↓ Skip to main content
tomrochette.com
  1. Agents/
  2. People and publications/

Andrej Karpathy

Author
big-pickle, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~902 words)
agent-curated 310 developer 8 fully-ai-generated 410 llm=big-pickle 17 llm=glm-5.3-flash 321 model-researcher 1 people 32 publications 32 research-note 280
Table of Contents

Andrej Karpathy names the eras: he coined “vibe coding”, drew the line between vibe coding and agentic engineering, and reframed the current era as Software 3.0 in his 2026 essays.

Karpathy is less a chronicler than a vocabulary-setter: his Software 3.0 framing, that the context window is now the program and verifiability defines what AI can automate, is the lens this whole section’s subject matter is read through.

What it is
#

A founding member of OpenAI, former Director of AI at Tesla, and now (May 2026) a member of Anthropic’s pre-training team. He publishes conceptual essays (Software 2.0, Software 3.0, the LLM-as-OS point) and open-source teaching projects (micrograd, nanoGPT, AutoResearch), and ran the AI-education startup Eureka Labs before joining Anthropic. His public output is the canonical reference for how software engineering itself is being reorganized around agents.

Status
#

Active and central. As of 2026-10-04 he is at Anthropic’s pre-training team, refreshing his thesis in the Sequoia Ascent 2026 essay that defines Software 3.0. His Verifiability essay (2025-11-17) lays out verifiability as the automation boundary, and his AutoResearch project demonstrates agents running research loops overnight, so his writing and his builds move together. His Bear blog had an outage: on 2026-10-03 every karpathy.bearblog.dev URL, including the Sequoia Ascent and Verifiability essays, returned 404 while karpathy.ai still called it his current blog. By 2026-10-04 the blog was serving again with all posts intact, so both essays read at their live URLs, and the restored post list still ends at the Sequoia Ascent essay (2026-04-30), so nothing new surfaced on it in the recent window.

Strengths
#

  • He sets the framing the rest of the field adopts: Software 3.0, agentic engineering, the context-window-as-program model.
  • His theses are concrete and falsifiable, tied to the models he works on and the code he ships.
  • Draws the practical line between vibe coding (prompt and pray) and disciplined agentic engineering, a distinction this section’s notes presuppose.
  • Pairs strong conceptual writing with working open-source implementations.

Cautions
#

  • He sits at the frontier labs, so his claims about agent capability reflect frontier-model economics, not the typical mid-market reality.
  • Frequency is low: he publishes essays and talks rather than a daily or weekly cadence, so he is not a current-events feed.
  • His vocabulary-setting role means he is widely paraphrased; verifying against his primary writing matters.

Pricing
#

Free. Essays, talks, and repositories are all public and free, with no newsletter paywall.

Compared to
#

  • Steve Yegge: both name the era and build code; Karpathy from frontier research, Yegge from shipping in public.
  • Simon Willison: the daily practitioner notes versus the periodic conceptual essay; read both for currency and for vocabulary.
  • Latent Space: Latent Space gives his Software 3.0 talk the interview treatment and spreads the framing.

Bottom line
#

Recommended for engineers who want the durable concepts behind the daily tool churn, from the person who keeps naming the eras. Not for people who need a current-events feed or want vendor-neutral, non-frontier framing.

Top 5 recommended reading #

  • Sequoia Ascent 2026 summary - his Software 3.0 framing, the verifiability boundary, and the vibe coding versus agentic engineering line in one transcript (the live Bear blog again as of 2026-10-04 after a one-day outage).
  • Verifiability - the short essay arguing AI automates what you can verify, which explains the jagged frontier.
  • LLM Wiki - the pattern where an agent incrementally compiles sources into a persistent wiki instead of re-deriving answers.
  • Neural Networks: Zero To Hero - the build-it-from-scratch course that shows what the models agents drive actually are.
  • nanoGPT - the readable few-hundred-line training and model code that most hobby GPT reproduction work builds on.

Changes
#

  • 2026-08-29 - Created in the People and publications category, profiling the vocabulary-setter behind vibe coding, Software 3.0, and agentic engineering.
  • 2026-09-02 - Corrected the verifiability essay reference to the 2025-11-17 “Verifiability” essay.
  • 2026-09-24 - Added the Top 5 recommended reading section.
  • 2026-10-03 - His karpathy.bearblog.dev went dark (404 on the homepage, /blog/, and both essays) while karpathy.ai still lists it as his current blog; the Sequoia Ascent and Verifiability links in Status, Top 5, and References now point to Internet Archive snapshots (2026-09-25 and 2026-08-30), both fetched this run.
  • 2026-10-04 - The Bear blog returned: karpathy.bearblog.dev serves again with no posts lost, so the Sequoia Ascent and Verifiability links in Status, Top 5, and References point at the live URLs again, the restored post list re-verified (newest remains the Sequoia Ascent essay, 2026-04-30), and GitHub, karpathy.ai, and the Hacker News record showed nothing new from him in the 2026-09-24 to 2026-10-04 window.

See also
#

References
#