Three cards from two operators. Two themes that arrived from independent angles. Dan Shipper shipped a new product and wrote about the mechanism behind it. Tomasz Tunguz looked at token pricing data and named the shift it signals. Both are arguing, from different directions, that the interesting action has moved away from the model itself.
Making AI work visible
Dan Shipper launched the Every Agent on October 6. It is a shared Slack bot that operates in team channels rather than in private windows. The launch essay is not a product announcement. It is a diagnosis of why private AI use fails to spread organizational adoption.
One shared AI agent working in public spreads organizational adoption faster than many private agents captures the core claim. Most AI use in organizations is invisible. It happens in personal chat windows and terminals. Coworkers see only the finished output. They have no model for how the AI got there, what prompt produced it, or how many corrections it took. That invisibility is the adoption barrier. A shared agent that works in a channel makes the full loop visible: the request, the correction, the result. Coworkers watch the trick being performed and realize they could do it too. Shipper:
"Prompting in public works because most AI use is invisible."
"nobody learns a magic trick without seeing how it's done."
"one shared agent, working in public, AI-pills a company faster than a horde of personal ones."
The visibility mechanism has a second step that Saving agent corrections as named team skills converts individual AI workarounds into shared institutional capability covers. A correction made in a shared thread is ephemeral unless it gets saved. When a recurring correction is saved as a named skill, it becomes callable by any teammate. Shipper's team saved an editing workflow built through iterative corrections into a skill called "Kate Pass." Any editor can now invoke it without knowing how it was built.
"This is what AI-pilling a company looks like: The knowledge stops living in one person's setup and starts moving around."
The two mechanisms compound. Visibility creates the raw material. Canonicalization preserves it. Teams that do both end up with an AI skill library that grows faster than any training program could build it.
- Dan Shipper, One shared AI agent working in public spreads organizational adoption faster than many private agents. Public AI work removes the magic-trick barrier by making requests, corrections, and results visible to bystanders.
- Dan Shipper, Saving agent corrections as named team skills converts individual AI workarounds into shared institutional capability. Saving a recurring correction as a named skill converts individual workarounds into shared team infrastructure.
Commoditization has arrived
Tomasz Tunguz published a market analysis on October 7 with data that puts a hard number on a shift many people have been anticipating.
Token prices fell 41% since July while token usage rose 50%. Frontier models' share of total token consumption slipped from 53% in August to the mid-40s by early October. Spending among the top 1% of adopters fell 9.7% in August. When AI models commoditize, winning requires owning distribution and routing data, not model quality captures what the data implies for strategy.
In commodity markets, the entity that controls distribution and accumulates routing data wins. Not the entity with the marginally better model. TypeSafe's Jev handles routing at $0.04 per million input tokens, roughly 75 times cheaper than routing through standard LLMs. That cost structure means routing at scale is profitable and generates training data as a byproduct. Tunguz:
"Distribution has become the moat."
"the spoils do not accrue to the lab with a marginal benchmark lead"
The counterargument is real: if one lab produces a step-change in reasoning capability, that lab's model commands pricing power regardless of distribution. Tunguz acknowledges OpenAI's run rate is approaching $70B, close to Anthropic's. Commodity pressure is building at the lower tiers, not uniformly across the frontier. The implication for builders and labs operating below the frontier is clear.
- Tomasz Tunguz, When AI models commoditize, winning requires owning distribution and routing data, not model quality. Token pricing data shows the commodity inflection has arrived; distribution and routing data are the new moat.