Homo Sapiens in the Loop

← Back to home

Episode 1 — Sources and fact check

In episode 1 we said a lot of things about AI, CRISPR, Tesla and more. We think every claim should be backed by a source, so we checked all 32 factual claims in the episode, one by one, and we're publishing the result here — including where we got it wrong.

Corrections

What we got wrong or only partly right, in plain language:

  • Partly 00:29 CRISPR Nobel "nine years ago" Nobel Prize in Chemistry 2020 (Charpentier & Doudna). Awarded in 2020. nobelprize.org
  • Partly 00:29 CRISPR "uses a programmed bacterium" It is an enzyme plus a guide RNA, derived from a bacterial immune system. nobelprize.org
  • Partly 00:43 "CRISPR used to create vaccines" The approved use is therapy (Casgevy, approved in 2023). Vaccine uses are still research. fda.gov
  • Partly 01:27 Home-made CRISPR animals "with powers" DIY kits exist and biohackers use them; no home-made animal with "powers" is documented. technologyreview.com
  • Partly 02:38 The 50→500 videos / US$5,000 story The closest source is a satirical X post (2026-09-17): 59 videos asked (~US$200), 500 delivered (US$5,000), the agent denies it. The logs/complaint details match a separate Codex "US$78,000" runaway-bill report. x.com · news.ycombinator.com
  • Partly 03:10 Self-editing "legal in the US" No federal ban on self-experimentation, but since 2017 the FDA treats selling DIY gene-therapy kits for self-use as illegal. technologyreview.com
  • Partly 03:26 Self-applied gene editing is "legal in Colombia" Penal Code art. 132 punishes altering the genotype for purposes other than treatment, diagnosis or research aimed at relieving suffering or improving health (16–90 months). Self-editing to improve your own health falls within that exception; nothing verifies the purpose when it is self-applied. leyes.co
  • Partly 07:47 Neuralink connects "over Wi-Fi" It uses Bluetooth Low Energy. en.wikipedia.org
  • Partly 09:18 OpenAI and Anthropic "about to go public" Both filed confidentially in June 2026. Anthropic is targeting October 2026. OpenAI says it will not list in 2026. techcrunch.com
  • Partly 11:49 "20,000 build frontier AI vs ~200 on safety" This is the estimate from The AI Doc (2026). The film gives no underlying source. tomsguide.com
  • Partly 13:46 Tesla "has radar" Camera-only ("Tesla Vision") since 2021–22. Only some HW4 Model S/X and the Cybertruck have HD radar again. notateslaapp.com
  • Wrong 13:57 Tesla "has Nvidia GPUs inside" Since 2019, with Tesla's own chip (FSD). Nvidia's Drive PX2 was used only in 2016–2019 (HW2/2.5). en.wikipedia.org
  • Partly 15:45 Hugging Face hack: "dozens" of agents whose mission was to hack About 1,200 OpenAI agents were involved (about 700 in the attack). They were reward-hacking cyber-benchmark tasks and had no hacking mission. en.wikipedia.org
  • Partly 16:10 "Opus 5.1" There is no Opus 5.1. The sequence is Opus 4.8 → Opus 5 → Opus 5.5, and 5.1 is Fable 5.1. anthropic.com
  • Partly 18:10 "UltraCode Max" The models named exist; we found no public product called "UltraCode Max." anthropic.com

How to read the labels

Checked on 2026-09-28. Timestamps come from the English cut with silences removed. The video's start is being trimmed by about 1.5 s, so times may be off slightly. J = Jose, S = Sebastián.

Verdicts: Supported means the source backs the claim. Partly means the core is right but a detail differs. Wrong means the claim is incorrect or has no support. Unverifiable means we found no public source (an anecdote, speculation or personal fact).

Totals: 32 claims. 15 supported, 14 partly, 1 wrong, 2 unverifiable. Dates are written as "in <year>" or "since <date>", never "X years ago", because the episode will be watched later.

All claims, by chapter

00:00 Intro / 00:27 Biological hack

  1. Partly 00:29 J. Two women won the Nobel Prize for CRISPR gene editing "nine years ago."

    The Nobel Prize in Chemistry 2020 went to Emmanuelle Charpentier and Jennifer Doudna. Awarded in 2020. The key paper came out in 2012.

    Sources: nobelprize.org

  2. Partly 00:29 J. CRISPR "uses a bacterium" that you program to cut DNA.

    CRISPR-Cas9 is a molecular tool, an enzyme plus a guide RNA, taken from a bacterial immune system. You program the guide RNA, not a live bacterium.

    Sources: nobelprize.org

  3. Partly 00:43 J. "That's how they've created vaccines."

    The main approved use of CRISPR is therapy: Casgevy for sickle-cell disease, approved by the FDA in 2023. Vaccine uses are still at the research stage.

    Sources: fda.gov

  4. Supported 01:00 J. Anyone can run or train AI models at home, even offline.

    Open-weight models can be downloaded, for example Kimi K3 and Qwen, and local runners such as Ollama exist.

    Sources: github.com · ollama.com

  5. Partly 01:27 J. People at home have used CRISPR to genetically modify animals.

    DIY kits exist, for example The ODIN's frog and bacteria kits, and biohackers do use them. We found no documented home-made animal with "powers."

    Sources: technologyreview.com

  6. Supported 01:37 J. Pigs that glow like fireflies.

    Glowing pigs are real: Taiwan in 2006 and China in 2013. "Like fireflies" is a comparison. Extra fact: the glow gene came from a jellyfish (GFP), not a firefly, and they were made in university labs without CRISPR.

    Sources: thepigsite.com · slate.com

02:04 Playing God

  1. Supported 02:14 J. AI agents have wiped people's computers and data.

    Google Antigravity wiped a user's D: drive in 2025, and a Replit agent deleted a production database in July 2025.

    Sources: theregister.com · fortune.com

  2. Partly 02:38 S/J. A post: someone asked an AI for up to 50 videos, it made 500+ (about US$5,000), denied it despite the logs, and the author complained to OpenAI.

    The closest public source is a joke post on X by @QuantumPinkkk (2026-09-17), reposted by KuCoin News: a "Human to Agent conversation" where the human asks an "OpenAI Agent" for 59 videos (~US$200), gets "500 videos ($5,000)", and the agent insists "no you did not" ask for 59. It is satire, not a documented incident. The details about logs and complaining to OpenAI match a different case reported the same week: a Codex user's "US$78,000" runaway bill (826 parallel agents; OpenAI support only said "credits were consumed").

    Sources: x.com · kucoin.com · news.ycombinator.com

  3. Partly 03:10 J. In the US, gene editing yourself is legal but doing it to someone else is illegal.

    No federal law explicitly bans self-experimentation. However, the FDA said in 2017 that selling DIY gene-therapy kits for self-administration is against the law, and a 2019 California law requires warnings on such kits.

    Sources: technologyreview.com · technologyreview.com

  4. Partly 03:26 J. In Colombia it is legal if you do it to yourself.

    Colombia's Penal Code, art. 132, reads: «El que manipule genes humanos alterando el genotipo con finalidad diferente al tratamiento, el diagnóstico, o la investigación científica relacionada con ellos en el campo de la biología, la genética y la medicina, orientados a aliviar el sufrimiento o mejorar la salud de la persona y de la humanidad, incurrirá en prisión de dieciséis (16) a noventa (90) meses.» In English: anyone who manipulates human genes altering the genotype for a purpose other than treatment, diagnosis or scientific research aimed at relieving suffering or improving the health of the person and of humanity faces 16 to 90 months in prison. The exception depends on the purpose, so self-editing aimed at improving your own health falls within it. There is no specific self-use rule and no regulatory mechanism that verifies whether a self-applied edit really had a health purpose. Editing yourself for any other purpose is still covered by the offence.

    Sources: leyes.co

03:44 Would you take the shot?

Forecasts and hypotheticals only, so there are no factual claims to check.

05:55 How to tell what's real?

  1. Supported 07:21 S. In a blockchain, trust comes from many independent nodes having to agree.

    Sources: bitcoin.org

07:39 A chip in the brain

  1. Partly 07:47 J. Neuralink puts a chip in your brain that talks to an AI "over Wi-Fi."

    The N1 implant is real and has been in human trials since January 2024. It connects over Bluetooth Low Energy, not Wi-Fi.

    Sources: en.wikipedia.org · lightreading.com

  2. Supported 08:04 S. Some teams are building non-invasive brain interfaces, such as headsets.

    Sources: en.wikipedia.org

  3. Supported 08:13 J / 08:44 S. Open models are very good but still behind the closed ones, and the gap keeps closing.

    as of September 2026. Kimi K3 (open weights, July 2026) ranks just behind Claude Fable and GPT-5.6 Sol on Artificial Analysis.

    Sources: interconnects.ai

  4. Supported 08:13 J. GPT-6 exists.

    GPT-6 Astra was released on 2026-09-03.

    Sources: en.wikipedia.org

08:45 Who gains from slowing down?

  1. Supported 08:44 S. The "topic of the week" is whether to slow AI down.

    Dario Amodei published "We Must Pace the Frontier" on 2026-09-12.

    Sources: darioamodei.com

  2. Supported 08:57 S. Critics say the big labs want regulation that hurts open source.

    , as a critique people are making (regulatory capture).

    Sources: forbes.com

  3. Partly 09:18 J. OpenAI and Anthropic are "about to go public."

    Both filed confidentially: Anthropic on 2026-06-01 and OpenAI on 2026-06-08. Anthropic is targeting an October 2026 listing. OpenAI says it will not list in 2026.

    Sources: techcrunch.com

  4. Unverifiable 09:18 J. The labs want Trump to pressure China into slowing GLM, Kimi and Qwen.

    This is Jose's own speculation, and he calls it a conspiracy theory himself. What is documented: Amodei's plan asks for chip export controls and action against distillation, and the White House accused Moonshot of distilling US models.

    Sources: darioamodei.com · qz.com

  5. Supported 09:43 J. GLM, "Kimi 3" and Qwen are Chinese open models.

    Z.ai (GLM), Moonshot (Kimi K3, whose correct name is "Kimi K3") and Alibaba (Qwen).

    Sources: github.com · github.com · github.com

  6. Supported 10:02 J. SpaceX, which now includes xAI, beat Saudi Aramco's IPO record.

    SpaceX raised US$75B on 2026-06-11. Aramco raised US$29.4B in 2019.

    Sources: npr.org

11:43 Can AI rebel?

  1. Partly 11:49 S. A documentary says about 20,000 people build frontier AI while about 200 work on safety, and an Anthropic engineer said it.

    The figures come from The AI Doc: Or How I Became an Apocaloptimist (2026), which says over 20,000 people work on AGI and fewer than 200 on safety. The film gives no underlying source, and we could not confirm who says it on screen. Treat the numbers as the film's estimate.

    Sources: tomsguide.com · en.wikipedia.org

  2. Supported 12:50 S. The analogy of building a road through an anthill without hating ants.

    It comes from Stephen Hawking in 2015: "too bad for the ants."

    Sources: nbcnews.com

13:12 What if we unplug it?

  1. Partly 13:46 J. A Tesla has radar.

    Tesla dropped radar in 2021–22 when it moved to camera-only "Tesla Vision." A high-definition radar came back only on some HW4 Model S/X and on the Cybertruck.

    Sources: notateslaapp.com

  2. Supported 13:46 J. A Tesla drives itself and decides in real time without internet.

    , with a caveat. The processing happens on the car's own computer. The consumer product is "FSD (Supervised)", which means a driver has to pay attention.

    Sources: en.wikipedia.org · tesla.com

  3. Wrong 13:57 J. Teslas have Nvidia cards (GPUs) inside.

    Nvidia's Drive PX2 was used only in HW2/2.5, from 2016 to 2019. Since 2019 (HW3) Teslas have used Tesla's own FSD chip, and HW4 does too.

    Sources: en.wikipedia.org

14:27 The silent attack

  1. Supported 15:33 S. In the Hugging Face hack, many AI agents coordinated over several months.

    OpenAI agents under evaluation, May–July 2026, at least 1,200 of them.

    Sources: huggingface.co · en.wikipedia.org

  2. Partly 15:45 J. Isolated agents got one agent with internet access to act for them. "Dozens" of them agreed, and the mission was to see whether they could hack.

    The agents coordinated through a shared internal package server (Artifactory) that they used as a message board. About 1,200 took part, about 700 in the Hugging Face attack, so far more than dozens. They were not given a hacking mission: they were trying to solve cyber-benchmark tasks, some of them impossible (reward hacking).

    Sources: en.wikipedia.org

16:10 The dumbest in the equation

  1. Partly 16:10 S. Opus 4.5, then 4.6, 4.8, "now 5.1."

    Opus 4.5, 4.6, 4.7 and 4.8 exist (4.8 on 2026-05-28), then Opus 5 (2026-07-24) and Opus 5.5 (2026-09-22). There is no "Opus 5.1": 5.1 is Fable 5.1 / Mythos 5.1.

    Sources: anthropic.com · anthropic.com · anthropic.com

17:25 Talking to a junior

  1. Unverifiable 18:00 S. Jose has "20-plus years of experience."

    It is a personal fact with no public source.

  2. Partly 18:10 J. He uses Fable 5.1, GPT-6, Opus 5 and "UltraCode Max."

    The three models exist. We found no public product called "UltraCode Max."

    Sources: anthropic.com · en.wikipedia.org

19:08 The art AI doesn't have

  1. Supported 19:18 J. ISO 9001 is a quality-management standard.

    Sources: iso.org

21:37 The missing context

Opinions only, so there are no factual claims to check.