Logan Kilpatrick from Google DeepMind: Building for 100m developers
Logan Kilpatrick on why 'code flying across the screen' in vibe-coding tools is useless even to trained engineers, how DeepMind's org merger actually happened, and why he still wakes up every day treating Google's AI lead as the fight of his life.
Logan Kilpatrick leads developer relations for the Gemini API and AI Studio at Google DeepMind, after previously working on developer relations at OpenAI. We talk about his goal of getting Google's tools in front of 100 million developers, what's actually missing from most vibe-coding products today, and what changed inside Google over the last 18 months that turned its AI story around.
We also get into the CLI-versus-IDE debate as Windsurf's team joins Google, why he thinks the market fragmenting into dozens of successful AI coding tools (not one winner) is actually a sign of how big developer demand really is, and what he's most excited about over the next six months.
The Q&A below is experimental — extracted from the episode transcript with AI. If you spot a mistake, let us know at jack@scalingdevtools.com.
You've said you want to get Google's tools to 100 million developers. How do you even approach a goal that ambitious?
It won't just be us doing this — it's an ecosystem-wide thing. There's something like 30 million professional developers writing code today, and the narrative online feels split in two: half the world thinks you shouldn't learn to code anymore, just prompt your way there, and the other half says you still need to learn. I'm in the you-should-still-learn-to-code camp, but I think vibe coding and AI-accelerated coding are this perfect storm for taking the drudgery out of building software. My own experience learning C++ was boring specifically because the early, interesting stuff is always at the tail end of the learning curve — and I remember having to hunt down a professor or a tutor, or hope someone had already asked your exact question on Stack Overflow. Now you have a personal tutor available on demand, twenty-four hours a day, that can reformulate the problem however you need. That really is going to be a powerful enabling factor for bringing in the next 100 million people who can learn to code.
You're reaching vibe coders as well as classic developers — pretty different audiences. How are you thinking about both?
Our core products, AI Studio and the API, have historically been very traditional-developer-centric, and now we're building a vibe-coding product inside AI Studio to bring in that next group. I think the missing link in most current vibe-coding products is that they don't do the thing that's actually useful to the learner. Look at what the outcome of most of these tools is optimized for: build the software for you, deploy it, and hopefully help you monetize it. That's all genuinely useful, but it misses the more important question — how do you bring people along with you as they build these complex systems, in a really incremental way? I use v0 constantly and love it, but when there's code flying across the screen while it works, it's entirely useless to me — even as someone who studied computer science and has shipped software at Apple, ChatGPT, and now Google. There's a real design problem here around showing the right level of detail and hiding the wrong things, and I think there's a huge amount still to be done in that space, for traditional developers too, not just vibe coders.
You just launched Gemini CLI, and a lot of the Windsurf team joined Google. How are you thinking about CLI tools versus IDEs versus vibe-coding products?
It feels like there's a genuinely different user persona for each of those. My own take, which might just be about how I personally learned to code — I was never a Vim or Emacs person, I use the terminal out of functional necessity, but I've never found it a great way to actually develop, so I've personally never been excited by CLI tools. I've basically just accepted we'll have both, because some people love them and some don't. Gemini CLI is crushing it, and it'll be interesting to see what the Windsurf folks build now that they're part of DeepMind. It's also struck me how the narrative keeps resetting: a couple of years ago it felt like GitHub Copilot had already won the whole market, then maybe eight or nine months ago it felt like Cursor had won it, and now there's Windsurf, Cognition, Amp from Sourcegraph, Cline, and fifty other tools all doing genuinely well. That's actually solidified in my mind just how big the developer market really is, and how different developers' needs are from each other — there's room for a lot of these companies to build really successful products.
It felt for a while like Copilot won, then Cursor won. Now there are dozens of tools doing well. What does that tell you?
What's remarkable is Copilot is doing well at the same time VS Code has built a bunch of AI features natively, at the same time lots of people are using Cursor and other tools. It's a genuinely cool moment for the ecosystem, and all of us as developers are the beneficiaries — I love the competition, because it means better tools for me as a developer. The flip side is that at bigger companies, Copilot is often the only mandated option, and once you're in a large company you're typically much less able to just go try Cursor even if you wanted to. Google does sell a lot of enterprise infrastructure already — BigQuery is a great example of something ubiquitously used across the ecosystem — but this AI moment has reset a lot of the conventional wisdom that just because you're Google with huge distribution, you automatically win the market. It's genuinely competitive across every dimension, from the models to the products to developer tooling, and if we're not laser-focused on winning across all of it, from support to error messages in the API to how easy it is to adopt Gemini, it's very easy to get lost in how much is happening in the ecosystem.
You left OpenAI for Google, which surprised a lot of people at the time. What actually changed at Google over the last 18 months?
There's definitely a large degree of coincidence, but I think there are a couple of real dimensions to it. One is that being organizationally set up to succeed matters enormously, especially at a big company — pre-LLM, Google was doing a lot of things and was largely successful across search, Docs, and long-term speculative bets like Waymo that have started paying off, plus YouTube and Google Cloud both becoming huge businesses. AI had already been underlying a lot of that; Transformers came out of Google and were already powering search. But organizationally, Google had many different teams doing overlapping LLM work — Google Research, Google Brain, and DeepMind, which had the original mission of building AGI and had done a lot of the foundational RL work. In 2023, the decision was made to bring those teams together into a single consolidated DeepMind, and that consolidation is a big part of why the pace of model progress accelerated. Since then, DeepMind has gone from doing foundational research to building and deploying the models to actually owning product experiences too — our developer team, the Gemini API and AI Studio, sits inside DeepMind, right next to the researchers training the models, and so does the Gemini app team. All of that had to happen to put us in a position to be successful, and getting thousands of people aligned takes real time — it's much more a human problem than a technology problem, even in an era where everything feels like a technology problem.
What are you most excited about over the next six months?
A few things. People are finally figuring out how to build AI products that people actually like — around 2023, most AI products weren't great, mostly because the models were changing so fast that nobody had settled on what a good customer experience even looked like yet. I think the models are already good enough today that you can largely build the thing you want, so the constraint becomes whether you have the right context to build it. On the model side, we keep unlocking new dimensions of scaling — pretraining scaling, then about a year and a half of post-training scaling, and now RL scaling, which xAI has talked publicly about betting heavily on. On top of that, tool use is becoming a new scaling dimension of its own, as models turn into systems that get access to more hosted tools and tool chains, and there's a lot of headroom left in training models to be better at those complex tool interactions. From a builder's perspective, AI coding tools keep working better, millions of new people are using vibe-coding products, and companies like Lovable are announcing some of the fastest paths to $100M ARR ever — that's all incrementally positive for the whole ecosystem.
Any developer tools outside AI coding tools that you personally find exciting?
I'm really excited about the Live API we've been building — we released it back in December and have kept improving the underlying models and function calling since. The model can see what's on your screen, hear what you're saying, and take proactive action based on your system instructions. Some of the early traction is around copilots that help people use complex products — imagine going into something like Photoshop, which has a thousand buttons you have no idea what they do, and just clicking one to talk to the model about what you're trying to do while it draws bounding boxes around the right buttons and walks you through the workflow. I think about this with my grandma using different websites — she could just click a button and say what she's trying to do. That's a fairly basic use case, but I think we'll increasingly feel this kind of AI co-presence as the Live API gets more adoption. I also want this for pair programming — every IDE just has a button, and behind it is an on-demand, twenty-four-hour pair programmer that can see your screen with permission, talk to you, and run commands for you. That's not some futuristic version of the world — it's literally the version we're living in right now; people just need to build the product experience around it.
You're often the direct point of contact when people have issues with Google's AI products. How do you keep that sustainable?
It's honestly not that difficult for me, partly because of how I started out — during undergrad I worked twenty-five hours a week at the Apple Store on top of going to school full time, first selling iPhones and eventually becoming a technical expert helping people troubleshoot. You'd get everyone from someone who genuinely didn't understand what their phone was doing to actual engineers from the core Apple teams, and you had to be able to talk to both. That gave me thousands of iterations at exactly this kind of interaction, and it trained a reward model in my brain that, to my detriment sometimes, isn't very scalable — X does make it somewhat scalable, but it can become a bad crutch, where spending a lot of time answering individual complaints on Twitter is usually a symptom that there's some other underlying problem I should actually be fixing. So I try to reflect on that balance: actually answering someone's question versus stepping back to solve the broader problem space so it's easier for everyone, not just the person who happened to tweet at me.
Links & references
- Google DeepMind ↗
- Logan Kilpatrick's X ↗
- Logan's podcast, Around the Prompt ↗
- NotebookLM ↗
- Gemini CLI ↗
- Veo ↗
This episode is brought to you by WorkOS.