1 06:08:26,100 --> 06:08:31,080 Hey, everyone. I'm Ishan, and I work on anti-gravity at Google DeepMind. 2 06:08:32,120 --> 06:08:34,660 Today I'm going to talk about the future of building products, 3 06:08:34,660 --> 06:08:36,400 what judgment looks like, 4 06:08:36,600 --> 06:08:38,820 and some lessons that we've learned 5 06:08:38,820 --> 06:08:40,880 from shipping antigravity at Google. 6 06:08:44,040 --> 06:08:47,320 So, antigravity is a tool to get work done with agents, 7 06:08:47,560 --> 06:08:49,040 and we're trying to make the best interface 8 06:08:49,040 --> 06:08:51,420 for developers and knowledge workers 9 06:08:51,420 --> 06:08:52,900 to do high-intelligence tasks. 10 06:08:55,920 --> 06:08:57,980 The software engineering we used to know 11 06:08:57,980 --> 06:08:59,180 has completely changed. 12 06:08:59,620 --> 06:09:00,920 We can do things now 13 06:09:00,920 --> 06:09:03,160 that we wouldn't have even dreamed of a year ago. 14 06:09:04,200 --> 06:09:06,480 A sentence can turn into an operating system, 15 06:09:06,660 --> 06:09:09,200 and my voice can create a AAA game. 16 06:09:11,100 --> 06:09:12,580 So what's left for us? 17 06:09:14,520 --> 06:09:15,820 The tools we use have changed. 18 06:09:16,540 --> 06:09:17,800 We used to use the code editor. 19 06:09:17,980 --> 06:09:19,100 It's been around for decades. 20 06:09:19,300 --> 06:09:22,380 But now, every year, if not more often than that, 21 06:09:22,600 --> 06:09:25,140 there's a new paradigm for how to do our work. 22 06:09:25,240 --> 06:09:27,180 It redefines what our work even looks like. 23 06:09:28,880 --> 06:09:31,220 And on the side of building products, 24 06:09:31,220 --> 06:09:32,700 That's changed as well. 25 06:09:34,060 --> 06:09:36,820 We used to spend a lot of time up front, 26 06:09:37,000 --> 06:09:38,920 figuring out the right thing to make, 27 06:09:40,560 --> 06:09:41,760 because building was expensive. 28 06:09:42,000 --> 06:09:43,660 It took a lot of time, effort, and resources, 29 06:09:43,800 --> 06:09:46,660 so we ideated more and invested more in that phase. 30 06:09:47,640 --> 06:09:48,940 This is completely flipped, 31 06:09:50,300 --> 06:09:52,360 where now you should still understand the problem 32 06:09:52,360 --> 06:09:54,960 you're trying to solve and how to go about it, 33 06:09:55,120 --> 06:09:58,020 but if you want to make something, it can exist just like that. 34 06:10:00,460 --> 06:10:05,580 You can try it, feel it, get feedback, and understand the nuances that you might not have 35 06:10:05,580 --> 06:10:08,640 realized about it before, and use that to inform how to move forward. 36 06:10:09,740 --> 06:10:13,380 This is exciting, but it's also a little bit deceptive. 37 06:10:14,080 --> 06:10:18,840 It's easier than ever to just turn our brains off and blindly accept what's made for us. 38 06:10:19,660 --> 06:10:24,980 So if anything can be made by anyone, then it's more important than ever to know what 39 06:10:24,980 --> 06:10:26,420 what the right thing to even make is. 40 06:10:28,080 --> 06:10:31,740 Everyone knows their own definition for what taste or judgment looks like. 41 06:10:34,140 --> 06:10:35,060 And I think that's OK. 42 06:10:35,260 --> 06:10:35,680 That's good. 43 06:10:36,640 --> 06:10:39,980 Because we all work in different ways, and we're all working towards different things. 44 06:10:41,280 --> 06:10:45,400 But by the end of this talk, I'm hoping that I can help you form a little bit of a clearer 45 06:10:45,400 --> 06:10:48,640 idea on what judgment in building products looks like to you. 46 06:10:50,400 --> 06:10:54,580 I've gone from the scale of working at startups to now at Google DeepMind, where the scale 47 06:10:54,580 --> 06:10:55,920 is, it feels pretty incredible. 48 06:10:56,400 --> 06:10:59,660 We're making frontier models, and on the product side, 49 06:10:59,680 --> 06:11:02,140 we're able to use this internal frontier 50 06:11:02,840 --> 06:11:06,640 to build the interfaces that will use the next steps 51 06:11:06,640 --> 06:11:07,700 everyone will see in intelligence. 52 06:11:08,600 --> 06:11:10,060 We're very close to the research, 53 06:11:10,100 --> 06:11:12,780 and we can experiment and iterate really fast 54 06:11:12,780 --> 06:11:16,100 and ship to over 100,000 people inside Google. 55 06:11:18,120 --> 06:11:20,540 These people at Google are using antigravity internally 56 06:11:20,540 --> 06:11:22,340 in really ambitious and unique ways. 57 06:11:22,340 --> 06:11:26,900 And this creates a really cool feedback loop where we can continue building and do awesome things. 58 06:11:27,940 --> 06:11:32,240 I want to walk you through three big lessons that we've learned from this cycle. 59 06:11:34,620 --> 06:11:39,520 So I think we've all seen this first pattern of the tools we use scaling up with intelligence. 60 06:11:41,040 --> 06:11:45,180 We used to build products in a pretty relatively straightforward way. 61 06:11:45,400 --> 06:11:51,000 We had to identify a problem, understand it deeply, and make something that people want with a clear target. 62 06:11:51,920 --> 06:11:53,260 Now that target is moving. 63 06:11:53,500 --> 06:11:56,640 The ceiling is going up for what's even possible to make. 64 06:11:57,200 --> 06:12:00,120 And with that, the possible value that you can create 65 06:12:00,120 --> 06:12:02,380 in the products that you build is changing as well. 66 06:12:05,400 --> 06:12:08,800 So I'm always coming back to a few thoughts whenever I'm 67 06:12:08,800 --> 06:12:10,460 working of what we can achieve now 68 06:12:10,460 --> 06:12:14,200 and what's just past the edge of what we can achieve. 69 06:12:14,680 --> 06:12:16,540 And we've seen this look like two main ways 70 06:12:17,420 --> 06:12:18,100 building at DeepMind. 71 06:12:20,180 --> 06:12:24,400 As our models got better earlier in the year, they got better at communicating with and 72 06:12:24,400 --> 06:12:29,100 orchestrating each other, we created systems for teamwork to do more ambitious work. 73 06:12:29,260 --> 06:12:32,260 We stepped up in intelligence to let users do more. 74 06:12:32,540 --> 06:12:33,660 At the time, it was awesome. 75 06:12:33,720 --> 06:12:38,340 You could create operating systems and really cool things we've never seen before, but it 76 06:12:38,340 --> 06:12:40,720 was also pretty prescriptive because the models needed it at the time. 77 06:12:41,480 --> 06:12:43,880 More recently, this has looked like our version of agent swarms. 78 06:12:44,000 --> 06:12:45,640 You can actually try this now in anti-gravity. 79 06:12:45,640 --> 06:12:49,580 But this has been creating a ton of value inside Google for us. 80 06:12:50,400 --> 06:12:54,520 People can be more ambitious, there's more agents working together, and we're able to 81 06:12:54,520 --> 06:12:56,980 relax the constraints to achieve more. 82 06:12:58,380 --> 06:13:03,580 You can bet that with the next models coming out, we'll see even more cool patterns and 83 06:13:03,580 --> 06:13:04,220 things we didn't expect. 84 06:13:05,040 --> 06:13:07,260 But I think we all already kind of knew this. 85 06:13:07,840 --> 06:13:09,660 We've been seeing this for the past few years. 86 06:13:09,760 --> 06:13:14,940 All the tools we're using are just adding more and more new, cool ways to scale up with 87 06:13:14,940 --> 06:13:15,220 intelligence. 88 06:13:16,760 --> 06:13:20,180 But the second way I've seen evolving with intelligence look 89 06:13:20,180 --> 06:13:23,860 is a little bit trickier, because being ambitious does not 90 06:13:23,860 --> 06:13:24,880 always mean adding more. 91 06:13:26,360 --> 06:13:28,800 Let's go back in time a little bit and walk together 92 06:13:28,800 --> 06:13:31,720 through a decision in the product. 93 06:13:32,560 --> 06:13:35,320 The way we used to use AI tools looked very much 94 06:13:35,320 --> 06:13:37,620 like a structured turn-based format, like this. 95 06:13:38,100 --> 06:13:40,740 I would ask for some small scoped change, 96 06:13:40,860 --> 06:13:42,140 for example, editing a few files. 97 06:13:42,320 --> 06:13:43,680 The model would do it, and then we'd 98 06:13:43,680 --> 06:13:45,680 move on to the next thing, step by step. 99 06:13:46,020 --> 06:13:49,600 It made sense at the time to be able to cue instructions 100 06:13:50,180 --> 06:13:52,820 so the next things I wanted could be delivered to the model 101 06:13:52,820 --> 06:13:53,420 one at a time. 102 06:13:53,760 --> 06:13:56,280 They might not have been smart enough to get everything at once, 103 06:13:56,420 --> 06:13:59,100 but it made sense to give it limited context progressively 104 06:13:59,740 --> 06:14:02,040 it could get the work done bit by bit. 105 06:14:03,500 --> 06:14:04,380 Fast forward to now. 106 06:14:07,100 --> 06:14:08,400 There's a lot more going on. 107 06:14:09,060 --> 06:14:11,060 Long running agents, asynchronous tasks, 108 06:14:11,060 --> 06:14:13,280 external triggers, and everything else we're doing. 109 06:14:14,240 --> 06:14:15,620 Things are looking more complicated, 110 06:14:15,700 --> 06:14:16,860 and the agents are also smarter. 111 06:14:17,240 --> 06:14:19,100 So maybe the decision we made in the past 112 06:14:19,100 --> 06:14:20,420 might no longer be correct. 113 06:14:21,080 --> 06:14:24,180 Maybe a better pattern looks like treating the agent more 114 06:14:24,180 --> 06:14:25,780 as an entity that can make better decisions, 115 06:14:25,980 --> 06:14:28,660 and giving it user intent immediately, just like this. 116 06:14:30,600 --> 06:14:32,960 Maybe it'll choose to orchestrate to a bunch of worker agents, 117 06:14:33,080 --> 06:14:35,340 or do the task after we're done the current task at hand. 118 06:14:35,540 --> 06:14:37,780 It's smarter, so it can decide for itself. 119 06:14:39,180 --> 06:14:46,540 The decision we made in the past can easily become obsolete as intelligence grows, and you need to challenge your prior assumptions. 120 06:14:49,360 --> 06:14:56,560 If 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. 121 06:14:56,840 --> 06:15:00,500 If you build too far for the future, it could be visionary but impractical. 122 06:15:01,120 --> 06:15:06,760 It's important for us to ideate quickly and work towards building in the right time window 123 06:15:07,500 --> 06:15:08,240 create the most value. 124 06:15:08,960 --> 06:15:12,900 Challenge your past assumptions and rebuild your foundations to create good systems that 125 06:15:12,900 --> 06:15:13,560 scale with intelligence. 126 06:15:16,160 --> 06:15:21,060 With these foundations, we've seen another really awesome pattern evolve or come out 127 06:15:21,060 --> 06:15:27,160 at Google where we see the value in creating a platform for ambition. 128 06:15:28,080 --> 06:15:30,220 So, something like this might sound familiar. 129 06:15:31,400 --> 06:15:33,500 I used to use design tools in a very manual way. 130 06:15:33,740 --> 06:15:35,760 I'd create frames, build user interfaces manually, 131 06:15:37,380 --> 06:15:39,240 and create things by hand. 132 06:15:39,940 --> 06:15:44,100 And the tool, if we look at how it changed from before to now, 133 06:15:44,180 --> 06:15:46,700 it looks pretty similar, but the way I use it has completely flipped. 134 06:15:47,060 --> 06:15:49,260 Maybe now agents are creating hundreds of visualizations 135 06:15:49,260 --> 06:15:52,520 for me to go over very broadly before I go deeper into something. 136 06:15:56,800 --> 06:16:01,240 building software now doesn't always look like giving the user value in a very clear and 137 06:16:01,240 --> 06:16:05,900 deterministic way users can create their own leverage and make room for their own ambition 138 06:16:06,760 --> 06:16:12,720 let me walk you through how this has looked for us in our product there's a lot of room for 139 06:16:12,720 --> 06:16:18,540 extensibility people can build specialized tools that work for them and this can often end up 140 06:16:18,540 --> 06:16:19,600 leading back into the product. 141 06:16:20,340 --> 06:16:21,400 I'll show you some examples. 142 06:16:22,500 --> 06:16:25,300 Personal software can come up where I can make a dashboard 143 06:16:25,300 --> 06:16:29,000 where agents will scan across feedback channels and bug 144 06:16:29,000 --> 06:16:31,740 fixes to show me all this information in a way that just 145 06:16:32,300 --> 06:16:32,980 I like. 146 06:16:34,700 --> 06:16:38,260 Maybe it'll look like software for my team, where designers 147 06:16:38,260 --> 06:16:41,120 can create the tool that works just for them, where they can 148 06:16:41,120 --> 06:16:42,700 collaborate visually on demos. 149 06:16:44,080 --> 06:16:45,620 This can also branch out even more. 150 06:16:45,840 --> 06:16:47,860 The science team at Google is doing some awesome things 151 06:16:47,860 --> 06:16:49,000 with their science plugins. 152 06:16:50,040 --> 06:16:51,680 We're creating this platform for ambition 153 06:16:51,680 --> 06:16:53,960 for a wide range of users to create their own value 154 06:16:53,960 --> 06:16:54,900 and create their own leverage. 155 06:16:55,780 --> 06:16:57,940 And this can actually feed back into the product 156 06:16:57,940 --> 06:16:59,280 you're making at its core. 157 06:17:00,600 --> 06:17:02,680 The agent swarm example I mentioned earlier 158 06:17:02,680 --> 06:17:05,560 actually started as a plugin through our extensibility system. 159 06:17:06,140 --> 06:17:08,660 It's created a ton of value internally for us at Google. 160 06:17:09,400 --> 06:17:12,360 And by giving the users surface area 161 06:17:12,360 --> 06:17:14,160 to explore and create value for themselves, 162 06:17:14,440 --> 06:17:17,180 we were able to learn from it and iterate in the product 163 06:17:17,180 --> 06:17:18,580 to make the core better. 164 06:17:20,800 --> 06:17:22,220 Let your users be ambitious 165 06:17:22,760 --> 06:17:24,620 and watch it feedback into your product 166 06:17:24,620 --> 06:17:26,480 to create more value for everyone. 167 06:17:29,040 --> 06:17:30,660 The last principle I want to go over 168 06:17:30,660 --> 06:17:32,380 kind of applies to all of this. 169 06:17:35,900 --> 06:17:37,520 There's a lot of great talks at this conference 170 06:17:37,520 --> 06:17:39,080 about software factories. 171 06:17:39,380 --> 06:17:40,480 You know, it's really exciting. 172 06:17:40,680 --> 06:17:43,060 We can give our agents ways to build more, 173 06:17:43,280 --> 06:17:45,900 verify more reliably, and ship faster. 174 06:17:46,400 --> 06:17:47,180 It's exciting. 175 06:17:47,960 --> 06:17:52,240 But if all the parts of what software engineering used to look like get abstracted away, 176 06:17:52,540 --> 06:17:57,140 then how do we define what the remaining value in our human judgment is? 177 06:17:58,280 --> 06:18:00,800 I'll argue that human judgment is more important than ever. 178 06:18:01,960 --> 06:18:04,660 If I ask an agent to, say, make this slide better, 179 06:18:05,260 --> 06:18:09,360 maybe it'll add some visualizations and a bunch of dashboards and cool colors and everything. 180 06:18:09,620 --> 06:18:11,540 Maybe it'll rewrite the whole slideshow in Rust. 181 06:18:14,100 --> 06:18:19,200 So it'll solve every problem, but does every problem need to be solved? 182 06:18:20,600 --> 06:18:24,740 Take, for instance, what the agent manager looked like around a year ago when Antigravity 183 06:18:24,740 --> 06:18:25,680 first launched. 184 06:18:27,320 --> 06:18:32,340 If we want to make it better to an agent, that might look like adding a lot more detail. 185 06:18:32,800 --> 06:18:33,820 Ship everything. 186 06:18:34,100 --> 06:18:38,920 Add more detail into every single step the agent is taking and give the user knobs for 187 06:18:38,920 --> 06:18:39,320 everything. 188 06:18:40,440 --> 06:18:41,780 It might be better in some ways. 189 06:18:44,420 --> 06:18:47,520 But it's easy to blindly accept the first thing that we're given. 190 06:18:48,320 --> 06:18:49,820 What does better even mean to us? 191 06:18:49,960 --> 06:18:52,520 If we're trying to make a better agent-manager interface, 192 06:18:52,760 --> 06:18:54,400 maybe that looks a little bit different. 193 06:18:55,080 --> 06:18:57,540 Maybe we need to ship fast, 194 06:18:57,560 --> 06:19:00,040 but understand at each step of the iteration process 195 06:19:00,040 --> 06:19:03,140 what better actually means to us and what the next direction is. 196 06:19:04,920 --> 06:19:05,700 Let's start here. 197 06:19:08,340 --> 06:19:11,080 Better might mean simplifying the interface, 198 06:19:11,500 --> 06:19:14,400 only showing the user exactly what the latest thing the agent 199 06:19:14,400 --> 06:19:15,660 is doing semantically. 200 06:19:15,660 --> 06:19:18,000 So they're not shown a wall of text and tool calls. 201 06:19:18,020 --> 06:19:18,960 That might not actually matter. 202 06:19:19,760 --> 06:19:22,160 Maybe it looks like simplifying the whole interface 203 06:19:22,160 --> 06:19:24,580 to give the user less cognitive load 204 06:19:25,180 --> 06:19:27,900 so that they can multitask easier 205 06:19:28,360 --> 06:19:30,440 and be able to achieve more in a different way. 206 06:19:35,800 --> 06:19:40,160 So this kind of brings us closer to what antigravity looks like today. 207 06:19:40,920 --> 06:19:45,320 All these steps along the process have brought us closer to what good means. 208 06:19:46,680 --> 06:19:49,860 And we're making a lot of improvements fast, so I'm excited for you all to see it. 209 06:19:51,560 --> 06:19:53,520 Now's the time to try many things. 210 06:19:54,080 --> 06:19:54,920 Be ambitious. 211 06:19:55,400 --> 06:19:59,200 It's more exciting than ever, and it'll actually get more exciting as the models get better. 212 06:20:03,400 --> 06:20:08,700 Try many directions, but take a moment to think. Use your human judgment. Just because 213 06:20:08,700 --> 06:20:12,740 you're finished creating something first, which can be done in an instant, it doesn't 214 06:20:12,740 --> 06:20:17,720 actually mean you're done. Think about what you're actually trying to achieve and how 215 06:20:17,720 --> 06:20:24,360 the thing you're making feels. Iterate quickly, identify many directions you can go in and 216 06:20:24,360 --> 06:20:28,060 try them, but arrive at the right thing and go deep into that. 217 06:20:30,860 --> 06:20:34,800 So make time to think, make room to think, and use your human judgment. 218 06:20:36,460 --> 06:20:42,040 To summarize, some of the lessons we've learned involve scaling up with intelligence, which 219 06:20:42,040 --> 06:20:43,140 doesn't always mean adding more. 220 06:20:43,540 --> 06:20:45,140 It could mean rebuilding your foundations. 221 06:20:46,040 --> 06:20:49,840 Take these foundations and enable your users to be ambitious in new ways you might not 222 06:20:49,840 --> 06:20:50,660 have expected before. 223 06:20:51,600 --> 06:20:56,180 And at the end of all this, when we're shipping faster and faster, make room to think and protect 224 06:20:56,180 --> 06:20:58,100 the human side of software engineering. 225 06:21:00,520 --> 06:21:03,620 At Google, this has led to some awesome things by creating a platform for ambition. 226 06:21:04,540 --> 06:21:09,000 Improvements to YouTube, better video decoders, and memory optimizations. 227 06:21:11,120 --> 06:21:14,920 As for DeepMind and anti-gravity, it's a more exciting time than ever, with new models coming 228 06:21:14,920 --> 06:21:15,160 out. 229 06:21:16,540 --> 06:21:19,200 And on the product side, we're also moving quicker than ever. 230 06:21:19,200 --> 06:21:24,240 These are just some of the future launches we've announced in the last two weeks. 231 06:21:28,040 --> 06:21:31,680 And we're continuing to use the principles we've covered today to work on the future 232 06:21:31,680 --> 06:21:32,880 of interfaces for higher intelligence. 233 06:21:33,180 --> 06:21:35,940 I'm excited for you all to see it, and I'll leave you with this. 234 06:21:37,420 --> 06:21:40,080 This new age of software engineering is actually worth celebrating. 235 06:21:41,060 --> 06:21:45,540 Agents might be doing more and more, but it's also more important than ever to solve real 236 06:21:45,540 --> 06:21:46,960 problems and use good product thinking. 237 06:21:48,620 --> 06:21:51,120 Protect your judgment and build things that people want. 238 06:21:52,500 --> 06:21:53,120 Thank you.