Andrej Karpathy

See why he is an Architect

• In an essay published on 11 November 2017, called “Software 2.0”, Andrej Karpathy described neural networks as a different way to write software. Instead of writing every rule by hand, people train a model on data. He wrote that “the optimization can find much better code than what a human can write”. Karpathy named the pattern before building a way to teach it. [2]
• At a CVPR workshop keynote in June 2021, Karpathy said Tesla had started shipping cars without radar. “We deleted the radar and are driving on vision alone in these cars,” he said. He argued that the camera-based system was about 100 times better than radar, and that radar had started to “contribute noise” instead of helping. [3]
• In 2020, Karpathy released micrograd, a back-propagation engine in about 100 lines of plain Python. In January 2023, he released nanoGPT, which trains a working GPT-2-class language model in about 300 lines for the training code and 300 more for the model. Both projects are small enough for 1 person to read in a single sitting and understand how the pattern works. [8]
• On 16 July 2024, Karpathy announced Eureka Labs, a school built around a Teacher + AI model. A teacher designs the course, while an AI assistant personalises it for each student. Its first course, LLM101n, asked students to build a working chatbot-style language model from scratch in Python, C and CUDA. Karpathy created a company to help people learn the pattern, not only to demonstrate it. [5]
• On 19 May 2026, Karpathy joined Anthropic's pretraining team under Nick Joseph. He joined the team to use Claude to speed up training for future models. He said he remained deeply committed to education and planned to return to it, after pausing his work running Eureka Labs to return to frontier research. [6]

What you can learn from his approach

What to watch for with the Architect approach

• The June 2021 keynote announcing Tesla's vision-only approach came shortly before reports of sudden, unexplained braking rose from 99 at the end of 2021 to 314 by February 2022. The National Highway Traffic Safety Administration opened a formal investigation covering about 695,000 Model 3 and Model Y cars. [7]
• NHTSA linked the timing of those reports to Tesla's mid-2021 move from radar-plus-vision sensing to vision-only sensing. This was the same change Karpathy had presented as an improvement. [7]
• NHTSA closed the investigation on 2 July 2026, 4 years after it opened. It cited Tesla's 2022 firmware updates, a sharp fall in reports and a finding of “low hazard to drivers and occupants”. The design was not a lasting failure, but it still required years of real-world data and federal scrutiny before the investigation could close. [7]

3 practices you can use

1. 1

   Build a complete small version first. Before you teach or scale an idea, make the smallest version that proves it works. micrograd shows the core idea behind a modern neural network in about 100 lines that a beginner can read in 1 sitting. [8]

2. 2

   Write down the pattern. Publish a clear explanation of the pattern before you build a company around it. “Software 2.0” was a freely available essay, not a product pitch, and it appeared years before Eureka Labs. [2]

3. 3

   Explain difficult decisions clearly. When you make a hard technical decision, explain the reasoning in enough detail for other engineers to assess it. Karpathy's CVPR keynote explained why Tesla removed radar, not only that it had done so. [3]

Avoid calling a pattern finished too soon

Sources

Every numbered claim on this page links to one of these.

1. [1] Andrej Karpathy
2. [2] Software 2.0
3. [3] Tesla AI chief explains why self-driving cars don't need lidar
4. [4] karpathy/nanoGPT
5. [5] After Tesla and OpenAI, Andrej Karpathy's startup aims to apply AI assistants to education
6. [6] OpenAI co-founder Andrej Karpathy joins Anthropic's pre-training team
7. [7] Tesla Phantom Braking Investigation Closed by Feds
8. [8] Andrej Karpathy's Nano Repos: 120,000 Stars and One Teaching Philosophy