Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Claude BROKE Wall Street Overnight...
TL;DR
Wes Roth says there was no AI “turnaround” — just a straight line people refused to follow — His core argument is that Anthropic and OpenAI revenue, capabilities, and enterprise usefulness have been compounding all along, while journalists wrongly reframed that steady trajectory as a sudden post–“Claude Code” breakthrough.
The real story isn’t Claude Code itself — it’s the deployment stack around frontier models — Roth argues tools like Claude Code, Codex, OpenClaw, Hermes, and other “harnesses” are what let models actually do valuable work inside companies, much like a Formula 1 car amplifies the driver.
Anthropic just teamed up with Wall Street heavyweights to build an AI deployment machine — He highlights a new $1.5 billion joint venture backed by Anthropic, Blackstone, Hellman & Friedman, Goldman Sachs, Apollo, General Atlantic, GIC, Leonard Green, and Sequoia to push enterprise AI into finance at scale.
OpenAI is reportedly building the same playbook, only bigger — Roth says OpenAI is raising $4 billion from 19 investors for a “development company” at a $10 billion valuation, aimed at broad deployment across finance, manufacturing, and healthcare.
The bottleneck was never raw model capability — it was getting AI installed in messy real businesses — Using examples like Nvidia’s Voyager Minecraft agent and Palantir’s forward deployed engineer model, he explains that production AI needs custom scaffolding, domain knowledge, and engineers embedded with customers.
Palantir’s forward deployed engineer model is the template frontier labs are now copying — Roth points to Palantir’s approach of embedding real engineers inside customer organizations, especially banks, hospitals, and governments, and ties that model to the company’s stock move from about $19 at IPO to $6 and then a roughly 640% five-year return.
The Breakdown
The “AI bubble” crowd got the story wrong
Wes opens in full rant mode: the people calling AI a bubble, fake, or “just fancy autocorrect” weren’t seeing the actual trend. His point is blunt — when AI threatens your job or status, denial kicks in, and for journalists especially, that can distort how they cover what’s happening.
Claude Code didn’t change the world overnight — the line was already going up
He pushes back on the idea that some recent product suddenly flipped AI from hype to reality. Claude Code, Codex, OpenClaw, and Hermes are useful, he says, but they’re scaffolding around models, not the fundamental cause of progress; the real story is a capability curve that has been climbing steadily, even if people only noticed once the outputs got impossible to ignore.
His favorite framework: just follow the straight line
Roth leans hard on a chart of software-engineering task performance and says the simplest forecast beats most expert commentary: extrapolate the line. He quotes ex-Google researcher Julian Schrittwieser — known for work on AlphaGo, AlphaZero, MuZero, and AlphaFold — to argue that extending trend lines often gives you a better model of the future than listening to pundits who keep inventing reasons the trend somehow doesn’t count.
Then he gets to the actual news: Anthropic and Wall Street are joining forces
The big development is Anthropic launching a joint venture to deploy enterprise AI, with Blackstone, Hellman & Friedman, and Goldman Sachs as founding partners. Roth frames this as a huge signal because these firms, plus backers like Apollo Global Management, General Atlantic, GIC, Leonard Green, and Sequoia, represent the kind of institutional muscle that can force AI into the core of major industries.
OpenAI is reportedly building the same machine at a larger scale
Anthropic’s vehicle is valued at $1.5 billion with a $300 million commitment from Anthropic, Blackstone, and Hellman & Friedman. Roth says OpenAI is preparing a parallel structure called the “development company,” reportedly seeking $4 billion from 19 investors at a $10 billion valuation — same idea, just aimed more broadly across sectors like manufacturing and healthcare, not only finance.
Why AI adoption has felt slower than the hype suggested
Here he gets practical: AI is amazing in demos, but deploying it into real workflows is hard. He compares frontier-lab research tricks to a lagging real-world rollout cycle, saying an idea can show up in a paper and still take 12-plus months to become something a normal business can actually use reliably.
The Minecraft story that explains modern AI products
Roth revisits Nvidia’s Voyager agent, which used GPT-4 and GPT-3.5 to play Minecraft, self-improve, and build up skills through text-based state descriptions before multimodal systems were mature. His broader point is memorable: modern tools like Claude Code, Cursor, OpenClaw, and even AlphaEvolve are all versions of the same pattern — a model plus a harness, like a driver inside a Formula 1 car.
Palantir’s “forward deployed engineer” model is the unlock
The deployment gap, he says, gets solved when the people who know the models work directly with the people who know the customer’s business. That’s why he spotlights Palantir’s forward deployed engineers — embedded engineers shipping real code inside banks, hospitals, and governments — and argues Anthropic and OpenAI are now copying that exact model to create sticky, long-term AI installations across the economy.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
