Andrej Karpathy
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 10
- Karpathy's viral line: 'The hottest new programming language is English'
TechRadar's quote-of-the-day column revisits OpenAI co-founder Andrej Karpathy's viral post on X: "The hottest new programming language is English" — a reference to natural language becoming the main way to direct AI in writing code. According to the article, Karpathy added in a follow-up that many examples in the scientific literature show natural language can wield AI for complex programming, and the piece links this to tools that later became widespread, such as GitHub Copilot and Claude Code. The article also argues that as vibe coding and instruction-based development spread, developers increasingly review rather than write code, potentially reducing demand for juniors and permanently changing the career pipeline — these are the outlet's commentary, and no new model release or data is presented.
TechRadar · 🔥 9 - Databricks details how its AI research team uses AI Runtime for agent-led training
Databricks published a post describing how its AI research team uses its own AI Runtime (AIR) to train and evaluate the models behind products such as Genie, Smart Routing and AI Functions. The team cites AIR capabilities including GPU preflight checks and health monitoring, cordoning off unhealthy nodes, the AIR CLI with custom Docker containers, automatic MLflow experiment tracking, governed data access through Unity Catalog FUSE mounting, and workload prioritization and preemption to share GPUs across projects. Databricks says agent-led "autoresearch" workflows have driven an 8x increase in the number of training jobs for its AI Functions research, with models ranging from a few billion to nearly a trillion parameters trained on NVIDIA H100 or B300 GPUs, and the largest runs using hundreds of GPUs over days or weeks.
Databricks Blog · 🔥 4 - Karpathy: only ~5,000 people can access frontier AI; 6 billion never touched an LLM
Andrej Karpathy said in a post that only about 5,000 people have access to frontier internal AI systems, roughly 6 billion people (about 75% of the population) have barely touched a large language model, another 1–2 billion occasionally use free ChatGPT-style products, and about 20 million (around 0.2%) use frontier models deeply in areas such as math and coding. He argues frontier capability only shows up on large, complex, long-horizon problems, and that reinforcement learning's need for verifiable feedback is why math and code advance first while open-ended writing lacks clear signal. The article also cites a16z data showing the top 1% of paying users generate 19.5% of consumer AI revenue, says ChatGPT has about 810 million weekly users versus roughly 35 million paying ones, and notes METR's estimate that agent capability doubles about every seven months, while an early-2025 randomized experiment found experienced developers familiar with their own codebase took 19% longer on tasks when using AI tools.
36氪 人工智能 · 🔥 0 - Karpathy: AI gap has become a sharp funnel, only ~5,000 see the frontier
OpenAI co-founder Andrej Karpathy said on X that the AI understanding gap he flagged earlier is widening, and is now less "two groups speaking past each other" than a "sharp funnel." In rough napkin estimates he splits users into four tiers: about 6 billion people (~75% of the population) who have barely encountered LLMs; roughly 1–2 billion casual users of free-tier ChatGPT-like products; some 20 million math-and-code professionals who see firsthand that agents can finish projects that used to take weeks or months; and about 5,000 people with internal access to frontier-grade systems. He attributes the funnel to three factors: impact scaling with problem size and horizon, the jaggedness of capability concentrated in digital, verifiable, economically valuable domains, and free/paid/internal access differences — adding that the outside world has mainly seen a "preview," while human review and comprehension are starting to fall behind.
科技新报 · 🔥 0 - Google pilots coding interviews that let candidates use Gemini
Google is piloting a software engineer interview change in which candidates in a new “Code Comprehension” section can use Gemini to read, debug and optimize existing code, while interviewers assess “AI fluency” — how candidates prompt, verify AI output and debug, according to a report republished by 36Kr from Machine Heart. A Google spokesperson confirmed Gemini is the tool offered to candidates, and VP of recruiting Brian Ong said the change aims to better reflect how engineers work in the AI era; the pilot covers junior and mid-level US roles in the second half of this year, with Google Cloud and Platforms & Devices among participating organizations, and expansion depending on results. The article also cites Google CEO Sundar Pichai saying 75% of new code at Google is AI-generated and reviewed by engineers, Andrej Karpathy saying roughly 80% of his coding is now done by agents while his hand-writing ability has declined, and similar AI-in-interview moves by Canva and Meta, plus Codility allowing interviewers to inject hidden instructions into candidates’ AI coding tools.
36氪 人工智能 · 🔥 0 - Karpathy: the outside world has only seen the preview of AI
Andrej Karpathy, an OpenAI co-founder now at Anthropic, described a widening gap between what the public sees of AI and what internal frontier systems can already do, splitting LLM users into four tiers. He estimates about six billion people (roughly 75% of the world) have barely touched LLMs, one to two billion use them casually as a better search engine or writing aid, about 20 million professionals in math and code use frontier models for work, and roughly 5,000 people have access to internal frontier systems. That last group sees swarms of thousands of agents collaborating over weeks on software mega-projects, minting zero-days, running cyber attacks and defenses at machine speed and advancing science and mathematics: "The external world has seen the preview," he writes, while human review and comprehension fall behind. Karpathy says capability is peaking in digital, verifiable and economically valuable domains, and cites Sam Altman's similar claim that a coding tool did in 17 minutes what a startup built in three months.
The Decoder · 🔥 0 - Karpathy: Only ~5,000 people worldwide see frontier labs' strongest internal AI
Andrej Karpathy sketched an AI capability "funnel": of roughly 8 billion people, about 6 billion have barely touched large models, a billion-plus use free chat or search occasionally, and only around 20 million use frontier models for complex projects — while just ~5,000 (about 0.00006%) can access the strongest internal systems at frontier labs. He described the numbers as his own rough estimates, and the article says Microsoft, a16z and OpenAI reports point to similar divides: Microsoft put generative-AI use among ages 15-64 at 18.8% globally in Q2 2026, a16z said only 4.5% of eligible US consumers pay for ChatGPT, Gemini or Claude, and OpenAI said top enterprises produce 8.3x more tokens per person than average firms. The piece also notes Karpathy's own shift toward letting agents write most of his code and says he joined Anthropic's pretraining team in May.
虎嗅 AI · 🔥 0 - Karpathy: a four-tier 'cognitive funnel' is splitting AI perception
Andrej Karpathy offered a "napkin math" estimate that roughly 75% of the world's population (about 6 billion people) has barely touched an LLM, while another 1–2 billion are occasional users of free ChatGPT-style products and therefore see AI as a "clumsy chatbot." He argues that some 20 million math and coding professionals are experiencing "agent vertigo" from paid frontier agents — especially OpenAI Codex and Claude Code — while at the very top only about 5,000 insiders have seen thousands of agents collaborate autonomously for weeks, hunt zero-days and push frontier mathematics. He attributes the split to three factors: the ambition and scale of your problems, the verifiable rewards available in code and math that suit reinforcement learning, and the access gap between free, paid and internal tiers. The population figures are his own estimates, not audited statistics.
36氪 人工智能 · 🔥 0 - Karpathy: the gap in shared understanding of LLM capability is widening
Andrej Karpathy posted on X that the gap in shared understanding of LLM capability is widening, and the post was reshared on Hacker News, where it drew 100+ points and a comment thread. The only source material available is the post's title and a "Comments on Hacker News" note; the post's specific arguments, examples and data are not included. It therefore reads as a discourse item about how model capability is judged and communicated, rather than a new factual announcement.
Hacker News · AI（100+ 分） · 🔥 0 - Karpathy-inspired CLAUDE.md guidelines for Claude Code trend on GitHub
An open-source project called andrej-karpathy-skills is drawing attention on GitHub Trending for using a single CLAUDE.md file to improve Claude Code's behavior, derived from Andrej Karpathy's observations about common LLM coding pitfalls. It lays out four principles — think before coding, simplicity first, surgical changes, and goal-driven execution — and can be installed via the Claude Code plugin marketplace or by downloading the CLAUDE.md into a project. The page also promotes the author's new open-source Multica platform for running and managing coding agents with reusable skills.
GitHub Trending（每日） · 🔥 0
Experience and discussion from the community
Share my Andrej Karpathy experienceAsk about Andrej Karpathy
Nobody has shared their experience with Andrej Karpathy yet.