Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
I'm done. I'm f***ing done.
TL;DR
Mo Bitar rage-quit the AGI rabbit hole after a 12-hour binge — what started as a 2 a.m. “what if AGI is possible?” thought ended with him furiously reading at his in-laws’ house and declaring he’s “noping the f*** back out.”
David Deutsch’s views were the breaking point — Bitar says he was “flabbergasted” to find that the quantum computing pioneer doesn’t even treat animals as intelligent or conscious, which made the whole AGI debate feel definitionally incoherent.
His core objection is that AGI talk has no stable definitions — if people can’t agree on intelligence, consciousness, or personhood, he argues, then claims about generally intelligent computers are built on vapor.
For coding, Bitar thinks specialization beats general intelligence anyway — his example is a mid-level engineer with 10 years in a codebase outperforming a fresh PhD, because real software work depends on deep context, business history, and team politics.
He calls current LLMs “a trick, not AI” — in his framing, they are statistical systems that convincingly imitate intelligence long enough to fool executives, not systems with theory, understanding, or genuine cognition.
His new rule is brutal: stop extrapolating and judge only what exists now — no more “12 to 18 months” narratives, no benefit of the doubt, just evaluating present-day LLMs on present-day performance.
The Breakdown
A family weekend gets hijacked by AGI panic
Bitar opens with a great self-own: he’s at his in-laws’ house, surrounded by food, family, and “a child who I am told is mine,” but he’s stuck in the corner doom-scrolling AGI takes. The trigger was a 2 a.m. thought that AGI might actually be possible in theory, and by the next day he’d burned 12 hours reading himself into total exhaustion.
One foot into the “cult of AGI,” then immediate retreat
He describes briefly entertaining the idea that AGI could just be an engineering problem solvable in a couple hundred years, then instantly recoiling. The phrase he uses is memorable and sets the tone for the whole video: he accidentally dipped one foot into the “cult of AGI” and is now backing the hell out.
David Deutsch becomes the final straw
The real snap comes when Bitar encounters David Deutsch’s views on animals, intelligence, and consciousness. He’s stunned that “the father of quantum computing” and author of The Beginning of Infinity could, in Bitar’s telling, deny animals are even intelligent or conscious, and that pushes him to ask: if the baseline is that messy, how are people seriously debating machine general intelligence at all?
His big complaint: nobody can define the thing they’re selling
From there, he turns the rant into a conceptual takedown. His argument is that terms like intelligence, consciousness, personhood, and AGI are being thrown around as if they’re well-defined, when in his view “there is f***ing nothing” solid underneath them.
Why AGI doesn’t even solve the coding problem people point to
Bitar then shifts from philosophy to software work and argues that coding is the wrong use case for AGI hype. His example is concrete: a mid-level engineer who has lived inside a codebase for 10 years will beat a newly hired PhD because the job is about soaked-in context — not abstract general reasoning, but knowing the code, the business, the politics, even which designer will quit over a button color.
Context, not cleverness, is the real bottleneck
That leads to his sharper claim that current LLMs fail at coding because they can’t “hold a damn thing in [their] tiny ass head.” For him, the missing ingredient isn’t more scale or more generalized intelligence, but deep, durable immersion in a problem space — something he insists an AI cannot feel, understand, or become engrossed in.
“LLMs are a trick”: the Turing test rant
In the back half, Bitar draws on Deutsch again to attack the idea that convincing imitation proves intelligence. He says if a system can pass by tricking people, then the test is broken, and compares LLMs to a polished lie: add a billion parameters and it’s still just a bigger lie.
No more forecasting — only the present counts
He ends with a personal policy shift: no more extrapolating, no more giving the technology the benefit of future progress, no more “12 to 18 months” optimism. From now on, he says, there is only now — and right now, LLMs are persuasive statistical tools that CEOs may be gullible enough to overtrust, not evidence of real AI.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
