Google DeepMind

“All of these variables fed into a system, and they were able to take 20 to 25% of the electricity out of the equation, serving the exact same workload.”

“So I think that that's actually a really important capability that we didn't have with Genie too as well, right? Because we relied on image prompting. And so, there was some transfer issue, like, where you rely on imagine to generate the image, and that often …”

“So we had this Genie II project that was much more sort of like, three environments that it could generate, and it wasn't super high quality. It felt like it coming from Genie I, but it wasn't the same quality as things like VO II, which you know, the Saints o…”

“And I think that that really is the property of scale and breadth of training. So this is very much like an emergent thing. I don't think there's anything Like really specific we do for this, right?”

“We weren't sure, like, how, how big it's gonna be but we definitely felt, I felt definitely that we have something that was kind of like for a long time coming, basically being able to generate environments in real time. I think a lot of work that was done in,…”

“Gini Free is pretty much a research preview, right? It's not something we are releasing at this point.”

“We designed it to be an environment rather than an agent, right? So Genie three is very much like an environment model. Like we don't see it as like an agent itself that can like think and act in the world. It's more just a general purpose sort of simulator in…”

“And then for Genie three, we basically went much more ambitious on the same sort of approach, right? And we made it like a headline goal for ourselves. It's like, can we make the memory be what it is, right? We said we want minute plus memory and real time and…”

“We obviously made a choice that we want VO three and Gini three to be separate projects this year, right?”

“Really what we think with Genie three is it's the best of both, right? Because you're taking a real world data driven approach, right? But then you've got the ability to learn in simulation. So It kind of combines the good parts of each of those. And so that's…”

“There was not like, there is no like fundamental limitation, but Currently the current design we limited to one minute of this type of memories.”

“One thing that we didn't want to do, and we didn't want to build an explicit representation, right? So there are definitely methods that are able to achieve consistency, and they did that through an explicit, some three, you know, their nerves, and other metho…”

“And it, it's at the point where, like, a human who is not an expert, like, will watch it and think it looks real, right?”

“We just worked with Russ Lovegrove on fine tuning a model on his sketches so that he can then create something new out of that, and then we design an actual physical chair that we, like, have a prototype of.”

“So when we're developing this model, we actually started out doing character consistency evals and faces we didn't know, and it doesn't tell you anything. And then we started testing it on ourselves and quickly realized like, okay, this is what you need to do …”

“We haven't exactly figured out how to enable both of those yet.”

“Every model can cherry pick images that look perfect. So like now I think the real question is like how expressible is this model and what's the worst image you would get given what you're trying to do? So I think by raising the quality of the worst image We r…”

“We're probably not going to go build a software for an architecture firm. My dad is an architect and he would probably love that. But I don't think that's something that we will do, but somebody should go and do that.”

“Really what an artist wants when they want to do something is they want the intent to be understood. And I think that, that these AI models are getting better at understanding the intent of users. So often when you ask text queries now, the model gets what you…”

“What we saw was that we budgeted like, you know, a comparable amount of queries per second as we had for our previous models that were on Ella Marina. And we had to keep upping that number as people were going to Ella Marina to use the model.”

“For this first release, the model's not as good as text rendering at, as we would like it to be, and that's something that we want to fix in the future”

“I've seen examples where people give it, like an image of HTML code and have the model render the webpage.”

“Yeah, I think in that case, I wouldn't spend a ton of time building a custom interface for making this picture of 46 people. It seems like the kind of thing that we can solve.”

“Once you get into real long conversations, like it starts to follow your instructions a little bit worse”