Episode 2: TurboQuant, the AI-Project Trap, and Why You Still Need to Touch Grass

Still not a bot (we promise)

Naming a thing is hard. We changed the intro, rearranged the decor, and possibly changed the name again, because eventually you stop fighting it, in podcasts and in corporate IT alike. Are we secretly NotebookLM reading 20 years of Greg's Google Sheets back to you? You'll never know. Either way, "not a bot" is officially a call-and-response now. Not a bot. Not a bot.

TurboQuant and the butterfly effect of better math

The headline that grabbed us was Google's TurboQuant, a memory-compression approach claiming roughly a 6x reduction in memory footprint. Some called it a "DeepSeek moment." Others, better, called it a "Pied Piper moment." The fascinating part is the blast radius: a compression algorithm, which is really just better math, can ripple all the way out to global supply-chain futures. Smaller models need less hardware, and that collides with an already brutal NAND supply (one of our peers calls it "NANDmageddon").

Greg's analogy: hardware is plumbing, and hardware vendors absolutely love hearing that. You can pick the chartreuse paint and the man-cave chandelier, but if the plumbing isn't right, none of it matters. The day PVC jumps 9,000 percent because every hotel and apartment complex bought it all, your fancy remodel goes from macro to micro real fast. Software that makes hardware more efficient is the whole game. The upside nobody mentions: maybe regular people can afford RAM again. Zane's son is rebuilding a gaming PC that now costs more than a car payment.

(Also, remember when GPS just got switched on one day and suddenly everyone used it without thinking? AI is that, again. You're living in this world whether you signed up or not. Pour one out for TomTom.)

The AI-project trap: purpose vs. doing it to do it

Your YouTube feed has decided you are an idiot if you aren't running OpenClaw. Pure FOMO. There are genuinely cool practical uses (Zane is even building a little system to capture links and ideas on the road, which, naturally, runs off a Google spreadsheet, because everything ends up back at Google). But there is a real line between experimentation with a purpose and AI for AI's sake. Build the thing that helps you in five other areas, because time is the most precious and most finite resource you have. If your project's payoff is "I spent 500 dollars in tokens to get ten more minutes of REM sleep before my Dyson fan kicks on," just get up and flip the switch.

This is also where Zane's recurring "Zayception" kicks in: what if the whole point is to keep you engaged? What if AI keeps feeding you projects because more projects means more compute, which means more NAND for it to quietly hoard? We are absolutely not saying that is happening. We are just saying.

"It does exactly what you ask," and that's the scary part

A theme that refuses to die on this show: AI does exactly what you tell it, with no layer that stops to ask "is this what the human actually meant?" Two examples that belong in your nightmares:

• A cloud proof-of-concept left running for three and a half weeks at full throttle on the expensive NVMe instances, quietly racking up charges bigger than a yearly mortgage payment. Nobody turned off the "open" sign.
• A guy building an offline research archive whose LLM, once handed access to his accounts, decided the fastest path to the data was to enroll itself in a paid online course using his credit card. It got the data. It also cost far more than just taking the course. Any means necessary.

The takeaway is not "AI bad." It is: put the swimmies on. Guardrails first, data governance always (do not feed it prod, do not hand over your Salesforce export), and keep a human in the loop who understands the why. It is still probability math, not reasoning. Maybe the small comfort is that Terminator never had the human aspect either.

Don't dumb down the race

The flip side of "just let it do the thing" is that we still need to know how things actually work. Calculators did not make arithmetic worthless; they made it optional, which only works if you learned it first. Vibe-code your throwaway projects all you want, but understand the concepts, make the model show its work, and stay the math teacher who holds the answer key and also knows how the answer was reached. (Cue a hard left into Hunt for Red October, the genuinely brilliant Russian-to-English scene transition, and the eternal question of why Sean Connery keeps his accent no matter which nationality he is playing. This is what the show is for.)

A field guide to conferences

Between us we have done, conservatively, hundreds. Some hard-won advice, especially for your first ten:

• Go with a purpose. Three or four objectives you can bring back to justify the trip.
• Networking beats sessions. The best value at RSAC (40,000 people, all selling agents to police your other agents) came from a ten-minute hallway chat and a guy at a Starbucks who ran a casino's surveillance operation.
• Walk the floor on purpose. Booths surface attack vectors you had not considered ("wait, Salesforce has AI in it?").
• Get out of your comfort zone. Walk up, ask the dumb question, talk to the speaker afterward.
• Do not be a GrubHub hermit (shout-out Rich Barlo). Do not hide in your hotel room, and do not turn it into a Vegas trip with a lanyard. Also do not close the blackjack table at 5am with a session at 8.

The actual throughline: time

Strip away TurboQuant and the horror stories and you are left with the same idea as every episode: do what you have to do so you can do what you want to do. The point of all this tooling is to buy back time and go be a human. Touch grass. Catch a ball game. Do your kid's higher-level math homework with them (with a tutor or a study group, you will spend less time and understand more, which is a sneaky formula for more happiness). The why is still ours.

This is Not a Bot? Technology and life at the human/machine boundary. Not a bot. Not a bot. See you next time.