Machine-generated source transcript; recognition errors remain. Timestamps and subtitles use the selected source recording’s clock, not clip-relative time. Section links use cited source moments; sections without a verified moment return to the talk start.
6:08:26Hey, everyone. I'm Ishan, and I work on anti-gravity at Google DeepMind.
6:08:32Today I'm going to talk about the future of building products,
6:08:34what judgment looks like,
6:08:36and some lessons that we've learned
6:08:38from shipping antigravity at Google.
6:08:44So, antigravity is a tool to get work done with agents,
6:08:47and we're trying to make the best interface
6:08:49for developers and knowledge workers
6:08:51to do high-intelligence tasks.
6:08:55The software engineering we used to know
6:08:57has completely changed.
6:08:59We can do things now
6:09:00that we wouldn't have even dreamed of a year ago.
6:09:04A sentence can turn into an operating system,
6:09:06and my voice can create a AAA game.
6:09:11So what's left for us?
6:09:14The tools we use have changed.
6:09:16We used to use the code editor.
6:09:17It's been around for decades.
6:09:19But now, every year, if not more often than that,
6:09:22there's a new paradigm for how to do our work.
6:09:25It redefines what our work even looks like.
6:09:28And on the side of building products,
6:09:31That's changed as well.
6:09:34We used to spend a lot of time up front,
6:09:37figuring out the right thing to make,
6:09:40because building was expensive.
6:09:42It took a lot of time, effort, and resources,
6:09:43so we ideated more and invested more in that phase.
6:09:47This is completely flipped,
6:09:50where now you should still understand the problem
6:09:52you're trying to solve and how to go about it,
6:09:55but if you want to make something, it can exist just like that.
6:10:00You can try it, feel it, get feedback, and understand the nuances that you might not have
6:10:05realized about it before, and use that to inform how to move forward.
6:10:09This is exciting, but it's also a little bit deceptive.
6:10:14It's easier than ever to just turn our brains off and blindly accept what's made for us.
6:10:19So if anything can be made by anyone, then it's more important than ever to know what
6:10:24what the right thing to even make is.
6:10:28Everyone knows their own definition for what taste or judgment looks like.
6:10:34And I think that's OK.
6:10:35That's good.
6:10:36Because we all work in different ways, and we're all working towards different things.
6:10:41But by the end of this talk, I'm hoping that I can help you form a little bit of a clearer
6:10:45idea on what judgment in building products looks like to you.
6:10:50I've gone from the scale of working at startups to now at Google DeepMind, where the scale
6:10:54is, it feels pretty incredible.
6:10:56We're making frontier models, and on the product side,
6:10:59we're able to use this internal frontier
6:11:02to build the interfaces that will use the next steps
6:11:06everyone will see in intelligence.
6:11:08We're very close to the research,
6:11:10and we can experiment and iterate really fast
6:11:12and ship to over 100,000 people inside Google.
6:11:18These people at Google are using antigravity internally
6:11:20in really ambitious and unique ways.
6:11:22And this creates a really cool feedback loop where we can continue building and do awesome things.
6:11:27I want to walk you through three big lessons that we've learned from this cycle.
6:11:34So I think we've all seen this first pattern of the tools we use scaling up with intelligence.
6:11:41We used to build products in a pretty relatively straightforward way.
6:11:45We had to identify a problem, understand it deeply, and make something that people want with a clear target.
6:11:51Now that target is moving.
6:11:53The ceiling is going up for what's even possible to make.
6:11:57And with that, the possible value that you can create
6:12:00in the products that you build is changing as well.
6:12:05So I'm always coming back to a few thoughts whenever I'm
6:12:08working of what we can achieve now
6:12:10and what's just past the edge of what we can achieve.
6:12:14And we've seen this look like two main ways
6:12:17building at DeepMind.
6:12:20As our models got better earlier in the year, they got better at communicating with and
6:12:24orchestrating each other, we created systems for teamwork to do more ambitious work.
6:12:29We stepped up in intelligence to let users do more.
6:12:32At the time, it was awesome.
6:12:33You could create operating systems and really cool things we've never seen before, but it
6:12:38was also pretty prescriptive because the models needed it at the time.
6:12:41More recently, this has looked like our version of agent swarms.
6:12:44You can actually try this now in anti-gravity.
6:12:45But this has been creating a ton of value inside Google for us.
6:12:50People can be more ambitious, there's more agents working together, and we're able to
6:12:54relax the constraints to achieve more.
6:12:58You can bet that with the next models coming out, we'll see even more cool patterns and
6:13:03things we didn't expect.
6:13:05But I think we all already kind of knew this.
6:13:07We've been seeing this for the past few years.
6:13:09All the tools we're using are just adding more and more new, cool ways to scale up with
6:13:14intelligence.
6:13:16But the second way I've seen evolving with intelligence look
6:13:20is a little bit trickier, because being ambitious does not
6:13:23always mean adding more.
6:13:26Let's go back in time a little bit and walk together
6:13:28through a decision in the product.
6:13:32The way we used to use AI tools looked very much
6:13:35like a structured turn-based format, like this.
6:13:38I would ask for some small scoped change,
6:13:40for example, editing a few files.
6:13:42The model would do it, and then we'd
6:13:43move on to the next thing, step by step.
6:13:46It made sense at the time to be able to cue instructions
6:13:50so the next things I wanted could be delivered to the model
6:13:52one at a time.
6:13:53They might not have been smart enough to get everything at once,
6:13:56but it made sense to give it limited context progressively
6:13:59it could get the work done bit by bit.
6:14:03Fast forward to now.
6:14:07There's a lot more going on.
6:14:09Long running agents, asynchronous tasks,
6:14:11external triggers, and everything else we're doing.
6:14:14Things are looking more complicated,
6:14:15and the agents are also smarter.
6:14:17So maybe the decision we made in the past
6:14:19might no longer be correct.
6:14:21Maybe a better pattern looks like treating the agent more
6:14:24as an entity that can make better decisions,
6:14:25and giving it user intent immediately, just like this.
6:14:30Maybe it'll choose to orchestrate to a bunch of worker agents,
6:14:33or do the task after we're done the current task at hand.
6:14:35It's smarter, so it can decide for itself.
6:14:39The decision we made in the past can easily become obsolete as intelligence grows, and you need to challenge your prior assumptions.
6:14:49If we separate the timeline of what we can achieve into three clear parts, if you build too far in the past, it'll be obsolete.
6:14:56If you build too far for the future, it could be visionary but impractical.
6:15:01It's important for us to ideate quickly and work towards building in the right time window
6:15:07create the most value.
6:15:08Challenge your past assumptions and rebuild your foundations to create good systems that
6:15:12scale with intelligence.
6:15:16With these foundations, we've seen another really awesome pattern evolve or come out
6:15:21at Google where we see the value in creating a platform for ambition.
6:15:28So, something like this might sound familiar.
6:15:31I used to use design tools in a very manual way.
6:15:33I'd create frames, build user interfaces manually,
6:15:37and create things by hand.
6:15:39And the tool, if we look at how it changed from before to now,
6:15:44it looks pretty similar, but the way I use it has completely flipped.
6:15:47Maybe now agents are creating hundreds of visualizations
6:15:49for me to go over very broadly before I go deeper into something.
6:15:56building software now doesn't always look like giving the user value in a very clear and
6:16:01deterministic way users can create their own leverage and make room for their own ambition
6:16:06let me walk you through how this has looked for us in our product there's a lot of room for
6:16:12extensibility people can build specialized tools that work for them and this can often end up
6:16:18leading back into the product.
6:16:20I'll show you some examples.
6:16:22Personal software can come up where I can make a dashboard
6:16:25where agents will scan across feedback channels and bug
6:16:29fixes to show me all this information in a way that just
6:16:32I like.
6:16:34Maybe it'll look like software for my team, where designers
6:16:38can create the tool that works just for them, where they can
6:16:41collaborate visually on demos.
6:16:44This can also branch out even more.
6:16:45The science team at Google is doing some awesome things
6:16:47with their science plugins.
6:16:50We're creating this platform for ambition
6:16:51for a wide range of users to create their own value
6:16:53and create their own leverage.
6:16:55And this can actually feed back into the product
6:16:57you're making at its core.
6:17:00The agent swarm example I mentioned earlier
6:17:02actually started as a plugin through our extensibility system.
6:17:06It's created a ton of value internally for us at Google.
6:17:09And by giving the users surface area
6:17:12to explore and create value for themselves,
6:17:14we were able to learn from it and iterate in the product
6:17:17to make the core better.
6:17:20Let your users be ambitious
6:17:22and watch it feedback into your product
6:17:24to create more value for everyone.
6:17:29The last principle I want to go over
6:17:30kind of applies to all of this.
6:17:35There's a lot of great talks at this conference
6:17:37about software factories.
6:17:39You know, it's really exciting.
6:17:40We can give our agents ways to build more,
6:17:43verify more reliably, and ship faster.
6:17:46It's exciting.
6:17:47But if all the parts of what software engineering used to look like get abstracted away,
6:17:52then how do we define what the remaining value in our human judgment is?
6:17:58I'll argue that human judgment is more important than ever.
6:18:01If I ask an agent to, say, make this slide better,
6:18:05maybe it'll add some visualizations and a bunch of dashboards and cool colors and everything.
6:18:09Maybe it'll rewrite the whole slideshow in Rust.
6:18:14So it'll solve every problem, but does every problem need to be solved?
6:18:20Take, for instance, what the agent manager looked like around a year ago when Antigravity
6:18:24first launched.
6:18:27If we want to make it better to an agent, that might look like adding a lot more detail.
6:18:32Ship everything.
6:18:34Add more detail into every single step the agent is taking and give the user knobs for
6:18:38everything.
6:18:40It might be better in some ways.
6:18:44But it's easy to blindly accept the first thing that we're given.
6:18:48What does better even mean to us?
6:18:49If we're trying to make a better agent-manager interface,
6:18:52maybe that looks a little bit different.
6:18:55Maybe we need to ship fast,
6:18:57but understand at each step of the iteration process
6:19:00what better actually means to us and what the next direction is.
6:19:04Let's start here.
6:19:08Better might mean simplifying the interface,
6:19:11only showing the user exactly what the latest thing the agent
6:19:14is doing semantically.
6:19:15So they're not shown a wall of text and tool calls.
6:19:18That might not actually matter.
6:19:19Maybe it looks like simplifying the whole interface
6:19:22to give the user less cognitive load
6:19:25so that they can multitask easier
6:19:28and be able to achieve more in a different way.
6:19:35So this kind of brings us closer to what antigravity looks like today.
6:19:40All these steps along the process have brought us closer to what good means.
6:19:46And we're making a lot of improvements fast, so I'm excited for you all to see it.
6:19:51Now's the time to try many things.
6:19:54Be ambitious.
6:19:55It's more exciting than ever, and it'll actually get more exciting as the models get better.
6:20:03Try many directions, but take a moment to think. Use your human judgment. Just because
6:20:08you're finished creating something first, which can be done in an instant, it doesn't
6:20:12actually mean you're done. Think about what you're actually trying to achieve and how
6:20:17the thing you're making feels. Iterate quickly, identify many directions you can go in and
6:20:24try them, but arrive at the right thing and go deep into that.
6:20:30So make time to think, make room to think, and use your human judgment.
6:20:36To summarize, some of the lessons we've learned involve scaling up with intelligence, which
6:20:42doesn't always mean adding more.
6:20:43It could mean rebuilding your foundations.
6:20:46Take these foundations and enable your users to be ambitious in new ways you might not
6:20:49have expected before.
6:20:51And at the end of all this, when we're shipping faster and faster, make room to think and protect
6:20:56the human side of software engineering.
6:21:00At Google, this has led to some awesome things by creating a platform for ambition.
6:21:04Improvements to YouTube, better video decoders, and memory optimizations.
6:21:11As for DeepMind and anti-gravity, it's a more exciting time than ever, with new models coming
6:21:14out.
6:21:16And on the product side, we're also moving quicker than ever.
6:21:19These are just some of the future launches we've announced in the last two weeks.
6:21:28And we're continuing to use the principles we've covered today to work on the future
6:21:31of interfaces for higher intelligence.
6:21:33I'm excited for you all to see it, and I'll leave you with this.
6:21:37This new age of software engineering is actually worth celebrating.
6:21:41Agents might be doing more and more, but it's also more important than ever to solve real
6:21:45problems and use good product thinking.
6:21:48Protect your judgment and build things that people want.
6:21:52Thank you.
