text stringlengths 0 2.68k |
|---|
I don't know, I think it's one of the most well received things I put on the internet in the last 15 years, or something. Yeah, it was really, really nice; a really nice response to it. And I think the nicest one was somebody -- basically, it was saying that this is the first non-hype thing that makes this approachable... |
**Adam Stacoviak:** You said the word, or at least the acronym, AGI, implying general intelligence, artificial general intelligence... What makes you feel like that brute force nature was AGI, or you felt the AGI? |
**Thorsten Ball:** I mean, I was half kidding, right? Like, just making fun of this. I'm not saying that -- |
**Adam Stacoviak:** Okay. I was just checking your literalness there on that. |
**Thorsten Ball:** Yeah, no, it's -- I mean, half kidding, because... I mean, we could talk about what does it even mean to be intelligent, and whatnot. But what I said is "Look, you have this model that can have tools, and then you give it a problem." For example, I can build \[unintelligible 00:10:40.01\] I mean, tha... |
And we could talk about what AGI is or what it isn't, but I don't have another word to describe this as to say it did something smart here; it looked at what it's doing, and it looked at the feedback it got back from what it's doing, and it acted on that feedback, and tried to achieve this goal. And that's not -- it's ... |
**Jerod Santo:** \[12:06\] It looks like it. |
**Thorsten Ball:** Yeah, it looks -- yeah, you know. |
**Adam Stacoviak:** What did the transcript describe? I know that you alluded to this transcript. I've never read one of these transcripts. What is -- can you describe the transcript? What is it? What details are in there? Can you allude to like the thinking part of this? |
**Thorsten Ball:** Oh, you mean the transcripts of like -- |
**Adam Stacoviak:** Right, you said you saw it do this and you're like "How did this happen?" and you looked at the transcript. When that transcript was revealed to you, what did you see? |
**Thorsten Ball:** I mean, the transcript is just the conversation. So every time you talk to an LLM, at the basic level, you send text in, and you get a completion back, right? So if you say what numbers are in the flag of the US, and you say "blue, red", and then it will come back and complete with white, right? And ... |
And the funny thing is that with tool calling, you add another element to this. So I described this in the article, that tool calling sounds super-fancy, it sounds like there's a lot of stuff going on, but in some sense the way I describe it in the blog post is you're having a conversation with a friend, and you say "H... |
**Adam Stacoviak:** Okay... \[laughs\] |
**Thorsten Ball:** Yeah, it sounds -- |
**Adam Stacoviak:** Yeah, I'm digging that analogy. |
**Thorsten Ball:** It's a weird conversation starter, right? Like, you don't get people excited. But with tool calling, you basically start a conversation with the LLM and you say "In the following conversation, when you feel the need to, say, read a file, or list files, or..." What else? "Run a terminal command, respo... |
So what you say to the model is "You are a coding assistant. You have access to the following three tools: read file, list directory, run terminal command. Here's the conversation with the user." And then the user says "What's in the readme file?" And then the model thinks "I'm going to wink..." |
**Jerod Santo:** He's still using air quotes... |
**Thorsten Ball:** Yes. \[laughs\] I'm air quoting everything. But then the model comes back and says "Let me read that file. Like, that's the thing that I want to do." And then how it works on a practical level is that you send that up to the provider, to Anthropic, to Google, OpenAI, and the response comes back, and ... |
\[15:52\] So if we go back to the first example, you would ask -- you have three files; list file, read file, run terminal command. And you say "What kind of project is this?" That's what you ask. And then the model just -- that's why I keep saying, just like us, just like us... What it will do is "Well, let me list th... |
**Jerod Santo:** What's interesting about it is that it's a very basic algorithm, right? |
**Thorsten Ball:** Yeah. |
**Jerod Santo:** It's like "Loop until you have a solution." |
**Thorsten Ball:** Yeah. |
**Jerod Santo:** And really, that's kind of what we do as human engineers, or we give up. And that's the difference. This thing's not going to give up. So it very much is a brute force. But if you come to me with a problem and you say "Jerod, I've got to solve this thing. I'm going to read from a file", I'm going to pi... |
If that doesn't work, I'm going to get another idea, I'm going to try that. If that doesn't work... Until I've exhausted all of my ideas. And then what? Then I go ask a friend, or I go out to Stack Overflow, or now I go to an LLM and get more ideas. "Okay, I need more ways of doing it until I eventually get there." And... |
**Thorsten Ball:** On that, the simplicity of this algorithm, what we kept saying last few months is -- what we've seen over, say, the past year is that a lot of tooling or a lot of stuff that has been built around this model has collapsed into the model. Meaning, say a year ago, they weren't that good at tool calling.... |
And the funny thing is if, I don't know, if you ask a hundred engineers, half of them would say "It's just a for loop." And the others would say, with a smile on their face, "It's just a for loop. Like, this is crazy." |
**Jerod Santo:** Right. |
**Thorsten Ball:** \[20:02\] It's all in this model. You just give it output of five commands and then say "What should I do next?" And it goes and tries 15 other things, because it knows -- based on the previous conversation, it then thinks the next best step is to do the following. And it's -- again, I'm not going to... |
**Adam Stacoviak:** It's bananas! |
**Jerod Santo:** My mental model for that, which is not a super-complicated thing to think about, but I compare it to -- tool calling specifically, I compare it to like shelling out of a programming language. It's like, you know, Elixir has all these things you can do in it... When it comes time to tag an MP3 with ID3 ... |
**Thorsten Ball:** I mean, that's how they've been trained, right? A year ago, one of the big topics was hallucinations. And that's because - to use your analogy here - the model was only inside the programming language. It couldn't shell out. It only knew what was in its standard library. I'm going to break down the a... |
**Jerod Santo:** \[laughs\] All metaphors break down eventually... Some really quick. |
**Thorsten Ball:** \[laughs\] Yeah. It didn't know that there's a world outside, in some sense. So if you would ask it, like, "What's in this directory?" I mean, we've all tried this... People tried this, and without telling it what's in the directory, it will come up with something. It will then say "In this directory... |
**Jerod Santo:** Right. So this unlocked a huge opportunity, which of course, Sourcegraph is trying to jump on, and other people are. We were talking before we started recording, Google just got into the game... We know OpenAI is in the game... We know that Anthropic is in the game... There are open source players of t... |
**Thorsten Ball:** \[23:43\] My thought is that Amp was built in February, which seems like an eternity ago, when basically this phase shift happened where suddenly, with Claude 3.7, people started to realize that these models are really good at tool calling. That you can quickly get something running, hence the blog p... |
And we started working on this, and we're amazed by how well it works. Like, Quinn and I quickly started building Amp with Amp, and just all day long sending each other messages, "This is amazing. It just did this, it just did that." And then nobody's going to believe me, but we actually started working on this before ... |
And I think that's what people are now discovering. Like, this is powerful stuff. And is it for the enterprise? Is it for individual devs? I'm an individual dev, I love using it. I know a lot of other individual devs love using it. |
When it comes to enterprise. I think it's just -- our expertise at Sourcegraph of working with large scale customers and some of the best software companies in the world gives us customer trust. It gives us the ability to build something for their need. We know what their codebases look like. We've seen how many thousa... |
And of course, it sounds ridiculous now, because we're in times where individual devs spend hundreds of dollars a month to use these tools, and you two have also been around a while, you know how crazy that is, that even two years ago if I would have said to you "An individual dev for this side project will spend 50 bu... |
And I think we've accepted this change, and that this is now how you are productive and how this stuff works... And if you want to say "Well, the individual dev cannot afford this, and it's costing maximum five bucks and you get only that many tokens and whatnot, or requests", that's not what Amp is, you know? Amp is -... |
And the other thing is, on a super -- coming up from the level of product principles, or product vision, on a purely practical level, we are a CLI application; Amp is in the CLI. We are in a VS Code extension, which works in Cursor, Windsurf, and VS Code, obviously, and Codium, and it even works in the - what's it call... |
\[28:16\] And what also is different to the others is that we have a server component, so all of your conversations, you can share with your team. They can see how you talk to the agent. You can share links to these conversations, you have a leaderboard, you see how many tokens everybody burns, and you see how many lin... |
And when we show these customers - or potential customers - we show them "Look, with Amp you can share the threads, and you can share the prompts, and you can see what the results are", they go "Perfect. Then I can send this around and can show others. This is how I would prompt it. This is the trick that I use. This i... |
So yeah, that's roughly the overview. And the other meta thing to mention here is that we specifically started on Amp with the assumption that every week a model might get stronger and better, and stuff might collapse into it again, and we need to be prepared for changing our product again. Like, if you get Beyang on, ... |
And we tried to embrace this from the start, by saying - our products get out of the way of the model. The picture I use is built light scaffolding, wooden scaffolding around the model, so when the model gets better and bigger and stronger, the scaffolding falls away, and you again get access to the raw power of this m... |
And yeah, that's the meta thing - keep it simple, be able to move fast, move as fast as you want, be able to -- I should have mentioned this at the start... We don't have a model selector. We pick the best model for the job, that we think is the best model for the job right now, and we are prepared to change this. So i... |
**Adam Stacoviak:** Yeah. When we talked to Steve, one thing that stuck out, I suppose, was the copy - the web copy, when I say this, the word copy - on ampcode.com. And it said -- I just can't believe some of the words that was written here... And this is Quinn, apparently, because I asked Beyang who wrote this, and h... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.