Here’s what’s on our plate today:

• 🧪 Sourcegraph, AI code sprawl, and the new value of comprehension.

• ⚡ Quick Bits: Delve freezes demos, TikTok driving risk, Musk’s chip plant.

• 🧠 Roko’s Brain Snack for Builders on search, context, and traceability.

• 📊 Poll on what matters most as AI writes more of the codebase.

Let’s dive in. No floaties needed…

The Laboratory

How the AI coding boom is redefining the role of companies like Sourcegraph

TL;DR

• Code search is infrastructure, not a feature: Sourcegraph built universal code search for enterprises drowning in millions of lines of code across hundreds of repositories, serving 800K+ developers and indexing 54B+ lines, and the rise of AI-generated code is making that capability more critical, not less.

• The company split itself in two: In late 2025, Sourcegraph separated its enterprise code search business from its AI coding agent (Amp), with co-founders Quinn Slack and Beyang Liu leaving to run Amp while Dan Adler stepped up as CEO of the search-focused company, a move that reflects a broader market bifurcation between tools that generate code and tools that help people understand it.

• AI creates code faster than humans can comprehend it: 84% of Sourcegraph’s large enterprise customers saw codebases grow after adopting AI tools, and with 41% of all code written in 2025 being AI-generated, the gap between what gets produced and what gets understood is widening rapidly.

• The ‘quality tax’ is real: Larger pull requests, rising bug counts, and overwhelmed code reviewers are the hidden costs of AI-accelerated development, creating structural demand for comprehension tools that help teams manage the complexity AI leaves behind.

• The existential question nobody has answered yet: As AI-written code approaches the majority share of enterprise codebases, organizations face a future where most of their software was created by systems that cannot explain their reasoning, reviewed superficially by humans who did not write it, and maintained by engineers with no institutional memory of why it was built that way.

The modern world is a technological marvel built to evolve as humans continue making scientific breakthroughs. Powered by silicon chips and binary (the only language computers truly understand), this technological marvel requires an incomprehensible amount of programming code to transform computers from blank, multipurpose electronic devices to machines that can follow human instructions.

A simple smartphone operating system contains roughly 12M lines of code; social media platforms can have up to 62M lines of code, and Big Tech companies like Google depend on an estimated 2B lines of code for their operations. Then there are developers writing millions of lines of code every day to introduce new features, secure existing infrastructure, fix bugs, and develop new capabilities. The result is an incomprehensible volume of computer code that needs ot be preserved and studied for future use.

According to a 2025 Menlo Ventures report, code has become AI’s first true ‘killer use case,’ with half of developers at top-performing organizations now using AI coding tools every day. The result is that companies are producing more code, faster than at any point in software history.

But writing code was never the hardest part of building software. The harder problem, and the one that becomes more acute with every passing year, is finding and understanding the code that already exists.

The company that made code searchable

Consider what it looks like inside a large technology company. An organization like Uber or Lyft might maintain hundreds or even thousands of code repositories (the centralized locations where code is stored and versioned), spread across multiple hosting platforms, written in a dozen programming languages by hundreds of engineers over many years.

When a critical security vulnerability is discovered, as happened with the Log4j flaw in late 2021, the engineering team needs to locate every instance of the affected code across the entire system, understand what each instance connects to, and deploy a fix, often within days rather than weeks. Without a way to search all of that code from a single interface, engineers are left opening repositories one by one, piecing together context from memory and documentation, and hoping they have not missed something.

This is the problem Sourcegraph was built to solve. Founded in 2013 by Stanford graduates Quinn Slack and Beyang Liu, the company created what it calls universal code search: a platform that indexes entire codebases and makes them searchable the way Google makes the web searchable.

The idea was partly inspired by Liu’s experience using Google’s internal code search tool during a summer internship. Google had the resources to build that tool for itself, but most companies did not, and as software systems grew larger and more fragmented, the need became universal.

The platform now serves over 800K developers and has indexed more than 54B lines of code, according to TechCrunch. It connects to every major code hosting service, including GitHub, GitLab, Bitbucket, and Perforce, and supports more than 75 programming languages. Early enterprise customers included Uber, Lyft, and Dropbox, and today the company counts 4 out of 5 FAANG companies, major global banks, and government agencies among its clients.

However, much like the rest of the technology industry, Sourcegraph is now facing new challenges with the advent of artificial intelligence.

How AI forced Sourcegraph to rethink its business strategy

As large language models became capable of generating, debugging, and explaining code, the developer tools market was reshaped in ways that caught even well-positioned companies off guard.

GitHub Copilot, launched in 2021 through a collaboration between GitHub and OpenAI, normalized the idea of AI as a coding partner. By mid-2025, Copilot had reached 20M users and was deployed at 90% of Fortune 100 companies. Within months, the AI coding tools market reached $7.37B in 2025, with projections pointing toward $30.1B by 2032.

Sourcegraph responded to this shift by building its own AI coding assistant that used its search engine to give language models access to the full context of a codebase rather than just the file a developer happened to be editing. The logic was that AI assistants with deeper context would produce more accurate results, especially in large enterprise systems where understanding the broader codebase matters as much as generating the next line of code.

What Sourcegraph did not anticipate was that the success of its coding assistant would create tensions inside the company that would prove difficult to resolve.

Why Sourcegraph split into two companies

Sourcegraph’s main business was enterprise infrastructure, with stable, security-focused tools used by large organizations that prioritize reliability over new features. Its AI assistant products, however, required fast experimentation, frequent updates, and competition with rapidly evolving developer tools. Trying to serve both markets inside one company meant the needs of each were too different for either to get full attention.

To resolve the tension without compromising on either of its business propositions, Sourcegraph formally split into two independent companies in late 2025. Sourcegraph, under new CEO Dan Adler, would focus entirely on code search and code understanding for the enterprise, while Amp Inc., led by co-founders Slack and Liu, would operate as an independent research lab building frontier AI coding agents.

The divide between generating code and understanding it

The split in the company was not an irrational reaction to a changing market; rather, it was the acceptance of a broader industry shift, one that matters well beyond the specifics of any single company.

On one side of that divergence are the tools that generate code: Copilot, Cursor, Claude Code, Amp, and a growing roster of AI-powered editors and agents. These products are where the majority of venture capital, media attention, and developer enthusiasm are concentrated. They compete on the quality of their underlying models, the speed of their suggestions, and how seamlessly they fit into a developer’s daily workflow.

On the other side is what might be called the comprehension layer: tools that help humans and AI understand, search, navigate, and govern the code that already exists. This is where Sourcegraph now sits, alongside categories like security scanning (Snyk), infrastructure monitoring (Datadog), and search infrastructure (Elastic). While these products do not garner the same media and investor attention, they tend to become more important, not less, as the systems they serve grow more complex.

The interesting thing to note is that the relationship between these two layers is not competitive; it is symbiotic and increasingly self-reinforcing.

AI coding tools speed up development, but they also make codebases larger and harder to understand, increasing the need for tools that help teams keep track of what they have built. Sourcegraph’s own data reflects this loop. The company says 84% of its large enterprise customers saw steady growth in lines of code after adopting AI tools, and that growth led to more search activity as engineers tried to understand what the AI had generated and how it fit into existing systems.

The wider industry shows the same trend. According to MIT Technology Review, about 65% of developers now use AI coding tools weekly, and roughly 41% of code written in 2025 was generated by AI. But bigger pull requests, more bugs, and heavier review workloads have followed, a burden many engineers now call the ‘quality tax,’ creating growing demand for tools that manage the complexity AI leaves behind.

However, while companies like Sourcegraph continue to enhance their platforms, the question beneath all these developments is one the entire software industry will eventually need to answer: what happens when AI writes the majority of a company’s codebase?

What happens when AI writes most of the code

The current trajectory suggests this threshold is approaching. If 41% of code was AI-generated in 2025 and adoption curves continue to accelerate, several industries could reach 50% by late 2026 or 2027. When that happens, organizations will face a situation they have never encountered before: the majority of their software was written by systems that cannot explain their reasoning, reviewed (often superficially) by humans who did not write it, and maintained by a combination of AI agents and future engineers who have no institutional memory of why it was designed that way.

Sourcegraph is betting that this scenario makes code understanding infrastructure more valuable, not less. The argument is straightforward: if humans are writing less code and comprehending less of what exists, the tools that provide visibility, searchability, and context across an entire codebase become essential rather than optional. Post-split, the company has leaned fully into this positioning, describing itself as the platform that gives both humans and AI agents “the context to understand and evolve the world’s largest, most complex codebases.”

Whether the company’s bet pays out remains to be seen. In the meantime, for Sourcegraph and others operating in this space, the challenge is to figure out how to keep pace with the rate at which AI can assist and write the instructions that run the modern world.

Quick Bits, No Fluff

• Delve freezes demos: Delve paused product demos after allegations it fabricated compliance evidence, while Insight Partners scrubbed a post about its $32M investment. 

• TikTok behind the wheel: UK researchers warn a new wave of distracted driving is being fueled by drivers taking quick looks at TikTok and other apps on the road. 

• Musk’s chip factory: Elon Musk says he is building a Terafab chip plant in Austin, aiming to make AI hardware manufacturing a bigger part of xAI’s stack. 

Brain Snack (for Builders)

Wednesday Poll

Meme Of The Day

The Toolkit

• Mirage: Browser-based 3D design app that lets you prototype interactive product shots and visuals directly in your editor.

• Speechmatics: Speech recognition engine for real-time, multilingual transcription and voice analytics across audio and video.

• Superhuman: Rebranded Grammarly platform offering an AI assistant that drafts, edits, and manages communication across email and work apps.

Rate This Edition