Hard Fork · Max Read

With Erin Griffith, Mike Isaac and Jasmine Sun

The AI boom needs a better scorecard

Our interpretation: judge economics, commitments, control and decision quality—not one headline metric.

A loss is not a cash-burn figure · 04:28

Anthropic’s first-half 2026 net loss was $80.7 billion, heavily affected by financing revaluations; its operating loss was $693 million.

These are different accounting measures, not interchangeable cash-burn figures or proof that costs disappear.

Capacity comes before certainty · 05:31

Griffith describes the timing risk of compute commitments: excess capacity can sit unused, while insufficient capacity can lose customers.

Erin Griffith’s explanation of the trade-off.

Revenue growth does not remove the risk that capacity and demand arrive at different times.

Shares are not control · 13:38

The proposed S-1 structure gives founders 50.1% of general voting power while a separate trust appoints a board majority.

Voting power, economic ownership and board-appointment rights differ. The filing describes a proposed offering, not a completed IPO.

Tokens are not outcomes · 46:36

Isaac argues that token use and code volume alone are weak productivity measures. More output does not necessarily mean durable value.

Mike Isaac’s argument, rather than a universal productivity result.

Test human oversight, too · 48:52

In a 50-analyst experiment, choices tracked randomly nudged AI advice.

One task, one model and deliberately strengthened agent preferences limit generalization. This does not show that every human-AI workflow behaves the same way.

Sources:

Exact Hard Fork episode on Apple Podcasts: https://podcasts.apple.com/us/podcast/are-we-in-an-a-i-bubble-were-about-to-find-out/id1528594034?i=1000798873825

Anthropic SEC S-1, filed October 26, 2026: https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm

Anthropic research: How AI agents shape experts’ judgment calls: https://www.anthropic.com/research/agents-judgment-calls

Anthropic filing announcement: https://www.anthropic.com/news/s1-sec

Source context:

Five selected ideas from Hard Fork’s October 30 episode with Max Read, Erin Griffith, Mike Isaac and Jasmine Sun. Accounting and control context are linked to the S-1; the judgment experiment is linked to the original study. Time anchors refer to podcast audio, not synchronized video.

Keep in mind:

First-half 2026 net loss and operating loss measure different things; neither should be substituted for cash burn. Compute commitments retain timing risk.

The S-1 describes a proposed offering. Founder voting power is distinct from economic ownership and the trust’s board-appointment rights.

The 50-analyst experiment used one task and one model with intentionally nudged agent preferences; do not generalize it to all AI-assisted work.