What happened
Andrej Karpathy, formerly of OpenAI and Tesla, told a Sequoia Capital audience that by December 2024 he felt the most behind he had ever been in his life, and defined the difference between what he named バイブコーディング（vibe coding） and more serious agentic engineering.
Why it matters
He says he may not have needed to build any of the apps he made, framing the shift as a paradigm change he learned about through his own failure, and the strongest AI can write 100,000 lines of code yet cannot walk to a car wash — so the capability is jagged rather than uniformly strong.
What to watch
The test is whether users stop over- or under-estimating these systems and instead carry a map of where the gaps are, since thinking can be outsourced but understanding cannot, and agents can produce hard-to-explain bugs such as matching funds by email address.
WHO IT HITSSoftware engineers and engineering hiring managers are the ones who would feel this most directly: the talk recasts programming as deciding what to paste into an agent, and suggests interviews move from puzzle questions to building a Twitter clone that can withstand attacks.
Summaries like this, in your inbox every morning.
The talk is notable less for a product announcement than for the person delivering it. Karpathy built his career at OpenAI and Tesla, so his account of feeling the most behind he had ever been in his life in December 2024 carries the weight of someone who was at the front of AI coding and now describes learning the shift through his own failure. He even says he may not have needed to build any of the apps he made.
He is also the person who coined バイブコーディング（vibe coding）, and in this talk he draws a line between that looser approach and more serious agentic engineering. That distinction pairs with his warning that thinking can be outsourced but understanding cannot — an argument that the ease of generating code does not automatically build the user's comprehension.
The concrete examples keep the talk from becoming abstract. An agent that matches funds by email address is the kind of hard-to-explain bug that shows where the gaps sit, and the image of the strongest AI writing 100,000 lines of code while failing to walk to a car wash is how he describes the unevenness. How much of this lands as practical guidance may hinge on whether teams adopt it as a map of where the gaps are rather than a reason for either overconfidence or dismissal.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
ElevenLabs completed a $300 million employee tender at a $22 billion valuation, and co-founder Mati Staniszews…
On October 15, OpenAI announced the Rosalind Academic Plan, giving approved US university labs free access to…
A Qiita article compared automated red-teaming methods for LLM apps — GCG, PAIR, and TAP — using mock implemen…
Liquid AI released d1-ultra on October 15, 2026, an upgraded open-weights decision model that averaged 76.7 ac…
Circlia started a proof-of-concept for TASTE SIGNAL: visitors to Holland Village Cafe FUKUOKA taste cookies an…
In a webinar hosted by French Mathematical Society president Isabelle Gallagher, Fields Medal winner Hugo Dumi…
