Re:Genie 4 is now available
by vchiriac on Wednesday November 18, 2026 @02:41PM
(#50551089)
Deepmind is creating so much that, at least to me, feels done for the love of the craft instead of earning potential, that I cannot but appreciate their work. Like, how could this end up being profitable?
[ Parent
| on HN ]
The theory is this will ultimately give robots foresight in the physical world, like we can imagine an action before taking it.
Re:Genie 4 is now available
by throwaway384034 on Wednesday November 18, 2026 @02:53PM
(#50551229)
Sounds really bad for latency. Seems better to reason in the abstract or produce a mathematical model.
Aye basically use it to make synthetic training data.
Re:Genie 4 is now available
by throwaway384034 on Wednesday November 18, 2026 @03:44PM
(#50551886)
Training in a physics engine like MuJoCo with deterministic control seems more reliable.
Once a model gets good enough, etch it directly onto silicon and get 50,000 tokens per second.
That’s.. actually insane. Just imagining a Claude operating a robot arm with a tool call that leveraged something like this to “imagine” scenarios before acting.
My god we are so cooked.
Re:Genie 4 is now available
by ReptileMan on Wednesday November 18, 2026 @02:55PM
(#50551267)
> love of the craft
The feeling of AI recently is Lovecraftian indeed.
Build realistic worlds -> Unleash robots in them to have more realistic, synthetic training environments. You know, just like the current world us humans live in. Turtles all the way down and all.
Re:Genie 4 is now available
by SpaceManNabs on Wednesday November 18, 2026 @03:27PM
(#50551683)
it is for building world models for robots to interact with without killing entire the state of the run (or destroying the robot)
Re:Genie 4 is now available
by sail0rm00n on Wednesday November 18, 2026 @03:49PM
(#50551934)
We've seen Waymo training and potentially other use cases: https://waymo.com/blog/2026/02/the-waymo-world-model-a-new-f...