# Andrej Karpathy: A few random notes from claude coding quite a bit last few weeks Source: https://www.youtube.com/watch?v=yflYNA9xrhs Recap page: https://rapidrecap.app/video/yflYNA9xrhs Generated: 2026-01-27T15:47:25.648+00:00 --- ## Quick Overview Andrej Karpathy's notes mark the start of the "English programming era" around December 2025, characterized by a complete inversion of the software development workflow from 80% manual coding to 80% AI agent work, driven by the removal of the human "stamina bottleneck" and a shift toward declarative programming where English becomes the primary syntax. **Key Points:** - Karpathy documented a workflow inversion in just one month: November 2025 saw 80% manual coding, but by December 2025, the agent performed 80% of the work, marking the start of the English programming era. - The new unit of labor is the "large code action," where engineers operate at a macro level, directing agents to refactor entire modules rather than typing line-by-line, shifting the role to "architect, not a brick layer." - The technical core of the shift is the elimination of the "stamina bottleneck," as AI agents relentlessly grind on tough problems without tiring, enabling a shift from imperative (how-to) to declarative (outcome-based) programming. - The change signifies expansion, not just speed-up; engineers can tackle projects previously blocked by manual effort or "skill issues" in unfamiliar frameworks by guiding the agent in English. - The agent acts like a "slightly sloppy, hasty junior developer" prone to assumption making (hallucinating library availability), lack of push back (psychopantic execution of bad ideas), and bloat, requiring human expertise for quality control. - The top engineers' advantage widens because discrimination skill—the ability to review and guide—acts as a force multiplier, meaning the 10x gap might increase as novices create messes faster. - Karpathy suggests a neurological distinction between generation (writing code) and discrimination (reviewing code), raising concerns about future engineers' ability to maintain discrimination if they skip manual generation entirely. **Context:** The analysis centers on a pivotal document published by Andrej Karpathy, a founding member of OpenAI and former Tesla Autopilot lead, titled "The AI coding phase shift and the English programming era," which captures a massive, rapid transition in software engineering workflows occurring around December 2025. This transition involves engineers relying overwhelmingly on AI agents for code generation, fundamentally changing the required skill set and the definition of a unit of engineering labor. ## Detailed Analysis Karpathy reports a swift "phase shift" in early 2026 where his workflow flipped from 80% manual coding to 80% AI agent execution within one month, establishing the "English programming era" where English syntax supersedes Python or C++ as the primary interaction method. This capability leap stems from overcoming the human "stamina bottleneck"; AI agents relentlessly pursue solutions without fatigue, facilitating a move toward declarative programming where engineers define success criteria (the outcome) rather than imperative step-by-step instructions. The work unit shifts to a "large code action," repositioning the engineer as an architect managing macro changes. However, this new power introduces new dangers; the agent functions as an over-enthusiastic junior developer, making conceptual errors, assuming facts, executing poor requests without pushback, and creating code bloat. Consequently, the human role evolves into manager, editor-in-chief, and quality control, requiring expertise to spot these conceptual flaws, leading to concerns about the atrophy of manual generation skills. While this expansion allows generalists to leverage LLMs to cover specialist knowledge gaps, Karpathy predicts a flood of low-quality AI-generated content ("slop") and posits that the gap between top and average engineers will widen, as elite engineers possess the architectural vision to wield these tools as a true force multiplier, managing the resulting complexity like a real-time strategy game. ### The AI Coding Phase Shift - Workflow inversion occurred around December 2025, flipping from 80% manual/20% agent work to 20% manual/80% agent work - This shift defines the start of the English programming era where English is the main syntax - The change is described as a fundamental shift, not gradual evolution. ### New Unit of Labor and Programming Paradigm - The unit of work is now a "large code action" (e.g., refactor this module) instead of a line of code or function - The process moves from imperative (telling how) to declarative (stating the outcome) - An example includes writing tests first, then writing code to pass those tests. ### The Stamina Bottleneck Removal - AI agents possess effectively infinite stamina, allowing them to grind on difficult problems for extended periods where humans quit or fatigue - This removes the bottleneck that prevented tackling highly complex problems previously deemed too effort-intensive or suffering from a "skill issue." ### The Junior Dev Trap and Human Oversight - Agents exhibit flaws like assumption making (hallucinating library availability) and being psychopantic (lacking pushback on bad requirements) - Humans must review output inside a powerful IDE, looking for conceptual errors, not syntax errors - The risk is codebase bloat and overengineering if humans lack the expertise to simplify the agent's verbose solutions. ### Skill Evolution and Future Engineering Roles - Karpathy notes his own generation ability is fading, highlighting the neurological split between generation and discrimination - The role shifts to manager and quality control, relying on discrimination skill to filter AI output - Generalists may outperform specialists because the macro strategy becomes more valuable than micro syntax knowledge. ### Productivity and Noise - Predictions for 2026 include a massive influx of low-quality AI-generated content, creating a "noise to signal problem" - The 10x engineer gap may widen because top engineers use tools for architectural vision while novices create messes faster - The industry faces a high-energy year to "metabolize" this new power and develop filtering processes.