AI's safety reckoning: a forecaster's 70% doom vs. a 'moral panic,' Anthropic at the center
Last week's fight over confident narratives reached AI's central question. Former OpenAI forecaster Daniel Kokotajlo told Steven Bartlett there's a ~70% chance development goes 'horribly wrong'; days later All-In's hosts called AI regulation a 'moral panic' and cast Anthropic's safety push as regulatory capture — with energy and token costs running underneath, and Kevin Rudd sketching the industrial state on the other side of them.
Jul 13 – Jul 19, 202610 episodes5 shows269 claims on the record37 new predictions
01
70% doom or 'moral panic'? The AI-safety fight comes into the open
Two shows staked out opposite poles on AI's central question. On Diary of a CEO, former OpenAI forecaster Daniel Kokotajlo — who risked ~$2M in equity rather than sign a non-disparagement clause — put ~70% odds that AI ends up ruling the world, arguing OpenAI and Anthropic are racing to automate their own research and that none of these people should be trusted with that much power. Three days later, All-In's David Sacks framed the safety movement as a 'moral panic' and cast Anthropic's regulatory push as capture: brand yourself as a safe AI company, ban unsafe AI, profit. Both circle the same company — Anthropic, which Sacks pegs near a $1T valuation headed for $3T. The disagreement isn't about the stakes; it's about whether regulation protects the public or entrenches the incumbents writing it.
“None of these people should be trusted with that much power, basically.”
The boom's binding constraint is energy — and China built its state around owning it
If the safety fight is about whether to slow AI down, two All-In guests argued physics may decide it anyway. Former Intel CEO Pat Gelsinger called energy capacity the natural governor on the buildout, a 'silver lining' that keeps a bubble from running away, while warning that Taiwan's fabs sit on under three weeks of energy reserves. Chamath Palihapitiya put numbers on the squeeze: by 2050 the US will be 'two and a half California's worth of energy in deficit,' and frontier tokens run about $56 per million against roughly 50 cents on Chinese models. On Ezra Klein's show, Kevin Rudd described the state on the other side of that gap — one deliberately holding down household consumption to fund manufacturing and own global supply chains, financed out of Chinese savings. Rudd's read on the endgame is not subtlety but attrition, and Ezra Klein notes the leverage already worked once: rare earths broke a US trade war.
“there is a silver lining here that guarantees we don't get too far ahead of ourselves in terms of bubble, and that is energy capacity.”
AI stops augmenting and starts doing the work — the services economy gets eaten
Across two All-In interviews, European founders described AI crossing from assisting professionals to replacing their work. Legora's Max Junestrand said legal is a trillion-dollar market that is only '4% software, 96% service, which is bananas' — and that after Opus 4.5 and 4.6, agents can now run end-to-end case strategy, breaking the billable hour. Lovable's Anton Osika, whose platform sees a million new products a week, put it bluntly: you don't need to be an engineer to create software. And ElevenLabs' Mati Stanislawski, at $600M ARR with no product managers, added voice to the list of newly-softwared industries. It's the concrete version of Kokotajlo's warning that job displacement, when it comes, arrives suddenly and late — except here it is already reshaping law and code.
“So it means there's 4% software, 96% service, which is bananas.”
Amid the AI noise, two guests on why trust beats logic
While the tech shows argued about compute, two guests made the week's quietest, most portable case: people move on trust and attention, not argument. On Invest Like the Best, veteran fundraiser John Kim — who helped build General Catalyst and now raises for a $500M+ science venture — reduced persuasion to 'desire minus fear' and warned that the logic is an output of a successful pitch, not the input. Money, he said, moves at the speed of trust. On Diary of a CEO, Harvard's Alison Wood Brooks gave the mechanics: for most people the single best fix is to ask far more questions, and the benefits are almost entirely driven by follow-ups. One was closing billion-dollar rounds; the other, second dates — and both landed on the same finding.
“it gets back to money moves at the speed of trust.”
UCLA's Alan Castel told Huberman that a walking group grew hippocampal volume ~1% in a randomized trial, and that how old you feel predicts longevity better than your biological age does. HUBERMAN LAB →
A Huberman Lab Essentials traced lithium to John Cade's 1940s guinea-pig experiments and stressed that bipolar disorder — carrying a 20-30x elevated suicide risk — needs medication, not talk therapy or fish oil alone. HUBERMAN LAB →
Boston College's Thomas Seyfried argued to Steven Bartlett that cancer is a mitochondrial metabolic disorder, not a genetic disease, and that ketosis lets standard chemo work at far lower doses — which mainstream oncology rejects. DIARY OF A CEO →
Kokotajlo's 'Plan A' floats a Citizen's Dividend starting near $25,000 a person per year and growing toward ~$10M each in a fully automated economy. DIARY OF A CEO →
Deal wave: Apple's 41-page suit accuses OpenAI of poaching 400+ employees and stealing trade secrets for a consumer device, while Stripe, Advent and Block bid ~$53B for PayPal and its 439M accounts. ALL-IN →
Kevin Rudd names a date for the Taiwan risk: he worries Xi could miscalculate in 2028, when Taiwan's January election, Xi's fresh party mandate and a US presidential race all land in the same year. EZRA KLEIN →
The record
Every episode this week
The issue above is written from these structured records — nothing else. Each card is the full breakdown: claims, predictions, quotes, all timestamped.
A topical anti-glycation (CML-clearing) cream would be the first market for the enzyme and a trillion-dollar (up to two-trillion-dollar) cosmetic market.
The week's episodes are diarized into structured records; one editorial pass writes this issue from those records when the week's ingest lands. A grounding gate greps every quote above against its episode record — a quote that isn't verbatim fails the build.
the data behind this issue
# this issue, as structured data — stories, citations, receipts
curl 'https://aipod-two.vercel.app/digest/2026-w29.json'# the records it was written from
curl 'https://aipod-two.vercel.app/api/episode/<id>'# any id in episode_ids