00:00It's really a treat to be here in India, and it's incredible to see the country's leadership in advanced AI.
00:06I was last here a little over a year ago, and it's striking how much progress has happened since then.
00:12We've gone from AI systems that struggled with high school-level math to systems that can do research-level mathematics
00:18now and derive novel results in theoretical physics.
00:22It's also striking how much progress India has made in its mission to put AI to work for more people
00:27in more parts of the country.
00:28And India's leadership in sovereign AI, building on infrastructure, SLMs, and much more, has been great to watch.
00:35More than 100 million people in India use ChatGPT every week.
00:40More than a third of them are students.
00:42India is also the fastest-growing market now for Codex, our coding agent that works to help people develop software
00:48faster and better.
00:51India, the world's largest democracy, is well-positioned to lead an AI, not just to build it, but to shape
00:57it,
00:58and decide what our future is going to look like.
01:00And it's important to move quickly.
01:03On our current trajectory, we believe we may be only a couple of years away from early versions of true
01:09superintelligence.
01:12If we are right, by the end of 2028, more of the world's intellectual capacity could reside inside of data
01:19centers than outside of them.
01:21This is an extraordinary statement to make, and of course we could be wrong, but I think it really bears
01:28serious consideration.
01:31A superintelligence, at some point on its development curve, would be capable of doing a better job being the CEO
01:36of a major company than any executive, certainly me, or doing better research than our best scientists.
01:43As we prepare for this possibility, we are guided by three core beliefs.
01:49Number one, we believe that democratization of AI is the only fair and safe path forward.
01:57Democratization of AI is the best way to ensure that humanity flourishes.
02:02On the other hand, centralization of this technology in one company or country could lead to ruin.
02:08The desirable future a couple of decades from now has got to look like a world of liberty, democracy, widespread
02:15flourishing, and an increase in human agency.
02:19Some people want effective totalitarianism in exchange for a cure for cancer.
02:23I don't think we should accept that tradeoff, nor do I think we need to.
02:28AI should extend individual human will.
02:31We'll probably need superintelligence to help us figure out the new governance mechanisms to ensure that this happens fairly at
02:38scale, and to avoid problems like extremely unbalanced compute, access, or something else.
02:45Second, we believe that AI resilience is a core safety strategy.
02:50We don't mean that this is the only safety strategy.
02:53We will continue to need to build safe systems and solve difficult technical alignment challenges.
02:58But increasingly, we need to start broadening how we think about safety to include societal resilience.
03:04No AI lab, no AI system, can deliver a good future on their own.
03:09For an obvious example, there will be extremely capable biomodels available open source that could help people create new pathogens.
03:17We need a society-wide approach about how we're going to defend against this.
03:22And third, the future of AI is not going to unfold exactly like anyone predicts, and we believe that many
03:29people need to have a stake in shaping the outcome.
03:33The development of AI has already held many surprises, and I assume there are bigger ones to come.
03:39We understand that with technology this powerful, people want answers.
03:43But it's important to be humble about what we don't know, and always remember that sometimes our best guesses are
03:48wrong.
03:50Most of the important discoveries happen when technology and society meet, sometimes have some friction, and co-evolve.
03:58For example, we don't yet know how to think about some superintelligence being aligned with dictators in totalitarian countries.
04:05We don't know how to think about countries using AI to fight new kinds of war with each other.
04:10We don't know how to think about when and whether countries are going to have to think about new forms
04:14of social contracts.
04:16But we think it's important to have more understanding in society-wide debate before we're all surprised.
04:22Of special note, and related to all three points, we continue to believe that iterative deployment is a key strategic
04:29insight, and that society needs to contend with and use each successive new level of AI capability, have time to
04:37integrate it, understand it, and decide how to move forward.
04:40This has been working surprisingly well so far.
04:44If we are right, and systems continue to improve at this pace, it's going to change the economics of a
04:50lot of things.
04:52A really great thing about AI progress is that it looks like many things are going to get much cheaper,
04:56and have much faster economic growth.
04:58We're already seeing what AI is doing for access to high-quality healthcare, education, and more.
05:04In the coming years, we expect to see robots make many products and physical goods cheaper as supply chains get
05:11automated.
05:12The limit to how far this cost reduction can go may only be government policy.
05:17But the other side of this coin is that current jobs are going to get disrupted, as AI can do
05:22more and more of the things that drive our economy today.
05:26It'll be very hard to outwork a GPU in many ways.
05:29It'll be easy in some other ways.
05:31For example, we really seem hardwired to care about other people much more than we care about machines.
05:38We're somewhat less concerned about the long-term future.
05:41Technology always disrupts jobs.
05:43We always find new and better things to do.
05:45The people of 500 years ago would have thought that our current jobs often look silly, like ways to entertain
05:51ourselves, create stress.
05:53And the people 500 years from now hopefully will look at us, hopefully look to us, like impossibly rich people
05:59playing games, trying to find ways to pass their times.
06:02But we should all hope they feel much more fulfilled than we do today.
06:07I'm confident we will keep being driven to be useful to each other, to express our creativity, to gain status,
06:13to compete, and much more.
06:15But the specifics of what we do day to day will probably look very different.
06:20Each generation has built on the work of the generations before, and with new tools, the scaffolding gets a little
06:26taller.
06:27This collective external lattice, the set of tools that we have built up around ourselves, is remarkable, and we are
06:33capable of doing things that our great-great-grandparents couldn't have dreamed possible.
06:39It is a moral imperative to make sure that our great-great-grandchildren can stay the same, and technology, and
06:45especially AI, is how we're going to get there.
06:49For a democratic AI future, it is not enough to just give people tools and wealth.
06:54We also need to give them agency and power.
06:58The vision that AI companies lay out fundamentally reduced to either unilateral control or decentralized power.
07:05Sharing control means accepting that some things are going to go wrong in exchange for not having one thing go
07:10mega-wrong, cemented totalitarian control.
07:14This is a fundamental trade-off of democracy, and it is one that we believe in very strongly as the
07:20way to give everyone collective agency over the future.
07:23Of course, this is not to suggest that we won't need any regulation or safeguards.
07:27We obviously do urgently, like we have for other powerful technologies.
07:31In particular, we expect the world may need something like the IAEA for international coordination of AI, and especially for
07:39it to have the ability to rapidly respond to changing circumstances.
07:43The next few years will test global society as this technology continues to improve at a rapid pace.
07:49We can choose to either empower people or concentrate power.
07:53Thank you very much.
Comments