text stringlengths 0 2.02k |
|---|
**Adam Stacoviak:** Almost none. |
**Eran Yahav:** Yeah. Tabnine is a developer-centric almost to a fault, in a sense. We have zero salespeople. We have almost no top-down sales at all, and everything is happening organically, from the developers up. |
**Adam Stacoviak:** Is that right? |
**Eran Yahav:** \[36:06\] Yeah. We have some community on Discord, and the usual thing that you have around community, but it's mostly, actually, a support channel, above all. |
**Jerod Santo:** Yeah. That's great. |
**Adam Stacoviak:** What's growth been like then in the last year? So let's say Copilot came out roughly a year ago, nine months ago? What's the timeframe? I don't even know . Time is weird even -- especially these days. |
**Eran Yahav:** Yeah. It's been trickling to more and more users; growth has been very, very strong. Very good. |
**Adam Stacoviak:** You're happy then. You're smiling, you're not frowning. |
**Eran Yahav:** No, I'm definitely not frowning. And actually, as this thing goes - you know, Copilot creates awareness... So even if it takes some piece of the pie, the overall size of the pie increases, and so Tabnine is actually happy. |
**Jerod Santo:** So one of the major differentiators comes down to really the ability to synthesize, or the ability to generate the code and have it be useful more often than not, or more often than my shallow experience with Copilot, or whatever... Because developers are kind of fickle, and we can have one bad generat... |
**Eran Yahav:** Correct. |
**Jerod Santo:** How do those data sets differ? Do you know how they're doing it? |
**Eran Yahav:** Yeah, I think I have a pretty good idea. So I think maybe one of the differences is that Tabnine has only been trained on source code with the permissive open source licenses. When we started, we were not sure what is going to be the legal implications of training on code that is GPL or not... And we sa... |
**Jerod Santo:** Smart, |
**Eran Yahav:** I don't want to worry about the ethical implications or the legal implications of training on GPL. And let's keep it clean. I think, in retrospect, that has been a pretty solid decision. |
**Adam Stacoviak:** I agree. |
**Eran Yahav:** I'm sure that a lot of people don't care, but some people do. And as developers, I guess we do. |
**Adam Stacoviak:** Yeah. So those licenses are MIT, Apache II, BSD 2-Clause, and BSD-3-Clause. So those are the four licenses, the permissive open source public licenses that you leverage against in terms of where you can source code from. |
**Eran Yahav:** Yeah, that's right. Again, the beautiful thing about Tabnine is that you're always in control. If you're a developer and you feel that you need a model that does something else, you want to train your own model for your own use, using some other datasets, then we facilitate that. You can do that. You ca... |
**Adam Stacoviak:** So let's say I got myself a copy of-- somehow the Windows source code was leaked, and I happen to have it. Could I train at my own will? Obviously, I'm breaking laws, and all these things if I want to... But is it my choice to do that? Is that what you're saying? Like, if I have the source code... |
**Eran Yahav:** Yeah. Yeah. If you have the source code, you can train. I wouldn't know that this is the Windows source code, right? We don't know what our customers-- |
**Adam Stacoviak:** Of course not. I'm just saying... \[laughs\] |
**Eran Yahav:** Yeah. We don't know what our customers are training on. So you could, in effect, do that. |
**Adam Stacoviak:** \[39:58\] I was trying to go as far as I could of potentially offending, but obviously, as a hypothetical, not a realization. But if I had a copy of source code that was not mine - whatever. I could train on any code I have, essentially. |
**Eran Yahav:** Yeah, you could, and we have no way of knowing what you're training on, and we actually don't want to know. But the beautiful thing is that, assuming that you've trained on something of value to you and obtained it legally, now the model is being able to generate code that is following those kind of pra... |
**Jerod Santo:** So I guess one advantage of being David in the David and Goliath story - besides the fact that David wins, so you've got that going for you... But these things will probably play out in courts of law, at some point. I mean, you're avoiding the whole GPL thing entirely by excluding it... But, I mean, if... |
**Eran Yahav:** I don't think so. I don't think there are going to be any lawsuits here. I don't know. I'm not a lawyer, so I have absolutely zero legal training. I don't know that this is something that David cares about deeply. David cares about generating value to developers in a way that respects developers. And th... |
**Jerod Santo:** Right. Well, I guess my point wasn't them getting sued. My point is if it plays out so that it's totally fine to use GPL code in your model -- I mean, you guys are limiting quite a bit the amount of data that you're using. |
**Eran Yahav:** Correct. |
**Jerod Santo:** So you would think Joe Blow over here watching the sidelines think, "I bet OpenAI's Codex is better out of the box than Tabnine, because they have way more data." Now, once you can start customizing it and training your own models and stuff, maybe not. But that first user experience... And maybe that's... |
**Eran Yahav:** The universal model -- again, there are so many trade-offs that play here: the size of the model, where inference runs, whether it's Tabnine cloud or Tabnine local... So I wouldn't go into the breakdown of this whole thing, but let me just say, there's enough data that is with clear permissive open sour... |
**Jerod Santo:** Sure, but like you're missing the Linux kernel. There's so much knowledge inside the Linux kernel. Wouldn't you want to be able to use that? |
**Eran Yahav:** I would like to be able to use it, but I restrict myself, because the trade-off, I think, is still that I have sufficient amounts of code with the permissive licenses. |
**Jerod Santo:** So even if it came to it, it was no big deal. All open source code, all licenses that are open source, you would leave it with the permissive...? |
**Eran Yahav:** At this point, I think that's the right choice. If down the road we are convinced that we're not infringing on people's work in a way that offends them, in a sense, that is unethical to them, we may change that decision. It also depends on the granularity of the predictions that you're making. If what y... |
Tabnine works on line completions and snippet completions, so it could work in both levels. We find that most developers actually get most of the value from the shorter completions, and the reason is exactly the tight loop that I mentioned earlier. The developer wants to say, "Tab through, tab through. Yeah, accept, ac... |
\[44:18\] And so when you give me a line as a developer, it's easy for me to make the snap judgment, "Yeah, that's the line I wanted." If you give me 30 lines, that's like, "Oh wait, you just made me read a bunch of code that was written by a machine, and can have subtle bugs somewhere.'' And now, as a developer, you m... |
**Jerod Santo:** So let's say I write a function name... I'm in JavaScript and I write function, "Upload to S3", and I hit tab. Tabnine's going to generate the code that gets that done, or it's going to give me, `var bucket = someBucket` Is it just going to give me the next line? |
**Eran Yahav:** Hopefully, it's going to give you `const`, I guess, and not `var` or `let` or whatever |
**Jerod Santo:** Sorry, I'm old school still. I'm old school. \[laughter\] |
**Adam Stacoviak:** He's linting you on the fly here, Jerod. \[laughter\] |
**Jerod Santo:** I'm pseudo-coding over here... Come on, help me out. |
**Eran Yahav:** Yeah. So Tabnine will likely -- even if it knows the entire snippet, it would unroll it for you line by line. So even if internally the model has predicted the entire snippet, it would, in fact, unroll it for you line by line, because of two reasons. One, it gives you more control, and basically, you ca... |
And so there is a sense of walking together with you through the snippet, line by line saying, "Oh, yeah, yeah, yeah, yeah... No, I want the timeout. No, I want some other default value. Oh, yes, do some error handling if you don't find it." And so it's actually this guided -- let's call it a guided walkthrough of the ... |
**Adam Stacoviak:** Yeah. This isn't Stack Overflow generation. I like that. And you can actually see this in action at tabnine.com, I believe it's /pro. And I'm not trying to advertise anybody to sign up, but you can try it if you want. There's a great video there that showcases Tabnine free versus Tabnine Pro. And yo... |
It's a little bit by a little bit, as you might do on your own. And that's interesting, because it does help you learn things. I might learn new things about a different function I haven't used before, because if this model's trained on my codebase - well, I might write a new page or new something in our application th... |
\[48:07\] I think it's very, very interesting. I think this is definitely interesting in comparison to just open source at large, training on an open source at large model, and giving me what the consensus has chosen as good... Or bad, potentially, because there's buggy code that gets generated, like you said. But it l... |
**Eran Yahav:** You asked earlier about trust in Tabnine, and we are seeing people trust us with code all the time. So a huge numbers of teams and companies are basically sharing the code with Tabnine to train models on. Obviously, we don't store this code anywhere. We don't want to. We train the model and the code goe... |
**Break:** \[49:03\] to \[52:22\] |
**Jerod Santo:** Eran, do you think this level of granularity is ideal for AI assistant, or do you think it's ideal for now? Is there a world in which I'm not writing my functions anymore? Or is this the best way to interact with this kind of a system? |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.