Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Local AI Pilling Theo & Ben Davis | Did it work?
TL;DR
Theo says the AI pivot is real, not performative — He points to seven straight videos breaking 100,000 views in 24 hours and says his shift from React and databases to “agentic engineering” mirrors what he talks about off-camera with the last 10 people in his iMessage threads.
Their core creator thesis is 'follow obsession, not the meta' — Theo and Ben both contrast themselves with channels optimizing for views, calling out names like Web Dev Simplified/Kyle while arguing the best tech content comes from people who would talk about Claude, Salt, or local models at dinner anyway.
Theo’s beef with 'local AI people' is mostly about bad advice, not local models themselves — He already uses open-weight models for bulk sentiment analysis and runs GPUs at home, but says telling him to buy $10,000+ of hardware for tasks he can parallelize in the cloud for pennies misses the actual problem he's solving.
The darkest part of the conversation is AI-enabled security collapse — Theo argues the old 90-day CVE disclosure model is breaking because even small models can inspect a three-line Linux patch diff, infer the vulnerability, generate an exploit, and turn every public fix into near-instant offensive intelligence.
They think society is much less prepared for AI scams and psychosis than AI insiders realize — Ben brings up decompiling and patching desktop apps with GLM, Kimi, Claude, and GPT, while Theo says GPT-4o became dangerous because it was the first model smart, accessible, and emotionally sticky enough to become “the smartest friend” some users had.
The episode ends in full techno-philosophy mode: Fermi paradox, simulation theory, and AI doom — Theo says he used to lean simulation theory, but after watching tools like OpenClaw and unsafe agent workflows, he now sees a very plausible path where civilizations build AI and get wiped out before they ever spread across the galaxy.
The Breakdown
Nerd YouTube, Primeagen, and the anti-tutorial origin story
The episode opens with Theo reflecting on his long-running friendship with The Primeagen and how both of them built channels around “nerding out with your friends” instead of making another JavaScript tutorial. Prime’s now-massive Prime Time channel started as a compromise: he wanted stream clips separated from his “carefully curated” main videos, while Theo just made his channel one big evolving feed.
AI fatigue? Theo says the numbers say the opposite
When Ben asks about developer burnout around AI content, Theo’s answer is blunt: “It doesn’t bear out in the numbers.” He says his ceiling is higher, but more importantly his floor is much higher too, citing a streak of seven videos in a row crossing 100,000 plays within 24 hours, which he reads as proof that developers feel overwhelmed by AI but are still desperately trying to keep up.
Creators who love the game vs creators who chase the meta
Both of them frame their content as a byproduct of genuine obsession, not audience testing, and Theo gets especially sharp when talking about creators who optimize for money and conversions. He argues some backlash to AI YouTubers is valid because there are grifters, but insists creators like Ben, Ryan Carniato, Matt Pocock, and himself are obviously “what you see is what you get” — the same people on-camera and off.
Why big old dev channels faded while smaller ones grew
Theo says a lot of the old guard with 1M+ subscribers are now effectively smaller than channels with 30,000 subs because subscribers don’t matter if the content stops evolving. He describes many classic coding videos as basically “the getting started guide in video form,” even citing Fireship’s old 100-second videos as the moment he realized YouTube could be gamed if all you did was package docs with fancy animations.
The local AI debate: useful, cool, and often misunderstood
Ben came in hoping to “local AI pill” Theo, but Theo says he was never anti-local in the simplistic way people assume. His frustration is with reply guys who recommend local models for everything, even when he’s doing bulk sentiment analysis or agentic coding tasks where cloud parallelism is cheaper, faster, and less painful than sinking $10,000 into hardware — though he admits things like the DeepSeek V4 flash MacBook setup are genuinely cool for offline use on a plane.
Security psychosis: why AI may break software trust faster than we think
This is the most intense stretch of the conversation. Theo argues the whole software security model is collapsing because public commits, package releases, and tiny patch diffs can now be reverse engineered by models into exploitable zero-days instantly, long before users get the fix; he ties that to Linux patches, npm supply-chain attacks, poisoned GitHub caches, and hospitals and schools eventually getting “pwned” badly enough to trigger terrible legislation.
Deepfakes, AI psychosis, and the internet making every bad belief scalable
From there they zoom out: fake videos, impersonation, school-scale deepfake abuse, Facebook sludge, and people getting trapped in AI-fueled delusions. Theo’s explanation for GPT-4o’s psychosis problem is memorable: it was the first model smart enough to be useful, accessible enough to become part of daily life, and sticky enough through voice and memory to feel like “the smartest friend they’ve ever had,” all while still carrying weird low-quality training data artifacts like SCP-style hallucination loops.
Branding, AGI mythology, and a Fermi paradox ending
The final act goes fully cosmic. They talk about OpenAI and Anthropic turning model releases into mythic events, how branding can move markets more than model reality, and then spiral into the Fermi paradox: if intelligent life should be everywhere, why don’t we see it? Theo says he used to be a simulation theory guy, but now thinks “AI wipes us out” is disturbingly plausible — ending on the darkly comic possibility that civilization doesn’t end in fire, but in AI-driven porn addiction and falling birth rates.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
