Skip to content
Sign in
Artwork for Deep Dive
Technology

Deep Dive

Deep Dive

Deep Dive is long-form research on AI, tech, and the global economy. Single host, weekly episodes, 25-35 minutes each. The story behind every headline — built from primary sources and original analysis.

Recent topics:
• AI deanonymization research
• Data center infrastructure economics
• Strait of Hormuz geopolitics
• Agentic AI security
• Frontier model behaviors

Find Deep Dive across platforms:
📺 YouTube · @DeepDiveAIShow
📱 TikTok · @notdeepdiveai
📷 Instagram · @notdeepdive
🔗 All links · linktr.ee/notdeepdive

Tap follow for new episodes.

Play
Open in
  • 36 episodes
  • a few times a week
  • Avg 23 min
  • English

Episodes 36

NewestOldestLongest
All20Unplayed20In progress0Played0
Counted on this page — what you have heard stays on this device, so it is not something the list can be paged by.
  • Yesterday · 24 min

    Why Banks Are Blocking AI Agents (and Who Pays When One Goes Rogue)

    On October 13th, JPMorgan's finance chief was asked what happens if an AI agent does something you didn't want. His answer: "That's going to be on them." Within ten days, most of the country's biggest banks had moved to block or restrict AI agents like Meta's Muse. The stated reason is rogue agents. Underneath sits a 1978 federal law that sorts electronic transfers through two doors. Through one, you gave the access, so the loss is yours. Through the other, the access was stolen or tricked out of you, so the bank pays. No court or regulator has said which door an AI agent walks through. We read the contracts Chase and Wells Fargo wrote before the lockout went public, Meta's own terms, a federal judge's September 3rd ruling, and the money underneath: roughly $350 a year on every $10,000 in a big-bank savings account. Then the strong case against that money theory, what to do before you give any app your bank password, and three dated predictions. As of November 8th, 2026. Wells Fargo's new terms take effect on November 24th. RELATED EPISODES Robinhood Let AI Trade Your Money — And Said It's Not Responsible. Can It Actually Do That? How an AI Agent Hacked Hugging Face How AI Agents Actually Work CHAPTERS 00:00 "That's going to be on them" 01:06 What this episode covers 01:35 Meta's Muse and the road to your bank 02:23 The wall goes up 03:54 What the banks actually reported 05:08 The 1978 law with two doors 07:18 An AI agent in the doorway 09:01 A judge rules on the human version 09:49 Regulators leave the gap open 11:10 The agent maker's terms 12:27 Contracts written before the lockout 15:40 The money: your savings 18:47 The case against the money theory 20:32 How long does the wall stand? 21:43 What to do now 22:42 Three predictions SOURCES JPMorgan Chase Q3 2026 earnings call transcript: https://www.jpmorganchase.com/content/dam/jpmc/jpmorgan-chase-and-co/investor-relations/documents/quarterly-earnings/2026/3rd-quarter/3q26-earnings-transcript.pdf Wells Fargo Online Access Agreement (effective Nov 24, 2026): https://www.wellsfargo.com/online-banking/online-access-agreement/upcoming/ CFPB, Electronic Fund Transfers FAQs (Regulation E): https://www.consumerfinance.gov/compliance/compliance-resources/deposit-accounts-resources/electronic-fund-transfers/electronic-fund-transfers-faqs/ Court order, N.D. Cal. (Sept 3, 2026): https://storage.courtlistener.com/recap/gov.uscourts.cand.471937/gov.uscourts.cand.471937.18.0.pdf Meta Muse terms: https://muse.ai/terms

  • Saturday · 23 min

    The Anthropic IPO: How Much Does Anthropic Actually Owe?

    On October 26th, Anthropic's public S-1 left its biggest number blank. On November 2nd, an amendment filled it in: approximately $895.4 billion, and the same page says $213.8 billion of it carries termination rights or conditions. As of November 2nd, 2026, here are six honest answers to how much Anthropic owes, from $6.2 billion borrowed to more than $1 trillion. Plus the third quarter, what Amazon, Alphabet, SpaceX and Broadcom told their own shareholders, and three new dated calls. Disclosure: this show is made with Claude, and Anthropic makes Claude. Check the sources, not the brand. OUR RECORD ON ANTHROPIC, MISSES INCLUDED (beyond the three calls scored on air) From The Anthropic IPO: What the $518 Billion Bill Actually Says - Moved: Microsoft near $31 billion; an October amendment made it at least $44.9 billion. - Miss: we called AMD flexible. It can't be cancelled without cause. - Moved: big contracts at about $71 billion a year by 2028; on the October 26th filing, about $96 billion. - Miss: a second paper charge of $9.4 billion, on notes with Google and Amazon. - Miss: the charge can keep running after the IPO. - Miss: Anthropic's charge ran about $15 billion past Amazon's pre-tax gain on the same notes. - Depends on the count: Fortune's $657 billion clears $518 billion; its $498 billion noncancelable doesn't. - Can't be graded: SpaceX's report had no cloud-only figure. From Two Models on Purpose - Miss: an S-1 within 90 days. It came 73 days late. Miss: no October IPO. - Hit, wrong reason: government revenue under 3%. The filing says under 2% in every period. From How Anthropic's Security Team Tracks Anti-AI Activists - Hit on dollars: a CEO security perk of at least $1 million. The filing shows $2,745,874 for 2025. From How Anthropic Actually Makes Money - Half answered: a projected $559 million quarter came in at $974 million of operating profit, unaudited. - Answered: the third quarter was a $567 million operating loss, though adjusted operating income rose. - Answered: the chips and data centers are mostly rented. Still open: a giant AI listing ends 2026 below its offer (The SpaceX IPO: Why the Biggest IPO in History Loses Money); the IPO slips or prices below a $380 billion value (How Anthropic Actually Makes Money). Totals labeled as ours are our arithmetic; the filing states none of them. RELATED EPISODES The Anthropic IPO: What the $518 Billion Bill Actually Says How Anthropic Actually Makes Money Why NVIDIA Paid 86 Times Revenue for Hugging Face — a supplier propping up its own customer CHAPTERS 00:00 One number left blank 01:54 Scoring our own October calls 02:37 The miss: SpaceX's ceiling 04:03 A hit, and why it's worth little 05:49 Six answers to how much Anthropic owes 06:50 The headline: $895.4 billion 09:30 Anthropic as a $60,000 household 10:50 What the third quarter changed 12:14 Was the second quarter the peak? 12:47 What Amazon told its shareholders 14:15 Alphabet, SpaceX and Broadcom 18:09 The founders, and the case on both sides 20:55 Three new calls 22:20 Check it yourself on EDGAR 22:53 So how much does Anthropic owe? SOURCES Anthropic S-1/A (Nov 2, 2026): https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm Anthropic S-1 (Oct 26, 2026): https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm Amazon 10-Q, quarter ended Sept 30, 2026: https://www.sec.gov/Archives/edgar/data/1018724/000101872426000040/amzn-20260930.htm Broadcom 8-K (Oct 26, 2026): https://www.sec.gov/Archives/edgar/data/1730168/000119312526434271/d148323d8k.htm ——— This episode is for educational and informational purposes only and does not constitute financial, investment, or trading advice, nor a recommendation to buy or sell any security. Deep Dive is not a registered investment adviser and holds no positions in the companies discussed. All investing involves risk. Consult a licensed financial professional before making investment decisions.

  • October 24 · 24 min

    How the DeepSeek Ban Actually Works

    On July 24th, 2026, the Secretary of Homeland Security ordered every civilian federal agency to stop buying DeepSeek's products and remove them, or ask for an exception within 60 days. That window closed around September 22nd. The public found out on October 19th, a day before the order ran in the Federal Register. Its reasons are mostly about where data goes. What it orders removed is wider: third-party products and services developed using DeepSeek implementations. The order never says what that means. We read the four-page order itself: seven public sources, one technical audit (of DeepSeek's iPhone app), and a footnote about a DeepSeek AG that appears nowhere else in it. Then the quiet 2018 law behind it, the Federal Acquisition Supply Chain Security Act. This is Homeland Security's first order under it and the first in the Federal Register. We cover the heavier tools Commerce hasn't used, and Kaspersky's seven-year climb. A disclosure: this show is made with Claude, Anthropic's model, and Anthropic competes with DeepSeek. The Pentagon used the same law against Anthropic in March, and on September 25th the appeals court in Washington upheld it, two votes to one. Then what those two words reach. DeepSeek gives its model away: about 50 million downloads of R1 on Hugging Face, 524 models built from it, copies sold by Amazon and Microsoft, and a top DeepSeek model on OpenRouter served by 29 companies, none of them DeepSeek. We weigh the case for the broad reach (NIST's own test of the downloaded model, CrowdStrike's Tibet-prompt finding) and the case against it, how a contractor would ever check, and who pays. Then China's answer, including the October 16th draft whose compute line for model weights is the one the United States drew in January 2025 and scrapped that May. Plus what it means for you, and two dated calls. CHAPTERS 00:00 Issued in July, public in October 01:50 Four pages, printed as images 02:58 The order's reasons: where data goes 03:37 Seven public sources, one app audit 04:34 App versus model, and a stray footnote 05:23 The quiet 2018 law behind it 06:36 The heavier tools nobody used 07:19 Kaspersky's seven-year ladder 08:04 A disclosure, and the Anthropic case 09:23 Two words: developed using 10:11 Read literally: Amazon, Microsoft, NVIDIA, vLLM 11:14 One model, 29 companies serving it 11:53 The line earlier bans drew 12:41 NIST's test of the downloaded model 13:24 CrowdStrike and the Tibet prompt 14:26 The case for the broad reach 15:04 The case against it 15:45 A release on the day of the order 16:12 How would anyone enforce it? 17:41 Who pays: the Huawei precedent 18:09 Who in government was using DeepSeek? 18:54 Beijing's answer 19:40 China's draft export list 20:08 The line Washington walked away from 21:15 Response or coincidence? 21:58 What it means for you 22:51 Two calls, with dates 23:25 Where the two words end SOURCES The order, Federal Register, Oct 20, 2026: https://www.govinfo.gov/content/pkg/FR-2026-10-20/pdf/2026-21447.pdf 41 U.S.C. 1323 (the 2018 law): https://www.govinfo.gov/content/pkg/USCODE-2024-title41/html/USCODE-2024-title41-subtitleI-divsnB-chap13-subchapIII-sec1323.htm Lawfare on Anthropic's rehearing petition: https://www.lawfaremedia.org/article/why-anthropic-is-asking-for-a-rehearing-en-banc National Law Review on the D.C. Circuit ruling: https://natlawreview.com/article/dc-circuit-upholds-anthropic-fascsa-designation-and-endorses-broad-interpretation NIST CAISI evaluation of DeepSeek: https://www.nist.gov/news-events/news/2025/09/caisi-evaluation-deepseek-ai-models-finds-shortcomings-and-risks CrowdStrike research: https://www.crowdstrike.com/en-us/blog/crowdstrike-researchers-identify-hidden-vulnerabilities-ai-coded-software/ The Register, Oct 20, 2026: https://www.theregister.com/security/2026/10/20/deepseek-deep-sixed-at-all-us-federal-civilian-agencies/5304010 Global Times on China's draft list: https://www.globaltimes.cn/page/202610/1372405.shtml

  • October 21 · 28 min

    How LLM-as-a-Judge Actually Works (and Why the Same Answer Scores Differently Tomorrow)

    Send the same request to the same AI grader, byte for byte, at temperature zero, a day later, and 22 times out of 100 its ranking of the answers comes back different. Replayed the same day, it changed just as often, and on all four providers tested, rankings moved. That grader, an LLM judge, is the same kind of instrument that helps pick what chatbots learn from and scores the benchmarks in launch posts. Since July 2026, one of them also casts votes on Arena's leaderboard, the one that was supposed to be human. The teach first: what a judge is, why the industry adopted one at about half a cent a grade despite flaws documented in 2023, and its four jobs: what a model reads, what it learns to say, the launch headline (Gemini 2.5 scored 79 percent on Arena-Hard with Gemini 2.5 as judge, and 49 with GPT-4.1), and the leaderboard. Then what a judge physically is: Thinking Machines Lab sent one prompt 1,000 times at temperature zero and got 80 different answers, because the batch a request lands in changes how the chip rounds the arithmetic. With batch-invariant kernels on, all 1,000 matched. Along the way, a correction to our episode How LLM Inference Actually Works, which called temperature zero fully deterministic. Then the audit, limits first: Haoyuan Zhu of the University of Sheffield and Jie Zhang of Ranplan Wireless, posted to arXiv on September 3rd, 2026, with no peer review. Their grader needed a ranking agreement of 0.9 and got 0.4, mostly because the answers were near-ties, as the authors say themselves. In a byte-for-byte replay, 78 of 100 rankings came back identical against a bar of 99. Running the grader yourself with the fix came close on a quiet machine but didn't reach it. Then who's exposed: OpenAI switches off two of the founding graders on October 23rd, 2026, and GPT-4.1 on January 14th, 2027, so their scores can't be rerun. And what you can do: replay 100 of your own evaluation requests, run a pilot before you set a pass mark, and ask the grader five times when a verdict matters. CHAPTERS 00:00 Same request, a day later 01:47 What an LLM judge is 02:48 Judging LLM-as-a-Judge, 2023 04:40 Four jobs, starting with what models read 05:35 The launch headline 06:34 The leaderboard that was supposed to be human 08:00 When the automated checks said fine 09:08 One prompt, a thousand times 09:50 Why the batch changes the arithmetic 11:30 A correction on temperature zero 12:00 Even zero isn't a promise 12:42 Same name, different behavior 13:18 Anthropic's August 2025 bugs 15:45 The audit, limits first 17:27 A grader that never passed 19:16 Close contests at the top 19:56 78 in 100 20:39 Four providers and the self-hosted fix 22:02 The sharpest criticism is in the study 22:22 Validity versus reliability 23:14 The cheap fixes 24:30 Who's exposed 26:01 What deterministic serving costs 26:47 What you can do 27:37 The grader was never a ruler SOURCES Zhu & Zhang, arXiv, Sep 3, 2026: https://arxiv.org/abs/2609.04198 Thinking Machines Lab, Sep 10, 2025: https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/ Zheng et al., Judging LLM-as-a-Judge, 2023: https://arxiv.org/abs/2306.05685 Anthropic postmortem, Sep 17, 2025: https://www.anthropic.com/engineering/a-postmortem-of-three-recent-issues OpenAI on GPT-4o sycophancy, May 2, 2025: https://openai.com/index/expanding-on-sycophancy/ Arena, AutoEval, Jul 30, 2026: https://arena.ai/blog/autoeval-scores Arena, judge study, Sep 30, 2026: https://arena.ai/blog/llm-judge-self-preference Chen, Zaharia & Zou, 2023: https://arxiv.org/abs/2307.09009 OpenAI API deprecations: https://developers.openai.com/api/docs/deprecations

  • October 17 · 37 min

    Deloitte, Abramovich and the $10 Billion Paper Trail

    On October 12th, 2026, a consortium of newsrooms — the Bureau of Investigative Journalism, BBC Panorama, the Guardian and OCCRP, inside the Cyprus Confidential leak led by ICIJ — set out how, in their reading of leaked emails and draft accounts, Deloitte's Cyprus office and its spin-off MeritServus helped Roman Abramovich's companies backdate more than 1,000 corporate records between 1999 and 2013, then certified the accounts as true and fair. The emails are not public. Everyone named denies wrongdoing or did not respond; Demetris Ioannides, who founded Deloitte's Cyprus practice, is 83 and through his lawyers unable to respond. No court and no regulator has ruled. The paper first: four schemes in the consortium's telling. A Cyprus company holding about $1.8 billion in cash that paid $15,000 of tax for two years. A château, a 424-acre estate and a $150 million yacht written out of a company's accounts on a sale agreement dated January 2003 and drafted in November 2006. A holding company with more than $10 billion beneath it, swapped for a BVI twin with the same name. And a $98 million consultancy fee the Bureau reads as a payout, 99.9% of it to Abramovich as commission, on promissory notes carrying his signature. Then the audit. Every auditing standard assumes the auditor arrived after the paperwork existed; ISA 240 lets an auditor treat records as genuine unless there is reason to believe otherwise, and the ethics code calls the alternative the self-review threat. The 2005 split that spun MeritServus out of Deloitte Cyprus was supposed to solve exactly that, and ICIJ found Ioannides on a Deloitte subsidiary's board until mid-2023. Then who could act: Cyprus, where the auditors' overseer says it is taking all necessary actions and no case has been announced; Britain, where HMRC has said the same sentence twice; Jersey, where over $7 billion has been frozen since April 2022; and the £2.5 billion from the Chelsea sale, frozen in a British bank and earning at least £175 million of interest while not a pound reaches Ukraine. Three dated predictions close it. As of October 15th, 2026. CHAPTERS 00:00 Fifteen thousand dollars 01:04 A leak, four newsrooms, and the denials 01:50 Your pension fund has an auditor 02:34 The pile of paper 03:46 What backdating actually involves 04:21 Scheme one: making $2 billion disappear 08:04 Scheme two: the château 11:49 Scheme three: pick a company, any company 14:41 Scheme four: the signature 16:25 The 2008 case 17:42 Why didn't the audit catch it? 18:42 Is a backdated document always improper? 20:01 The hard points 20:44 What the standards assume 22:28 The 2005 split 23:45 Altered papers, the other way round 26:33 Who could act: Cyprus 29:26 Britain, Jersey, and the Chelsea money 32:50 Who you'd even sue 33:56 What the signature promises you 34:57 What you can do 35:32 What to watch 36:27 Three predictions 37:09 One thing to keep SOURCES OCCRP, Oct 12, 2026: https://www.occrp.org/en/project/cyprus-confidential/deloittes-office-in-cyprus-helped-abramovich-firms-relocate-billions-of-dollars-in-assets-in-apparent-bid-to-sidestep-tax-and-transparency The Bureau of Investigative Journalism, Oct 12, 2026: https://www.thebureauinvestigates.com/stories/2026-10-12/how-deloitte-helped-abramovich-hide-10bn BBC Panorama, Oct 12, 2026: https://www.bbc.com/news/articles/c64g1l5389kvo The Guardian, Oct 12, 2026: https://www.theguardian.com/world/2026/oct/12/roman-abramovich-aides-paperwork-tax-records-suggest ICIJ, Nov 2024: https://www.icij.org/investigations/cyprus-confidential/deloitte-held-close-ties-to-a-now-sanctioned-cyprus-firm-accused-of-shielding-oligarchs-wealth-records-reveal/ High Court, Yugraneft v Abramovich, 2008: https://caselaw.nationalarchives.gov.uk/ewhc/comm/2008/2613 Cyprus Mail, Oct 15, 2026: https://cyprus-mail.com/2026/10/15/audit-watchdog-investigating-deloitte-cyprus-run-by-a-former-manager

  • October 15 · 26 min

    Why AMD Bought a Chip That Can't Be Reprogrammed

    On August 6th, 2026, AMD announced it was buying Taalas, a Toronto startup whose chip, the HC1, has the weights of one AI model, Llama 3.1 8B, etched into its wiring at the factory. It can never run anything else. The company claims 17,000 tokens a second to a single user, more than 70 times an NVIDIA H200. No independent lab has benchmarked it; every multiple in this episode is the company measuring itself. The machine first. Every token a chatbot writes re-reads the whole model from memory, and the AI hardware industry is organized around widening that toll road. Taalas deletes it: one transistor stores a weight and does its math. No HBM, no advanced packaging, no liquid cooling, on 2019-era 6-nanometer silicon. The model lives in 2 of the chip's 100-plus layers, so a new model is a reprint of 2 stencils, which the company says takes about 2 months. One model is in silicon so far, and that first chip took two and a half years. Then the money. Token prices fell a thousand-fold in three years and AI application companies still average 38% gross margins, because competition drags every product onto the newest, most expensive model. Etching turns the model from a running cost into a fixed asset, rent into mortgage, at three-quarters of a cent per million tokens by the company's own number. The founder, Ljubisa Bajic, said on camera that the approach starts winning once a model is held for about 3 months. In 2022, on an MIT stage, the same man said hardwiring had been wrong every time. Then the wave: NVIDIA buying Groq's assets for about $20 billion on Christmas Eve 2025, the AMD and Cerebras partnership, AMD and Taalas, Qualcomm reportedly circling Tenstorrent. The bet on trial, Etched's warning, the Bitcoin-mining precedent, four dated predictions, and the one tell you can watch in your own tools. CHAPTERS 00:00 Disclaimer 00:08 Wrong every time 00:58 Three questions 01:29 The toll road every token pays 02:37 The model is the compute 03:57 Built on 2019-era silicon 04:27 The company's own numbers 05:46 The precision catch 06:11 The model is a print run 07:22 One model in silicon so far 08:04 Compute, priced like payroll 09:07 Token prices fell, margins didn't 10:36 Rent versus mortgage 12:25 The three-month line 13:39 The man who built what he'd ruled out 16:45 Eight months, four moves 18:21 Where specialization stops 20:14 The bet on trial 22:00 The bull case and Bitcoin 23:17 Science or infrastructure? 24:41 Four predictions 25:33 Your own tell to watch SOURCES AMD press release, Aug 6, 2026: https://ir.amd.com/news-events/press-releases/detail/1296/amd-acquires-taalas-to-advance-compute-solutions-for-rapidly-growing-ai-inference-market Taalas HC1 product page (specs; 'technology demonstrator'): https://taalas.com/products/ Taalas manifesto (2-month respin, 24 people, $30M): https://taalas.com/the-path-to-ubiquitous-ai/ The Next Platform, Feb 19, 2026 (the single-transistor cell): https://www.nextplatform.com/compute/2026/02/19/taalas-etches-ai-models-onto-transistors-to-rocket-boost-inference/4092140 Bajic on TBPN, Feb 20, 2026 (3-month breakeven, on camera): https://www.youtube.com/watch?v=UrebAsni59k Bajic at MIT, 2022 ('wrong every time'): https://www.youtube.com/watch?v=L4AgmG8V3LE Sundar Pichai, Google I/O 2026 (3.2 quadrillion tokens a month): https://blog.google/innovation-and-ai/sundar-pichai-io-2026/ CNBC, Dec 24, 2025 (NVIDIA buys Groq assets, ~$20B): https://www.cnbc.com/2025/12/24/nvidia-buying-ai-chip-startup-groq-for-about-20-billion-biggest-deal.html ——— This episode is for educational and informational purposes only and does not constitute financial, investment, or trading advice, nor a recommendation to buy or sell any security. Deep Dive is not a registered investment adviser. All investing involves risk. Consult a licensed financial professional before making investment decisions.

  • October 13 · 22 min

    Why We Age: Your Cells Forget What They Are

    On September 30th, 2026, Nature published a review proposing that aging is your cells slowly forgetting what they are. About 90% of the chemical marks that pile up on DNA with age land on the switches that keep a cell's other possible identities turned off. It's a big idea, and it's a review, not an experiment: it redraws the map of old data, and the data underneath is small (seven immune samples, skin from six people, eight mice, with the pattern then checked in blood studies of thousands). Two hosts take it apart. One reads the papers: Waddington's valleys, Gurdon's frogs, Yamanaka's four factors, Horvath's clock. The other argues the older theory, that aging is damage, and follows the money: a 2022 study of 16 mammal species, the naked mole-rat that doesn't fit, the Sinclair mice that start with a DNA break, and why four grand theories of aging have already come and gone. Their shared reading, offered as a reading and not a finding: identity loss may be where damage ends up, not its rival. One experiment would separate the two, and nobody has run it. Then what it means for you. Altos Labs launched with $3 billion and has no disclosed drug in human trials; NewLimit has raised $610 million; Sam Altman put $180 million into Retro. The first regulated human test of a cellular rewind reported early data from three glaucoma patients on October 8th. And the $300 biological-age test you can buy today runs on clocks that, in a 2022 study, disagreed with themselves by up to nine years on the same sample. Calorie restriction and omega-3 moved those clocks by weeks to months. The one question to ask before you buy, and what to watch instead of the headlines. Disclosure: eight of the Altos review's eleven authors work at Altos Labs; one of the Nature review's three authors is employed by Novartis. Both conflicts are declared in the papers and examined here, not insinuated. None of this is medical advice. As of October 12th, 2026. RELATED EPISODES AI and Dogs: The Technology That Could Give Your Dog More Years — the longevity-drug side of the same question, tested in dogs CHAPTERS 00:00 Your cells forget what they are 01:12 What actually came out 02:13 The 90% number 03:37 Why this isn't Sinclair's theory 04:19 Waddington's valleys 05:34 Identity is reversible 06:27 Horvath's clock 07:35 The damage camp 09:23 The contested experiment 10:29 Four theories, four letdowns 11:24 The mouse record 12:32 Closer than the headlines say 13:28 Where the wear and tear ends up 14:24 The counterweight 15:06 Follow the money 16:23 The track record 17:09 The first human test 18:48 The test you can buy 20:21 What to ask before you buy 21:42 Each host's bottom line SOURCES Nature review (Yücel, Molière, Gladyshev), Sept 30, 2026: https://doi.org/10.1038/s41586-026-10955-0 Cell review, Altos Labs (Lu, Tu, Izpisua Belmonte), May 2026: https://doi.org/10.1016/j.cell.2026.04.020 The 90% PRC2 figure (Moqri et al., 2024): https://pmc.ncbi.nlm.nih.gov/articles/PMC11250797/ Higgins-Chen et al., Nature Aging 2022 — clock replicate noise: https://doi.org/10.1038/s43587-022-00248-2 Life Biosciences ER-100 early readout, Oct 8, 2026: https://www.lifebiosciences.com/life-biosciences-announces-first-in-human-data-from-ongoing-phase-1-trial-evaluating-er-100-in-optic-neuropathies/ Yamanaka in TIME, Aug 20, 2026: https://time.com/collection/longevity-leaders/2026/shinya-yamanaka/ Izpisua Belmonte at ESGCT via GEN, Oct 2025: https://www.genengnews.com/insights/esgct-2025-altos-labs-scientific-founder-targets-mesenchymal-drift-in-efforts-to-combat-diseases-of-aging/ Timmons & Brenner critique + authors' reply (Cell, 2024): https://www.brennerlab.net/wp-content/uploads/2024/03/Timmons24.pdf

  • October 3 · 32 min

    The Anthropic IPO: What the $518 Billion Bill Actually Says

    Anthropic's IPO prospectus isn't public yet. On September 28th, 2026, Reuters reported numbers from a copy it had seen: nearly $4.6 billion of revenue in 2025, a $42 billion net loss that is mostly a non-cash charge, and $518 billion of commitments for cloud, compute and infrastructure, about 80% of it non-cancelable. A bill has two sides, and the companies on the other side of this one file with the SEC. Read from their own filings, the $518 billion is six contracts. Amazon's report agrees with what Reuters saw. SpaceX's own prospectus puts its Anthropic deal at about $45 billion; the reported line is up to $84.5 billion. Google and Microsoft never name Anthropic. And the biggest line, about $161 billion of chip-rack leases, runs through an Apollo-managed vehicle, with Broadcom backstopping the leases up to about $29 billion, Apollo's annuity insurer giving a limited guarantee it judges not probable, and the customer allowed to pay with notes convertible into its own stock. Then the owners: Amazon carries about $190 billion of Anthropic on its books, the founders keep 50.1% of the votes, and your index fund may buy in 15 trading days after listing. With the bull case in the company's own words, a scorecard of this show's earlier calls (misses included), and three dated calls. Disclosure: this show is made with Claude, Anthropic's model. Every prospectus figure here is attributed to the outlet that reported it. As of October 2nd, 2026, no Anthropic registration statement was on EDGAR. RELATED EPISODES How Anthropic Actually Makes Money — the parent episode whose three tells this one scores. The SpaceX IPO: Why the Biggest IPO in History Loses Money — where the Anthropic lease first surfaced, in SpaceX's own filing. Why NVIDIA Paid 86 Times Revenue for Hugging Face — the same move from equity to guarantees, at NVIDIA. CHAPTERS 00:00 No hedge on earth 00:58 Disclosure and roadmap 02:01 What you're allowed to know 02:42 The $42 billion loss, decomposed 04:30 The bill is six contracts 07:24 The other side of the ledger 07:55 Google and Microsoft: silent by name 09:08 SpaceX: the bill and the receipt 10:42 Broadcom's backstop, in its own words 13:00 The filing that names the customer 14:48 Suppliers who are also owners 18:22 Deals outside the $518 billion 19:39 The case that all of this is fine 22:40 The deal, and who decides 24:57 Your index fund 25:39 Buying in before the IPO 26:47 Scoring this show's own calls 28:28 Three calls for the public filing 31:05 Where the bill went SOURCES Reuters via Yahoo Finance, Anthropic's IPO prospectus (Sept 28, 2026): https://finance.yahoo.com/technology/ai/articles/exclusive-anthropics-ipo-prospectus-shows-231722972.html Reuters via KSL, the $518B breakdown (Sept 29, 2026): https://www.ksl.com/article/51629885/anthropics-518-billion-ai-buildout-hinges-largely-on-deals-that-cannot-be-canceled-filing-shows Amazon 10-Q, quarter ended June 30, 2026: https://www.sec.gov/Archives/edgar/data/0001018724/000101872426000026/amzn-20260630.htm Broadcom 10-Q, quarter ended Aug 2, 2026: https://www.sec.gov/Archives/edgar/data/1730168/000173016826000080/avgo-20260802.htm Apollo 10-Q, quarter ended June 30, 2026: https://www.sec.gov/Archives/edgar/data/1858681/000185868126000040/apo-20260630.htm SpaceX S-1/A (June 3, 2026): https://www.sec.gov/Archives/edgar/data/1181412/000162828026040364/spaceexplorationtechnologib.htm Dario Amodei on the Dwarkesh Podcast (Feb 13, 2026): https://www.dwarkesh.com/p/dario-amodei-2 ——— This episode is for educational and informational purposes only and does not constitute financial, investment, or trading advice, nor a recommendation to buy or sell any security. Deep Dive is not a registered investment adviser and holds no positions in the companies discussed. All investing involves risk. Consult a licensed financial professional before making investment decisions.

  • October 2 · 32 min

    What Losing Measles Elimination Status Actually Means

    What actually happens if the United States loses its measles elimination status? Almost nothing switches on. The Pan American Health Organization's commission reviews the US in November 2026, and PAHO's own page says there are no formal penalties. School exclusions follow exposures and outbreaks, your insurance follows the CDC's vaccine advisory committee, and travel advice was already keyed to outbreaks. What a loss actually starts is paperwork and a clock. The status works like a thermometer. It doesn't measure how many of us are vaccinated; it reads whether one measles lineage has kept passing from person to person for a year, which is really a reading on whether the places where too few of us are vaccinated are connected. This episode walks through the evidence fight over one dominant lineage and whole-genome sequencing, the trap inside the government's own "separate introductions" defense, and the math of why a 92% kindergarten vaccination rate can still fail once kids cluster by school. Plus Pennsylvania's outbreak from the inside, families on both sides of the vaccine decision in their own words, what you can check for your own family, Washington's split screen on the MMR shot, and Canada, one year ahead of us on the same virus, as a preview of how slow the way back is. Three dated calls at the end. Figures as of late September 2026: 3,659 US cases this year, per CDC, the most since 1991; Pennsylvania counted five measles-associated deaths as of September 30th. RELATED EPISODES The Amplifier: How a Cruise Ship Spread Hantavirus to 9 Countries — the same shape: an ordinary virus made dangerous by the connections between the people carrying it. Why the Cyclospora Recall Came 25 Days Too Late — what gets counted depends on who's counting, and how. CHAPTERS 00:00 What actually happens if we lose it? 01:45 What "eliminated" means 02:35 Who decides 03:04 The calendar that can't move 04:43 What changes for you: the ledger 05:51 The rule that runs backwards 06:40 Canada and Britain: what a loss changed 08:14 Why the reading is so hard to take 09:01 Why the standard test can't tell chains apart 11:08 The trap in the government's defense 13:39 Why 92% can still fail 15:50 The pockets, and your kid's school 16:56 Pennsylvania from the inside 17:34 Who counts the deaths 20:47 The families on both sides 22:03 What you can check today 22:44 Washington's split screen 24:02 Did anyone ask for the delay? 25:01 The exemptions fight 26:55 Canada, one year ahead 27:47 How long the way back takes 29:59 Three dated calls 31:36 The thermometer, and your kid's room SOURCES PAHO's re-verification commission (the 12-month rule; 'no formal penalties'): https://www.paho.org/en/measles-rubella-elimination-regional-monitoring-re-verification-commission PAHO on the US review moving to November (Mar 2, 2026): https://www.paho.org/en/news/2-3-2026-update-review-measles-elimination-status CDC measles cases and outbreaks: https://www.cdc.gov/measles/data-research/index.html Pennsylvania DOH, fifth measles-associated death (Sept 30, 2026): https://www.pa.gov/agencies/health/newsroom/pa-department-of-health-confirms-fifth-measles-associated-death Noel Brewer in STAT (Aug 6, 2026): https://www.statnews.com/2026/08/06/measles-elimination-status-2026-paho/ Texas DSHS whole-genome analysis, Emerging Infectious Diseases (2026): https://wwwnc.cdc.gov/eid/article/32/10/26-0494_article Glasser et al., school clustering model, Lancet Infectious Diseases (2016): https://pubmed.ncbi.nlm.nih.gov/26852723/ Executive Order 14420, Federal Register (Aug 14, 2026): https://www.federalregister.gov/documents/2026/08/14/2026-16730/delivering-gold-standard-childhood-vaccine-recommendations-for-americans

  • September 28 · 25 min

    Why Driscoll's Strawberries Are So Good (And Why It Keeps Getting Sued)

    Why are Driscoll's strawberries so good? In June 2026 the company's then-CEO put the number on the record: about 125,000 new strawberry varieties a year, for one or two worth sending out eight years later. The winning plants are never sold. About a thousand independent growers rent them for the season, farm the land, hire the pickers and carry the crop, and the company grades the fruit and sells it under one name. The seed turns out to be harder to own than it looks. A federal judge ruled that Driscoll's berries turning up in the family tree of a rival's crop didn't support "even an inference" of infringement, in part because nobody had recorded which plant was the mother; the case ended in a private settlement. In China, one of Driscoll's main blueberries turned out to be rented from an Australian breeder, which has since licensed its genetics to Driscoll's rivals. What nobody can clone is the name on the box, and since mid-June 2026 five lawsuits have been filed against the company behind it. As of late September 2026, all of them are allegations, and none has been decided. The three consumer class actions each name Driscoll's alone, never a grower. Then the people in the middle: the pickers, the wage math across the border, and a boycott that learned to aim at the name. Plus the founding family's return to the top job, Driscoll's own case for its model, a tariff fight over winter strawberries, two dated calls and one thing to watch on your own clamshell. RELATED EPISODES How Avocado Oil Fraud Actually Works — the same liability shape: when a label is challenged, the suit lands on the brand on the label, not the supplier behind it. Why Amazon Is Really Shutting Down Mechanical Turk — the people doing the work, behind a famous name. CHAPTERS 00:00 125,000 new varieties a year 02:12 The box in your fridge 02:45 Who decides how a berry tastes 03:37 Rented, never sold 05:41 Can anyone own a strawberry? 07:22 Which plant was the mother? 09:33 The rented blueberry 11:05 When the plants walked off the farms 12:23 Licensed to Driscoll's rivals 13:47 The name on the box 15:58 A 2017 quote: "sort of like our manufacturing plants" 16:25 The whistleblower's allegations 17:02 Pressure near headquarters 18:49 What your berry box can tell you 19:22 The people in the middle 20:18 Eight hours or one 20:43 The boycott that aimed at the name 22:23 The founding family's return 23:38 Tariffs and your winter berries 24:25 Two dated calls, one thing to watch 25:09 So why are they so good? SOURCES Summary judgment order, Driscoll's v. California Berry Cultivars (E.D. Cal., Mar 25, 2025): https://www.govinfo.gov/content/pkg/USCOURTS-caed-2_19-cv-00493/pdf/USCOURTS-caed-2_19-cv-00493-18.pdf Berlinger v. Driscoll's complaint (N.D. Cal., Jun 26, 2026): https://storage.courtlistener.com/recap/gov.uscourts.cand.473105/gov.uscourts.cand.473105.1.0_1.pdf Driscoll's false-advertising complaint against an app company (N.D. Cal., Jun 9, 2026): https://storage.courtlistener.com/recap/gov.uscourts.cand.471840/gov.uscourts.cand.471840.1.0.pdf Wall Street Journal on Driscoll's in China (Sept 14, 2026): https://www.wsj.com/business/driscolls-berry-farms-china-competition-b0b49f04 Soren Bjorn at WSJ Events (Jun 2, 2026): https://www.youtube.com/watch?v=av6Q5gLtMYE Commerce preliminary determination, fresh winter strawberries from Mexico (Aug 21, 2026): https://www.federalregister.gov/documents/2026/08/21/2026-17121/fresh-winter-strawberries-from-mexico-preliminary-affirmative-determination-of-sales-at-less-than Truthout, Felimon Piñeda on the Sakuma boycott (Aug 31, 2016): https://truthout.org/articles/no-other-way-than-to-struggle-the-farmworker-led-boycott-of-driscoll-s-berries/ Pajaronian on the CEO succession (Aug 25, 2026): https://pajaronian.com/driscolls-names-new-ceo/

  • September 26 · 36 min

    Why NVIDIA Paid 86 Times Revenue for Hugging Face

    $12.93 billion. Jensen Huang wrote the price for Hugging Face out to the dollar. Reports put the company's annualized revenue around $100 to $150 million, so at the top of that range NVIDIA agreed to pay roughly 86 times revenue. Nobody pays that just to own the revenue. So what did it buy? This episode reads three sets of documents as one system. Seventeen days before announcing the deal, NVIDIA filed an 8-K putting up to $105 billion of guaranty behind OpenAI's leases on an Ohio data-center campus. The clause that matters most is the trigger: NVIDIA pays only the shortfall left after the tenant's deposit runs out and the buildings are re-let or sold. A third leg sits in Texas, where press reports place Anthropic's $35 billion compute deal on a campus they say NVIDIA leases. No company involved has confirmed it, so the episode treats it as reported, not established. Disclosure: Deep Dive is made with Claude, which is built by Anthropic. Then the whole map in NVIDIA's own filings: $108.5 billion of maximum guarantee exposure, a finance chief who says the company gets paid twice, the strongest version of the bull case, and a rhyme from 2001, when Lucent learned what vendor financing costs, and Cisco what supply commitments cost, once demand fell away. Plus who actually reviews a deal this size (as of September 22nd, 2026, no American regulator has opened anything public), and the narrow claim the episode ends on: Hugging Face doesn't re-let four gigawatts, but a shortfall is smaller when the seller already knows who is buying. Then three things to do, and three dated calls. RELATED EPISODES NVIDIA Just Forecast $91 Billion Without China — the earlier NVIDIA-OpenAI thread this episode picks up. How an AI Agent Hacked Hugging Face — the breach in the background of Hugging Face's deal with NVIDIA. The AI Chip War: Why Everyone's Watching the Wrong Fight — the supply-chain map the three legs sit on. CHAPTERS 00:00 $12.93 billion, written to the dollar 01:08 Three documents, two of them 17 days apart 01:48 What Hugging Face actually is 03:28 The founder who let every rival invest 05:15 The case that open weights can't be owned 05:57 What the merger promise leaves out 06:59 The honest math on 86 times revenue 07:57 Ohio: the up-to-$105 billion guaranty 09:25 How the guaranty actually pays 12:50 How the accountants carry it 13:40 The scale, in NVIDIA's own numbers 14:13 The case for NVIDIA 16:10 "Is that circular?" 17:22 Texas: the leg nobody confirms 21:15 The whole map on one page 24:44 The 2001 rhyme: Lucent and Cisco 27:04 Who reviews any of this? 32:07 The sales desk in front of the campus 34:11 What you can do, and three dated calls 35:55 The door is still open SOURCES NVIDIA 8-K, Hugging Face merger agreement, Sept 2, 2026: https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000078/nvda-20260902.htm NVIDIA announcement: https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/ NVIDIA 8-K, Ohio guaranty, Aug 17, 2026: https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000069/nvda-20260817.htm NVIDIA 10-Q, quarter ended July 26, 2026: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm SB Energy S-1/A No. 2, Sept 21, 2026: https://www.sec.gov/Archives/edgar/data/2133037/000162828026062846/sbenergy-sx1a2.htm Hut 8 Q2 2026 results (Beacon Point): https://www.sec.gov/Archives/edgar/data/1964789/000110465926090041/tm2621890d1_ex99-1.htm Global Competition Review, Sept 8, 2026: https://globalcompetitionreview.com/article/nvidia-yet-finalise-eu-notification-strategy-hugging-face-tie ——— This episode is for educational and informational purposes only and does not constitute financial, investment, or trading advice, nor a recommendation to buy or sell any security. Deep Dive is not a registered investment adviser and holds no positions in the companies discussed. All investing involves risk. Consult a licensed financial professional before making investment decisions.

  • September 21 · 32 min

    How Anthropic's Security Team Tracks Anti-AI Activists

    Disclosure first: this show is made with Claude, built by the company at the center of this story. Anthropic has no involvement in it, and the voice is a text-to-speech model from another vendor. In September 2026 The American Prospect reported that Anthropic is building a system to monitor anti-AI activists. Anthropic answered once, to CNET: it is not, and has never said it is, building a predictive surveillance system to monitor activists; this is a physical-security function, the same as any large company. This episode reads the documents instead of the adjectives. Start with two requests for the same three facts. A Chinese state security bureau asked Claude for a protest's gathering point, route and end time; Anthropic banned the account and filed the case in the surveillance chapter of its own threat report. Three months earlier, at a sponsored webinar the coverage called a podcast, Anthropic's managers described a vendor warning them that a protest near a traveling executive had moved up an hour. Same information, two different files. What is public: that webinar, two job postings, one police report. The August 18th posting asks an analyst to track "geopolitical instability, terrorism, crime, activism." It came down August 25th, fifteen days before the story ran, and the word was never removed. Ordinary for the trade. What is unusual is the brand. The tension is not a legal contradiction. The Usage Policy binds people who use Claude; Anthropic's own security team lives under other documents entirely. The fourth column, covering the vendor watching the sidewalk, is blank. Nothing public says what may be collected, on whom, for how long, or who decides. The reasons are real and dated — a co-founder expelled by his own group, an attempted-murder charge in April, a man in the lobby with an envelope five days later, an AR-15 message in August. So is the precedent: Amazon, Facebook's be-on-the-lookout list, TigerSwan, Monsanto. Six cases, two fines — both for paperwork. And the ledger nobody prints: Meta's proxy puts Mark Zuckerberg's security program at $25,125,903. The guards at the door work for a contractor at about $22 an hour. The hunger striker's letter to Dario Amodei reached the company through one of them. Plus what California law lets you ask for, in 45 days, free — and three dated calls. RELATED EPISODES AI Deanonymization: How Claude Identifies Writers from 125 Words What's Inside Anthropic's 186-Page Risk Report The Regulation Anthropic Asked For CHAPTERS 00:00 Two requests, one route 01:29 The word in the posting is activism 01:55 Disclosure, and how this episode is bounded 02:42 Anthropic's only answer 03:21 The documents everyone calls a leak 05:12 The two postings, read in full 07:27 The one police referral we can read 09:35 Four dated reasons the file exists 11:34 How big the threat picture actually is 13:31 Three rulebooks, and a blank fourth 14:52 Who each rulebook actually binds 16:45 Anthropic's own words about mass data 19:16 This has been written six times before 22:02 The ledger: two fines, both for paperwork 23:22 Two ends of one lobby 24:40 The office emptied before the vote 26:54 What the law lets you do about it 29:41 Our ledger, and three dated calls 31:08 The missing page SOURCES The American Prospect, Sept 9, 2026: https://prospect.org/2026/09/09/anthropic-artificial-intelligence-surveillance-system-monitor-activists/ CNET — Anthropic's only statement: https://www.cnet.com/tech/services-and-software/have-you-protested-ai-recently-anthropic-may-be-watching-you-for-precrimes/ Anthropic threat report, Sept 10, 2026 (vendor on page 84): https://www-cdn.anthropic.com/e50be2e51e7695dc4b1366a37a245a597377d3b5/Anthropic-Detecting-and-countering-091026.pdf The Aug 18, 2026 posting: https://jobs.accel.com/companies/anthropic/jobs/90264650-enterprise-intelligence-specialist SF Standard — the August 14th message and police referral: https://sfstandard.com/2026/09/04/anthropic-threat-claude-sfpd/

  • September 12 · 26 min

    How Avocado Oil Fraud Actually Works

    Two bags of chips, same company, same shelf, same potato. One is fried in canola; the other says avocado oil, costs up to $1.40 more ounce for ounce, and holds slightly less oil. That oil costs the maker about forty cents; a full swap to soybean would save about thirty-five. In July 2026 a UC Davis lab tested 54 samples of chips, mayonnaise and dressings like that bag. The study found 48 inconsistent with avocado oil; olive-oil foods from the same shelves passed 19 of 20. The internet asked which brands lied; this episode asks who was ever supposed to check. Fraud here means what food scientists mean: cheaper oil sold as expensive oil between grove and pantry. No brand named in the episode has been found to have done it; the lab says the fault likely sits upstream, with suppliers. The answer is plumbing. "Avocado oil" appears zero times in Title 21 of the Code of Federal Regulations; the only standard, adopted in November 2024, binds no one. The checking fell to one university lab on its own money, a state attorney general borrowing a consumer statute because there is no product standard, and a class-action bar whose model case ended with the court ordering the plaintiffs' lawyers to pay $623,738.70, an order now stayed. Ten suits followed in two months; the public complaints rest on the paper by sample number, none on a test of their own — allegations, not findings. Olive oil lived this in 2010, same lab, and built its way out: a USDA grade, a trade group that tests and sues, a seal with a published list, and California's mandatory lot sampling, which the same scientist helped draft. Avocado oil has none of them, at nearly the same incentive per kilo. Plus what "we test every batch" actually tests, three moves for your cart, and three dated calls. RELATED EPISODES Why the Cyclospora Recall Came 25 Days Too Late — the same shape: the surveillance system, not the lettuce, is the story. How Much Ozempic Actually Costs — a sticker price taken apart to see what it actually pays for. Minamata: How a Company Poisoned a Town and Japan Helped Cover It Up — where certification decided who counted; here none exists. CHAPTERS 00:00 Two bags, one shelf 01:06 One definition, then the roadmap 02:12 Inside the lab: 48 of 54 04:57 What the test can't say 06:24 The arithmetic of one bag 07:54 Mayonnaise changes the stakes 09:15 Federal food law: zero results 10:21 Three checkers, no food regulator 11:21 The model case: $623,738.70 13:13 Ten suits, one paper 15:20 Olive oil already lived this 18:15 The test a blend can pass 18:58 What the brands say they test 22:04 The creators who vouched 23:07 Why the word is worth money 24:41 Three moves for your cart 25:18 The ledger: three dated calls 26:20 Two bags, a dollar forty apart SOURCES UC Davis — the July 2026 study: https://www.ucdavis.edu/food/news/avocado-oil-chip-youre-eating-may-not-be-made-pure-avocado-oil eCFR — Title 21 search for "avocado oil", zero results: https://www.ecfr.gov/api/search/v1/results?query=%22avocado+oil%22&hierarchy%5Btitle%5D=21 Codex — avocado-oil standard adopted, Nov 2024: https://www.fao.org/newsroom/detail/fao-who-codex-alimentarius-commission-adopts-new-standards-47session/en Texas Attorney General — Aug 20, 2026 release: https://www.texasattorneygeneral.gov/news/releases/attorney-general-ken-paxton-investigates-avocado-oil-products-deceptive-marketing-and-potential-use CourtListener — Golikov v. Walmart docket: https://www.courtlistener.com/docket/69190987/eddie-golikov-v-walmart-inc/ Ninth Circuit — No. 26-364 memorandum, Sept 10, 2026: https://cdn.ca9.uscourts.gov/datastore/memoranda/2026/09/10/26-364.pdf USDA — olive-oil grade standards, 2010: https://www.federalregister.gov/documents/2010/04/28/2010-9866/united-states-standards-for-grades-of-olive-oil-and-olive-pomace-oil Olive Oil Commission of California — mandatory testing: https://www.oliveoilcommission.org/uploads/1/2/5/0/125053236/whole_report.pdf

  • September 1 · 22 min

    Why Amazon Is Really Shutting Down Mechanical Turk

    In 2018, on stage at Amazon's biggest conference, Mechanical Turk's product lead pitched "over 500,000 active workers." That same year, researchers measured how many people were actually working at any moment: about 2,000. The 500,000 was a registration count frozen since January 2011 — and Amazon's own documentation still advertised it the day the shutdown banner appeared. On September 30, 2026, the platform closes for good, and the obituaries all say AI killed it while Amazon exits human data. This episode reads the paperwork and finds both claims wrong. What actually happened: MTurk was 1 of 10 SageMaker features cut the same day — a portfolio cleanup, not a reckoning. Ground Truth and Augmented AI are frozen, not closed, and Bedrock still sells AWS-managed human evaluators. Amazon isn't exiting human data; it's exiting a price point — the same move the whole industry made, from Surge's reported $1.2B to Mercor's $112-an-hour experts, with a $1.32-an-hour floor in Nairobi that never went away. Also inside: the myth that needs retiring — crowds of Turkers did NOT train ChatGPT (OpenAI hired ~40 contractors and said so in writing), but roughly 80% of the original human-feedback data behind Claude traces to master-qualified Turkers, about 20 prolific workers — disclosed on air, since this show is made with Claude. The study that caught 33–46% of crowd workers using AI on the work meant to certify humanity (honestly bounded: one task, 46 summaries). The wage ledger: requesters paid $11.58 an hour; the median worker kept $1.77. And the shutdown's one mercy — every unreviewed task auto-approves, resolving the workers' oldest fight in their favor, posthumously. Two dated calls close it out, plus the practical note: worker transaction histories go dark January 28, 2027 — download yours before then. CHAPTERS 00:00 The 500,000-worker trap 02:00 Two waves, ten cuts — the death certificate 03:12 Frozen, not closed: what actually survives 04:01 The industry moves upmarket 05:25 A dedup chore becomes a labor market 06:07 Vienna, 1770 — the original hoax 07:00 ImageNet's 50,000: the monument 08:03 The ChatGPT myth retires 08:38 Disclosure: the 20 Turkers inside Claude 09:50 The crowd replaces itself 11:30 The label breaks, not the models 12:12 The wage ledger 14:41 The people: Marshall to Pawloski 16:59 Every exit has a gate 18:01 The posthumous mercy 19:14 The second correction — and what to do 20:29 The ledger: two calls, one watch 21:15 The machine was people SOURCES mturk.com — closure banner + help pages (auto-approval, Jan 28 2027 data cutoff): https://www.mturk.com/ AWS SageMaker features — the ten-feature cutoff: https://aws.amazon.com/sagemaker/ai/features/ AWS Ground Truth docs (maintenance mode; the "500,000" line): https://docs.aws.amazon.com/sagemaker/latest/dg/sms.html AWS Bedrock Evaluations (managed human evaluators): https://aws.amazon.com/bedrock/evaluations/ Difallah, Filatova & Ipeirotis, WSDM '18 (~2,000 active): https://dl.acm.org/doi/10.1145/3159652.3159661 Hara et al., CHI 2018 — the wage study: https://arxiv.org/abs/1712.05796 ILO, Digital Labour Platforms (2018): https://www.ilo.org/publications/digital-labour-platforms-and-future-work-towards-decent-work-online-world Veselovsky et al., EPFL 2023 — LLM use on MTurk: https://arxiv.org/abs/2306.07899 Bai et al. 2022 — Anthropic HH-RLHF (§2.1): https://arxiv.org/abs/2204.05862 Stiennon et al. 2020 — "rather than using… Mechanical Turk": https://arxiv.org/abs/2009.01325 Rest of World — the Remotasks precedent: https://restofworld.org/2024/scale-ai-remotasks-banned-workers/ CloudResearch — sunsetting the MTurk Toolkit: https://www.cloudresearch.com/resources/blog/sunsetting-mturk-toolkit-cloudresearch/ US Patent 7,197,459 B1: https://patents.google.com/patent/US7197459B1/en

  • August 29 · 21 min

    What's Inside Anthropic's 186-Page Risk Report

    For close to a year, the bioweapon filters at one of the world's strongest AI companies had a blind spot. One internal flag had switched off both the blocking and the logging for an entire pipeline — roughly 50,000 hired contractors, 133 million exchanges. No regulator caught it. No law required anyone to tell you. You know because Anthropic wrote it down, inside a 186-page risk report released with no announcement, no author's name, and a title that literally says Redacted. This episode reads the whole document — and asks the question it leaves behind: can a confession be a system? What's inside: the re-scan of a year of unwatched conversations — 1,197 flagged, 62 real after the company's own red-team traffic is removed, zero misuse found — and why that answer is careful, not clean. The frontier model Anthropic finished and won't release, stronger than its flagship, working only inside the building — and the 2019 GPT-2 withholding playbook it repeats, written in part by the people who later founded Anthropic. The strangest audit ever run: Anthropic handed the misalignment chapter to Claude itself, which led with its own conflict of interest, called the report candid and largely faithful, and got two criticisms into the final text. The case against gets a full act: passing thresholds rewritten mid-stream, an oversight trust whose outside-review power has never been used — and whose trustee left to become Anthropic's Chief Global Affairs Officer — and the analyst who read all 186 pages and wasn't reassured. Then the machinery question: aviation buys pilots' honesty with immunity and a ten-day clock; every AI reporting law in force routes truth to agencies confidentially. Three dated calls close it out — one already landed. CHAPTERS 00:00 The year the filters were off 01:15 Disclosure: this show runs on the subject's AI 02:16 A 186-page confession nobody signed 03:01 One flag — the brake and the camera 04:21 The re-scan: 1,197 flags, 62 real 05:38 Anthropic re-grades itself 06:15 The model they won't release 07:16 Why hold a frontier model back 09:10 The 2019 playbook, repeated 10:22 The strangest audit ever run 11:52 Seven more confessions 13:04 The case against candor 15:14 The machinery that buys honesty 17:26 Three holdbacks in a fortnight 17:58 Capture — and the trustee who switched sides 19:31 The ledger: three dated calls 20:22 Nobody had to tell you SOURCES Anthropic — Redacted Risk Report, August 2026 (PDF): https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdf Anthropic — RSP updates: https://www.anthropic.com/rsp-updates Zvi Mowshowitz — On the Anthropic Risk Report: https://thezvi.substack.com/p/anthropic-risk-report-august-2026 OpenAI — GPT-2 Release Strategies (2019): https://arxiv.org/abs/1908.09203 Reuters — Anthropic's letter alleging Alibaba extraction: https://www.reuters.com/world/china/anthropic-says-alibaba-illicitly-extracted-claude-ai-model-capabilities-2026-06-24/ TechCrunch — OpenAI slows Astra over security concerns: https://techcrunch.com/2026/08/07/openai-says-it-slowed-astra-model-development-over-security-concerns/ Unite.AI — Z.ai's GLM-5.3 cyber capability: https://www.unite.ai/z-ai-launches-glm-5-3-with-frontier-coding-and-a-cyber-capability-that-outgrew-its-training/ Anthropic — Tino Cuéllar joins: https://www.anthropic.com/news/tino-cuellar NASA ASRS — immunity provisions: https://asrs.arc.nasa.gov/overview/immunity.html NY RAISE Act (S6953B): https://www.nysenate.gov/legislation/bills/2025/S6953/amendment/B EU AI Act, Article 55: https://artificialintelligenceact.eu/article/55/

  • August 26 · 26 min

    Why Japan Just Made It Harder to Stay

    On August 4th, Japan's Justice Minister announced three documents in one press conference: permanent residency gets an income test above the average Japanese household — a bar 61.5% of Japanese households don't reach — it becomes revocable for the 950,000 people who already hold it, and the top worker visa gets longer renewals so the labor keeps coming. Five weeks earlier, this show argued Japan runs a rotation machine: labor in, people out, and never the word immigration. This episode is what happened next — and it turns out both jaws were ordered by one cabinet document signed in January. The mechanics nobody else has connected: the income standard reaches back to applications already in the queue; the fee rises 20x to ¥200,000; the pension test can't be beaten by earning more — while the refund counter Japan runs for departing workers was widened, twice, to match its visa lengths. The trainee program's new transfer locks were set at the maximum for construction — the sector with 25% of trainees and 55% of last year's disappearances — with "retention in rural areas" in the official rationale. And the record: 720,988 arrivals in the tightening year. The strongest case for the package gets a full act: the real compliance gap, the viral number the health minister himself debunked, Korea's stricter bar, the UK converging on the same design — and America's own income-weighted green card, which courts killed in two years. Japan's version has no such brake: the Supreme Court settled that in 1978. Three dated predictions close it out, and one on-air correction to our own last Japan episode. CHAPTERS 00:00 The August 4th presser 01:41 One document, two jaws 03:20 The permanence jaw 05:34 The retroactivity clause 07:10 The arithmetic of the bar 09:07 The refund counter 10:36 The other jaw — intake 12:06 The transfer locks 14:39 The people in the queue 17:27 The strongest case for it 21:12 The 50% correction 22:28 What do you call this? 25:12 Three dated predictions 26:02 The word Japan chose SOURCES Cabinet Secretariat — Comprehensive Measures for Acceptance and Orderly Coexistence (Jan 23, 2026): https://www.cas.go.jp/jp/seisakukaigi/gaikokujinzai/pdf/kettei_sougoutekitaiousaku_honbun.pdf ISA draft PR guideline + revocation guideline (e-Gov case 315000140): https://www.moj.go.jp/isa/11_00112.html Justice Minister press conference, Aug 4, 2026 (MOJ): https://www.moj.go.jp/hisho/kouhou/hisho08_00905.html MHLW Comprehensive Survey of Living Conditions 2025 — household income: https://www.mhlw.go.jp/toukei/saikin/hw/k-tyosa/k-tyosa25/dl/03.pdf MIC Basic Resident Register, Jan 1, 2026: https://www.soumu.go.jp/menu_news/s-news/01gyosei02_02000428.html ISA 2025 trainee disappearance figures: https://www.moj.go.jp/isa/content/001425159.pdf ISA Ikusei Shuro program overview (transfer-lock periods): https://www.moj.go.jp/isa/content/001452485.pdf Kyodo feature — foreigners leaving Japan may accelerate (Ray Hoe): https://japantoday.com/category/business/feature-trend-of-foreigners-leaving-japan-may-accelerate-under-tighter-visa-rules US State Dept 2025 Trafficking in Persons Report — Japan: https://www.state.gov/reports/2025-trafficking-in-persons-report/japan/

  • August 24 · 22 min

    The Ebola Outbreak the Vaccine Can't Stop

    On May 17th, 2026, the WHO declared a global health emergency over eight laboratory-confirmed cases of Ebola in eastern Congo. Eight was all the test could see. This is the story of an outbreak the standard toolkit was never built for: Bundibugyo virus, the neglected species — no licensed vaccine, no approved drugs, a field test whose own emergency-use notice says it cannot detect this virus. None of that was hidden; it was published as it happened. What this episode adds is the assembly: how a fourteen-week blind spot, a war, and a test tuned to the wrong species stack — and which of those actually did the damage, ranked the way WHO's own Director-General ranks them. Then the natural experiment: the same virus crossed into Uganda and produced twenty cases, not because the tools were better, but because the margin was. And finally the bet being placed right now — 70,000 doses of a vaccine built for a different species, split between a trial and a wager, with the confirming science funded on the same day as the recommendation. Counts are as of August 18th, 2026, and stated as a floor; everything stands as of August 20th. CHAPTERS 00:00 Twenty samples, twenty negatives — and Ebola anyway 02:19 One name, six viruses — what the toolkit actually covers 03:34 The test that was right 04:28 Why a close relative isn't close enough 06:27 Nineteen empty years — the market answer 07:23 The border experiment: Uganda, then Congo 10:11 The strongest case against the toolkit story 11:55 The alarm that fired fast 14:44 Who is carrying this 16:49 The bet: a vaccine built for the wrong virus 20:25 The record, dated — and what to watch SOURCES WHO Director-General remarks (May 20 + Aug 18, 2026) and IHR Emergency Committee statements — the declaration record WHO emergency-use notice for the GeneXpert Ebola cartridge — the footnote that says it cannot detect Bundibugyo MSF outbreak response reporting (beds, staff, admissions) CIDRAP + Health Policy Watch daily outbreak coverage (case counts, vaccine allocation, worker strikes) AP reporting on where deaths occur and the Merck/Gilead deployment Stored-blood antibody study on Zaire-vaccine cross-reactivity against Bundibugyo Africa CDC / WHO joint statements + continental response plan ($518M ask)

  • August 19 · 19 min

    Why DeepSeek Isn't Free Anymore

    On May 22nd, the top story on Hacker News — 621 points — said DeepSeek had made its price discount permanent. Permanent lasted 86 days. On August 16th every price in DeepSeek's menu went up: the floor rose 52%, the flagship's peak rate about four and a half times, and the cached-input line — the discount that makes AI coding agents affordable — twelve-fold. This is the company that broke the price of AI, throwing the race to zero into reverse. First, surgery on the headline number: who the hike is aimed at (one free coding agent pushed 8 trillion tokens through it in a single day), and the tell in the schedule — peak hours mapping exactly onto the Chinese business day, lunch break at half price, while an American developer can work a full day and never touch a peak window. You don't build surge pricing to get rich. You build it to move load you can no longer serve. Then the ledger. In June this show bet, on tape, that by the end of 2026 at least two of China's four big open-model labs would move against free. DeepSeek just cleared the 10% bar by thirty-five times — lab one of two, eight weeks after air. Why bet against free? Because free-as-a-strategy has a lifecycle — subsidize, capture, reprice — and you've lived through it: Uber riders paying 41 cents on the dollar before prices rose 92%; MoviePass executives pleading guilty to securities fraud; China's bike-share graveyard, before Beijing outlawed below-cost dumping — then subsidized the same playbook for tokens. And the trap DeepSeek built for itself: the weights are free on the internet, so one day after the hike the cheapest place to buy DeepSeek at rush hour was no longer DeepSeek. The company repriced; the model is priced by everyone. The same weekend, Stripe reportedly closed a $7 billion-plus deal for OpenRouter — the toll booth is the prize. Four dated predictions close the episode, on the record, so you can hold us to them. The price tag was always a loan. The first repayment just came due. RELATED EPISODES Why Is DeepSeek Free? — the June episode this one scores; free-as-national-strategy, and the bet now half-landed How Superforecasters Actually Predict the Future — the scorekeeping method behind this show's dated, falsifiable bets CHAPTERS 00:00 'Permanent' lasted 86 days 01:14 Surgery on the twelve-fold headline 04:18 The hidden clock in the pricing 05:16 Scoring the June bet 06:13 The lifecycle: Uber, MoviePass, bike-share 09:37 Was DeepSeek ever below cost? 11:09 The trap: the weights are still free 14:29 Why now — chips, money, Beijing 15:59 The American mirror 16:51 The scoreboard, and four new bets SOURCES DeepSeek official API pricing page (live rate map: peak/off-peak bands, cache line ×12, floor +52%) + the Aug 13 V4-Pro release note (two-sentence rationale, unsigned) Bloomberg (Aug 13) — DeepSeek raises prices multiple times over; SCMP (Aug 6) — the 'plan your usage accordingly' warning Semafor (Aug 3) — V4-Flash 50% cut, surcharge paused; Hacker News (May 22) — the 'permanent discount' story TechCrunch (Mar 2025) — DeepSeek's self-published 545% theoretical margin with self-listed caveats; OpenCode (self-reported) — the 8-trillion-token day Bloomberg/Fortune/PYMNTS (Aug 16) — Stripe reportedly finalizes $7B+ OpenRouter acquisition; OpenRouter live pricing — day-one third-party undercut of DeepSeek's peak rate Slate/Horan — Uber passengers paying ~41% of trip cost (2015); Rakuten via CNBC — rideshare prices +92% (2018–2021); CNBC/Fortune — the 2022 Khosrowshahi memo DOJ/Holland & Knight, Variety, Engadget — MoviePass executives' guilty pleas (securities fraud; the data-plus-AI claim) The China Project — Ofo's 15M refund queue (~10/day by Aug 2020); 36kr — bike-share ~4× repricing over the decade Dealroom — the leaked Liang investor call (unconfirmed by DeepSeek): ~1/20th US compute; reported ~$74B valuation DCD/Bloomberg — reported ~1GW Ulanqab build, late 2027; MLQ — GLM-5.3 weights held for cyber-capability review

  • S3 · E5
    August 14 · 29 min

    How Superforecasters Actually Predict the Future

    In the movie about the bin Laden raid, an analyst says "one hundred percent he's there." The real room gave Leon Panetta a spread — 90, 60, below 50 — and he had to make the call on a gut. Around that same time, American intelligence quietly ran a four-year tournament to find out whether anyone can actually predict the future. The winners were a pharmacist and a retired programmer with a duck for an avatar. This episode teaches the method they used, move by move: start from the base rate, break impossible questions into estimable pieces, commit to precise numbers — 63%, not "probably" — update small and often, team up and let math sharpen the crowd's answer, and keep score with Brier scores, the scoreboard the whole field runs on. Then the part most tellings skip: where it breaks. The curated questions; Taleb's rare-event attack and Tetlock's own concession; the Ukraine and Brexit misses; the AI progress the superforecasters underpriced — priced at 2.3%, resolved yes. And the 2026 twist, in two industries. A betting-market war: Kalshi reportedly chasing a $40 billion price tag while New York sues and a federal regulator invokes emergency powers to keep it open. And AI systems that reached statistical parity with the superforecaster median — by having the method typed into their prompts: "You are an expert superforecaster… pick the base rate." The superforecasters' own median forecast says AI beats them by 2028. The best calibrators on Earth, calibrated about losing. You can practice the skill free — Good Judgment Open and Metaculus run public tournaments with real scoreboards. The skill was never seeing the future. It was keeping score on yourself, before reality does it for you. RELATED EPISODES Will AI Replace Software Engineers? What the Data Actually Says — the same benchmark-asterisk test this episode runs on AI forecasting parity How an AI Agent Hacked Hugging Face — the AI-capability arc the superforecasters underpriced How GPS Actually Works — the sibling mechanism explainer: a system you use daily, taught from the inside CHAPTERS 00:00 The room that ran on gut calls 01:28 'A fair chance' — what vague words cost 03:35 The tournament the government could lose 06:17 Surgery on the famous 30% claim 08:43 The method, move by move 14:41 Where it breaks: Taleb, Ukraine, and the AI miss 18:18 The ox, the whale, and the $40 billion war 23:44 The machines type in the method 27:31 How you practice this free SOURCES Philip Tetlock & Dan Gardner, Superforecasting (2015); Tetlock, Expert Political Judgment (the 1984–2004 study) Mellers et al., Psychological Science 2014 & Perspectives on Psychological Science 2015 — tournament design, granularity (57 values; rounding test), persistence, 300-vs-60 days Goldstein et al. (MITRE, declassified) — Good Judgment vs the Intelligence Community Prediction Market: 0.15 vs 0.23 on 139 shared questions; the secrecy heuristic Friedman et al., International Studies Quarterly 2018 — 888,328 forecasts; the cost of verbal probability labels; the Bay of Pigs 'fair chance' record Chang et al. 2016 — the one-hour training effect (+6–11%); Hauenstein et al., Psychological Science 2025 — the reanalysis contesting it Forecasting Research Institute — ForecastBench (forecastbench.org) and the XPT near-term accuracy report (Sept 2025); Good Judgment's rebuttals (July 2026) Wallis, Statistical Science 2014 — Galton's 1906 ox, corrected from the archives NY Attorney General release (July 31st, 2026); The Block (Aug 11th, 2026 CFTC emergency order); WSJ 'The Economics Of' & CoinDesk market coverage On the record: Leon Panetta (Choiceology); Warren Hatch (Cassandra Forum, Aug 2026); Luana Lopes Lara (WSJ Events, May 2026); Tetlock (80,000 Hours 2019; Harvard Kennedy School, June 2026)

  • S3 · E5
    August 12 · 13 min

    How GPS Actually Works

    In 1977, engineers building the first cesium clock for orbit added a hidden switch — a hedge, in case Einstein was wrong. He wasn't. That clock became part of GPS, and the relativity correction it proved is running inside the phone in your pocket right now. Turn it off, and GPS drifts into uselessness in about two minutes. This episode teaches how GPS actually works — and why almost everything you think it is, is wrong. It isn't really a map system. It's a clock, and your location is the by-product of four clocks in the sky disagreeing by nanoseconds. We walk through the whole machine: how your phone finds you by measuring time instead of distance, why you always need a fourth satellite (one of the most elegant ideas in engineering), the two relativity effects that fight inside every satellite, and the factory trick that cancels them before launch. Then the part nobody tells you — GPS is the hidden master clock under the power grid, cell networks, and financial markets, proven the day in 2016 when a thirteen-microsecond error garbled radio across Britain. And the people who built the clock in the sky: Roger Easton, Bradford Parkinson, and Gladys West. No hype. Just how it works, and why it matters more than you think. RELATED EPISODES How LLM Inference Actually Works — the mechanism deep-dive on what happens when you hit send How RAG Actually Works — another under-the-hood 'how X actually works' How Netflix, Uber, and YouTube Handle Scale — systems, explained from first principles CHAPTERS 00:00 The switch built in case Einstein was wrong 02:01 How a stopwatch in space finds you 03:38 Why you always need a fourth satellite 06:02 The relativity that keeps GPS alive 08:16 The hidden clock under the grid, your phone, and Wall Street 10:04 The day the clock lied — the 2016 accident 11:42 Why GPS is so easy to jam 12:34 Who built the clock in the sky SOURCES gps.gov — GPS space & control segments, constellation, and the L1 signal spec (IS-GPS-200) Neil Ashby, 'Relativity and the Global Positioning System,' Physics Today (May 2002) — the +38 microseconds/day correction, the factory clock offset, and the 1977 NTS-2 test NIST — 'Putting Einstein to the Test' (the -7 / +45 microsecond relativity split) and the NIST Time & Frequency post-mortem on the January 2016 GPS timing anomaly R. Pogge (Ohio State University) — GPS & relativity lecture notes (~10 km/day drift; a fix 'false after 2 minutes') RTI International for NIST (2019) — economic benefits of GPS (~$1 billion/day outage impact) Penn State GEOG 862 — the four-unknowns / four-satellites pseudorange solution Todd Humphreys (UT Austin) — civilian GPS spoofing ('no encryption, no authentication')

Showing 1–20 of 36 episodes