May 21, 2025 · 32m · latent-space
DeepWiki: The GitHub Encyclopedia
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Silas Alberti of Cognition joins the Latent Space podcast to discuss DeepWiki, an AI-powered documentation and deep research tool that generates high-level architectural knowledge graphs for GitHub codebases. The conversation covers Cognition's custom infrastructure, graph-based code analysis, the future of cross-repository search, and their newly released open-source CUDA model trained with multi-turn reinforcement learning.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 5.5 Guest teaching 4.6 Guest disagreement 1.0 The hosts pushing back 1.4
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 10:44 Silas Defends Building Custom In-House Tooling
Silas playfully rejects the conventional SaaS approach of using off-the-shelf workflow orchestrators and vector databases, defending their culture of rolling custom infrastructure.
Hardest push from the hosts ▶ 15:28 Swix Challenges Deep Research DifferentiationSwix pushes Silas on whether DeepWiki's deep research mode provides measurable quality improvements over competing tools like OpenAI's GitHub research.
Biggest teaching moment ▶ 29:07 Silas Explains Multi-Turn RL Compiler DynamicsSilas breaks down why single-turn models become overly conservative while multi-turn models take aggressive optimization risks because they have execution steps to self-correct compiler errors.
The host holds their own ▶ 25:42 Alessio Details Real-World Code Search Failure ModesAlessio leverages his deep framework experience with Rails gems to demonstrate why naive cross-repo search fails on abandoned forks and outdated packages.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| DeepWiki Origin and GitHub Research Concept | 4 | 5 | 1 | 1 | Swix and Alessio probe the origins of DeepWiki and its compute costs. Silas explains their recency-weighted repository selection and indexing economics without conflict. | |
| Security, Rate Limiting, and Frictionless Access | 6 | 6 | 1 | 2 | Swix discusses VS Code extension architecture, ASTs, and dynamic linking challenges. Silas details how graph algorithms built by competitive programmer Gennady Korotkevich parse commit history and LSP graphs. | |
| In-House Infrastructure and Indexing Queues | 5 | 5 | 2 | 1 | The hosts inquire about orchestration tools, leading Silas to explain Cognition's in-house engineering culture and custom vector database implementation. Alessio suggests consumption patterns via MCP and deep links. | |
| Browser Integration and DeepWiki as Dynamic Documentation | 5 | 4 | 1 | 2 | Swix brings up OpenAI's GitHub deep research feature and discusses code embedding models and eval limitations. Silas describes why DeepWiki's high-level system mapping provides more grounded code exploration. | |
| Live Demonstration with smol-podcaster Codebase | 6 | 3 | 0 | 0 | Alessio shares a live walkthrough of DeepWiki indexing his smol-podcaster repository. Silas highlights how automated architectural summaries solve common developer onboarding friction. | |
| Interactive Docs, Sandbox Execution, and Onboarding Friction | 6 | 4 | 1 | 2 | Alessio argues for interactive code execution within documentation, leading into a technical discussion about dev container limitations and sandbox configuration hurdles during Devin onboarding. | |
| Cross-Repository Deep Research across GitHub | 6 | 4 | 1 | 2 | Silas proposes cross-repository deep research across all of GitHub. Alessio and Swix counter with real-world edge cases like stale forks and semantic search limits previously encountered by Sourcegraph. | |
| Open Source CUDA Model Release and Multi-Turn RL | 6 | 6 | 1 | 1 | Silas details Cognition's open-source CUDA model release trained with multi-turn RL. Swix connects the verifier mechanics directly to OpenAI's reinforcement fine-tuning paradigm. |
Statements from this episode (17)
Alberti: Cognition is based in SF after early nomadic hacker houses
“We're actually mainly based here in SF. I mean, yeah, the founding mythology is, I think we were a nomadic company in the beginning. So we had hacker house in different places. I think the first one was actually in Burning Game. Then it was in New York, then b…”
Silas Alberti: DeepWiki began as a free open-source side project
“I mean, it was supposed to be this like open source focus, like free site project almost, but it became bigger than expected.”
Silas Alberti: DeepWiki pairs AI codebase documentation with a research agent
“And I think it has like two components. It has the wiki itself, which is basically like AI generated Got documentation for any code base on GitHub, and then there's, like, the deep research agent that leverages both this, like, wiki, but also, like, any code f…”
Silas Alberti: DeepWiki scaled infrastructure for tens of thousands of developers
“I think the particular thing that we had to make happen for DeepWiki was, like, scaling up the infra so we could support, like, tens of thousands of open source developers at the same time. And I think also just, like, Making it be, there's like no sign up req…”
Alberti: DeepWiki compute spend may break $1M soon
“So we think we should be now at like the high, like tens of thousands of repos. So I guess you can like do the rough map. Like we might be approaching quite significant numbers of compute spend here. I mean maybe we'll like break the million soon.”
Alberti: 80% to 90% of DeepWiki users view pre-indexed popular repos
“So I would say, like, 80%, 90% of users just, like, check out one of the popular repos.”
Alberti: Folder structure graphs are poor representations of codebase architecture
“There are some, the simplest one is a folder structure graph, but unfortunately, it's, like, actually a pretty bad one because people sometimes just, you know, like, put all the components in this folder and all the servers in this folder.”
Alberti: Cognition built its own in-house vector database
“This culture of like building everything in-house. It's also kind of funny, like we, there's like a rack component to this. And I've been actually, I personally advocate, I can just like use a vector database and make life easy. But then there's a really stron…”
Alberti: DeepWiki experienced spikes of 2,000 indexing jobs queued
“Sometimes it's like random spikes. I don't know how they happen. Like 2000 indexing drops in the queue. Suddenly it was like the other day.”
Alberti: Devin updates codebase indexes on every commit for paid users
“If you like pay for Devin, for example, like we actually update it on literally every commit incrementally.”
Alberti: Over 1,000 GitHub projects added DeepWiki badges to their repositories
“There was more than a thousand projects already had The like deep wiki batch, which I thought was like super awesome to see like way more than I expected.”
Alberti: DeepWiki auto-updates documentation for free repositories displaying its badge
“So we just decided, okay, how about we just detect whether you have the wiki batch and then we'll keep it updated for you.”
Alberti: Small numbers of manually curated evals beat large eval sets
“I think usually like small numbers of very high quality evals are the way to go and like we curate them manually and make sure they're really good.”
Alberti: Setting up dev environments is Devin's highest onboarding hurdle
“So Devon actually needs, like, a full dev environment to be set up. And I think you guys probably both went through this when you onboarded with Devon. It's, like, probably, like, the highest friction onboarding point that we still need to, like,”
Alberti: Wiki pages are a better abstraction than pure RAG for codebase search
“If you just do like pure, like rack on like such a big base code files, it'll just be like pretty bad at a certain point, you know, on a single code base. Sure. I can see it, but Tens of thousands of code bases. It's tougher. But I think actually the wiki page…”
Alberti: Multi-Turn RL Enables Aggressive Code Optimization Over Single-Turn Models
“Basically the single-turn model that was just trained on, like, getting the best result after one turn. It would basically be a little bit, like, too careful, because it couldn't risk writing, like, non-compiling code, whereas, like, the multi-turn model would…”
Alberti: AI App Companies Must Encode Product Needs Into RL Feedback
“I think that's almost how I view like the future of application layer companies. Cause I mean, yeah, you see like the different, the labs are also now creating these like RL platforms and you can soon like customize models with RL on your personal, like on you…”