Will Manidis published something uncomfortable last week. In his Substack newsletter 'Minutes', he argues that the current LLM boom's defining product isn't intelligence — it's the feeling of productivity. He calls the resulting systems 'tool-shaped objects', and he means it as indictment, not observation.
His central metaphor is the Chiyozuru kanna, a $3,000 Japanese hand plane requiring days of setup to produce shavings of exceptional quality. It's economically worthless compared to a power planer, and Manidis argues that's the point: the kanna doesn't exist to plane wood. It exists so the ritual of mastery can happen. GPU clusters, orchestration layers, token budgets, agentic dashboards — the most sophisticated tool-shaped objects ever assembled, consuming capital at extraordinary scale, with the primary output of that spending being the experience of spending it.
The FarmVille comparison is his bluntest instrument. No matter what input you give, the number goes up, the screen fills with evidence of effort. Manidis describes engineering teams building agent systems of 'breathtaking complexity' whose primary output is the system itself: agents generating logs, other agents analyzing those logs, reports populating dashboards, the whole apparatus sustaining its own operation.
He uses the viral spread of Matt Shumer's AI-generated essay 'Something Big is Happening' — reportedly reaching 40 million readers — as a worked example. AI-generated content about AI, amplified by an audience performing engagement with ideas, with no durable output that mattered. The loop closed on itself.
Manidis also targets the 'token budget' concept, popularized by investor Miles Grimshaw as a compensation proxy for AI value, arguing it naturalizes a metric that measures consumption rather than output. 'It is, in most cases, a cloud,' he writes.
The critique maps uncomfortably onto a sector that has made orchestration complexity a growth strategy. Executives at several agentic framework vendors declined interview requests for this piece. One spokesperson, speaking on background, pushed back on the framing: the orchestration layer is infrastructure, they argued — a means to an end — and the measure of its value is what enterprise customers build on top of it. 'You don't judge a database by how elaborate its query planner is,' the spokesperson said.
That's a reasonable defense, but it sidesteps Manidis's actual claim. He isn't arguing that orchestration tools are badly made. He's arguing that the market rewarding them is primarily a market for the sensation of building.
Lily Chen, who studies enterprise software adoption at Carnegie Mellon, has tracked similar dynamics in previous platform cycles. 'There's a consistent pattern where the tooling ecosystem monetizes faster than actual use cases mature,' she said. 'You saw it with blockchain, with big data. The difference here is the tooling is genuinely more capable, which makes the displacement from real output harder to see.'
Manidis stops short of dismissing LLMs outright. They are genuinely useful in constrained, well-defined deployments, he concedes, and will likely have real productivity effects over time. His claim is narrower: diffusion into the real economy will be slower, and look structurally different from the consumption narrative currently driving institutional AI budgets. The market for feeling productive, he concludes, reliably dwarfs the market for being productive. He thinks that's what's being built right now.