How to Build a Self-Improving Company with AI

Predictions and Expectations:

• The speaker expects AI to break the assumption that hierarchically organized companies are the optimal economic unit.
• The speaker believes AI can reimagine what a company is as a set of recursive, self-improving loops.
• The speaker expects companies to get to Series A and Series B with about 5x more revenue per employee than they did 18 months ago.
• The speaker predicts organizations will soon be constrained on token usage rather than headcount.
• The speaker states middle management is "done" and will not be needed for coordination problems in the future.
• The speaker believes sales conversations will require a human being in the room for the next 20 years.
• The speaker believes high-stakes, high-emotion moments (e.g., co-founder breakups) will always require human intervention.
• The speaker expects the company to start self-improving even when humans are sleeping.

Timelines and Milestones:

• The speaker references a period of "two or three weeks ago" regarding recent tweets as context for current ideas.
• The speaker notes the YC user manual was written "five to 10 years ago" most of it, contrasting it with recent updates.
• The speaker mentions recording every Slack message and DM for "the last three or four months."
• The speaker plans to update the user manual "every single month."
• The speaker predicts models will get smarter "in a month or two."

Technology and Product Direction:

• The speaker proposes a future organizational structure based on "recursive self-improving ai loops" consisting of a sensor layer, policy layer, tool layer, quality gate, and learning mechanism.
• The speaker expects to replace internal software dashboards with "on-demand software" that can be regenerated as models improve.
• The speaker plans to make the entire organization legible to AI by recording all emails, Slack messages, and office hours.
• The speaker expects to utilize diarization to aggregate and synthesize recordings into "breadcrumbs" for AI context.
• The speaker predicts that "every function can generate" software, making the software itself ephemeral while data and skills remain valuable.
• The speaker expects to use agents to run A/B tests, research best practices, and deploy changes without human intervention.

Market and Industry Outlook:

• The speaker expects the current phase to be one of maximum experimentation to determine what is possible with new intelligence.
• The speaker warns that measuring token usage via leaderboards will likely be gamed once tied to promotion or firing.
• The speaker believes "token usage" will become a key metric for evaluating employee effectiveness directionally.

Company Plans:

• The speaker states the intent to record every conversation, including introductions, to ensure AI legibility.
• The speaker describes a specific plan to regenerate the YC user manual from 100,000 hours of recorded office hours, resulting in a 150-page document.
• The speaker plans to train agents to act as a "chief product officer" and "chief technology officer" to triage customer suggestions and implement them overnight.
• The speaker intends to delete internal software and regenerate it based on original instructions as models evolve.
• The speaker plans to position humans at the edge of the "company brain" to interface with the real world.

Financial Guidance:

• The speaker cites a metric of companies reaching demo day with "about 5X more revenue per employee" than 18 months ago and expects this to continue through Series A and B.
• The speaker suggests measuring everyone's "token usage" as a blunt, though potentially gamed, metric for organizational performance.

Risks and Caveats:

• The speaker is "not sure" if anyone currently has a truly self-improving company in every function, inviting challenge to that claim.
• The speaker warns that using token usage as a performance metric creates a risk of gaming the system if tied to employment outcomes.
• The speaker notes that AI cannot yet access all "novel situations," ethical considerations, or high-stakes moments, limiting full automation.

Confidence and Disagreement:

• The speaker expresses high confidence in the direction that AI will break hierarchical organizations, stating "I genuinely believe" an AI can produce more code than an entire team.
• The speaker expresses strong agreement with the "IC (Individual Contributor) only" model and states "middle management is over."
• The speaker explicitly disagrees with a specific third role proposed by Jack Dorsey, stating "I actually don't like the third one, so I deleted it."
• The speaker hedges on the completeness of current implementation by saying "I might be wrong" regarding whether a fully self-improving company exists in every function.