CEO & Member of Technical Staff @ampcode, founded @sourcegraph
San Francisco
- Tweets 12,753
- Following 3,920
- Followers 16,882
- Likes 31,602
Pinned Tweet
Every human has Freedom of Intelligence. Nobody & no gov't should infringe on our right to use available intelligence.
Companies shouldn't be compelled to offer models freely, but they should be shunned if they engage in intelligence censorship, esp. if they lobby gov't for it.
12
8
116
20,163
my daughter typed this and told me I can't delete it, so I guess I need to commit it. good thing I taught her line comment syntax.
136
149
5,932
230,984
We made Amp Free.
It's powered by great tokens and tasteful ads.
Agentic coding is now free for everyone.
193
144
1,587
896,600
Slack is threatening to shut off a big nonprofit teen coding community @hackclub (which is amazing and which I’ve helped for years) with just 3 days to migrate off. This would strand thousands of of the world’s smartest future coders. They are asking for $50k (a surprise 60x nonprofit price increase) but won’t even give assurances of more time or future service if that’s paid.
Anyone who can help? Feels like a mistaken decision on their part by someone who isn’t seeing the whole picture.
132
163
1,173
556,370
Microsoft just started enforcing their longtime license restriction on VS Code forks using more MS-published language extensions.
Play this out a few more steps…
47
59
948
157,040
If you saw how people actually use coding agents, you would realize Andrej's point is very true.
People who keep them on a tight leash, using short threads, reading and reviewing all the code, can get a lot of value out of coding agents. People who go nuts have a quick high but then quickly realize they're getting negative value.
For a coding agent, getting the basics right (e.g., agents being able to reliably and minimally build/test your code, and a great interface for code review and human-agent collab) >>> WhateverBench and "hours of autonomy" for agent harnesses and 10 parallel subagents with spec slop
My pleasure to come on Dwarkesh last week, I thought the questions and conversation were really good.
I re-watched the pod just now too. First of all, yes I know, and I'm sorry that I speak so fast :). It's to my detriment because sometimes my speaking thread out-executes my thinking thread, so I think I botched a few explanations due to that, and sometimes I was also nervous that I'm going too much on a tangent or too deep into something relatively spurious. Anyway, a few notes/pointers:
AGI timelines. My comments on AGI timelines looks to be the most trending part of the early response. This is the "decade of agents" is a reference to this earlier tweet nitter.app/karpathy/status/188254… Basically my AI timelines are about 5-10X pessimistic w.r.t. what you'll find in your neighborhood SF AI house party or on your twitter timeline, but still quite optimistic w.r.t. a rising tide of AI deniers and skeptics. The apparent conflict is not: imo we simultaneously 1) saw a huge amount of progress in recent years with LLMs while 2) there is still a lot of work remaining (grunt work, integration work, sensors and actuators to the physical world, societal work, safety and security work (jailbreaks, poisoning, etc.)) and also research to get done before we have an entity that you'd prefer to hire over a person for an arbitrary job in the world. I think that overall, 10 years should otherwise be a very bullish timeline for AGI, it's only in contrast to present hype that it doesn't feel that way.
Animals vs Ghosts. My earlier writeup on Sutton's podcast nitter.app/karpathy/status/197343… . I am suspicious that there is a single simple algorithm you can let loose on the world and it learns everything from scratch. If someone builds such a thing, I will be wrong and it will be the most incredible breakthrough in AI. In my mind, animals are not an example of this at all - they are prepackaged with a ton of intelligence by evolution and the learning they do is quite minimal overall (example: Zebra at birth). Putting our engineering hats on, we're not going to redo evolution. But with LLMs we have stumbled by an alternative approach to "prepackage" a ton of intelligence in a neural network - not by evolution, but by predicting the next token over the internet. This approach leads to a different kind of entity in the intelligence space. Distinct from animals, more like ghosts or spirits. But we can (and should) make them more animal like over time and in some ways that's what a lot of frontier work is about.
On RL. I've critiqued RL a few times already, e.g. nitter.app/karpathy/status/194443… . First, you're "sucking supervision through a straw", so I think the signal/flop is very bad. RL is also very noisy because a completion might have lots of errors that might get encourages (if you happen to stumble to the right answer), and conversely brilliant insight tokens that might get discouraged (if you happen to screw up later). Process supervision and LLM judges have issues too. I think we'll see alternative learning paradigms. I am long "agentic interaction" but short "reinforcement learning" nitter.app/karpathy/status/196080…. I've seen a number of papers pop up recently that are imo barking up the right tree along the lines of what I called "system prompt learning" nitter.app/karpathy/status/192136… , but I think there is also a gap between ideas on arxiv and actual, at scale implementation at an LLM frontier lab that works in a general way. I am overall quite optimistic that we'll see good progress on this dimension of remaining work quite soon, and e.g. I'd even say ChatGPT memory and so on are primordial deployed examples of new learning paradigms.
Cognitive core. My earlier post on "cognitive core": nitter.app/karpathy/status/193862… , the idea of stripping down LLMs, of making it harder for them to memorize, or actively stripping away their memory, to make them better at generalization. Otherwise they lean too hard on what they've memorized. Humans can't memorize so easily, which now looks more like a feature than a bug by contrast. Maybe the inability to memorize is a kind of regularization. Also my post from a while back on how the trend in model size is "backwards" and why "the models have to first get larger before they can get smaller" nitter.app/karpathy/status/181403…
Time travel to Yann LeCun 1989. This is the post that I did a very hasty/bad job of describing on the pod: nitter.app/karpathy/status/150339… . Basically - how much could you improve Yann LeCun's results with the knowledge of 33 years of algorithmic progress? How constrained were the results by each of algorithms, data, and compute? Case study there of.
nanochat. My end-to-end implementation of the ChatGPT training/inference pipeline (the bare essentials) nitter.app/karpathy/status/197775…
On LLM agents. My critique of the industry is more in overshooting the tooling w.r.t. present capability. I live in what I view as an intermediate world where I want to collaborate with LLMs and where our pros/cons are matched up. The industry lives in a future where fully autonomous entities collaborate in parallel to write all the code and humans are useless. For example, I don't want an Agent that goes off for 20 minutes and comes back with 1,000 lines of code. I certainly don't feel ready to supervise a team of 10 of them. I'd like to go in chunks that I can keep in my head, where an LLM explains the code that it is writing. I'd like it to prove to me that what it did is correct, I want it to pull the API docs and show me that it used things correctly. I want it to make fewer assumptions and ask/collaborate with me when not sure about something. I want to learn along the way and become better as a programmer, not just get served mountains of code that I'm told works. I just think the tools should be more realistic w.r.t. their capability and how they fit into the industry today, and I fear that if this isn't done well we might end up with mountains of slop accumulating across software, and an increase in vulnerabilities, security breaches and etc. nitter.app/karpathy/status/191558…
Job automation. How the radiologists are doing great nitter.app/karpathy/status/197122… and what jobs are more susceptible to automation and why.
Physics. Children should learn physics in early education not because they go on to do physics, but because it is the subject that best boots up a brain. Physicists are the intellectual embryonic stem cell nitter.app/karpathy/status/192969… I have a longer post that has been half-written in my drafts for ~year, which I hope to finish soon.
Thanks again Dwarkesh for having me over!
39
65
831
183,771
Anthropic just published best practices for using Claude Code and coding agents generally.
The key is you must be open to adapting your prompts to "empathize" with the model. Every vendor tiptoes around this suggestion because it sounds like blaming the user.
17
67
813
122,248
Even amazon dot com is down. I don't recall that happening in past AWS outages. This must be a big one.
31
59
783
82,653
+1 to this. @DavidSacks is fantastic for @sourcegraph as an investor, advisor, and board member. He has supported us in risky, bold bets, and he knows how to get to the root cause of any problem really well and help solve it.
I’ve recommended Sacks and Craft to many other founders, and they all agree here.
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.
16
26
512
124,534
pushed and committed it. it's on the main branch forever.
slight twist though: it's on someone else's PR branch that was squash-merged, and they are the git author (I am the Co-Authored-By).
6
1
411
15,186
"There's a good chance that by the end of the year, people aren't using IDEs anymore." — Boris Cherny (Anthropic)
Yes. The IDE is shrinking into a debugging and review tool for most devs most of the time. And the code host...who cares? Everything is changing, there is no moat.
58
28
394
183,923
I love OpenAI and want to make things right here.
We picked AGENT.md in Amp (ampcode.com/news/AGENT.md) and then OpenAI picked AGENTS.md for Codex. One of these should win! But there's a big reason why it should be AGENT.md...
16
28
345
100,532
Hello, Cody.
Cody answers code questions and writes code for you by reading your entire codebase and the code graph.
Often magical, often frustratingly wrong...but getting better quickly. Rolling out to opt-in customers & Discord community members now.
about.sourcegraph.com/cody
10
66
312
156,374
From the vault: @dhh's advice to 17-year-old me about building great opinionated products: make it the best, make a stir, make dictatorial decisions.
11
16
325
34,883
MCP done right — What are the best-designed MCP servers, the ones with good tools (at the appropriate level of abstraction) and descriptions, good response text, good impls of prompts and resources, etc.? Please help me add to this list:
- Playwright
- PostgreSQL
- Sentry
- Linear
25
17
323
45,333
Today we followed through on a commitment made on May 17: @AmpCode would switch to AGENTS not AGENT dot md if OpenAI could get the domain. They did, and now Amp and several other tools have aligned on a standard agent filename.
Huge thanks to OpenAI and all the other agentmakers!
17
34
310
36,758
Amp: “Whoa. Another S-tier. Holy shit. ... They're devs that love agents and they're just building the best thing."
14
25
290
106,381
I just spoke code. 🤓 The founders of @serenadeai came by and showed me their super cool voice-to-code input tool for @code @AtomEditor etc. It solves a huge (literal) pain for people with RSI. Also great for voice shortcuts (“add import lodash”) and coding while feeding 👶🍼.
14
73
272
Wrong. CSS has been in use for hundreds of years. Why do you think so many old buildings are misaligned?
8
10
248
Amp Free: now with no training required, so you can use it at work.
The Internet’s best business model (advertising) delivers once again: a $0 coding agent that meets the same stringent infosec standards of Amp’s paid `smart` mode.
22
23
256
68,720
I'm still Sourcegraph CEO, so I'll choose to be a bit vaguer than @mitchellh, but this is very real. There's a very large company that you all know that would regularly threaten us directly, often in front of other people:
"I'm the guy at ____ who is cloning Sourcegraph" (they tried and gave up)
"You are the bad guys because you charge for your product, but we can just give away the same for free" (they do charge for their product)
"We will make you irrelevant" (they didn't)
Then when we publicly pointed out the ways in which our product was better, to combat the FUD their salespeople were spreading, it apparently massively offended some people there. They felt we were being aggressive and complained to us and many mutual connnections.
I have many friends who work and worked there who were embarrassed by their behavior. It was pretty frustrating in the moment, but I also realized, as Mitchell did, that it meant that we were doing something right.
In hindsight, having grown as CEO since then, I would do 2 things differently:
- Be 100x more direct in communicating why our product was better. The only thing that matters is helping our customers and the market make the right decision. Drawing that comparison is not aggressive nor is it giving "airtime" to our (far larger, well-known) competitor.
- Model the right reaction more strongly for our own team: the competitor VP's feelings don't matter. I don't want my team worrying about that, and as CEO they look to me to see whether they should worry.
My marketing strategy with Terraform early on was "we ship support for new cloud features faster than the clouds themselves." And I'd setup timers and things to show how long it took us vs. them. It worked pretty well, a lot of people liked that.
I did this because the biggest negative feedback I got early on was that Terraform would never be able to keep up with cloud providers. So I wanted to actively show the opposite.
One day Amazon complained and forced me to take down a tweet about it because I hurt the CloudFormation team's feelings (allegedly). They soft-threatened a partnership we were negotiating so I complied. I was pissed. It was the only time I can remember ever "censoring" myself at someone's request.
I remember telling our partnership VP to tell them they can resolve it by simply shipping the feature faster. He said he couldn't tell them that (obviously). I was just frustrated.
I stopped doing the "we're faster than them" marketing after that because I didn't want to piss them off or risk our partnership anymore. Also, them getting mad about it was enough admission to me that we had defeated that point.
You can find some of these tweets if you look at my history. There weren't any personal attacks, no inflammatory words, just stuff like: "We implemented feature X in Y days. Cloudformation still doesn't support it."
I could absolutely see how that would feel bad if you were on that team. But like, we're businesses competing and there were no personal attacks here. I don't think its a shady tactic to point out what you do well and what your competition does poorly (so long as you're not being deceptive about it). It still makes me frustrated thinking about it.
Anyways, it's nice to be so far post-infra now that I can just be honest and transparent about it all.
13
5
253
38,472
Amp Free just got a BIG upgrade: ~65% faster, *much* smarter model (ads pay well, we 5x'd ALL rate limits!).
Happy weekend coding. Will keep this if our tests and your feedback are all positive.
(`amp update` or update Amp extension to use.)
16
25
240
24,189
Imagine your CEO, CFO, and Head of Eng looking at a list of your company’s top AI coding agent users, and you’re at the very top in usage and spend.
Is that a good thing or bad thing at your company?
48
7
211
200,697
In this case Microsoft did legitimately contribute a lot with core VS Code, and they built the language extensions themselves. So it’s different IMO. It’s not some open ecosystem they barged into. At the same time, few realize that VS Code is not “fully” open source.
5
206
11,775
Congrats to our @sourcegraph team on a great quarter!
We smashed our ARR goals for both self-serve and enterprise, and we shipped Cody Enterprise GA, the leading code AI (on context, LLM choice, and scale).
Highlights:
- Hit 188% of our Cody Pro (self-serve) DAU and revenue goal
- Beat enterprise ARR by so much that we're already ahead of our entire H1 plan
- 67% of customers deployed on Sourcegraph Cloud
- Earned our 20th+ customer paying us $500k+ ARR
Thank you to our customers and team! And if you want to join our mission to make it so everyone codes, we're hiring.
17
12
203
17,088
Many coding agents have a "modes" feature that (kinda?) made sense in a RAG or early agentic era. Now they're adding subagents like Amp (search subagent, general subagents, the oracle) and Claude Code (which just shipped a really cool custom subagents feature).
Modes are heavyweight, non-composable UI dropdown menus. Subagents are invoked through natural language ("use the oracle to ...", "use the issue subagent to ...", or implicitly), are perfectly composable, and fit neatly into the conceptual model of agentic tool-calling.
It turns out modes and subagents basically serve the same purpose. Subagents are a strictly better solution.
So, prev-gen coding agents that have an existing "modes" feature now face a tough choice when they add subagents: keep both modes & subagents around (which is confusing and complex), rip out modes (which is painful to their userbase because they put a lot of effort into creating modes), or try to stretch the "modes" concept to include subagents as well (which will be confusing too). I don't envy them.
On the Amp team, we live in mortal fear of being in this unenviable position when we've mis-judged where agentic coding will be in the future and need to make painful product changes. We've made all of these mistakes on products we built before Amp (and we make and will make tons of mistakes on Amp). This is why we have an incredibly high bar to add new UI concepts, it's why we prioritize team members with strong intuition, and it's why we build for where models and the median dev will be in 6-12+ months, not today. In this case, I believe this orientation led us to the right decision (no modes, only composable subagents).
We are fortunate to be able to build for incredibly smart and open-minded/forward-thinking devs who share their feedback often and through basically every channel except carrier pigeons. We've gotten to know and understand so many of you all well. We simply couldn't see any of you or ourselves really loving "modes". But when @thorstenball built subagents and then the oracle, and said that to use them, just say so ("use the oracle to ...") instead of selecting them from some dropdown, it felt right to us and all of you, and it felt like it's where the models were going, too. Props to him there, and thanks to our awesome users for seeing the future.
Just wanted to share a bit behind the scenes about how we think about this stuff and why we might seem extreme or rigid on things.
...and stay tuned for this weekend, when I'll be admitting I was wrong about one of our most controversial Frequently Ignored Feedback (FIFs) and will be changing it, thanks to a ton of feedback that it turns out I did not ignore.
21
16
205
43,414
Amp can now use 432,000 tokens of context: longer threads, more iterations, and more replies.
13
16
199
60,018
A top-5 US bank purchased Amp Enterprise for ~10% of their devs on Tuesday.
To my knowledge, very soon this will be the biggest frontier coding agent deployment in any large US bank.
17
4
188
20,181
When I first saw @artman’s post about Fixing Bugs First at Linear, I nodded a bit, and then a lot. I'm so glad @thorstenball and I radically adhered to this from day 1 on @AmpCode. Here's what I've learned:
• Fix Bugs Now (not First): The overhead to file an issue, prioritize the issue, assign the issue, and update the issue is often greater than the time it takes to fix it. Just Fix Bugs Now, right when you hit them or hear about them; don't bother filing an issue. The reporter need not be the fixer, but you must confirm it's going to be Fixed Now before you let go of it.
• Backpressure: You will never be drowning in bugs if you truly Fix Bugs Now because you won't have time to build new features that would introduce new bugs. If you're drowning in bugs, it means you're insufficiently adhering to Fix Bugs Now, and new stuff is sneaking through. Fix that. Cut scope immediately if needed, no matter how painful.
• Go Direct: When fixing a bug, devs talk directly to customers. No middlemen. Fixing Bugs Now for a real person you're talking to is rewarding and effective; anything else is grunt work that requires a lot more motivation and management pep talks. Plus, customers love talking to devs.
• The 15-Minute Rule: From the moment a customer reports a bug, you need to be able to ship a fix to them within 15 minutes. Some fixes take longer, of course, but let's say it's a simple one-line fix: 15 minutes to the customer's screen. If it takes longer, you break the customer's flow, so they won't bother reporting most bugs (since they wouldn't benefit from the fix in their current task). Also, you kill the reward loop for your own devs and kill their focus, since they'll keep mentally tracking the bug until it's resolved. And if it takes days to ship a fix, then you'll need a process for tracking and communicating fixes, and so on, and all of a sudden the overhead dominates the time to code the fix.
• Repeat Repeat Repeat: Fixing Bugs Now is so radically different that your team won't believe that you really, truly mean it at first. Or even after you've repeated it 10 times. You need to say it in every single team message for it to stick. Even knowing you need to repeat it more than you think, you won't repeat it enough.
• The Stick: Yes, praise your team for Fixing Bugs Now. But for 80% of devs, the reality is you'll need to (at least once) give them direct, constructive feedback when they stray (with good intentions) from Fixing Bugs Now. You need to use the stick for it to stick: "I know you had good intent in building xyz feature while ignoring bugs, but that is not how we build here, and I need you to not do that again."
• Sorry and Thanks: Bugs are a fact of life, but that doesn't make them OK. Call me old-fashioned, but I think you should apologize to customers when they hit a bug and thank them for raising it. Knowing you appreciate it, customers will give you more feedback.
(Caveat: We build for devs, we have an amazing team who we trust, and we started Fix Bugs Now on day 1 at @AmpCode. This probably doesn't apply to existing products with existing bug lists or for non-dev products. I imagine what works for us will also change over time.)
A few months ago, we changed the way we address bugs @linear. We prioritized bugs over everything else. If you have bugs assigned when you wake up in the morning, you don't do anything else before they are addressed.
This approach felt scary and quite radical, but our theory was this: Every software product has defects and the amount of defects being found is somewhat constant. We want to build a product that is of the highest quality, so eventually all bugs will be addressed. This means that the amount of work required to tackle bugs is the same regardless of whether you address the immediately or whether you prioritize feature work and tackle bugs in bulk between feature work.
And this made it obvious that the better alternative is to prioritize bugs over everything else. The cost for doing this was that we had to spend a month or so to bring down the backlog of bugs, but once the backlog was eradicated, there was no additional cost to this approach.
Now, if we or a customer finds a bugs, it usually gets fixed the very next day.
You certainly should consider doing this at your software company.
11
35
286
45,461
The response to Amp Free far exceeded our expectations. We're getting people off the waitlist as fast as we can, first in first out.
Turns out, a lot of people want a free coding agent, and devs are OK with ads if they get something good in return and if the ads are tasteful.
The early financial metrics are also looking really good and sustainable:
• A lot of you are great customers for our advertisers ("This feels like a winner" said one advertiser after seeing a "bunch of conversions already").
• A lot of you are using `smart` mode alongside `free` and paying us.
I am now building a new feature that relies on a really great best-in-class service, where the vendor agreed to make it free for Amp Free users (and pay-as-you-go for paid Amp). Amp Free could become a way for you to try all the best dev tools/services/infra for free, kinda like how a lot of services (like GitHub) are free when used on public repos. Stay tuned!
Some of my favorite reactions:
• nitter.app/ThePrimeagen/status/19…
• nitter.app/GergelyOrosz/status/19…
• nitter.app/paraga/status/19785502…
• nitter.app/skastr052/status/19785…
• nitter.app/RidgetopAI/status/1978…
• nitter.app/tarproductions/status/…
• nitter.app/fatih/status/197856528…
We made Amp Free.
It's powered by great tokens and tasteful ads.
Agentic coding is now free for everyone.
18
14
184
28,365
This is why a model selector doesn't make sense.
1. Make a tool-calling coding agent
2. Make the tools better
3. Some tools become subagents under the hood
4. Different subagents need different models
5. Maybe even their own models
6. So what would a model selector even select?
26
7
165
33,107
Sonnet 4.5 is now the primary model in Amp. I used it a lot this weekend while coding, and we saw incremental yet meaningful improvements that make it an obvious upgrade.
Based on our eval/integration work and my personal experience:
• It Just Works; keep doing what you were doing with Amp, and it'll just work better (this is a good sign that you are on the model-product frontier and not saddled with a heavy harness that dulls the models)
• Less "You're absolutely right!" (too bad, you all loved the the rainbow animation @rockorager added)
• Less verbose assistant messages
When testing new models, we start from scratch with a blank system prompt. We experimented with system prompt changes for 4.5 but ultimately found that our Sonnet 4 system prompt worked best for now.
We try to keep our model scaffolding light so that we let the model "express itself" instead of trying to beat out its bad tendencies because, however tempting that is, it does not work. Likewise, if you found that you needed to make big changes to your AGENTS dot md files (or your own product's system prompt) for Sonnet 4.5, then I would recommend stepping back and having a lighter touch overall. Reach out if we can help here.
Happy coding!
20
8
166
27,190
12 hours after we shipped, and Amp is still the only agentic coding tool to support Sonnet's larger context window out of the box, with no special API keys needed.
Why were we first? We work hard, ship fast, and benefit from the pricing and craftmanship principles of Amp:
• Our competitors who have fixed-price-per-month pricing plans know that usage above 200k tokens means they lose even more money (negative gross margins) fast, so they are busy trying to make the numbers work while our team is making Amp better. We pass through the costs, and Amp is the only structurally gross-margin-positive coding agent, so our customers can choose when it's right for them and don't expect it to be subsidized. This is win-win.
• We realize that most of the immediate benefit comes from having ~200k additional tokens, and there are diminishing returns to another 600k tokens. After all, the first 200k tokens have to be good for the next 200k to be good, and the first 400k tokens have to be good for you to even want to use an additional 600k. That's why we shipped support for the additional 200k token support ASAP. The other 600k tokens? We'll support that soon, when we feel confident about the quality and when Anthropic has more capacity.
• I suspect our competitors were waiting on Anthropic to provide more capacity for the full 1M tokens because they thought a headline with anything less would look dumb. But we know our users are smart and that this is what actually ships the most benefit the soonest. (Also, I think "432,000 Tokens" is a cool headline!")
Much more to say here, but lots of customers have been asking how we got this first. It's our love of shipping fast and the principles around pricing and craftsmanship (doing the right thing, not the fake hype-y thing) that we obsess over. OK, back to coding!
15
16
162
308,140
We're all learning a new way to code, with agents.
Now you can see how other people actually use them.
12
19
159
46,829
2 WTF crazy things happened during my Cody demo:
1. Someone held a tinfoil Faraday cage over my laptop
2. Someone cut my Ethernet cable
But NOTHING could stop my demo of Cody's experimental local inference for code autocomplete+chat (using @ollama) at @replicate OSS AI night!
11
14
149
35,685
We’ve been working with OpenAI over the last few weeks on GPT-5 in Amp.
We’re still evaluating where and how to use GPT-5 (after all, it took 6 months to figure out 3.5 Sonnet’s true capabilities).
Join us and try GPT-5 in Amp now:
30
22
153
17,672
Making it 10x faster to write code won't 10x dev productivity. Most dev time is spent reading, fixing, and maintaining code—not writing code.
12
10
141
22,713
Amp:
- There's a big market of early adopters who will spend $100++/month on an agentic coding tool if it delivers.
- That cohort is where the future is built.
- The adoption cycle is on hyper-speed. Whatever early adopters use today, the median dev uses in 6 months (not years!).
Everyone else is trying to figure out how to make their coding agent the cheapest.
We only care about being the best.
8
7
149
19,108
Congratulations to Anthropic on Claude 4! We’ve been using Amp with Claude 4 internally for weeks. Amp is built for where the models are going, and our early Claude 4 access has informed product decisions around sub-agents, planning, etc.
Pending final checks, Claude 4 Sonnet/Opus will be Amp’s primary model.
7
10
146
34,149
(1/3) We open-sourced @srcgraph today because:
1) If @srcgraph is deployed inside your company, you probably use it frequently for code search + intelligence. (Many well-known companies have 100s/1000s of devs who use @srcgraph daily.)
Sourcegraph is now open source 🚀
about.sourcegraph.com/blog/s…
github.com/sourcegraph/sourc…
3
32
140
We were SO close to converging on a standard, but that pesky "S".
We came out with AGENT.md in Amp (ampcode.com/news/AGENT.md) and then OpenAI came out with AGENTS.md. :P
8
138
14,370
Great @sourcegraph board meeting today, hit 140% of H1 plan, DAU/WAU growth is accelerating (daily chats 3x and enterprise DAU 2x over last 3 mo), a brand new team member just piped $100k, and one AE signed 3 deals today. Let's go!!
8
8
134
6,565
At some point you realize being right is worth zero. You need to have a high % of being right AND convince people that you are right at the time, or else it doesn't matter.
10
7
136
7,454
The best devs are devs+support+sales+PM+marketers. They win hearts and minds, build, and follow through. That feedback loop is so valuable.
13
7
136
7,692
Amp uses Parallel's Search API for web search and retrieval for many users, and soon (next <7d) for all users.
The API/tool design is great, the results are high quality, and the team is awesome.
We shipped the Parallel Search API today. For use by agents, our API meaningfully outperforms products and technologies built over years and decades with substantial resources by very talented teams. Why?
6
5
144
61,264
Deep Search is how code search should be.
Shipping today in Sourcegraph Code Search, after heavy use inside our customers Indeed, Stripe, and more.
4
12
133
14,802
Just shipped support for multiple AGENT․md files in Amp.
ALT Screenshot of Amp with multiple AGENT.md files in use
9
9
131
14,760
Advice for new grads: join a company where you can use the best coding agents to their fullest
14
13
129
26,449
Story of a software startup:
First they ignore you.
Then they laugh at you.
Then they fight you.
Then someone posts on Blind saying "the product sells itself".
ALT A review of Sourcegraph on an anonymous job review site (called "Blind") where one of the pros listed is that "the product sells itself"
6
3
126
this is your weekly reminder that cody is open source: github.com/sourcegraph/cody
25
124
17,187
Today I asked a candidate who happened to be born in Europe: "What is some evidence that you are among the best at what you do?"
He paused for about 10 seconds and then said,
"That is the most American thing I have ever heard anyone say."
5
1
124
11,415
brb, gonna go submit an unsolicited bid today to buy GitHub Copilot for $25B
Exclusive: AI startup Perplexity makes an unsolicited longshot offer to buy Google’s Chrome browser for $34.5 billion on.wsj.com/4oz7E3W
20
1
127
11,565
Everyone else is trying to figure out how to make their coding agent the cheapest.
We only care about being the best.
13
8
128
32,693
What do you all think...should dev tools CEOs code? :)
This is the GitHub contribution chart of a CEO.
It's that of Quinn Slack (@sqs), cofounder & CEO of @Sourcegraph. Not a typical one: and neither is Quinn, or Sourcegraph.
Today's podcast epsiode is with Quinn, and how efficient scaleups operate in 2024:
newsletter.pragmaticengineer…
27
4
121
26,072
Good example of how someone smart uses a coding agent
Writing high quality initial code and prompting an agent to metaphorically "complete the owl" is some of the highest value you can get out of an agent. For example, the given thread took the agent less than 10 seconds. It would've taken me minutes. ampcode.com/threads/T-13e7e7…
3
1
127
33,377
Amp now uses Claude Sonnet 4.
We're exploring more Amp improvements now made possible by Claude 4, including better batch (parallel) tool use, sub-agents/sub-tasks, planning, and more, including use of Claude Opus 4.
Congratulations to Anthropic on Claude 4! We’ve been using Amp with Claude 4 internally for weeks. Amp is built for where the models are going, and our early Claude 4 access has informed product decisions around sub-agents, planning, etc.
Pending final checks, Claude 4 Sonnet/Opus will be Amp’s primary model.
5
13
122
19,274
Huge thanks to the team at @stripe for rolling out @Sourcegraph for all their devs to code better with less toil. We're Stripe customers/fans as well, so we're thrilled it's mutual. To all Stripes: send your feedback and complaints our way (in our shared channel)!
6
15
119
8,605
opencode is cool, and I use it to try out various models sometimes. I like that it's configurable!
But you gotta realize there's a difference between a product like Amp supporting a major new model capability in a first-class way and actually putting in the hard work to make it really good for all users, vs. making a swiss army knife product that lets you tweak the values in config and override stuff to support anything.
That's why I said "out of the box" and "with no special API keys needed" (Anthropic's tier 4 API keys are needed for the Sonnet long context).
I really hope you realize and embrace the difference. Amp and opencode are the polar opposites, the macOS and Linux. I use both myself, and I suspect a lot of our users do too...and they want both kinds of agentic coding tools, too. Happy to be your friendly foil, and also happy for you to not realize the difference between our products and continue to angrily retweet my banger poasts. :)
14
3
121
24,966
If you use Amp and share a great product idea that's aligned with our vision, we'll just ship it. If you report a bug, we'll just fix it.
ASAP. No planning. No backlog. No estimation. No roadmap yapping calls. Just shipping. Hold us to it!
10
17
121
16,972
Just had one of the most hilariously awkward moments of my professional life.
I accidentally crashed an exclusive CIO dinner at a fancy restaurant. Turns out I was supposed to be at the dinner party in the /other/ private room.
In hindsight there were signs in the first hour...
3
1
115
On a quirky family trip to SF/SV right after the dot-com bust, my parents took us to Sand Hill office parks of famous VCs and the Apple HQ, and our Christmas card photo was at SLAC. I still remember all the tech billboards back then. And now...look what's on the 101 near SFO!
4
12
115
4,811
Amp users are the best! @gvrooyen just published the senior eng guide to Amp he has been using internally. Self-recommending.
3
12
122
5,935
bored? try clicking on the @AmpCode CLI's orb
ampcode.com/manual#getting-s…
13
2
117
6,220
When you are speaking to a customer about your product, you should thank them for being a customer.
Even if you're a 100,000-person company and the customer is a 1-person startup.
3
6
117
7,256
The definitive explanation of VS Code’s non-open-sourceness and sly tentacles is from @GeoffreyHuntley at ghuntley.com/fracture.
1
3
110
14,672
"How Long Contexts Fail" by @dbreunig is a good read for users and builders of agentic coding tools.
The section on context clash and the negative impact of a long and winding thread is particularly interesting.
Amp has a feature "Share Thread with Support" that lets users choose to share a specific thread with us to get advice on their prompts and context. In these, we often see quality issues when the user starts with an underspecified (often unintentionally) initial prompt and refines their intent over the course of the conversation.
If you realize that your initial prompt was underspecified, you should try reverting the thread and starting from scratch with a prompt that includes the missing context and anticipates the agent's mistakes. See "Build one to throw away" from @thorstenball's "How I Use Amp" post.
We hope to be able to codify this practice in the product. I hope we can, because even if it's true that models perform worse with meandering context, it feels like something we as the tool builders can take more responsibility for.
3
12
112
9,583
Next time you get asked for a roadmap, tell the truth: AI changes so fast that a roadmap is impossible. Cite this paragraph from @calvinfo re: OpenAI.
AI products that claim to have a roadmap sound Enterprise Ready and Sophisticated(TM) but will diverge from it or fall behind.
5
11
113
10,679
Products that warn against their primary use:
• Cigarettes: don’t smoke me
• Car infotainment systems: don’t look at me while driving
• Hammocks: don’t fall asleep in me
• Q-tips: don’t stick me in your ear
• Coding agents: don't let me run commands without explicit approval
16
9
114
7,014
We set out to make Amp Tab better than Cursor Tab and usable in VS Code or any fork thereof.
It’s also free (too cheap to meter).
If you’re using VS Code, try it. If you’re using Cursor/Windsurf with Amp/CC/Codex, you might be able to ditch the fork and monthly sub.
We flipped the switch: Amp Tab is now on by default for new installs.
It’s our free completion engine for manual edits in VS Code. It’s fast, knows your recent changes and compiler errors, and suggests cross-file updates when needed.
Already on Amp? Run `Enable Amp Tab` from cmd/ctrl+shift+p.
17
13
111
16,282
Token-by-token, the new `rush` mode in @AmpCode is 67% cheaper and 50% faster than `smart`.
Prompt-to-result (which is what you actually care about), it depends. We've been liking it for small, well-defined tasks. Try it and let us know what you think!
10
14
119
29,808
something about using bun + jj + a CLI agent and switching back to emacs/vim, you start getting crazy thoughts like maybe we're in a post-GitHub/VS Code era, feels very free and radical
21
4
112
9,253
Anthropic's API is down, but @AmpCode is still 100% up because it falls back to using GCP Vertex for Claude inference (absolutely identical model and quality) during Anthropic outages.
Happy coding!
Amp is still up. It falls back to (equivalent) Claude on GCP Vertex when Anthropic's API is down.
We care deeply about making Amp the very best agentic coding tool, and we're really happy that most Amp users never noticed any downtime.
8
5
109
10,247
1 million tokens in Amp. But...
Amp is still better, faster, and cheaper if you start a new thread per task. Don't use one big thread per day.
8
12
105
22,988
Autocomplete is coming to Amp's prompt input field to save you time typing symbol/file names. Guess:
(A) Uses an LLM to suggest completions –OR–
(B) Just insta-suggests terms found in open/recent files (like in Emacs dabbrev)
20
8
109
7,827
Just used Amp to make a macOS menu bar widget to show network latency and connectivity status. Took about 30 minutes, on slow airplane WiFi.
Lines of code I read or wrote: zero ø 0 nil
9
3
106
14,407
So happy to welcome @sarahdingwang and @a16z to our mission of bringing @sourcegraph universal code search to every dev & company!
a16z.com/2021/07/13/investin…
17
6
105
You can use Claude Opus 4.1 in Amp now with the following hidden settings in your editor or CLI:
"amp.internal.primaryModel": "anthropic/claude-opus-4-1-20250805",
"amp.anthropic.provider": "anthropic",
We still recommend and personally use Sonnet 4 (the default). Stay tuned.
Today we're releasing Claude Opus 4.1, an upgrade to Claude Opus 4 on agentic tasks, real-world coding, and reasoning.
ALT Bar chart comparing software engineering accuracy between two versions of Opus 4. The first bar shows Opus 4 (May 2025) at 72.5% accuracy and the second bar shows Opus 4.1 (Aug 2025) at 74.5% accuracy.
8
15
107
17,202
My daughter Amelia’s first flight! Thank you to @united for a great flight and flight deck pic.
3
3
97
In an earnings call, the CEO of a big Sourcegraph customer just said Cody makes their best devs 40% better. Love to hear it!
8
6
104
7,266
Need someone to join the Amp team and help us build ads infra & user flows for Amp Free.
Backend, frontend, shipping many times per day, and working with customers. This role is San Francisco only. Lots of hard work and ambiguity. DM me.
11
18
108
27,826
Given (a) the growth of Amp & Claude Code in the last 2mo, (b) lots of insta-regretted year-long seat-based purchases of previous-gen AI coding tools, (c) all the pricing changes, (d) the overall crazy pace, everyone's picking usage-based pricing or month-to-month, not annual.
7
6
106
8,396
GPT-5 is also similarly faster in Amp now. Thank you to @OpenAI for the TTFT improvements, esp. on queries with 100k+ tokens.
Good news! GPT-5 should now be significantly faster in Cursor.
Big thank you to OpenAI for rolling out improvements to caching and API latency! P95 is ~2x faster versus before.
3
5
103
9,406
There's a lot of stuff people learned about software engineering 2005-2025 that now just ain't so anymore.
If you're new to SWE and never learned it, or if you have already discarded it, you're way ahead of everyone else.
12
5
101
12,751
Heard from someone who found a critical vuln in a popular site using Amp, and another who fought off a DDoS attack with Amp’s help.
We want to make Amp great for security and incident response, not just coding: Amp optimizes for quality (not cost), never locks you out at critical times, and has robust (automatic) fallback to Vertex for Claude inference in case of Anthropic downtime, when you might need it most.
What can we do to make Amp better for security and ops/SRE stuff?
5
17
102
12,518
RIP, Charlie Munger. You are an inspiration and teacher to me and so many others. I learned so much from what you wrote and from your Daily Journal annual meetings in Pasadena.
3
2
98
5,220
2024 has been a big year for @Sourcegraph as we accelerate human devs and automate the soul-sucking parts of coding with AI agents. Huge thank you to our customers, team, and community!
• Grew active users by 5x+ and chats per user per day by ~5x, and ARR growth is accelerating as we scale
• Earned big new enterprise customers with wall-to-wall usage, including Palo Alto Networks and Stripe
• …so that now 7/10 top public tech companies, 4/5 top private tech companies, 4/6 top US banks, and 14 US gov’t orgs are Sourcegraph customers
• AI coding agents built on Sourcegraph by Booking𛲔com, Palo Alto Networks, Priceline, and more are in production today with greater-than-human accuracy for low-level coding tasks and big migrations
• Only code AI to ship launch-day support for Claude 3 Opus and 3.5 Sonnet, GPT-4o, and Gemini 1.5 Pro, in partnership with all 3 leading LLM labs
• 1st code AI to default to Claude 3.5 Sonnet (which has now written 1b+ LOC for our users), thanks to our partnership with @Anthropic since mid-2022
• Published research advancing the state of the art of LLM code generation (for completions and agentic suggestions)
We keep our bar for talent incredibly high, and I’m proud to say our talent density has increased as we grow. To the very best people who want to build the future of coding and technology: we’re hiring.
There’s a lot of noise about different approaches to code AI out there, but we’ve been consistently right on key decisions, even when it was unpopular:
• Context is king. It sounded self-serving for us to say because of our big code search foundation, but now everybody realizes that code search is critical: cross-repo, code graph, history, and certainly beyond just basic vector search.
• Use state-of-the-art models, don’t train our own. OpenAI, Anthropic, Google, and others invest billions to build LLMs. Instead of trying to train our own model, which sounds mid sciencey and would let VCs collect 2% on $100m+ of dilution to our team’s incentives and focus, we commit to our customers that they can always use the best models, no matter who from.
• Enterprise focus. The biggest advances in code AI have occurred and will occur in the enterprise, since only there does the economy of scale exist that makes automation possible. Indie hackers are cool too, but they’re downstream of what’s forged in the enterprise. (This has always been true of dev tools, but few realize it.)
• AI coding agents need to iterate with ground truth in a tight loop, not just in an LLM slop doom loop. This approach works today in production for AI coding agents built on Sourcegraph by our customers including Booking𛲔com, Palo Alto Networks, Priceline, and more.
2025 is going to be the year of the AI coding agents, and there’s a clear battle between the horizontal agents (that actually work in the enterprise and verifiably automate a specific repetitive workflow) and vertical agents (that try to replace human devs end-to-end). We’re firmly in the horizontal agent camp.
Here's to many more customers building AI coding agents on Sourcegraph this year—and shipping more big improvements to search, AI assistance, and agent APIs for everyone who uses @Sourcegraph.
Thank you all, and happy coding in 2025!
10
10
98
8,318