally aware that risk is spiraling, and that IPO brings external oversight and market discipline to bear—not just capital. Nathan adds two points: first, official lab language ‘is literally what it says, just gift-wrapped,’ so ‘possibly soon’ essentially means ‘now’; second, OpenAI just proved AI can systematically find approximately 1000 vulnerabilities in the Linux kernel, showing the tech stack underlying AI training is stuffed with ‘low-hanging fruit’ explaining short-term acceleration.
— Prakash NarayananAccurate predictions without execution speed lose money
Thoughtful gave nine frontier models one week and $500 each to independently forecast roughly 800 U.S. midterm election races. Overall, the models were well-calibrated—candidates predicted to win at 80% actually won close to 80%, matching top human forecasters and even outperforming some market estimates. But the actual trading agents making real bets lost money: delays between polling updates and vote counts meant agents reacted too slowly; sometimes they hallucinated false information outright. Predicting well does not translate to trading wins.
— Brian ZhouLocal ballot measures concealed their true target: presidential politics
Forty-one local data center referendums appeared on midterm ballots, and 40 of 41 voted to restrict or oppose. Prakash sees the real fuel: Trump's approval rating, the ‘cheated’ feeling from LLC shells buying land cheap and flipping to Meta, and widespread anxiety that ‘the wealthy get rich on AI while everyone else loses jobs.’ Voters were not judging data center merits; they were using data centers as a handy club to hit the president. His forecast: any future builder wanting to open a data center will face land, labor, and construction costs at Google and Meta scale.
— Prakash NarayananThe real bottleneck is liquid cooling infrastructure, not silicon
Luxor's hardware trading desk shows 8-GPU B300 nodes at spot prices of $550,000 to $650,000, with individual trades routinely moving thousands of units worth $50 million. The market has inventory; the constraint is the data center itself. Most racks today fit only two B300 nodes. Liquid cooling capacity is sold out through Q1 2027, with many sites quoting delivery in Q3 2027 or even 2028. By contrast, long-term GPU rental runs $4.2 to $4.5 per GPU per hour, far below the $8 to $9 spot index, making node purchase and deployment more economical.
— Mike San MiguelRegulation by administrative pressure instead of by law
Mike sketches a specific compliance chain: most OEMs skip KYC, Luxor internally rejects no-KYC projects, buyers more than 50% Chinese-owned are blocked, and shell structures in the Caymans and BVI must be penetrated to find the end buyer. But Prakash points to a deeper problem: reports say Oracle canceled a $7 billion compute sale to Tencent under national security pressure—not pursuant to any written law. His critique: if Congress has the authority, legislate clearly instead of relying on companies to ‘read between the lines’ and self-censor.
— Prakash NarayananReal investment alpha is taste in what to see and what to cut
Theory Ventures does not assign each investor one or two agents; it runs thousands of agent tasks in parallel, with only 5 to 10 percent needing results immediately and most able to wait until the next meeting. The firm explicitly says the goal is not to replace investors but to boost efficiency 10 to 100 times over. Bischof rejects ‘company brain’ products selling internal knowledge as ‘smart search’; he wants research agents that automatically wire scattered information nodes across the business. Real scarcity is the judgment to know what to look at and what to ignore—still a domain where humans exceed machines.
— Bryan BischofEnterprise software will optimize for agent users, not humans
Prakash sketches a future where enterprise software becomes agent-first: Salesforce for agents only, Ramp for agents only, with massive parallel agent swarms coordinating capital to build these companies. Humans, consuming far less than agents, will free-ride on the overflow. Nathan expresses clear unease: he wants sensory satisfaction—fresh fruit—not API abundance; he does not want to become a ‘human integration layer too obedient to machines.’ Both note that frontier labs discuss ‘curing cancer’ seriously but no one dares say ‘better Doritos’; the lack of conversation about which mountains to climb is a quiet risk.
— Prakash NarayananMETR confirms: acceleration comes from algorithm and data, not hardware budgets
METR's independent assessment of Anthropic R&D acceleration is hedged throughout—‘we don't know,’ ‘perhaps 1.3 to 3x’—except for one sentence: ‘we are confident this acceleration is not due to faster training compute growth.’ Nathan reads this as: current acceleration roots not in more chips or data center buildout, but in algorithm and data quality itself being propelled forward by existing intelligence. If governments want to intervene in this trajectory, they should target frontier labs' internal R&D practices, not blunt restrictions on data center construction.
— Nathan LabenzIn their own words · checked verbatim
I think within the company, there is a sense that you need more controls. More people have to have exposure and more control over this
Prakash Narayanan3:46
It's almost certainly the case by analogy to Linux and by analogy to math that there is a lot of low hanging fruit, and a lot of low hanging fruit means acceleration.
Nathan Labenz4:31
we're seeing a lot of these instances of, for example, models making up information, parsing information wrong, and these sorts of, like, other broader issues, which partly explains why traders, the actual trading agents didn't do too great.
Brian Zhou23:05
So I don't think actually data centers were a big issue, but I think they were an issue to tar the president with.
Prakash Narayanan29:12
Like, if you if you wanna make laws, just make laws. Right? But if you don't have the political willpower to make those laws and you try to slide this under the gate, like, why? Why? Like, you have a mandate now. Right? You have a mandate in congress. Like, do something.
Prakash Narayanan44:10
Our goal is not to replace an investor with an AI system. Our goal is to make that investor 10 to a 100 times more effective and efficient by using AI tooling and building the AI tooling that will make that possible.
Bryan Bischof50:13
You're gonna get 50,000,000 Patrick Hollis. You're gonna get 50,000,000 Elon Musks. You're gonna get, like, humans who give them capital to deploy. You're gonna get agents coordinating investments into businesses built by agents for agents.
Prakash Narayanan1:11:03
we are confident that the acceleration was not due to an increase in the growth rate of training compute.
Nathan Labenz1:20:57
Figures
| METR's estimate of AI R&D acceleration | approximately 2x (range 1.3-3x) | 1:20:57 |
| Vulnerabilities OpenAI found in Linux kernel | approximately 1000 | 5:35 |
| AI agents in BallotBench | 9 | 15:26 |
| Data budget per BallotBench agent | $500 | 15:26 |
| Local referendums opposing data centers | 40 of 41 voted to restrict or oppose | 28:57 |
| Long-term GPU rental vs spot price | $4.2-4.5/GPU/hour on contract, $8-9/GPU/hour on spot index | 34:46 |
| Oracle's reported canceled Tencent compute sale | $7 billion | 44:10 |
Glossary
- Responsible Scaling Policy (RSP)
- Anthropic's internal risk-tier and safety-threshold framework
- BallotBench
- Benchmark test where AI agents independently forecast results of ~800 U.S. midterm election races
- Move 37
- AlphaGo's counterintuitive move in its match against Lee Sedol; metaphor for moments when AI displays superhuman judgment
- HGX node
- NVIDIA's standard 8-GPU server mainboard configuration, the basic unit of current AI compute trading
- PostTrainBench
- Benchmark measuring whether AI agents can independently perform post-training on foundation models
How to listen
Engineers and operators tracking AI acceleration pace, GPU spot pricing, and export compliance; also investors wanting to understand how AI-native venture firms conduct due diligence.
If you want just the hard information, skip the closing riff comparing AI ambitions using fruit metaphors and whether AI should make 'better Doritos'.