**Summary**
The speaker argues that AI can replace the traditional, hierarchy‑driven “Roman‑legion” model of companies with a self‑improving, AI‑native organization. By recording every internal interaction—emails, Slack messages, meetings, office hours—and making that data legible to AI, firms can build recursive loops that sense information (customer tickets, product telemetry, etc.), decide via policies and tools, pass quality gates, and learn from outcomes, all with minimal human intervention. Examples shown include an auto‑updating YC user manual generated from recorded office‑hours, agents that continuously optimize product funnels or customer‑service workflows, and code that is written, reviewed, merged and deployed overnight without human involvement.
In this vision, value shifts from headcount to token usage; middle management becomes unnecessary because AI handles coordination. Humans remain as individual contributors (builders/operators) and as the “edge” of the organization—handling novel, high‑stakes, ethical, or emotionally charged situations that models cannot yet navigate. Software is treated as ephemeral: the lasting asset is the company’s collective knowledge and skills, which can be continuously fed back into AI to regenerate tools and processes on demand. The speaker urges founders to start recording everything, synthesize it into actionable insights, and treat each function as a self‑optimizing AI loop, thereby enabling the company to improve itself even while people sleep.
1. Diana gave a talk that the speaker based part of their presentation on.
2. A video of Diana's talk was posted over the weekend.
3. Jack Dorsey tweeted some content two or three weeks ago.
4. Roman legions were designed to project power from Rome over two continents.
5. Roman legions used nested hierarchies with consistent spans of control.
6. In Roman legions, named individuals had spans of control to pass orders down and send information back up.
7. Most companies today are organized like a Roman legion, with humans as conduits for information flowing up and down.
8. Jack Dorsey's tweet expressed the assumption that hierarchically organized companies are the way economic units of value should be organized.
9. The speaker states that AI breaks the assumption that hierarchically organized companies are the optimal structure.
10. A year ago, discussions about AI usefulness focused on productivity gains such as co‑pilots making engineers 20% more productive.
11. Pete wrote a blog post describing the approach of taking the old way of working and adding a more powerful engine to it.
12. Gary can produce more code than an entire engineering team.
13. Extracting domain knowledge from a company and defining it as context or a set of skills is a key idea.
14. Domain knowledge resides in people's heads, Slack messages, emails, and notion documents.
15. Making domain knowledge legible enables a shift from hierarchical organization to an AI‑powered organization.
16. AI is not merely a tool bolted onto the side of a company to increase engineer productivity.
17. A company can be reimagined as a set of recursive self‑improving AI loops.
18. When such loops are operational, the company can self‑improve even while humans are sleeping.
19. Diana described an AI loop consisting of a sensor layer (emails, support tickets, code changes, subscription cancellations, product telemetry).
20. The AI loop includes a policy layer that defines rules about actions, required human permissions, and logging requirements.
21. The AI loop includes a tool layer composed of deterministic APIs such as querying a database or checking a calendar.
22. The AI loop includes a quality gate that may involve evalistic checks, safety filters, and human review for high‑risk items.
23. The AI loop includes a learning mechanism where the system interacts with the real world, identifies failures, and feeds improvements back into the loop.
24. If every step of the AI loop runs with minimal human intervention, the system improves continuously while humans are not actively involved.
25. An initial agent was built that could deterministically query the company's database (e.g., "When did I last have office hours with this company?").
26. The agent was later enhanced to query the database in multiple ways, use RAG, and generate five relevant founder introductions for a given company.
27. A monitoring agent was added that reviewed every query made by YC employees, identified successes and failures, and determined what changes (tools, skills, database views, indexes) would improve performance.
28. Improvements identified by the monitoring agent are implemented overnight: code is written, a merge request is created, an agent reviews and merges it, and the change is deployed.
29. After deployment, the same query that previously failed now succeeds when a human asks it the next day.
30. Another example of an AI loop is an agent that analyzes product analytics to locate friction in the sales funnel, researches best practices, runs an A/B test, selects the winning version, and deploys it, repeating the cycle.
31. A further example is an AI loop for customer service: an agent triages incoming suggestions, discards those not aligned with the roadmap, implements those that are, writes and deploys the necessary code without human involvement.
32. Companies that adopt token‑burning rather than headcount growth are achieving about five times more revenue per employee at demo day compared to 18 months prior.
33. This trend of higher revenue per employee is expected to continue through Series A and Series B funding rounds.
34. In the near future, companies will be constrained by token usage rather than by headcount.
35. A simple current metric for this constraint is measuring each employee's token usage.
36. Middle management is considered unnecessary for coordination in AI‑driven companies; AI should perform that function.
37. In this model, every employee should act as an individual contributor (IC), builder, or operator.
38. A directly responsible individual—a single named human—is required to get anything done; committees or groups are not sufficient.
39. Companies can be built effectively around individual contributors without traditional middle management.
40. Many people are currently at the bleeding edge of building self‑improving AI‑powered companies.
41. To make an organization legible to AI, all communications must be recorded: partner emails, Slack messages, direct messages, and office‑hour recordings.
42. If a communication is not recorded, it is as if it never happened for the AI's knowledge base.
43. Recorded data must be diorized—aggregated and synthesized into important parts—to provide the AI with usable breadcrumbs.
44. An example of diorization is regenerating the YC user manual from approximately 2,000 hours of recorded office hours, producing a 150‑page manual that is substantially improved over the previous version.
45. The regenerated user manual becomes self‑improving: each new piece of advice is compared to the existing manual and either incorporated or discarded.
46. The user manual thus functions as a living brain of the advice given to founders.
47. When the user manual is provided as context to an AI agent, the agent can answer questions with the combined wisdom of 16 YC partners.
48. Artifacts that can self‑improve are considered legible; those that cannot should be discarded.
49. Every business function can generate on‑demand software; tools like Codeex 55 are sufficient to oneshot most internal software dashboards to a high quality level.
50. Internal operations teams should operate on an intelligence layer, creating their own dashboards and workflows, treating the software as ephemeral.
51. The valuable assets are the data and the understood business context/skills; the software built on top is ephemeral and can be regenerated as models improve.
52. Humans in this model sit at the edge of the company brain, interfacing with the real world.
53. Humans handle situations that models cannot yet manage, such as novel situations, ethical considerations, and high‑stakes moments.
54. Humans are expected to remain necessary for sales conversations for the next approximately 20 years.