コンテンツにスキップ

Cognition AI 社の CEO、Scott と「Devin」について深堀り

  • URL: https://www.youtube.com/watch?v=UzBLeOBm2_0
  • 文字起こし日: 2026-02-24 02:54

内容概要

  • Cognition AIのCEO、Scott氏が、世界初の自律型ソフトウェアエンジニア「Devin」について語る
  • Devinはソフトウェアエンジニアの作業を支援し、多くのユースケースで使用されている
  • Devinは社内のSlackチャンネルと連携し、バグ報告や機能リクエストに対応
  • Devinはコードレビューの第一段階を担当し、人間が問題を解決するのを容易にする
  • Devinは非同期で動作するため、複数のDevinを同時に実行可能
  • Devinは検証可能なタスク、例えばバグ修正やフロントエンドタスクに最適
  • Devinはコードの10%の計画段階とレビュー段階を人間が担当し、残りの80%をDevinが自動で行う
  • 社内ではDevinを使いこなすのが非常に上手い人がたくさんいる
  • Webnal社では、全てのコードの39%がDevinによって書かれている
  • 長期的な未来として、Devinは人間の指示を理解し、ソフトウェアを構築できるようになることが目標

文字起こし

Scott： Sure yes so I'm Scott I'm currently there the CEO and co founder of a company called Cognition AI and 質問者: ソフトウェアのソフトウェアエンジニアの仕事って言うのは実装じゃなくて何を作るべきかっていうの考えるって言うことにシフトしていくんじゃないかってお話ありましたけど、それに達成するまでにあとどれくらいの時間がかかると思いますか？

男性: え、どうも、安野貴浩です。え、本日はスペシャルゲストに来ていただきました。 あの、簡単にまずは自己紹介をしていただけますか。

Scott： Sure yes so I'm Scott I'm currently the the CEO and co founder of a company called Cognition AI and what we built is Devin which is the first autonomous software engineer

質問者: はい。というわけでデビンの海の親と言ってもいいわけですよね。

Scott： I feel like Devin is uh you know we we we we all look to Devin as our uh as our guide actually as well on the Devon team because Devon actually is by far the biggest contributor to the Devon repository too.

質問者: はい。ということはそのデビンの開発にデビンってどれ、どういう風に使われてるんですか。その社内での使われ方こうすごい気にしてます。

Scott： Yeah yeah so we use Devin for a lot of use cases internally. Um the uh uh Devin makes actually several hundred commits every month to to production in the Devin repositories. Um and our most common use cases through slack usually. So we have a lot of channels where we talk about um different things. We have channels for bug reports, we have channels for crashes, channels for uh feature requests, channels for front-end issues or all these different things.

Scott： And so now every time somebody posts something to one of those channels, they always tag Devon as well and Devin does a first pass. うん。 Scott： And so sometimes Devin is right. Uh it is correct from the beginning. Sometimes uh it it needs uh a little bit of guidance to to get its to be correct. But it we find that it's much easier to go and solve all these issues uh with the help of Devon's first pass. And of course the nice thing is that Devin is fully asynchronous so you can have lots of Devins running at the same time you can do other work while your Devins are running or or things like that.

質問者: 素晴らしいですか。あの、 デビンがたくさん仕事をしてくれた後に、そのデビンがした仕事を人間がレビューしてマージする。 レビューアーのマージが難しい、あのたくさんあの負荷負担が大きくなってしまうのではないかということを懸念して、あの感じているんですけれども、そのところってどうやって解決していますか。

Scott： Yeah so the the review process, I think Devin is best for tasks that have some verification. Uh that you can verify. Um and so for that reason a lot of bugs or front-end tasks or small features are are very good for that.

質問者: そのあのたくさんのソフトウェアテストをあらかじめ用意してあるという理解でいいでしょうか。

Scott： Yeah so so you can prepare tests but you can also uh do quick testing manually. And so one way to describe it is um you know the whole idea of this Devin is uh you help for the 10% at the beginning where you're planning out the task for Devin and you help for the 10% at the end where you review the code and you make sure it does things correctly, but the 80% in the middle, Devin does it all autonomous. Great. 質問者: そうですね。なるほど。Thank you。 うん。 質問者: あのたぶんコグニッションの社内ではデビン使いと言うかです、デビンを使うのがめっちゃくちゃ上手い人がたくさんいるんじゃないかなと思うんですけど、 うーん。 質問者: その例えばま、色々なユーザーとも会ってると思うんですけど、この人はものすごいデビムを使うのがうまいなって思ったユーザーいます？

Scott： Yes a lot a lot. Uh I mean I think um even even in Japan for example. We've met with a lot of different customers who who are really excited about it. Webnal for example is one of them. I think they told us that actually 39% of all of their code now is written by Devon. はい。 質問者: 39%？ Scott： Yeah なんである。 質問者: Webnarはそんなに多くのことデビにやらせることができてるんですか。

Scott： Yeah so it it does take some real investment I think to get to the point where uh you can understand how to make Devon really really useful. And also you have to teach Devon uh the details about your code base in the beginning some as well. Um but but once you have the kind of the the right workflows uh and the right setup, I think they found that it's it's much faster for them to do a lot of work with the help of Devin. うん。 Scott： Um and similarly, you know, the I think that in the US for example, yeah we have a lot of teams who do even more than 40% of all of their code with Devin. Um and in some cases it's engineers who use Devin. In some cases uh it's engineering leads. In some cases product managers. Um but there there are a lot of different workflows that we've seen that have worked really well.

質問者: これやっぱりプログラムの言語であるとか、そのプロダクトの性質によってこの最大何%くらいまで使えるのかって言うのは結構変わって来るんですかね。 例えば、あのタイプスクリプトとかあるいはPythonとかはかなり、あの強いイメージがあるんですけど、弱い言語もあるよ。 気がしていてま、そういうフレームワークとか言語による得意不得意ってのはあるんですか。

Scott： Yeah I think the languages that are more common uh are are generally the ones that Devin is best at because there's more data out there. Um and so Python and TypeScript are quite common obviously and so I think there's a lot of data for Devin for that. Um but often it actually depends I think more on the use case than on the language. And so if it is uh kind of like small fixes or new features that you can quickly verify, then it's uh it's quite good for Devin. Whereas if it's something that requires a lot of context and kind of like high level decision-making, then perhaps you don't want to use Devin. うんうんうん。

質問者: いいですか？ 男性: どうぞ。 質問者: あの、多分今後どんどんデビンは進化していって、そういうハイレベルな、あのものもできるようになっていくと思うんですけど、そのために何かAIってどういう風に強くなるといいと思っているのかお伺いしたいです。なんかその、なんか推論能力が高まるのか、エージェンティックななんか能力が高まっていくのか、なんかどんな方向を思い描いてるのか聞いてみたいです。 うん。 Scott： Yes yeah so yeah so I I I I think the uh Scott： I I obviously I think there's still a long long way to go. Um I think um I I I think the the long term future I think is a point where you can truly just talk to your computer and tell your computer what to do. I think if you think about what a software engineer really does, you know I think the most important part is to decide what to do and how to build it and and then tell the computer what to do and and so on. Um but obviously currently there's also a lot of implementation detail that you have to handle. And I think more and more of that detail will go away over time. うんうん。 うん。 Scott： So to to to answer your question. I I think the the main two things I think we need to uh uh to to really solve or to make progress on are one is I think um on the capability side, I think there's actually a lot of just real world work that needs to be done uh teaching teaching Devin about like a lot of the details of real world engineering because it's very messy in the real world.

うん。 Scott： Uh and the second thing I would say is for the product interface, I think there's still a lot to solve. And uh even now it's for I mean Devin is in your slack but it is also in GitHub, it's in linear. And I think over time um Devin needs to be able to to work with basically all of the same systems that humans use or or to to have a very clean way to work alongside humans. Um and I think there's still a lot that we are uh discovering with that. And even the new Devin 2.0 with the kind of uh the NIDE, the cloud IDE interface is is already uh kind of a new version of that. うんうん。 男性: So you like do you think like the neck is like more like the interface between like Devin I mean AI and real world AI and human rather than the capability of AI itself?

Scott： So I I think the raw problem-solving of AI is very strong. Uh people have seen a lot of the the the numbers on, you know math problems or or very kind of uh like programming Olympiad problems for example. Um I think the um uh the reality is that a lot of our typical work is very high context. You know if you are solving a bug for example then usually what you do is you, you know, you run the server locally, you try to reproduce the bug yourself. Um you maybe go and click around, you look at the different files, maybe you read documentation. All all all of these different things, you test the code and and so a lot of it is creating a good training loop that mirrors the real world complexity of what software engineers do day to day.

Scott： Yeah so so it is a capabilities problem but it's more about um uh you know tool use, like using your tools well and also um like ingesting a lot of context well I would say rather than pure you know problem solving. うんうんうん。 質問者: もう1つあるクエスチョンいいですか？ あの、先ほど長期的な未来に音声でコンピューターに、あの指示するビアボイスって言う話だったんですけれども、一方で現時点ではキーボードを使ってるプログラマー多いかなと思っているんですね。 現時点でどの程度の割合の人がボイスで音声でそのレビンに音声の中でレビンに指示をしているのかっていうことと、将来的に本当に音声が100%になるのかキーボードがどれくらい残るのかと言うところの、話を、あの聞けると面白いかなと思います。

Scott： Yeah so so it's uh I I I mean less so voice in particular. I think more I was just saying that uh the goal would be able to just kind of express here is exactly what I want to build and then have the agent build it for you. Um it can be through voice or other means too. I mean maybe it'll be through through Neuralink for example.

質問者: あの、先ほどそのソフトウェアのソフトウェアエンジニアの仕事って言うのは実装じゃなくて何を作るべきかって言うの考えるって言うことにシフトしていくんじゃないかってお話ありましたけど、それに達成するまでにあとどれくらいの時間がかかると思いますか？

Scott： Yeah I mean I think that um already today I I think uh an engineer who doesn't use AI is is slower already than an engineer who does use AI. うんうん。 Scott： I think perhaps by the end of this year it may be as big a difference as twice as fast with AI. And then soon you know in the next few years it could be five or 10 times as fast. Scott： Um and so the skills will change for for software engineering but uh I think it's of course exciting because there's actually much more than 10 times as much software that we could build. Um and there's there's always so many ideas of things that we could build or or things that we could do so. Yeah so so I think um I think the skills in software engineering will change and so for example I think uh you know, a lot of the the details of knowing certain uh obscure syntax or something will be less important because you will be able to have AI uh to help you with those things.

質問者: あ、じゃあどこにフォーカスするようになっていくんでしょうかね。 Scott： Yeah so so on the other hand, I think um a lot of the the the core problem solving and and kind of logical reasoning is going to be much much more valuable. And I think actually it's it's always true, you know it's uh at the start like 80 years ago programming, there was punch cards um and then there was assembly and there was you know Pascal and C uh and it's already changed many times. But I think still it has always been the case that the most important skill of a programmer is the ability to take a problem and understand this is exactly what the problem is. Here's how I want to build my solution. Here are the details of the solution. Here are the edge cases. Um and to just be uh basically the decision maker of what should you build. And also of course I think it's uh especially since these are the tools are getting better so quickly, I think it's very valuable um to learn how to use the tools well. Um and to just spend time understanding them because they will continue to to grow and grow.

Scott： But yeah for for example, I mean, um you know 30 years ago there were uh less than a million software engineers in the whole world. And you know between 30 years ago now actually a lot of the implementation is already much easier, you know we have cloud, we have Python, we have TypeScript and all this thing. And and in fact actually there's now 30 million software engineers. You know there's much much more software. And and maybe one way to say it is um since the whole idea of programming is to be able to uh to tell your computer what to do uh either verbally or you know in in typing or whatever but the as as the computers get more and more powerful, of course it will still be up to us to decide what to do. And so I think that the most important skill will be kind of this fundamental uh um we can call it like architect like technical architect work and then you know also a little bit like product management work, but really just deciding what to do in in very specific detail and and how to build things and and what is the solution that you want to build. なるほど。 質問者: そうあのエンジニアというものが今、考えてられてるエンジニアと言う仕事の枠から出て、もっと広い範囲の事をこう見たり考えたり作っていくようなことがエンジニアに、今のエンジニアもっと求められるようになっていくのかなと言う風に思うんですけれども、そうはどうでしょう？ Yeah I think today already you know many of these details are owned by the engineers typically not by any managers or or product managers or things like that because the the very specific details obviously of the implementation uh uh you know the the engineers are the ones who are closest to to the decisions that get made. And so I think those core engineering decisions and trade-offs, you know I think those will still be made by humans and that'll be a lot of the um uh that that that will be the main the main way like the main thing that humans really excel at. 将来にわ 質問者: も人間が得意であり続けると思いますか。

Scott： I I think there's a I I think there's of course a spectrum. the way that I think about it is uh I think the way humans have worked with AI is always been AI does kind of the the tedious or the repetitive parts and humans do the the kind of uh complex or high context or decision-making parts right. Um and and the only thing that's changed is every time there's new developments, we kind of redefine which part is the tedious part and which part is the right and we always kind of move this. But but you know I I I think it's at the end of the day humans should be the ones who are deciding what to build right? So there will always be a part here that is uh that is up to humans. 質問者: あの、日本のお客さんがすごいあの、増えてきているって話聞いたんですけど、あの、に日本のマーケットどうですか？

Scott： Yeah um yeah I mean it's really exciting uh we uh it it it wasn't even available to Japanese it wasn't even available to to Japanese customers until uh a few months ago. Um and and over the last few months so we've seen already you know over a thousand uh different uh companies using Devin in Japan alone. Um and so a lot of the reason that I'm here in Tokyo is uh uh is is to get to meet with uh with many of these teams and companies and uh to learn more about what we can do to to make Devin better for for それで言うと日本のお客さん の傾向とか、特にこういう使われ方が多いなとか、あのそういう学びはありましたか？

Scott： Yeah I would say um the uh especially in the last couple weeks since Devin 2.0 we've seen a lot of folks who have told us that they they're using Devin Search and Devin Wiki much more now. うんうん。 I love that functions. Yeah. Good good. yeah. Um and certainly there's a lot of uh specific workflows that we see and so you know one example which I thought was funny is uh uh now a lot of teams uh we hear from a lot of teams that they onboard their new engineers with Devin Wiki. Uh and so basically Devin is the mentor for the new engineer initially because Devin knows the whole code base. うん。 Yeah similar it's actually also been pretty interesting to hear about a lot of use cases from uh product managers rather than engineers.

Uh and so often people will ask questions or run analytics or make small changes as PMS um and and just have have Devin do that for them. それで言うと、あのま例えばこのAIコーディングの領域って色々なコンペティターがいますよねと。 え、ま例えばカーソルであるとかWindows ハーブであるとか、アクライって言う所、あるいはえオトノマスAIで言うとオールハンズAIとか、そういった所がありますけどそういったコンペティターはどう見てますか？ Yeah so so we know actually a lot of these companies because many of them are also in the Bay area. Um and uh what I would say is yeah right now you know there there are very few autonomous agents uh in the market right now. I think um I I think a lot of the AI IDs for example are are great products, but they're very synchronous right. Um and for us uh we have always focused on kind of this asynchronous and autonomous experience. Um and I think yeah I I I mean I think you know we we were working on this experience and focusing on this even a long time before people really uh talked about agents or people thought agents were possible even I would say. Um and I think that uh uh now that uh now that agents are are gaining a lot more traction, I think people are thinking much more about agents and it's it's very cool to see honestly. I think uh actually I would say that our product experience is is actually very complimentary with with many of these tools. We we know a lot of uh teams who use both uh um both Devin and they use the AI IDs and for example we actually have a big partnership with Microsoft who uh who runs GitHub co-pilot. うんうん。 その非同期的なアシンクロナイズドな、あのAIっていう分野で、えっと長期的にそのコグニション社があの、どれくらいの優位性を持ち続けられるんだろうなと言うのを思っていて、1つあるのは例えばそのファンデーションモデルを作っているオープンAIとか、あとはそのクラウド、アンスロピックとかですね。そういった所もこういったレイヤーに参入して来うる可能性あるなとも思ってるんですよね。 で、えっとコグニション社としてはそういうファンデーションモデルには手を出してない中でえ今後も出さないって言うことなのか、あのそれとも何でしょうね？ この、この領域の色んなプレイヤーがいて他の人達がしみ出して来る可能性どう見てるのか、あその中でどういう風に長期的に勝ち続けられると思っているのかって言うのを聞いてみたいです。 やっぱ圧倒的に体験のクオリティが良いと。で、なんでデビン、あるいはコグニション社って誰よりも早くこの質の高い体験をローンチできたんですか？