Deep Dive

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.CHAPTERS00:00 Issued in July, public in October01:50 Four pages, printed as images02:58 The order's reasons: where data goes03:37 Seven public sources, one app audit04:34 App versus model, and a stray footnote05:23 The quiet 2018 law behind it06:36 The heavier tools nobody used07:19 Kaspersky's seven-year ladder08:04 A disclosure, and the Anthropic case09:23 Two words: developed using10:11 Read literally: Amazon, Microsoft, NVIDIA, vLLM11:14 One model, 29 companies serving it11:53 The line earlier bans drew12:41 NIST's test of the downloaded model13:24 CrowdStrike and the Tibet prompt14:26 The case for the broad reach15:04 The case against it15:45 A release on the day of the order16:12 How would anyone enforce it?17:41 Who pays: the Huawei precedent18:09 Who in government was using DeepSeek?18:54 Beijing's answer19:40 China's draft export list20:08 The line Washington walked away from21:15 Response or coincidence?21:58 What it means for you22:51 Two calls, with dates23:25 Where the two words endSOURCESThe order, Federal Register, Oct 20, 2026: https://www.govinfo.gov/content/pkg/FR-2026-10-20/pdf/2026-21447.pdf41 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.htmLawfare on Anthropic's rehearing petition: https://www.lawfaremedia.org/article/why-anthropic-is-asking-for-a-rehearing-en-bancNational Law Review on the D.C. Circuit ruling: https://natlawreview.com/article/dc-circuit-upholds-anthropic-fascsa-designation-and-endorses-broad-interpretationNIST CAISI evaluation of DeepSeek: https://www.nist.gov/news-events/news/2025/09/caisi-evaluation-deepseek-ai-models-finds-shortcomings-and-risksCrowdStrike 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/5304010Global Times on China's draft list: https://www.globaltimes.cn/page/202610/1372405.shtml

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.CHAPTERS00:00 Same request, a day later01:47 What an LLM judge is02:48 Judging LLM-as-a-Judge, 202304:40 Four jobs, starting with what models read05:35 The launch headline06:34 The leaderboard that was supposed to be human08:00 When the automated checks said fine09:08 One prompt, a thousand times09:50 Why the batch changes the arithmetic11:30 A correction on temperature zero12:00 Even zero isn't a promise12:42 Same name, different behavior13:18 Anthropic's August 2025 bugs15:45 The audit, limits first17:27 A grader that never passed19:16 Close contests at the top19:56 78 in 10020:39 Four providers and the self-hosted fix22:02 The sharpest criticism is in the study22:22 Validity versus reliability23:14 The cheap fixes24:30 Who's exposed26:01 What deterministic serving costs26:47 What you can do27:37 The grader was never a rulerSOURCESZhu & Zhang, arXiv, Sep 3, 2026: https://arxiv.org/abs/2609.04198Thinking 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/2306.05685Anthropic postmortem, Sep 17, 2025: https://www.anthropic.com/engineering/a-postmortem-of-three-recent-issuesOpenAI on GPT-4o sycophancy, May 2, 2025: https://openai.com/index/expanding-on-sycophancy/Arena, AutoEval, Jul 30, 2026: https://arena.ai/blog/autoeval-scoresArena, judge study, Sep 30, 2026: https://arena.ai/blog/llm-judge-self-preferenceChen, Zaharia & Zou, 2023: https://arxiv.org/abs/2307.09009OpenAI API deprecations: https://developers.openai.com/api/docs/deprecations

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.CHAPTERS00:00 Fifteen thousand dollars01:04 A leak, four newsrooms, and the denials01:50 Your pension fund has an auditor02:34 The pile of paper03:46 What backdating actually involves04:21 Scheme one: making $2 billion disappear08:04 Scheme two: the château11:49 Scheme three: pick a company, any company14:41 Scheme four: the signature16:25 The 2008 case17:42 Why didn't the audit catch it?18:42 Is a backdated document always improper?20:01 The hard points20:44 What the standards assume22:28 The 2005 split23:45 Altered papers, the other way round26:33 Who could act: Cyprus29:26 Britain, Jersey, and the Chelsea money32:50 Who you'd even sue33:56 What the signature promises you34:57 What you can do35:32 What to watch36:27 Three predictions37:09 One thing to keepSOURCESOCCRP, 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-transparencyThe Bureau of Investigative Journalism, Oct 12, 2026: https://www.thebureauinvestigates.com/stories/2026-10-12/how-deloitte-helped-abramovich-hide-10bnBBC Panorama, Oct 12, 2026: https://www.bbc.com/news/articles/c64g1l5389kvoThe Guardian, Oct 12, 2026: https://www.theguardian.com/world/2026/oct/12/roman-abramovich-aides-paperwork-tax-records-suggestICIJ, 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/2008/2613Cyprus Mail, Oct 15, 2026: https://cyprus-mail.com/2026/10/15/audit-watchdog-investigating-deloitte-cyprus-run-by-a-former-manager

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.CHAPTERS00:00 Disclaimer00:08 Wrong every time00:58 Three questions01:29 The toll road every token pays02:37 The model is the compute03:57 Built on 2019-era silicon04:27 The company's own numbers05:46 The precision catch06:11 The model is a print run07:22 One model in silicon so far08:04 Compute, priced like payroll09:07 Token prices fell, margins didn't10:36 Rent versus mortgage12:25 The three-month line13:39 The man who built what he'd ruled out16:45 Eight months, four moves18:21 Where specialization stops20:14 The bet on trial22:00 The bull case and Bitcoin23:17 Science or infrastructure?24:41 Four predictions25:33 Your own tell to watchSOURCESAMD 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-marketTaalas 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/4092140Bajic on TBPN, Feb 20, 2026 (3-month breakeven, on camera): https://www.youtube.com/watch?v=UrebAsni59kBajic at MIT, 2022 ('wrong every time'): https://www.youtube.com/watch?v=L4AgmG8V3LESundar 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.

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 EPISODESAI and Dogs: The Technology That Could Give Your Dog More Years — the longevity-drug side of the same question, tested in dogsCHAPTERS00:00 Your cells forget what they are01:12 What actually came out02:13 The 90% number03:37 Why this isn't Sinclair's theory04:19 Waddington's valleys05:34 Identity is reversible06:27 Horvath's clock07:35 The damage camp09:23 The contested experiment10:29 Four theories, four letdowns11:24 The mouse record12:32 Closer than the headlines say13:28 Where the wear and tear ends up14:24 The counterweight15:06 Follow the money16:23 The track record17:09 The first human test18:48 The test you can buy20:21 What to ask before you buy21:42 Each host's bottom lineSOURCESNature review (Yücel, Molière, Gladyshev), Sept 30, 2026: https://doi.org/10.1038/s41586-026-10955-0Cell review, Altos Labs (Lu, Tu, Izpisua Belmonte), May 2026: https://doi.org/10.1016/j.cell.2026.04.020The 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/s43587-022-00248-2Life 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