“LLM Knowledge Bases – Something I’m finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest… You rarely ever write or edit the wiki manually, it’s the domain of the LLM.” – Andrej Karpathy – Previously Director of AI at Tesla, founding team at OpenAI, PhD at Stanford
The traditional model of knowledge management-where researchers manually write, edit, and maintain wikis and reference systems-assumes that human curation is the primary value-add in organizing information. This assumption is collapsing. As large language models become capable of synthesizing, organizing, and updating information at scale, the bottleneck in knowledge work is shifting from content creation to content validation and strategic direction-setting.1
The Automation of Knowledge Curation
Andrej Karpathy’s observation about using LLMs to build personal knowledge bases reflects a fundamental change in how researchers interact with information systems.1 Rather than researchers serving as the primary authors and editors of their knowledge repositories, LLMs now function as the active agents in knowledge synthesis, with humans adopting a supervisory role. This inversion-where the LLM becomes the domain of the wiki and humans become the validators-represents a departure from decades of knowledge management practice.
The practical implication is significant: researchers can now maintain comprehensive, up-to-date knowledge bases across multiple domains of interest without the time investment traditionally required for manual curation. An LLM can continuously aggregate new research, synthesize findings, identify connections across disparate sources, and organize information according to specified schemas-all without human intervention in the day-to-day maintenance cycle.
Context: The Broader Transformation of Knowledge Work
Karpathy’s commentary arrives amid a broader recalibration of how AI is reshaping professional work. In early 2025, he articulated a vision of “Software 3.0,” where natural language becomes the primary programming interface and LLMs generate code with minimal human input.2 The knowledge base observation extends this logic: if LLMs can generate functional code from high-level specifications, they can equally generate and maintain structured knowledge from domain parameters and update directives.
This shift reflects Karpathy’s firsthand experience across multiple roles:
- As a founding member of OpenAI, he witnessed the emergence of increasingly capable language models
- As Director of AI at Tesla (2017-2022), he led teams managing vast datasets and neural network training pipelines, where information organization at scale was operationally critical3
- Upon returning to OpenAI in February 2023, he contributed to the development of GPT-4, which demonstrated substantially improved reasoning and synthesis capabilities4
His observation about LLM-driven knowledge bases is not theoretical speculation but a reflection of practical experimentation with tools that have reached a capability threshold where they can reliably perform knowledge synthesis tasks.
The Capability Threshold: Why Now?
LLMs have long been capable of generating text. What has changed is their ability to maintain consistency, follow complex organizational schemas, and integrate new information without degrading existing knowledge structures. Earlier language models could produce plausible-sounding content but lacked the coherence and reliability required for mission-critical knowledge systems. Current models demonstrate sufficient consistency and reasoning capability to serve as the primary authoring layer in knowledge management systems.
The shift also reflects improved prompt engineering and system design. Rather than asking an LLM to write a wiki article once, researchers can now:
- Define a knowledge base schema and update protocols
- Feed the LLM new research papers, data, or domain updates
- Allow the LLM to integrate new information into existing structures
- Reserve human effort for validation, strategic direction, and exception handling
This represents a qualitative change in the human-AI division of labor within knowledge work.
The Validation Problem and Human Oversight
Karpathy’s framing-“you rarely ever write or edit the wiki manually”-does not imply that human oversight becomes unnecessary. Rather, it suggests that human effort shifts from content generation to content validation and strategic curation. A researcher using an LLM-driven knowledge base must still:
- Verify factual accuracy of synthesized information
- Identify and correct hallucinations or misinterpretations
- Ensure the knowledge base reflects current understanding in the field
- Make strategic decisions about what information to prioritize or exclude
The time savings come from eliminating the mechanical work of writing and organizing, not from eliminating judgment. In fact, this model may increase the proportion of time researchers spend on higher-order validation and strategic thinking, even if total time investment decreases.
Implications for Research Velocity and Knowledge Accessibility
If researchers can maintain comprehensive, current knowledge bases with minimal manual effort, several downstream effects become possible:
- Faster literature synthesis: New researchers entering a field can access organized, synthesized knowledge rather than conducting manual literature reviews
- Cross-domain pattern recognition: LLMs can identify connections across knowledge bases in different domains, potentially surfacing insights that siloed manual curation would miss
- Reduced knowledge decay: Knowledge bases maintained manually often become outdated as researchers move to new projects. LLM-driven systems can be continuously updated with minimal friction
- Scalability of expertise: A single researcher can maintain knowledge bases across multiple domains of interest, rather than specializing narrowly
These effects compound over time. As knowledge bases become more comprehensive and current, their value as research tools increases, creating incentives for broader adoption and integration into research workflows.
The Broader Pattern: From Execution to Direction
Karpathy’s observation about knowledge bases fits within a larger pattern he has articulated about the transformation of knowledge work under AI. In 2025, he described developers increasingly functioning as “virtual managers” overseeing AI collaborators, focusing on architecture and decomposition rather than syntax.2 The same logic applies to researchers: they become directors of knowledge synthesis rather than executors of knowledge curation.
This mirrors his earlier reflection that “the profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse and between,” with the potential for individuals to become “10X more powerful” by leveraging AI as a collaborator rather than a tool.2 The knowledge base example demonstrates this principle in practice: a researcher directing an LLM to maintain and synthesize a knowledge base can cover more intellectual ground than one manually curating information.
By March 2026, Karpathy had extended this observation further, noting that coding agents had undergone a discontinuous capability jump-“basically didn’t work before December and basically work since.”5 The implication is that similar discontinuities may occur in other domains, including knowledge management, as LLMs cross capability thresholds that make them reliable collaborators rather than experimental tools.
Strategic Considerations for Knowledge-Intensive Organizations
The normalization of LLM-driven knowledge bases has implications for how organizations structure research, documentation, and institutional knowledge:
- Knowledge infrastructure: Organizations may need to invest in systems that integrate LLMs into knowledge management workflows rather than treating LLMs as external tools
- Validation frameworks: As LLMs become primary knowledge authors, organizations need robust processes for validating and correcting synthesized information
- Researcher skill evolution: Researchers will need to develop competency in directing LLMs, defining knowledge schemas, and validating synthesis-skills distinct from traditional research training
- Knowledge accessibility: LLM-maintained knowledge bases can be queried and synthesized in natural language, potentially democratizing access to domain expertise
The transition from manual to LLM-driven knowledge curation is not merely a productivity improvement. It represents a fundamental shift in how knowledge work is organized, who performs which tasks, and what skills are required to operate effectively in knowledge-intensive domains.
References
1. https://x.com/karpathy/status/2039805659525644595?s=20 – https://x.com/karpathy/status/2039805659525644595?s=20
2. Quote: Andre Karpathy | Quantified Strategy Consulting – 2026-01-21 – https://globaladvisors.biz/2026/01/21/quote-andre-karpathy/
3. Andrej Karpathy – https://karpathy.ai
4. The Professional Journey of Andrej Karpathy – Perplexity – 2024-12-02 – https://www.perplexity.ai/page/the-professional-journey-of-an-OvR1nmNIQNS5gJPAtPMk5w
5. Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era … – 2026-03-20 – https://www.youtube.com/watch?v=kwSVtQ7dziU
6. Tesla’s Former AI Director Andrej Karpathy who said he feels behind … – 2026-02-28 – https://timesofindia.indiatimes.com/technology/tech-news/teslas-former-ai-director-andrej-karpathy-who-said-he-feels-behind-as-programmer-now-says-software-programming-has-changed-due-to-/articleshow/128849256.cms
7. Andrej Karpathy: Architect of an AI Revolution – Klover.ai – 2025-06-12 – https://www.klover.ai/andrej-karpathy/
8. Andrej Karpathy — AGI is still a decade away – Dwarkesh Podcast – 2025-10-17 – https://www.dwarkesh.com/p/andrej-karpathy
9. OpenAI cofounder says he hasn’t written a line of code in … – Fortune – 2026-03-21 – https://fortune.com/2026/03/21/andrej-karpathy-openai-cofounder-ai-agents-coding-state-of-psychosis-openclaw/
10. Andrej Karpathy: Tesla AI, Self-Driving, Optimus, Aliens, and AGI – 2022-10-29 – https://www.youtube.com/watch?v=cdiD-9MMpb0
11. Andrej Karpathy – It will take a decade to work through the issues … – 2025-10-17 – https://news.ycombinator.com/item?id=45619329
12. Andrej Karpathy talks meaning of life and leaving Tesla with Lex … – 2022-10-29 – https://www.teslarati.com/andrej-karpathy-tesla-lex-fridman/
13. Andrej Karpathy Academic Website – Stanford Computer Science – https://cs.stanford.edu/people/karpathy/
14. No Priors Ep. 80 | With Andrej Karpathy from OpenAI and Tesla – 2024-09-05 – https://www.youtube.com/watch?v=hM_h0UA7upI
15. Fave Tweets – Andrej Karpathy – https://karpathy.ai/tweets.html
16. A Survival Guide to a PhD – Andrej Karpathy blog – 2016-09-07 – http://karpathy.github.io/2016/09/07/phd/
