Keynote from Y Combinator AI Startup School • San Francisco • 17 June 2025

---

Why this talk matters

In “Software Is Changing (Again)” Andrej Karpathy argues that large-language-model agents mark a third great epoch of software:

Wave	Interface	What engineers write	Compute target
1.0	Compilers/CPUs	Imperative code (C++, Java)	Deterministic machines
2.0	ML frameworks	Data + model architecture	Neural nets
3.0	LLMs-as-OS	Prompts & feedback loops	Stochastic “people spirits” running in the cloud

Karpathy’s core claim: “The hottest new programming language is English.”  

---

What “agentic AI” means in practice

Property of LLM agents	Engineering consequence
Jagged intelligence – super-human recall, sub-human arithmetic	Add verify loops & type-checkers around every call
Utility / fab / OS triple-role – expensive to build, cheap to consume	Treat model labs like cloud vendors; architect for fail-over & rate limits
Programmed in natural language	Shift effort from syntax → prompt design & feedback heuristics
Partial autonomy sliders	Design UIs that expose diff views and “confidence knobs” (Cursor, Perplexity, etc.)
Agents as first-class users	Ship llms.txt, structured markdown, and self-describing APIs so bots—not just humans—can consume your product

---

Five new pillars of Software 3.0 engineering

1. Prompt-oriented architecture
   • Store prompts under version control, test them like code.
2. Guard-railed execution
   • Wrap every model call in validators (regex, unit tests, type checks).
3. Tight generate/verify cycles
   • Latency budgets must include detours for automatic critique or human review.
4. Agent-friendly infrastructure
   • Documentation readable by both humans and parsers; think OpenAPI plus llms.txt.
5. Observability for cognition
   • Log token streams & system messages; inspect why an agent acted, not just what it returned.

These practices operationalise Karpathy’s warning that LLMs are “people spirits”—creative but error-prone；they demand DevOps-for-cognition, not just DevOps-for-code.  

---

Opportunities unlocked by agentic AI

Opportunity	Example today	Why it’s viable now
One-person MVPs (“vibe coding”)	MenuGen demo: full UI via prompts	LLM handles boilerplate; human supplies vision
Autonomous research & data pull	Perplexity Deep Research agents	Cheap in-context search + summarisation
Continuous code-base refactoring	Internal GPT-powered linters deleting dead paths at Tesla	Model can reason over millions of LOC
Agent-to-agent protocols	DeepWiki, GitIngest	Structured docs allow bots to traverse knowledge

---

Risks & open questions

• Reliability debt –  Hallucinations trade raw speed for debugging overhead.
• Platform centralisation –  Cloud LLMs resemble 1960s mainframes; open-weight models may rebalance power later.  
• Skill displacement –  Traditional “middleware” coding shrinks, but prompt engineering, evaluation and agent-safety grow.
• Security surface –  Agents that read/write code can exfiltrate secrets or commit “prompt injection” supply-chain attacks.

Karpathy advocates a “keep AI on the leash” stance—tight scopes, incremental release, human-in-the-loop—mirroring comments he reiterated in press interviews.  

---

How to prepare your team

1. Inventory high-leverage prompts –  Identify workflows where English beats code.
2. Layer tests –  Write property-based tests that re-prompt until they pass.
3. Instrument everything –  Token-level logs + embeddings let you analyse failures post-hoc.
4. Upskill in evaluation –  Learn metrics like graded self-consistency (GSC) or LM-confidence via entropy.
5. Pilot “agent-first” modules –  Start with side-cars (doc summariser, migration assistant) before core logic.

---

Verdict on the keynote

Strengths

• Clear, memorable mental model (Software 1.0-2.0-3.0) that frames LLMs as a new compute substrate.
• Concrete engineering anecdotes (Autopilot diff-slider, MenuGen) that ground the hype.
• Honest about limitations—“jagged intelligence” demands guardrails.

Gaps

• Little discussion of on-device LLM inference and its impact on edge privacy.
• Tooling section glosses over non-code disciplines (design, legal) that will also face agentic disruption.

Overall: A must-watch thesis for technologists. It’s less a product roadmap than a design brief for building robust, human-aligned AI agents—and a wake-up call that the next decade of software will be written as much in English as in Python.