Andrej Karpathy's talk on “Software 3.0”, with LLMs enabling programming via natural language, the decade of agents, LLMs as “fallible people spirits”, and more
Karpathy has previously framed LLMs as lowering barriers to software creation in a broader case for a new software cycle. This keynote turns that thesis into a practical vocabulary for founders: natural-language programming, agentic systems, and models that remain unreliable.
The talk also follows his move to build an AI-supported education venture, Eureka Labs, making the emphasis on how people learn to work with models—rather than simply consume their output—especially salient.
First-order effects
AI Startup School attendees get a concrete framing for building with LLMs: use natural language as a programming interface while treating model output as fallible and subject to human judgment.
The keynote elevates agent-building as a near-term product focus for startups, while explicitly cautioning against treating LLM behavior as deterministic software.
Second-order effects
Teams pursuing agentic products are pushed toward workflow design, evaluation, and human review, not just prompting or model selection; that favors products embedded in repeatable work.
Natural-language creation can widen who can prototype software, intensifying competition around execution quality and operational reliability rather than access to coding alone.
Third-order effects
If this pattern holds, software development shifts from writing every instruction directly toward supervising, constraining, and validating model-generated work—a core form of workflow-native AI.
The “fallible” framing suggests durable demand for governance and verification layers around agents, even as their capabilities improve; autonomy and accountability will advance at different speeds.
The trend: This is one data point in the shift from LLMs as chat interfaces to agentic, human-supervised systems embedded in software and business workflows.
Nice - my AI startup school talk is now up! Chapters: 0:00 Imo fair to say that software is changing quite fundamentally again. LLMs are a new kind of computer, and you program them *in English*. Hence I think they are well deserving of a major version upgrade in terms of
Andrej Karpathy says self-driving felt imminent back in 2013 — but 12 years later, full autonomy still isn't here “there's still a lot of human in the loop” he warns against hype: 2025 is not the year of agents; this is the decade of agents [video]
Some of the links: - My slides as keynote: https://drive.google.com/... - Software 2.0 blog post from 2017 https://karpathy.medium.com/ ... - How LLMs flip the script on technology diffusion https://karpathy.bearblog.dev/ ... - Vibe coding MenuGen (retrospective) https://karpathy…
Andrej Karpathy says that when state-of-the-art LLMs go down, “it's actually kind of like an intelligence brownout in the world. The planet just gets dumber the more reliance we have on these models, which already is like really dramatic.” [video]
Andrej Karpathy: Software Is Changing (Again) Key learning points from this brilliant lecture from yesterday. 🚀 The Shifting Software Map For 70 years code flowed in one style, then neural networks arrived and rewrote large patches of logic. Karpathy divides eras into [image]
🧵 @karpathy dropped the most compelling lecture at the AI startup school :) Here's my notes summarized with Opus. Main thesis: We're not in the “year of agents” — we're in the DECADE of agents. Here's the historical arc that explains why... ==== The Software Evolution Story [imag…
Everyone's hyped about autonomous agents (which fail so often. don't complete the entire task. something breaks) But Andrej Karpathy makes a dead-simple point: we've got the loop backwards. Al is fast at generating. Humans are slow at verifying. So speed doesn't come from
In case you forget @karpathy's vocabulary that will run the next decade: - autonomy slider - vibe coding - agents as an iron man suit - his pop culture references - demo-to-product gap - semi-autonomous agents [video]
.@karpathy shouted out my work on @exolabs at @ycombinator AI SUS! “we use LLMs similarly to mainframes in the '70s - compute is timeshared by having a slice in the batch dimension. models will compress over time, and with this we'll be able to run more on-device” [image]