Three of Andrej Karpathy's open-source repositories, nanoGPT, nanochat, and micrograd, have accumulated more than 120,000 GitHub stars combined, making him one of the most-followed technical educators in machine learning. Two years to the day after founding Eureka Labs, that body of work has brought him to Anthropic's pretraining team.
The Nano* Philosophy: Teaching AI by Removing Everything Extra
The nano* series rests on a single conviction: the fastest route to genuine understanding is a complete, working implementation that fits on one screen. Karpathy introduced micrograd in 2020 as a compact, scalar-valued autograd engine: a minimal backpropagation implementation in pure Python with no external dependencies. The repository now has more than 16,000 GitHub stars and is widely cited as the most approachable entry point to gradient-based learning in the field.
nanoGPT, released in January 2023, applied the same reduction to transformer language models. Its README describes the project as "the simplest, fastest repository for training/finetuning medium-sized GPTs." It accumulated more than 60,000 GitHub stars and became one of the most-forked machine-learning repositories on the platform, enabling a generation of practitioners to learn the transformer architecture through direct implementation rather than reading papers alone.
nanochat arrived in October 2025, extending the lineage to a full ChatGPT-class system. By January 2026, the repository documented a training run that reproduced a GPT-2-class conversational model for roughly $73 on a single eight-GPU H100 node in three hours, putting full-stack AI model training within reach of individual researchers on a limited budget. nanochat surpassed 43,000 GitHub stars. In February 2026, Karpathy published microgpt on his personal blog: a single Python script of 200 lines, no dependencies, that both trains and runs inference on a GPT from scratch. He described it on his blog as "the culmination of multiple projects (micrograd, makemore, nanogpt, etc.) and a decade-long obsession to simplify LLMs to their bare essentials."
Eureka Labs: Two Years of Scaling the Thesis
On July 16, 2024, exactly two years before today, Karpathy announced Eureka Labs on X, describing it as "a new kind of school that is AI native." The company's thesis was that post-2022 generative AI made it viable to build AI tutors capable of scaling a single domain expert's teaching capacity to an unlimited number of learners, at a quality level that prior ed-tech could not match.
The daily brief
Get the free daily brief with the day's AI news, new AI startups and funding rounds every morning.
Join 64,000 other subscribers.
The company raised approximately $20 million in seed funding from Conviction, the venture firm led by Sarah Guo, alongside Sam Altman and a group of AI-adjacent angel investors, according to reporting by TechCrunch and VentureBeat. The first product, LLM101n, was a 17-chapter undergraduate-level curriculum for building a "Storyteller" language model end to end in Python, C, and CUDA. The course repository accumulated 36,000 GitHub stars before product development was paused.
Eureka Labs ran for 22 months as an independent company. When Karpathy announced the transition to Anthropic in May 2026, he framed it carefully, stating that he "remain[s] deeply passionate about education and plan[s] to resume my work on it in time," per his X post on May 19, 2026. The team and technical assets remained intact at the point of his departure.
Before Eureka Labs launched any formal product, Karpathy's YouTube lecture series, including a four-hour walkthrough titled "Let's reproduce GPT-2 (124M)," established an informal curriculum that LLM101n was designed to formalize at scale. That sequence of lectures pre-dates the company but represents the direct intellectual precursor to the Eureka Labs model.
Pretraining as the Next Classroom
On May 19, 2026, Anthropic announced that Karpathy had joined its pretraining team, reporting to Nick Joseph, who leads pretraining research at the company. The stated mandate was to build a team that uses Claude to accelerate the next generation of pretraining research, applying the company's existing model as a research instrument rather than solely as a product, per CNBC. In his own announcement on X, Karpathy wrote: "I think the next few years at the frontier of LLMs will be especially formative."
The hire arrives as Anthropic moves toward a public market debut. The company recently scheduled investor meetings for an October 2026 Nasdaq listing, as reported by this publication. Adding a figure with Karpathy's technical credibility to the pretraining organization reinforces its positioning as a research-first company ahead of that process.
The connection between the nano* educational series and a role in frontier pretraining is not coincidental. Karpathy's teaching has always been a proxy for his research instincts: every problem reduced to its minimal reproducible form, understanding verified through working code, then rebuilt upward. At Anthropic, that instinct is now directed at the most compute-intensive phase of LLM development, pretraining, where scale and rigor must coexist. The question his appointment implicitly poses is whether the nano* sensibility, clean, from-scratch, auditable, transfers to resource envelopes measured in hundreds of millions of GPU-hours.
What it means
The nano* series represents a sustained bet that comprehension in AI is a precondition for progress, not a by-product of it. With more than 120,000 combined GitHub stars across three repositories, a 17-chapter open curriculum, and a 22-month education startup behind him, Karpathy arrives at Anthropic with a track record of making frontier concepts legible at scale. Whether that ability shapes how the next Claude generation is trained, or whether Eureka Labs eventually resumes on the other side of a pretraining chapter, the two threads of his career have always been the same thread: build it from scratch until you understand it.
Sources
CNBC: Anthropic hires OpenAI co-founder Andrej Karpathy, former Tesla AI leader
TechCrunch: OpenAI co-founder Andrej Karpathy joins Anthropic's pre-training team
VentureBeat: Ex-OpenAI and Tesla engineer Andrej Karpathy announces AI-native school Eureka Labs
GitHub: karpathy/nanoGPT
Andrej Karpathy: microgpt (personal blog, February 2026)
Editorial standards: every claim is sourced. Tips: [email protected]
