In my experience as a founder, @DavidSacks has been nothing but helpful and supportive through a tumultuous period for startups. I hear the same from founder friends (many YC co’s) that he has invested in. Outside of tech, one thing I appreciated was he joined @garrytan in supporting the effort to recall the SF school board at a time when many were afraid to criticize those ideologies. As a result of that effort, Lowell has returned to meritocratic admissions and I heard they’re even ending the school lottery system that pushes many families into private schools in SF. You can certainly critique a man’s actions, but calling him evil? This world needs less of that right now.

Was chatting w/ someone about how coding at night feels way more productive than coding during day. Even early morning, when you're in theory more clearheaded seems less fun than late at night. Probably not universal, but strong anecdotal support. Is there any science on why? 🤔

Cody combines LLMs like GPT-4 and Claude with @sourcegraph's deep understanding of code. The result is an AI coding assistant that's much more factually accurate and attuned to the patterns in your codebase. Now we're open sourcing it! Here's why: about.sourcegraph.com/blog/o…

of course that's your contention. you're a first year ai influencer who just got back from neurips, probably just finished reading the toolformer paper, lemme guess—you trained your own foundation model and are now pivoting to agents. you'll probably be convinced of that until next month when you can't get the thing to pass the squirrel test, and then you're gonna be talkin' about vector databases and cross entropy and multimodal RAG over semi-structured data—you wanna plagiarize the whole @jerryjliu0 talk for us?

Two weeks ago, we open-sourced a new Go concurrency library, conc. It now has 5.3k stars. Here's the technical writeup on the motivations and design decisions behind conc: about.sourcegraph.com/blog/b…

We've been using @AnthropicAI's new language model, Claude, to build an in-editor coding assistant called Cody that helps you understand code and reduces day-to-day sources of programmer toil. Here's a sneak peek 👇

Premature abstraction is probably worse than premature optimization.

The best coding agent is now the free coding agent.

The OpenAI Cookbook contains tons of useful prompt tips and examples, but it can be onerous to read through them all. Instead, here's how you can turn these docs into a well-informed chatbot in 1 minute.

1 month into that monolith➡️microservices migration: "Really excited about how much faster we'll move after this bold new architecture" 3 months: "Taking a bit longer than expected, but what project doesn't?" 6 months: "You know, it's the journey that counts." 1 year later:

Reminder that Amp Free no longer requires sharing data for training. To my knowledge, it's the only free coding agent that has a sustainable economic model (we show ads for other dev tools). Try it for your side projects! ampcode.com/free

If you'd like to play with @AnthropicAI's new 100k-token model using @langchain, Cody offers a great way to learn new libraries and APIs. This illustrates another advantage Cody has over Copilot: freshness. Cody uses @sourcegraph to fetch context from current code. Its knowledge base is not limited to code written before the training cutoff date (2021 for many models). Cody is open source and is available today to everyone: docs.sourcegraph.com/cody/qu…

The bottleneck in coding is now the time it takes a human to review agent-generated code. But existing review UIs don't work well for agents. So we built a new one into Amp.

The future of code will be built on two foundational technologies

Apparently, Science™ backs me up on this one, and all those "sit upright" exhortations are misinformation: news.bbc.co.uk/2/hi/6187080.…

Instead of making me learn a new DSL, why can't you just provide a simple library/API that lets me describe what I want in a well-known language that already has great dev tools?

Wow. We just enabled GPT-4o in Cody and the first zero-shot code generation in a big existing codebase just...works? No red squigglies! It is *really* good at learning from the context Cody provides from our specific codebase.

This is why LLM portability matters anthropic.com/index/100k-con…. If you’re using Copilot, you have a 2-year old model with 2k tokens of context that doesn’t know anything past 2021. If you’re using Cody, you can use Claude, GPT-4, and the latest, greatest LLMs as they come online, with knowledge of your private code as it changes daily. And did I mention Cody is free and open source? 👉 docs.sourcegraph.com/cody/ex…

The Mythical Man Month gets so many things right that modern “dev productivity” experts get wrong. A true classic.

And here is Cody's non-confidential prompt. It's public and open source, along with the rest of Cody. You can view it, improve it, and upstream changes to it. You can even ask Cody about its own source code. This is the power of open dev tools 🙂

Cody now has a mechanism for pulling in context from *outside* the codebase! Introducing OpenCtx, a protocol for providing relevant technical context to humans and AI. This builds on Sourcegraph's foundation as the world's best code search and connects our code graph to entities like issues, designs, technical docs, production logs, how UI components are rendered, and beyond. As AI advances, we think it's important to have an open specification of how context is provided and integrated. Think of this as an "LSP for technical context" that makes it possible to integrate data from many tools across a diverse set of vendors. It will make it much easier to surface relevant context wherever you need it to accelerate software development—while preserving the freedom to choose the best tool at every point in the software development lifecycle. If you've ever wished you could jump immediately from a line of code to (a) the production logs that hit that line, (b) a rendering of the UI component implemented by it, (c) the PRs that touched it, (d) the design docs that mentioned it, (e) whatever other discussion or documentation is relevant in some way, then you've dreamed of the future that OpenCtx is seeking to realize. We currently use OpenCtx in two ways at Sourcegraph: (1) as a experimental context provider for Cody, for context-aware AI code generation and (2) as annotations within source files that tie symbols in the code to the relevant technical context that lies outside of source code. We think (2) is as important as (1), because we believe human developers will remain the heart and soul of the software development process. We will build *much* more on top of this protocol and look forward to building with our community to integrate many other tools into it, as well!

Prompt engineering means exploring textspace until you find an input token sequence that is (1) well represented in the training set and (2) precedes the type of output you’d like to see. Two conditions seem necessary for a good prompt: 1. Low perplexity 2. Nearby (in embedding-space) to emblematic training data Surprising no one has made a prompt eng tool that does embeddings search over the training set of a given model. This would greatly reduce the time it takes to find a good prompt and could be a huge advantage for a particular model.

Amp Free now uses a new model, 65% faster and also smarter, just in time for your weekend coding projects. Try it here: ampcode.com/free

Other coding agents implement a "model selector" architecture, which allows only limited customization of the application harness for each model. Amp employs an agent-oriented architecture, where the composable building block is specialized agents.

How much of what is considered "best practice" in tech management is cargo-culted from Google without asking if Google succeeds *because* of the practice OR if Google succeeds in spite of the practice OR if the practice fits Google's business but not necessarily yours?

I think the world is starting to realize that the AI IDE paradigm is a relic of the chat LLM era, and the ideal agentic coding form factor is closer to a simple, minimal power tool that can be programmed and composed. Unix philosophy >> vertically integrated kitchen sink

A great primer on context management for coding agents and all the mechanisms we've built into Amp for this

babe wake up new yegge post just dropped sourcegraph.com/blog/rag-to-…

Some companies are positioning their coding agent as a human replacement. We think that's wrong. In our view—and actual experience—agents are far more useful for automating away the soul-sucking toil that takes up 99% of your time as a developer, enabling you to focus on the fun parts of creating software and improving the lives of your users. sourcegraph.com/blog/introdu…

There's a new subagent in Amp called the Librarian. Its job is to locate relevant context in all the libraries you might be using in open source or your private dependencies. Amp's accessible context now extends to the entire universe of code: ampcode.com/news/librarian

I fucking love working on dev tools

Anyone else building with LLMs feeling that the interesting stuff that actually moves the needle for user experience is at the search/RAG end of things, not so much the language model itself?

At Amp, our focus has been 100% on making the best coding agent. No flat-rate rug pull, no selling $1 of tokens for 50c, no hidden nerfing. Just be the best. Is that changing with Amp Free? Nope. With our ad partners, we will democratize agentic coding sustainably.

People are saying that SWEs will be replaced by PMs wielding AI. I actually think it's the opposite—what remains an essential part of the creative process is the architecture, algorithms, and data structures. Having a SWE background means you know what and how to compose and combine these. Wielding AI means you don't need to spend all your time line-smithing and can spend more and more of your time mapping your end user needs to the abstractions you design in code. I see a much more direct line to SWEs directing the entire software development process (design, architecture, code) than I do with folks in other roles.

Zig cross compilation is f***ing amazing for cgo. Thank you @croloris for writing this great blog post: dev.to/kristoff/zig-makes-go…

Should we just make a code host that's nothing more than "GitHub but fast"?

A coding agent for everyone—Amp is now free!

Anyone have a good computer networking 101 blog post series they'd recommend to someone who is a more junior engineer that wants to ramp up on the basics of the networking stack, from TCP/IP through TLS?

If you've never built an agent and are looking for a fun, effortless, interactive way to make one, paste this into your terminal: echo 'build me a simple agent using ampcode.com/how-to-build-an-…' | npx @sourcegraph/amp Then have the agent you just created explain how it works.

We've been communicating this to our customers and partners for months now. NNS with naive embeddings yields very noisy results and you're likely better off starting with a keyword-based approach. This simple "do the dumb thing first" insight is one of the reasons why @SourcegraphCody is months ahead of the competition in terms of context quality.

Amp Tab is code-graph aware, enabling it to suggest fixes across multiple files. So if you update an interface definition, it will auto-suggest updates to references to that interface. This video is in native VS Code:

My family arrived in the United States with very little savings. Public school accelerated learning programs afforded me the opportunity to pursue my curiosity in math. SF has done its students a huge disservice by eliminating these in the name of "equity".

“DevOps” was supposed to be about dev-ifying ops but it has now led to opsification of dev—focusing too much on the outer loop (the SDLC), using DORA as the measure of dev productivity, which means commits implicitly become the unit of dev productivity.

I really have to congratulate the Anthropic team on a fantastic release. It’s not just Claude 4, but the taste shown in API design and the preference for more Unix-y composable building blocks makes me very excited to continue building with them. Plus, they’re just really great people.

The Amp finder, the subagent that searches through code to surface relevant context, is now 50% faster with no effective loss in quality.

If you've updated Ghostty today on macOS, the Amp TUI will now show loading state use the native progress bar. It's the little things :)

PSA: if you don't like ads, you can still... pay for Amp. Genuinely don't know if the "i hate ads" people can't read or are just engagement farming ¯\_(ツ)_/¯

It’s funny how many people complain about coding agents ruining code quality when I just caught up with a friend who described how agents allow him to finally write clean code with nice abstractions. They’re tools. Once you master them, you can wield them toward your own ends.

We are past the peak of the coding AI hype cycle. Devs don't want AIAIAI, they want solid tools that tackle the toil and tedium that prevents us from shipping awesome stuff. For tools that use AI, the devil's in the details—there's a big gap between flashy demos and great UX.

Two types of devs right now: 1. "Holy cow you spent $500 on an agent??" 2. "Holy cow, I just closed a $50k project with $500 in tokens"

I’ll be presenting Amp at Code with Claude tomorrow, diving into some Amp internals and how we’re getting the most out of MCP and the latest Claude models. If you’re there, hit me up after my talk at 3pm—I’m sure it’s going to be a fun and interesting day 🙂

AI engineering in 2024

Twitter's algo is LIVE on @sourcegraph with FULL compiler-accurate code nav! 👇 sourcegraph.com/github.com/t…

Coding agents have a variable reward, slot-machine style addictiveness that I've only ever encountered elsewhere in social media apps.

Thank you to all our customers and users who brought @sourcegraph into their organizations and coding lives. Thank you to all the amazing team members who got us to this point. Thanks to our fantastic investors for funding us. And thanks to @ron_miller for the great reporting!

You know you're doing something right when 2 CEOs of >$1B AI coding companies have reached out to team members trying to ask what you're doing that they aren't

We're going to see a clearer separation of types of work that all previously got lumped under the umbrella of "software engineering". On the one hand, there's the work that pushes the envelop of innovation (in both UX and algorithms+architecture). Inner loop tools (like @SourcegraphCody) will always play a role in augmenting the human creative spark. On the other hand, there's the gargantuan pile of toil composed of tasks like quick bug fixes, enhancements, telemetry, test coverage, etc. Inner loop tools help drastically with these as well, but we'll see more full automation as time goes on—starting with tools like @cognition_labs and @codegen. The latter category sucks up 75-90% of dev time today, while the former is where 75-90% of the creative/user/business impact—and joy of building—is. Shame on those who warp the celebration of progress into the stoking of fears or cynicism aimed at people who do the hard work of building software. The future of code has never been brighter—and the market never wider. Full steam ahead!

Sourcegraph's open-source Go concurrency library, conc, was featured in the Best of Go 2023 by @golangweekly. Thank you to @camden_cheek and @bobheadxi for creating and releasing an excellent library!

Should you use an AI IDE, or is the IDE dead? For me, the answer is neither. I wrote up a short post about how I configure vanilla VS Code for agentic coding:

After the Apm Free launch, we've been getting a ton of questions to the effect of “Why no model selector?” I’ll answer this here so I can link to it if it keeps coming up. Amp is different than your typical model-agnostic harness. We view ourselves as product builders who have a contract with our users to ensure a great end-to-end experience. This also means ensuring whatever customization points we provide lead to a great experience. The behavior of an agent is the product of the model, the system prompt, the tools and tool descriptions, and the memory (AGENTS.md or something else). The model is a very important implementation detail of the agent, and arbitrarily swapping in different models will lead to very different behavior. Similarly, small adjustments to the system prompt and tool descriptions can have a substantial impact on how a given model behaves—ever wonder why "Grep" is capitalized in many agent tool descriptions while "glob" is not? Or see the recent report where a research team improved performance of another coding agent by 15% just by altering the system prompt and tool descriptions. We think it wouldn't be tractable to have arbitrary model swapping while maintaining the end-to-end product experience contract with our users. Our team cares deeply about user bug reports, so if a user reported a bug and it turned out to be due to the fact that they had swapped in a model we hadn't tested and tried out, it would suck up a lot of time and prevent us from pushing forward the core experience. So the architectural approach we've taken treats the agent—not the model—as the fundamental unit of composition. We have top-level agents and subagents that are designed to work well together. As the number of specialized agents and subagents proliferates, it becomes exponentially more difficult to ensure a solid user experience while supporting arbitrary model swapping. This is not to say we will never support model selection of any kind—where it makes sense to (e.g., when two models are roughly comparable or from the same provider), we will consider exposing that choice to the user. And this is also not to say we think our approach will be universally appealing. There are many users who want to experiment with dropping different models into different harnesses and seeing what works well. This is similar to the testing we do on a daily basis to evaluate new models as part of the process of improving our agents. But when we ship things to end users, we want to have confidence in the robustness and consistency of the experience we're shipping. We're still a very small team and need to remain focused on building what we think is the best core coding agent experience. Our pitch to you, the user, is that our entire focus is on crafting the best agents that Just Work. If that's not for you or you prefer an option with arbitrary swapping of models, tool descriptions, and other major implementation details, then that's alright—there are many of other options on the market and you don't have to use just one agent! Thank you to all of you who are giving Amp a try for the first time. We're looking forward to hearing your feedback and making Amp work well for you.

Here is a side-by-side comparison of experimental Cody autocomplete v. Copilot. Cody's completions are both faster and higher quality. Note: there are cases where Copilot performs better, but it's already hit-or-miss and we move fast. Open source and enterprise-ready today 🙂

I wanted to add a feature to Cody to "rewrite code in a more functional style", so naturally I asked Cody how to do that. It walked me through the files I needed to edit and generated the code using existing source as a reference point. I thought this would take at least half an hour, but together, Cody and I got it done in a little over 5 minutes.

oh nothing, just using the power of Cody code context to turn Claude 3 and GPT-4 into *library-specific* app generators (note: also works for private internal libraries because we're using special indexers rather than the memorized model training data)

The most surprising demonstration of technology I've seen in 2024

Across all our customers, @Sourcegraph now indexes over 120 billion lines of code 🤯

Unleash the power of code search from within your editor!

This is the way to use a coding agent if you’re a serious engineer:

Once you free yourself of the AGI nirvana/doom cult mind virus, you can start to reason about Transformers and Attention as what they are: useful new tools in the programmer's toolkit. And then it will be clear that RAG and context retrieval are not hacks, but crucial components that have a firm place in the new era's emerging technical system archetypes. But first you have to disabuse yourself of the fantasy. The best way to do that is to build.

Dad works at a chocolate company and said he was bringing home a chocolate bar for the holidays. The chocolate bar:

You know you’re winning when the competition starts copy-pasting from your landing page (and forgets to ctrl-F lol)

Sourcegraph continues to push the frontier of context-aware code AI. We launched the first code-search-enhanced coding assistant, cody.dev, and now we've partnered with @googledevs to explore how superlong context models will impact how devs work. One conclusion is that long context models may soon commoditize local RAG for code. The tools that work best for large codebases will need to combine long context models with precise code search and other long-range retrieval methods that pull in context from well beyond the horizons of the editor. Read more about our joint study here: * sourcegraph.com/blog/towards… * developers.googleblog.com/en…

We're prototyping a notebook-like interface for code investigations and explorations in @sourcegraph! Thinking it'll be great for onboarding, collaborative debugging, and personal note-taking. Anyone interested in early access?

Cody now generates release notes! Try it out here 👉 about.sourcegraph.com/cody

If agentic coding is the future, then any talented dev who is serious about investing in themselves and their craft should be looking for a place of employment that trusts them to use the best agentic coding tools to their maximum capacity.

Not to alarm folks, but it's been a whole day of agentic coding and I haven't seen the phrase "you're absolutely right" once... everything is changing!

Would a short blog post on how to configure VS Code for optimal use with agents be interesting?

Sourcegraph may have invented the model selector for chat-based coding assistants with Cody back in early 2023. That was controversial at the time ("why would I want to use this Claude thing instead of GPT-4?"), but end-user model choice made sense to us given the RAG architecture. Agents are very different than RAG bots, but most coding agents have cargo-culted the model selector from the RAG era. We thought from first principles again and decided against user model selection in Amp. Just not good for a solid end-user experience. This decision has also been controversial, but I suppose time will tell.

An interesting difference in the Chinese vs. American AI landscape: In China, there seem to be no pure model companies. The major players viewed the model frontier as a stepping stone to end user apps, and this strategy gave rise to a vibrant open weight ecosystem. In the U.S., the AGI narrative made investors pile into a handful of foundation model labs and may still hoodwink our politicians into a crippling regulatory regime that serves the special interests of a few at the expense of our national interest. Ironically, it is competition with China that is pushing more openness into the American ecosystem. You talk to American AI app and agent builders and you will hear quiet—and sometimes not so quiet—praise for Deepseek, Qwen, Kimi, GLM, and others. Tbh, I think the silver lining of the revival of Great Power Competition is that it will, out of necessity, push American leadership to resist the type of cronyism where winners are chosen in backroom dealing rather than in the open marketplace of ideas and innovation.

GitHub Universe is this week, so we thought it'd be a good time to review how 5-month-old Cody is now beating 2-year-old Copilot across a spectrum of common programming tasks. Hype is fine, but you know what's better? Real-world use cases 👇 about.sourcegraph.com/blog/c…

Super pumped to work with one of my engineering heroes, @steve_yegge—he's coming out of retirement to join @sourcegraph!!! about.sourcegraph.com/blog/i…

A mistake folks have made in 1st-gen LLM app UX is too much magic. Magic works well for wow effect in shallow demos, but for day-to-day use, explainability and visibility are essential, especially for tools that wish to integrate into the human brain's core iteration loop.

Just after our 4.0 launch, @sourcegraph now indexes and maps over 130TB of open-source code—and much more than that of private code

After coding professionally for more than a decade, Amp has made me an agent-first developer. >90% of the code I write now is written by Amp. There is definitely a meta to discover—AI doesn’t make everything trivial, but it does lower the entry barrier and raise the ceiling. If you invest in it just like any other skill, I think you’ll find it enables you to ship fast and reliably.

Anyone want to try Cody for Neovim?

We're excited to be partnering with @Cloudflare, @jamdotdev, @warpdotdev, @clerk, @LogRocket, @prisma, @intercom and @mixpanel on the Dev Starter Pack, a bundle that makes first-class dev tools available at steep discounts for startups! devstarterpack.io

As of today, 7 out of the 10 top public software companies by market cap use @Sourcegraph

When writing dev docs, remember: a single good usage example is worth a thousand words.

Model swappability is dead. Compare these subagent runs with the exact same prompt and toolsets, different model. What you care about is the behavior of the agent, which is the product of the model, prompt, tools, and the non-linear interactions between all of the above.

At @sourcegraph, we've been choosy about our investments in the model layer. Model training is costly and has slow iteration cycles compared to context improvements (which you need to do anyway). But we've uncovered a few key areas where finetuning has a big impact on user experience. One such area is language-targeted code generation. Today, we're releasing research notes for a new model finetuned from Mixtral 8x7b on a permissively licensed dataset of Rust code. For autocopmlete, we believe this new model is the best choice for Rust code generation, validated against both HumanEval and a real-world context-aware evaluation suite. Read about our learnings here: sourcegraph.com/blog/enhanci…

Would people read and subscribe to a “How (Open-Source) Stuff Works” newsletter? The idea is every month, we’d interview a maintainer and walk through the “life of a query” through a different codebase, documented in a @sourcegraph notebook

Last week, I had the honor of sitting down with @kelseyhightower. One of the questions I asked him was how the heck do you make heads or tails of all the new emerging technologies in deployment and infrastructure.

Cody now answers your questions about codebases on sourcegraph.com and can explain any file to you in plain English—or your human language of choice! Invaluable for reading through and understanding code.

Shooting for a major Amp announcement this week. Stay tuned...

"it's about the notes you don't play"

Graphcool framework just open sourced to much applause at #graphqlsummit by @_schickling!

wake up babe, new yegge post dropped! sourcegraph.com/blog/the-sel…

With all the amazing advancements happening now in LLMs, I think it's time to bring him back.

Some awesome perf work happening on code search at @sourcegraph. Bringing memory usage down while scaling up! We now have every OSS repo with more than 26 GH stars, and counting down...