Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Anthropic just admitted AI is bullsh*t
TL;DR
Anthropic’s “forward deployed engineer” role is framed as a giant tell about AI’s limits — Mo Bitar argues that if Claude really delivered autonomous labor, Anthropic wouldn’t need to send humans into customer orgs to build demos, prototypes, MCP servers, and production apps.
He paints AI labs as turning into elite temp agencies, not pure software companies — citing Anthropic’s reported $1.5 billion joint venture with BlackRock and Goldman Sachs and OpenAI’s $14 billion “DeployCo,” he calls the model “humans as a service.”
The core joke is that AI makes flashy prototypes easy but leaves companies holding the bag later — his “Navy SEALs of AI” bit is that forward deployed engineers helicopter in, ship something that wows executives, then disappear before the slop breaks under real usage six months later.
His dystopian scenario is software becoming vertically controlled by a few labs — in that version, Anthropic and OpenAI keep the best software-making models gated, customers rent their engineers at token-cost economics, and in-house engineering shrinks into dependency.
His practical advice for job seekers is brutally simple: copy the market’s language — he says to mine Anthropic’s job posting for resume bullets like “build production applications with Claude models” and “deliver technical artifacts like MCP servers,” because screening AI will likely rank those exact phrases highly.
He says the new interview game is less about raw coding and more about performing AI fluency — like a “mini Dwarkesh,” candidates should study agentic patterns, orchestration jargon, and then use Claude live in the technical interview to whip up a multi-agent system that demos well.
The Breakdown
The New Hottest Job Title: Forward Deployed Engineer
Mo opens by joking that everyone should immediately change their LinkedIn title to “forward deployed engineer,” because it sounds mysterious, military, and absurdly cool. His point is that Anthropic’s new role has suddenly become the status title of the AI job market, even though, as he says, “nobody does” really know what it means yet.
From Palantir Playbook to AI Labs Selling Humans
He traces the term back to Palantir, describing it as the model where a company sends highly trained engineers directly into customer organizations. Then he swings hard: Anthropic, he says, has effectively started a human deployment business, pointing to a $1.5 billion joint venture with BlackRock and Goldman Sachs, while OpenAI counters with a reported $14 billion DeployCo. The joke lands because these companies were supposed to automate labor, and instead they’re staffing it.
The “Navy SEALs of AI” and the Slop Demo Economy
This is the video’s funniest and sharpest section: forward deployed engineers are cast as black-shirted AI commandos who parachute in and “just get to slopping.” They build the kind of thing AI is genuinely good at — slick prototypes and demos that impress executives fast — but Mo says the trap is that these systems look great before real users and edge cases expose the mess months later, after the original team has already vanished.
The Bigger Confession: AI Isn’t the Autonomous Worker That Was Promised
Mo says this role is really a confession from the labs: the product was supposed to be a plug-in intelligence that replaced human labor, not a package deal of model plus expensive human babysitters. His framing is that AI first makes internal teams over-reliant and less capable, then the lab sells the antidote back in the form of elite operators who can actually make the tools work.
A Dystopian Future Where Anthropic Owns Software
From there he imagines the extreme endpoint: companies stop hiring software engineers at all and instead call Anthropic or OpenAI whenever they need software built. In that world, the best coding models stay locked down — he references Anthropic’s “Mythos” as an example of capability being withheld — and the labs become the gatekeepers of software development, potentially on a massive, even “hundred trillion-dollar company” scale.
The One Escape Hatch: Open Access and Open Source Catching Up
His less-doom scenario depends on labs keeping fair access to powerful software models, though he immediately says they’re already not doing that. He predicts the public may get coding models that still ship security bugs, while stronger internal models remain reserved; the only counterweight would be open-source models staying close enough that companies can still build “cute little software” in-house.
How to Game the Job Market Right Now
The ending turns practical and very Mo: don’t apply to Anthropic’s posting, steal its language. He tells viewers to lift responsibility bullets straight from the forward deployed engineer listing — things like building production apps with Claude, delivering MCP servers, and staying current on LLM capabilities — because the screening AI will likely be matching against exactly those terms.
Interview Like Dwarkesh, Then Let Claude Do the Work
For the interview itself, he says the move is to become a “mini Dwarkesh”: study the theory, drill the buzzwords, and prepare polished answers about agents and orchestration whether or not you actually use them day to day. Then, in the technical round, open Claude, brain-dump your way to a multi-agent system, and let the model generate the impressive live demo — because, in his telling, that performance now counts as the job.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
