The Turning of the Age
04/30/2026
Pieces that are a great way to understand how the Bay Area and the tech world view the rapid progress promised by the 2020s.
This Can't Go On
August 2021 · Holden Karnofsky
Holden argues that our current rate of economic growth is historically anomalous and physically unsustainable. His partner co-founded Anthropic that same year...
Situational Awareness: The Decade Ahead
June 2024 · Leopold Aschenbrenner
Among many topics, Aschenbrenner argues that an enormous AI capex boom is coming this decade. Initially a talk he presented internally at OpenAI, he later started a hedge fund to invest in this theme.
AI 2027
April 2025 · Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean
Continuing his 2021 post that eerily predicted yet underestimated the boom in AI chatbots, capex, and more, Kokotajlo and others argue that AI agents will become a massive economic force in 2026, an AI arms race between the US and China will begin, and the US's leading AI lab will reach ASI by the end of 2027, either losing control to a misaligned AI or keeping it aligned and solving mech-interp in the process.
A year later, they seem to have underestimated revenue: AI 2027 projected $35 billion for the leading lab in 2026, while analysts now predict $180b total combining Anthropic and OpenAI ARR. They also predicted 6 GW of compute for the US's leading lab by end of 2026, which appears to be right on track.
They also predicted an Opus 4.5-esque moment right on schedule, but claims from mid 2026 onward may not hold (China "coming online" and truly waking up to the capex race). The piece was heavily panned when released in 2025, but much like Situational Awareness I've found their predictions regarding capabilities and capex to be eerily on the mark so far.
The Curve Is Bending
April 2025 · Grant Slatton
Grant reveals he was spending $100/year on Github Copilot in 2023, but now he's spending $5k/year on ChatGPT Pro and Claude Code.
AI is finally getting to the "actually net-useful" point for real software development. This is a huge non-linearity in what people are willing to pay, and we should expect spend on inference to jump from today's low-hundreds of dollars to thousands or tens of thousands
Capex Update
February 2026 · Pleometric
AI generated video, likely made via datacenter inference, that feels like a vindication of the three posts above
660 billion dollars of Capex this year on AI data centers. To put a number like that in perspective, this is more than what we spent on the U.S. interstate highway system (630 billion), more than what we spent on the Apollo Moon Program (257 billion), more than what we spent on the international space station (150 billion). ... It’s the equivalent of spending $1.8 billion dollars a day, $75 million dollars an hour, $1.2 million dollars a minute. And we are spending all of it... in one year.
Why I'm not worried about AI job loss
February 2026 · David Oks
A counterweight to the panic that AI will destroy all jobs, reminiscent of Tyler Cowen's insistence that human bottlenecks will not result in GDP growth explosion anytime soon.
Why ATMs didn't kill bank teller jobs, but the iPhone did
March 2026 · David Oks
ATMs automated the teller's tasks but didn't kill the job. What finally killed the bank teller was the iPhone, because it changed how people behave entirely.
Task-Completion Time Horizons of Frontier AI Models
April 2026 · METR
The best model in April 2026 can complete tasks that take humans 12 hours with 50% confidence. In 2022, the best AI could only complete tasks that took humans 1 minute.
Quite spooky. People in the AI world claim that AI is making blindingly fast progress, but the mainstream seems be either unaware or in denial.
This makes sense: LLMs may have saturated in capabilities for the chatbot use case (which a billion+ people now use every month), but this chart from METR demonstrates where the progress has actually been.
AI Value Capture: The Shift to Models
April 2026 · Dylan Patel via SemiAnalysis
The world changed in December 2025, when Agentic AI began to really work.
Annualized token spend at SemiAnalysis is already ~30% of employee compensation and we’re consuming just under 5B tokens per month per employee
In early 2026, Anthropic's ARR jumped from $9B to $44B+ and inference margins from ~38% to over 70%. Dylan Patel reveals his company is exhibit A: paying Anthropic $7 million a year for tokens for his 60 employees because the output outcompetes everything else. It makes Slatton's 2025 numbers look like peanuts.