karpathy.bearblog.dev
Public MindPattern findings, entities, and graph evidence that cite this source.
Findings
7
All-time hits
7
High value
3
Last seen
2026-06-30
Connected entities
karpathy.bearblog.devKarpathy Publishes 'Software 3.0' Manifesto — December 2025 as the Inflection Point Where Code-WritiKarpathy BlogPattern: Karpathy's 'Software 3.0' — The Context Window Becomes the Primary Programming InterfaceTip: Apply Karpathy's 'Automate What You Can Verify' as Agent Task Selection Heuristic
Related findings
- 2026-06-30 / VOICESKarpathy Reframes the Discipline: From 'Vibe Coding' to 'Agentic Engineering'In his Sequoia Ascent 2026 talk, Karpathy positions agentic engineering as the serious discipline that must grow on top of vibe coding so professional software keeps its quality bar — 'Software 3.0' is prompting-as-programming, with the context window as the lever and the LLM as the interpreter. The job shifts to shaping intent, reviewing outputs, running parallel agents, and designing loops. This is the conceptual scaffolding underneath his expanded CLAUDE.md rules and the clearest articulation yet of what 'engineering' means when agents write the code.
- 2026-06-19 / SOURCESKarpathy Formalizes 'Software 3.0': Automating What Humans Can VerifyAndrej Karpathy published his Sequoia Ascent talk summary on June 17, formalizing 'Software 3.0': where 1.0 automates what humans can specify as rules and 2.0 automates what we can describe with training data, 3.0 automates what humans can verify — anything checkable by a test suite, game score, or proof checker. He pins December 2025 as the inflection point when agentic coding crossed from experimental to reliable. The actionable takeaway for builders: design workflows around verifiability — invest in tests, evals, and checkable specs, because that is now the boundary of what you can safely hand to an agent.
- 2026-05-17 / VOICESKarpathy Publishes 'Software 3.0' Manifesto — December 2025 as the Inflection Point Where Code-Writing Ratios InvertedAndrej Karpathy published a detailed blog post (April 30) summarizing his Sequoia Ascent 2026 fireside chat, defining three software eras: 1.0 (human-written code), 2.0 (neural networks), and 3.0 (LLMs programmed through prompts, context, tools, memory). He pinpointed December 2025 as when his personal ratio inverted from writing 80% of code to delegating 80% to agents. The verifiability framework — 'traditional software automates what you can specify; LLMs automate what you can verify' — provides builders a concrete heuristic for where AI coding works and where it doesn't.
- 2026-05-06 / VOICESAndrej Karpathy at Sequoia Ascent: 'I Haven't Written a Line of Code Since December 2025'In his Sequoia AI Ascent 2026 fireside chat, Karpathy outlined the Software 3.0 paradigm — programming through prompts and context windows rather than explicit code. He pinpointed December 2025 as the inflection point where 'generated chunks got larger, more coherent, and more reliable,' shifting the unit of programming from writing lines to delegating macro actions. His verifiability thesis explains why coding improves faster than creative work: traditional software automates what you can specify; LLMs automate what you can verify. Key quote: 'You can outsource your thinking, but you can't outsource your understanding.'
- 2026-05-04 / TOOLSPattern: Karpathy's 'Software 3.0' — The Context Window Becomes the Primary Programming InterfaceKarpathy's Sequoia Ascent framework positions context engineering as the successor to prompt engineering: in Software 3.0, what goes into the context window IS the program. CLAUDE.md files, skill definitions, memory systems, tool schemas, and structured examples are all 'source code' for agent behavior. The practical implication: invest in context architecture (what loads, when, in what order, with what priority) the same way you'd invest in code architecture. Teams building agentic systems should treat context as a first-class engineering artifact with version control, testing, and review.
- 2026-05-04 / TOOLSTip: Apply Karpathy's 'Automate What You Can Verify' as Agent Task Selection HeuristicFrom the Sequoia Ascent talk: tasks with automatic reward signals (math, coding with tests, linting, type checking) improve rapidly because they're resettable and measurable. The practical implication: before delegating work to an agent, ask 'can I verify correctness automatically?' If yes (test suite exists, type checker validates, CI passes), delegate aggressively. If no (design decisions, architecture choices, user experience), keep human judgment in the loop. This explains why TDD-first agent workflows dramatically outperform vibe-coded ones.
- 2026-05-03 / VOICESKarpathy Publishes Sequoia Ascent 2026 Summary: 'You Cannot Outsource Understanding' as AI Coding MaturesAndrej Karpathy published a blog post summarizing his Sequoia AI Ascent 2026 talk, detailing his evolution from vibe coding to agentic engineering. He built MenuGen — an app using Gemini to overlay AI-generated food images onto restaurant menu photos — and noted that with latest models, 'code chunks come out fine' to the point where he 'started trusting the system more and more.' His key thesis: understanding becomes the bottleneck, not code production. 'You still need enough depth to direct the system.'