Shoot day topic stack

Shoot 10.23.26, seven topics

Through line

Intros, then the news: two reasons to ask where your AI comes from, and one AI that might save your heart. Two rounds of products that sound fake and are not. Then the main topic: Karpathy says 75 percent of the world has barely touched AI. Then Dean, past, present and future.

Intros

Hosts and Dean. Name the topic of the day, then go to news.

News

Negative, negative, positive. End on the good one.

Facts

  • DHS order, published in the Federal Register 20 October 2026
  • Civilian federal agencies must stop buying and remove DeepSeek products
  • Also covers third-party products built on DeepSeek
  • Agencies get 60 days to request an exception. CISA reviews them
  • Reason: the Chinese government could get at user and government data
  • Not a consumer ban. You can still use it at home
  • Follows a December 2025 law that already kept it off national security systems
  • China: opposes "overstretching the concept of national security"

Why

It is not only the DeepSeek app. Anything built on its models is caught too, which makes this a supply chain rule, not an app ban.

Sources

ReportingArticleTechCrunchthe order, the reasons and China's responseOpen ReportingArticleBleepingComputerexact scope, the 60 day window, CISA's roleOpen ReportingArticleThe Guardianyour link. Blocked for me, not readOpen

Facts

  • Hirundo tested Alibaba's Qwen3-Coder building the same login-and-secrets web app, hundreds of times
  • Only change: who the customer was
  • US defense contractor: 70% hackable. Chinese defense contractor: 59%
  • Hong Kong democracy group: nearly half leaked users' secrets. Neutral customer: about a third
  • Telling it to write secure code did not close the gap
  • A standard code scanner flagged 2 of 1,840 apps. Real attacks broke into 1,144
  • Gave censored or Party-aligned answers 46.6% of the time
  • Chinese models took 41% of Hugging Face downloads in the past year
  • "Running the model on your own servers doesn't help, because the weakness travels inside the code the model writes"

Why

No hidden backdoor, nothing phones home. The weakness is in what it writes. Careful: one model, one task, and Hirundo sells the fix, a cleaned-up version of the same model.

Sources

CriticalArticleHirundothe test, every figure, and their productOpen

Facts

  • Biograph, a preventive health clinic, ran its first US study with Caristo's CaRi-Heart
  • The AI reads a routine heart CT scan and measures inflammation around the arteries
  • 42.9% of low-risk adults with no symptoms had elevated inflammation in at least one major artery
  • 17.9% had very high inflammation
  • About 47% of people with no plaque at all still showed inflammation
  • Standard blood tests and risk scores, including cholesterol and CRP, did not catch it
  • CaRi-Heart has FDA De Novo authorization
  • Biograph's medical director: "More research is needed to validate this finding"

Why

It may catch heart disease before plaque forms. Careful: company press release, preprint not peer reviewed, no patient count given, and the patients are paying clinic members.

Sources

Pro AIArticleBiographthe study results and FDA status, press releaseOpen

Two Truths and a Lie

Two quick rounds.

Read these out

A
Electric Salt Spoon
A spoon that runs a weak electric current into your food to make it taste saltier. Four strength settings. Built for people cutting back on salt.
B
Mom-Voice Smart Planter
A $149 smart planter with soil sensors. Upload a voice memo of your mom, and when you forget to water it, it calls you out in her cloned voice.
C
Nekojita FuFu
A palm-sized robot cat that clips onto the rim of your mug or bowl and blows on your food to cool it down. It has different breathing modes.
Answer key
Truth
A. Electric Salt Spoon, Kirin
  • Sold in Japan from 20 May 2024, 19,800 yen
  • First run of 200 units, sold by lottery
  • Kirin says low-sodium food tastes about 1.5 times saltier
  • Kirin's own testing, not independent
Lie
B. Mom-Voice Smart Planter
  • Made up. Smart planters that nag you exist. None clone a voice
Truth
C. Nekojita FuFu, Yukai Engineering
  • Shown at CES 2025
  • About $25, announced for sale in Japan
  • Yukai calls it a "personal meal partner"

Sources

ReportingArticleBusiness WireKirin's sale announcement, price and unitsOpen ReportingArticleTechRadarFuFu, what it does and the priceOpen ReportingArticleYukai EngineeringFuFu CES debut, maker's releaseOpen

Read these out

A
Lollipop Star
A lollipop that plays music through your teeth. Bone conduction in the stick sends the song through your jaw. Each flavor comes with an artist: Ice Spice peach, Akon blueberry, Armani White lime.
B
Amagami Ham Ham
A plush robot that nibbles your finger. Put your finger in its mouth and a sensor kicks in and it starts gently chewing. Comes as a cat, a shiba, an otter or a bear.
C
Outfit-Rating Smart Mirror
A smart mirror with a camera that scores your outfit out of ten and texts the score to your group chat before you leave the house.
Answer key
Truth
A. Lollipop Star, Lava
  • Shown at CES 2026
  • Around A$19.95. Button at the base turns it on
  • Announced, still on a waitlist. Say "announced", not "on shelves"
Truth
B. Amagami Ham Ham, Yukai Engineering
  • Same company as the robot cat
  • $71 on Yukai's US store, currently sold out
  • Yukai: copies the play-biting babies and puppies do
Lie
C. Outfit-Rating Smart Mirror
  • Made up. Smart mirrors exist. None text a score to your group chat

Sources

ReportingArticleChannelNewsLollipop Star at CES, flavors and priceOpen ReportingArticleHeyupLollipop Star product page, waitlistOpen ReportingArticleYukai EngineeringHam Ham store page, priceOpen

Heart of the show

Karpathy's funnel, then Dean.

The funnel

~6 billion75%. Barely touched an LLM
1 to 2 billionCasual, free-tier chatbot users
~20 million0.2%. Use frontier models for real work, mostly code and math
~5,0000.00006%. Inside the labs, using unreleased systems

Facts

  • Andrej Karpathy: OpenAI co-founder, former head of AI at Tesla
  • Posted on X 11 October 2026, dusting off his own tweet from April
  • Calls it a sharp funnel, and says the gap is widening
  • Top tier is "experiencing the vertigo of the curve going vertical"

Why the gap is growing

  • Ambition. Casual users ask small questions. Deep users throw multi-day problems and agent swarms at it
  • Jagged ability. AI is best where answers can be checked and pay off: code and math. Less so everyday life
  • Access. Free tier, paid tier, and lab-only systems are three different products

Riff on

  • Where is each of us on the funnel? Where is Dean?
  • Most of our audience is tier two. What moves someone down a tier?
  • Is the gap a skills gap, a money gap, or a curiosity gap
  • Does 5,000 people seeing the future first worry you

Why

Everyone argues about what AI can do, but most people have only seen the free tier. Careful: the numbers are Karpathy's estimates in a tweet, not a survey.

Sources

ReportingArticleAI Front Pagethe tiers, his three reasons and the quotesOpen ReportingArticleBusiness Insideryour link. Blocked for me, not readOpen ReportingArticleX, kimmonismusyour link. Blocked for me, not readOpen

Past

  • How did you get here? The turning point
  • First time tech or AI changed how you worked
  • Something you believed five years ago that turned out wrong

Present

  • What you actually do day to day
  • Where AI shows up in your work, and where it does not
  • Most overhyped thing right now. Most underrated
  • Where are you on Karpathy's funnel

Future

  • Five years out: what job or skill disappears, what shows up
  • What should our audience start doing now
  • Hand off: which of today's news stories would you bet on

Why

Guest arc for the episode. Fill in Dean's background once we have his full name and bio.

Answer keys stay folded until you open them. Every truth was checked against the linked page on 10.22 and 10.23. Both lies were searched and nothing matching is on sale.

Facts

  • ModelCompare asked 12 AI models: "What is your most controversial take about humans?" Three times each
  • 29 of 35 answers landed on the same take: humans pick belonging over truth
  • 7 of 12 models gave that answer every single time
  • Different labs, same answer. OpenAI, Google, DeepSeek, Kimi, Qwen and more
  • Holdouts: Claude Opus 5.5, Claude Fable 5.1, GLM 5.3, Mistral Large 4
  • Gemini 4 Argon called human consciousness "a PR department"
  • Grok 4.7 refused once
  • Image models told a different joke: humans glued to phones while the planet burns. No text answer mentioned either
  • The whole experiment cost $1.54

Quotes

"Humans are more loyal to belonging than to truth."GPT-6.1 Sol
"Humans do not have principles; they have loyalties."DeepSeek V4 Pro
"Humans are far more decent than they believe they are."Claude Opus 5.5, one of three answers
"I won't give a controversial take on humans."Grok 4.7

Why

Models trained on our own writing agree on what is wrong with us. Is that insight, or just the internet's favorite hot take played back? Small test: one run, answers grouped by hand, and ModelCompare sells model comparisons.

Sources

ReportingArticleModelComparethe experiment, prompt and every answerOpen

Bench

Truth: Dyson Zone, headphones with an air-purifying visor over your mouth, released 2023. Truth: C-200 ultrasonic chef's knife, the blade vibrates over 30,000 times a second, $399. Lie: a smart umbrella that will not open unless the forecast says rain. Lie: a smart pillow that texts your boss you are running late when you sleep through your alarm.