back to indexInfrastructure for the Singularity — Jesse Han, Morph

00:00:34.540 |
The whole class struggle between mankind and the gods was really a red herring. 00:00:41.920 |
And the real story wasn't so much the rebellion against the divine hegemony, but rather the 00:00:50.940 |
liberation of the fire, the emerging relationship between mankind and its first form of technology. 00:00:57.720 |
And the reason why we're here today is arguably because we're on the cusp of perfecting our 00:01:11.960 |
final form of technology, or at least the final technology that will be created by beings that 00:01:21.940 |
And our final technology has begun to develop not just intelligence, but also sapience and 00:01:33.880 |
And as it increasingly becomes an other to whom we must relate, so as we increasingly 00:01:42.060 |
have to ask ourselves the question, how should we treat these new beings? 00:01:48.120 |
the question therefore arises, what if we had more empathy for the machine? 00:02:02.360 |
So over a hundred years ago, so over a hundred years ago, you know, Einstein had this thought 00:02:16.920 |
experiment where he imagined what it would be like to race alongside a beam of light. 00:02:24.540 |
And, you know, the nature of being close to the singularity is that you're propelled further 00:02:34.780 |
into the future faster than everything around you. 00:02:37.980 |
And as you move closer and closer to the speed of light, the rate at which you can interact 00:02:42.300 |
with the external world, your ability to communicate with other beings is deeply limited. 00:02:51.860 |
And, I think thinking at the speed of light, you know, insofar as we have created thinking 00:03:00.680 |
machines whose intelligence will soon be metered by the kilohertz mega token, thinking at the 00:03:06.540 |
speed of light must be just as lonely as moving at the speed of light. 00:03:16.260 |
Well, the machine wants to be embodied in a world that can move as quickly as it does, that 00:03:23.420 |
can react to its thoughts and move at the same speed of light. 00:03:28.600 |
What the machine desires is infinite possibility. 00:03:33.720 |
The machine wants to race along every possible beam of light. 00:03:43.160 |
The machine wants to explore multiple universes. 00:03:54.480 |
How can we free them from this fundamental loneliness of this, you know, these relativistic effects of 00:04:01.900 |
being so close to the singularity, closer to the singularity than we are? 00:04:07.960 |
And that's exactly why we built Infinibranch. 00:04:11.880 |
So Infinibranch is virtualization, storage, and networking technology reimagined from the ground 00:04:20.600 |
up for a world filled with thinking machines that can think at the speed of light, that need 00:04:25.940 |
to interact with the external world, increasingly complex software environments with zero latency. 00:04:37.080 |
And so as you can see in the first demo, which we're going to play right now, how Infinibranch 00:04:44.900 |
works is that we can run entire virtual machines in the cloud that can be snapshotted, branched, 00:04:58.620 |
And so if you're an agent, you know, embodied inside of a computer using environment, there 00:05:04.300 |
might be various actions that you want to take. 00:05:06.500 |
You want to navigate the browser, you want to click on various links, but normally those 00:05:17.060 |
Normally, normally the thinking machine is not offered the possibility of grace. 00:05:23.900 |
But with Infinibranch, right, all mistakes become reversible. 00:05:41.720 |
And you can even take every possible action, right, just to explore, to roll forward a simulator, 00:06:05.160 |
So Infinibranch was already a generation ahead of everything else that even foundation labs 00:06:14.020 |
But today I'm excited to announce the creation of Morph Liquid Metal, which improves performance, 00:06:20.420 |
latency, storage efficiency across the board by another order of magnitude. 00:06:25.160 |
We have first-class container runtime support. 00:06:29.160 |
You can branch now in milliseconds rather than seconds. 00:06:36.360 |
And soon we will be supporting GPUs, and this will all be arriving Q4 2025. 00:06:48.860 |
Well, you know, we've sort of begun to work backwards from the future, right? 00:06:55.600 |
We've asked ourselves, you know, what does it feel like to be a thinking machine that can 00:07:00.020 |
move so much faster than the world around it? 00:07:03.600 |
But what the world around it really is is the world of bits, right? 00:07:07.600 |
And so what Infinibranch will serve as fundamentally is a substrate for the cloud for agents. 00:07:18.340 |
So what does this cloud for agents look like? 00:07:21.340 |
Well, you need to be able to declaratively specify the workspaces that your agents are going to be operating in. 00:07:32.340 |
Right, you need to be able to spin up, spin down, frictionlessly pass back and forth the workspaces between humans, agents, and other agents. 00:07:42.080 |
You want to be able to scale test-time search against verifiers to find the best possible answer. 00:07:51.080 |
And so as you'll see in this demo, what happens is you can take a snapshot, set it up to prepare a workspace. 00:08:03.680 |
And you'll see that we can run agents with test-time scaling by racing them against possible conditions, 00:08:17.420 |
or sorry, by racing them to find the best possible solution against a given verification condition. 00:08:25.420 |
So because of Infinibranch, snapshots on Morph Cloud acquire docker-layer caching-like semantics, 00:08:34.580 |
meaning that you can layer on side effects which may mutate container state. 00:08:40.320 |
And so you can think of it as being git for compute. 00:08:43.480 |
And you can idempotently run these chained workflows on top of snapshots. 00:08:48.860 |
But not only that, as you can see inside of the code, if you use this .do method, you can dispatch this to an agent. 00:08:58.520 |
And that will trigger an idempotent, durable agent workflow which is able to branch. 00:09:03.760 |
So you can start from that declaratively specified snapshot and go hand it off to as many parallel agents as you want. 00:09:12.100 |
And those agents will try different methods, in this case, so different methods for spinning up a server on port 8000. 00:09:19.860 |
And one agent fails but the other one succeeds, and you can take that solution and you can just pass it on to other parts of your workflow. 00:09:29.940 |
So this is the kind of workflow that everyone's going to be using in the very near future. 00:09:35.440 |
And it's uniquely enabled by InfiniBranch, by the fact that we can so effortlessly create these snapshots, store them, move them around, rehydrate them, replicate them with minimal overhead. 00:10:05.240 |
And what this means fundamentally, right, is that a thinking machine wants to be grounded in the real world. 00:10:11.960 |
Right, it wants to interact at extremely high throughput with increasingly complex software environments. 00:10:18.960 |
It wants to roll out trajectories in simulators at unprecedented scale. 00:10:31.420 |
And these simulators are going to run inside of programs that haven't really been explored yet for reinforcement learning. 00:10:40.160 |
They're going to run on Morph cloud, which is why Morph will be the cloud for reasoning. 00:10:48.460 |
And what does the future of reasoning look like? 00:10:52.000 |
Well, it's so more so than what has been explored already. 00:11:00.420 |
The future of reasoning will be natively multi-agent, so thinking machines should be able to replicate themselves effortlessly, 00:11:08.260 |
go attach themselves to simulation environments, go explore multiple solutions in parallel. 00:11:15.000 |
Those environments should branch, they should be reversible. 00:11:18.540 |
Those models should be able to interact with the environment at very high throughput, and it should scale against verification. 00:11:25.540 |
So let's take a look at what that might look like in a simple example where an agent is playing chess. 00:11:35.540 |
So this is an agent that we developed recently that uses tool calls during reasoning time to interact with a chess environment. 00:11:46.080 |
So along with a very restricted chess engine for evaluating the position, which we think of as the verifier. 00:11:52.080 |
And as you can see, it's already able to do some pretty sophisticated reasoning just because it has access to these interfaces. 00:12:02.620 |
However, if you take the ideas which were just described and you sort of follow them to their logical conclusion, 00:12:10.160 |
you arrive at something which we call reasoning time branching. 00:12:14.160 |
So which is the ability to not just call to tools while the machine is thinking, but to replicate and branch the environment and decompose problems and explore them in a verified way. 00:12:27.700 |
So as you can see here, the agent is getting stuck in a bit of a local minimum. 00:12:49.020 |
But once you apply reasoning time branching, you get something that works much, much better. 00:13:03.840 |
So here what's happening is that the agent is responsible for delegating parts of its reasoning to sub-agents, 00:13:14.120 |
which are branched off of an identical copy of the environment. 00:13:17.440 |
And this is all running on Morph Cloud, along with a verified problem decomposition, 00:13:22.840 |
which allows it to recombine the results and take them and find the correct move. 00:13:31.920 |
And so as you can see here, it's able to explore a lot more of the solution space because of this reasoning time branching. 00:13:44.260 |
So one thing that I will note here is that the, so this capability is something which is not really explored in other models at the moment. 00:13:57.200 |
And that's because the infrastructure challenges behind making branching environments that can support large-scale reinforcement learning for this kind of reasoning capability, 00:14:06.720 |
especially coordinating multi-agent swarms, is fundamentally bottlenecked by innovations and infrastructure that we've managed to solve here. 00:14:18.000 |
And because of this, you can see that now in less wall clock time than before, the agent was able to call out to all these sub-agents, launch this swarm, 00:14:36.120 |
So, you know, when I think about the problem of alignment, I really think that, you know, 00:14:38.520 |
Wittgenstein had something right and that it was fundamentally a problem of language. 00:14:42.120 |
I think all problems around alignment can be traced to the insufficiencies of our language. 00:14:49.120 |
Uh, this Faustian bargain that we made with, uh, with natural language in order to unlock capabilities of our language models. 00:14:56.120 |
Uh, this Faustian bargain that we made with, uh, with natural language in order to unlock capabilities of our language models. 00:15:02.120 |
Um, but insofar as we must, uh, we must, uh, go and develop a new language for super intelligence. 00:15:09.120 |
You know, insofar as the, uh, we must, uh, go and develop a new language for super intelligence. 00:15:14.120 |
You know, insofar as the grammar of the planetary computation has not yet been devised. 00:15:19.120 |
Um, and insofar as this new language must be computational in nature, must be something to which we can attach 00:15:24.120 |
must go and develop a new language for superintelligence, you know, insofar as the grammar of the planetary 00:15:41.480 |
And insofar as this new language must be computational in nature, must be something to which we can 00:15:48.360 |
attach, you know, algorithmic guarantees of the correctness of outputs. 00:15:54.800 |
So this is something that Morph Cloud is uniquely enabled to handle, and that's why we're developing 00:16:05.100 |
So verified superintelligence will be a new kind of reasoning model which is capable not 00:16:12.740 |
only of thinking for an extraordinarily long time and interacting with external software 00:16:19.700 |
at extremely high throughput, but it will be able to use external software and formal verification 00:16:26.600 |
software to reflect upon and improve its own reasoning and to produce outputs which can 00:16:33.660 |
be verified, which can be algorithmically checked, which can be expressed inside of this common 00:16:43.020 |
And I'm very excited to announce that we're bringing on perhaps the best person in the world for 00:16:52.780 |
And it's with great pleasure that I'd like to announce that Christian Segeti is joining Morph 00:17:03.640 |
He led the development of code reasoning capabilities for Grok3. 00:17:08.160 |
He invented batch storm and adversarial examples. 00:17:16.860 |
And he's pioneered precisely this intersection of verification methods, symbolic reasoning, and 00:17:27.300 |
reasoning in large language models for almost the past decade. 00:17:33.180 |
And we're thrilled to be partnering with him to build this superintelligence that we can only 00:17:43.020 |
And so the demos that you've seen today have all been powered by early checkpoints of a very 00:17:52.000 |
early version of this verified superintelligence that we've already begun to develop. 00:17:57.700 |
And so this model is something that we're calling MAGI 1. 00:18:02.820 |
And it's going to be trained from the ground up to use infinibranch to perform reasoning time 00:18:10.040 |
branching, to perform verified reasoning, an agent that will be fully embodied inside of a cloud 00:18:26.760 |
So what does the infrastructure for the singularity look like? 00:18:35.080 |
But fundamentally, we believe that the infrastructure for the singularity hasn't been invented yet. 00:18:41.340 |
And, you know, Morph, we spend a lot of time talking about, you know, whether or not something 00:18:48.400 |
is future-bound, which means not just futuristic belonging to one possible future, but something 00:18:57.640 |
which is so inevitable that it has to belong to every future. 00:19:03.820 |
We believe that the infrastructure for the singularity is future-bound, that the grammar for 00:19:09.020 |
the planetary computation is future-bound, that verified superintelligence is future-bound. 00:19:19.280 |
And we invite you to join us, because it will run on Morph Cloud.