Top Nvidia Pro Dishes On The Future Of AI

One of the biggest questions in our new high-tech world is how soon we’re going to get physical AI agents in the form of advanced robotics.

You could also debate whether those robots are going to be humanoid, or somehow different.

But either way, there’s a consensus that one way or another, these new companions are coming to our world soon.

Jim Fan is a researcher at Nvidia, and has been talking about how imminent physical AI is.

“Robots will not be trained in isolation,” he posted on X on Christmas Eve. “They will be simulated as an ‘iron fleet’, deployed in real-time graphics engines, and scaled across a huge cluster to produce the next trillions of high quality training tokens. The majority of embodied agents will be born in sim, and transferred zero-shot to our real world when they are ready. They will share a ‘hive mind’ that sends latent embeddings back and forth to coordinate a multi-agent physical task.”

However, in a LinkedIn post that goes deeply into the mechanics of AI agents, he also suggests that this kind of intelligence is largely going to be disembodied first.

“Before we have a million robots in the physical world, we will first see a billion embodied agents in virtual worlds,” Fan writes. “Gaming is the second major area I'm dedicated to in 2024. AI and Gaming are born for each other, and their happy marriage is just getting started.”

The Evolution of AI

One interesting bit in this post is where Fan calls game environments a “primordial soup for generalist AI to emerge.” Citing examples of Minecraft algorithms, he noted how agents are bounded by the complexity of their surroundings.

“There are a lot more games that require extremely advanced perception, agility, exploration, reasoning, and planning,” he writes. “We are just starting to scratch the surface.”

Attention Mechanisms

Fan also makes reference to some of the advanced strategy being used by today’s LLM engines.

I’ve written extensively about the idea of the modern transformer – as a critical part of LLM design. The transformer acts as a kind of “attention mechanism,” to allow the model to focus more on what’s important to humans, and less on what’s less relevant. That in itself cuts down on resource intensity for any given task , and creates enormous efficiencies for high-token systems.

“Tokens are actively selected by the agent itself through exploration,” he writes. “It (the agent) can choose to experiment with things that maximally reduce its internal uncertainties - kind of like how human curiosity works.”

It’s interesting to think about the AI engines having this drive, and how they would gain knowledge, or, as Fan puts it, “reduce their internal uncertainties” – which, in fact, sounds kind of like a fancy phrase for learning itself…

The New Non-Player Character

“I believe 2024 is an inflection point,” Fan continues. “The Digital Westworld is coming, and will transform the industry once and for all.”

As he describes how non-player characters will act, think about how we have viewed NPCs in the past – as stilted, evidently artificial players, in contrast to fully human ones.

“Games will feel truly alive,” Fan adds. “The characters will interact with humans and each other, form relationships, take consistent actions over their lifetime, and react in human-like ways. Each game will have infinite replay value, and each player will have (a) unique and tailored experience.”

If you’ve already heard a young gamer talking about forming relationships with non-player characters in today’s games, get ready, because this sort of thing is going to expand as the NPCs really become more human.

“As video games have evolved, the technology underpinning NPCs has had to evolve with them,” wrote Ilya Gelfenbeyn at Inworld in January of last year, commenting on this advancement. “The evolution of NPC behavior has been shaped by advancements in technology opening up new opportunities for more complex trait scripts, also known in the industry as a job system. Put simply, this means NPCs can be scripted to respond in a larger number of ways according to set variables.”

What’s Ahead

At the end of the day, we’re likely to see a lot of these advances as a sort of gamified reality. They’ll come in in the form of entertainment and explorative play, but they’re likely to go beyond that, to become vital parts of our experience as humans.

As for the physical robots, we’re likely to see that play out in a utilitarian way, too. People always talk about the butler robot, and when we’re going to get that. At the same time, population is declining, and labor is in demand. We’re soon going to have all of these automatons doing the work.

Anyway, I’ll bring you more as it develops through 2025.