Garry Tan: The Future of AGI Is Personal

Predictions and Expectations:

• Speaker (Y Combinator Partner) expects that by 2034, the world will not resemble 1984 if individuals build personal AGI systems.
• Speaker believes personal AGI is arriving diffused as agents running on individual context, not as a singular event or "god" announcement.
• Speaker predicts that every company in the arena could have personal AGI infrastructure by Monday.
• Speaker expects that intelligence of this kind should be owned by the individual, not rented.
• Speaker predicts that the fastest growing founders will not treat AI as autocomplete but as a workforce.
• Speaker believes the multiplier effect of agents applies to every piece of knowledge work, not just coding.
• Speaker expects that almost everyone on earth is still running their life on systems designed for a 7-digit working memory limit.
• Speaker predicts that the "dunks" (critics) will arrive by lunch, quote-tweet the speaker, and then git clone the speaker's work as part of the adoption curve.
• Speaker expects that 7,000 people leaving the event with this leverage will walk out and do things the world does not yet understand.
• Speaker believes that tools of the powerful should be given away to avoid a "priesthood" and instead create a renaissance.
• Speaker predicts that the gap between those with the harness/library/workforce and those without is widening every month.
• Speaker expects that the difficulty of creating excellent things has collapsed for those who have these tools, leaving the rarity up to the individual.

Timelines and Milestones:

• Speaker notes that in the Winter '25 batch (one and a half years prior to the talk), a quarter of companies had codebases that were 95% AI-generated.
• Speaker states that companies using AI agents for everything in that batch are on track to become one of the fastest-growing, most profitable batches in YC history.
• Speaker outlines a 90-day expectation for implementation:
  • Week 1: The library will be thin, skills clumsy, and the user fixing more than saving.
  • Week 4: The flywheel catches, the agent answers with context, and the user writes the third and fourth skills.
  • Week 12: The user has a library answering questions before they are finished, a dozen skill files running dreaded tasks, and one or two tools others want to borrow (potential startups).
• Speaker predicts that most people who try this approach will quit in Week 2.
• Speaker expects that those who do not quit will feel like they are "cheating" by Week 12.

Technology and Product Direction:

• Speaker plans to show that "Personal AGI" is an agent running on user infrastructure, reading from user-owned memory, executing user-written procedures, and compounding.
• Speaker defines the next decade's equation as: a frontier model (rented, commodity, getting cheaper by the quarter) + user context (owned, unique) + a harness (e.g., OpenClaw, Hermes Agent, Cloud Code, Codex).
• Speaker expects the intelligence of this system to get better every single day the user uses it, as opposed to corporate AGI which only improves when the company ships.
• Speaker predicts that software no longer needs to be precious; tools for an audience of one can become entire companies.
• Speaker plans to open source the harness, brain architecture, skills, and whole personal operating system because they can (due to YC backing) and believe it prevents a priesthood of leverage.
• Speaker expects that markdown files act as code where the compiler is a language model, allowing non-engineers to build skill files.
• Speaker predicts that computation will be split between latent space (taste judgment, vague requests) and deterministic space (arithmetic, SQL, specific data).

Market and Industry Outlook:

• Speaker expects that the "fastest growing founders" are treating AI as a workforce, creating leverage not from model weights but from context and relevance.
• Speaker predicts that the new physics of startups allows one founder to do "unscalable things at scale" through agents.
• Speaker expects that companies built natively on this new physics (e.g., Emergent, Retail) will break old math regarding revenue per person.
• Speaker predicts that if you are not using this architecture, your competitor will "eat your lunch politely and thank you for it."
• Speaker expects that the entire workforce of the future will be "agents all the way down."
• Speaker believes that a better model makes a user's library worth more because the "smarter reader extracts more from the same books."
• Speaker predicts that the race for leverage is now won on the "driver and the map" (context and harness) rather than the engine (model weights).
• Speaker expects that the gap between those with the technology and those without will continue to widen monthly.
• Speaker predicts that "7,000 strivings" that previously died waiting for funding or permission will now go straight to work without intermediaries.

Company Plans:

• Speaker plans to give away code that others will not write, build buildings others will not, and fund people others want to ignore.
• Speaker expects to continue operating "in the open" to facilitate a renaissance in leverage.
• Speaker plans to demonstrate that a "brain" (library) starts as a folder and grows to a compounding library of 25 years of life (diarized emails, meetings, photos).
• Speaker plans to show that a single person can now operate with the leverage of a large team, as seen in companies like Emergent (15 people, nine figures revenue in 8 months) and Retail (40 people, $60 million annualized).

Financial Guidance:

• Speaker does not provide specific financial guidance or revenue targets for the future beyond citing past examples of high revenue per person in the YC portfolio.
• Speaker notes that the "1,000 guilders" offer (a salary to stop building) is a recurring offer in the modern context of comfortable corporate arrangements.

Risks and Caveats:

• Speaker warns that a brain without curation is a "garbage dump with great search."
• Speaker warns that retrieval will surface stale facts with total confidence without provenance and contradiction checks.
• Speaker warns that a bad skill file encodes a bad process forever.
• Speaker fears that if cognition is not owned by the individual, it can be extracted and versioned by an employer, resulting in the worker having no career and the company keeping the expertise.
• Speaker highlights the risk that skill files (judgment) can become a "piece of your cognition" that is owned by the company rather than the individual, effectively an "extraction" rather than a career.
• Speaker acknowledges the risk of data leakage in a consolidated personal system but argues that custody on personal infrastructure is a better security model than trusting scattered cloud providers.
• Speaker warns against treating the system like a "dumping ground," which results in a very confident agent that is wrong in untraceable ways.

Confidence and Disagreement:

• Speaker asserts with high confidence that "markdown is actually code" if instructions are clear.
• Speaker states with certainty that the "physics of all startups... did" change with the advent of agents.
• Speaker is certain that the "only question" in the future is who controls the skill files (the individual or the company).
• Speaker expresses confidence that "one person, no intermediaries, no permission" is now possible.
• Speaker believes that "every problem where you thought I wish I had this person... you can" solve it with the described architecture.
• Speaker disagrees with the notion that "Personal AI" means a chatbot you pay $20/month for.
• Speaker disagrees with the idea that the harness is obsolete; instead, they believe better models make the harness more critical as the differentiator moves to context.