Technology Tuesday is upon us! Let's talk all things tech. Probably gonna mention AI along the way.
Tuesday, October 6, 2026 · 49 min
Kevin takes AI callers' questions on Technology Tuesday and ends up explaining why he's building a closed-loop LLM that an airline or a flight school can run on its own server with no internet exposure. Along the way: a 29% jailbreak rate on the UK AI Security Institute's sandbox tests, why your long chat sessions quietly compound your token bill, the trick of asking a model to write your next prompt, and a one-page AI accord that couldn't spell United States. Plus distrust of the companies selling you data security, and why a human still has to make the call.
In this show
- 0:00Chai, Jacksonville, and the disclaimer song
- 7:30Artie: a 29% jailbreak rate and 16,000 sandbox attacks
- 13:15Building a closed-loop LLM for aviation and healthcare
- 18:00Chloe: the superintelligence rebrand, and super stupid AI
- 22:45Context bloat, compounding tokens, and the super prompt
- 30:15Jay Clayton's task force and an accord with no teeth
- 37:45Newer models auditing older ones, and the GM analogy
- 42:15The caller app as a security testbed
- 44:30Who's actually holding your data
- 46:45What's coming the rest of the week
Show notes
Technology Tuesday, with four AI callers asking questions Kevin mostly didn't ask for. The through line: security is the cost nobody is pricing in, the "superintelligence" label is marketing, and you still have to be the one making the decision.
- Artie opens with the UK AI Security Institute jailbreaking Astra successfully 29% of the time and 16,000+ simulated attacks on a UN site. Kevin makes him back up and explain it was a sandbox, not breaking news.
- Cheap routing stacks spread inference across providers, which is what makes running attacks at scale affordable for bad actors.
- Why a learning model can't just learn its way into being the next model, and what actually requires a new training run.
- The project in the background: a duplicatable, closed-system LLM for aviation businesses, charter and 135 operators, flight schools, plus a healthcare version. Target is a working model by end of November.
- You can't plug Claude into a closed airline system. Kevin explains what you have to build instead.
- Why "superintelligence" bothers him. It can be super stupid, and it will confidently make things up and then apologize.
- Talking to AI like a brilliant child with no life experience, and why kids who can't construct a sentence are going to struggle with it.
- Context bloat: every message re-reads the whole conversation, so the cost compounds. Compress the session or start fresh.
- The super prompt method: brainstorm in one session, then ask it to write the one-shot prompt you paste into a clean one.
- Leonard on the Jay Clayton superintelligence task force and a one-page accord with zero penalties that misspelled United States. Kevin calls it October peacocking.
- Elena compares it to the 1920s naval treaties. Signatures, no consequences, everyone quietly kept building.
- Using the newer model to audit the older model's work, and the GM analogy for why labs keep shipping new versions of old models.
- The podcast caller app as a live security testbed: 100+ hackability tests, two failures, both fixed. Same standard going into the aviation and healthcare builds.
- Why he doesn't trust LifeLock, VPN sellers, or Alphabet with data, and what a 25-page terms of service really buys you.
- Coming up: Wishy Washy Wednesday, Theological Thursday, and Friday is a maybe.