Making everything one big CUDA graph helped wall clock consistency a lot on a laptop, but we still fought with power management. Looking forward to using an Nvidia Spark in the future. It isn’t in the repo code, but the single biggest win I have seen is an “action input” model instead of separate Q output heads. Requires 18x evaluations, but it works better across the board. Using SGD instead of Adam for the value function final layer allows it to be “scale invariant” and work across orders of magnitude different values without reward clipping or categorical representation.

thank you 🙏

@yacineMTB @hugs

Thanks for sharing the GH repo, any chance to see your code is a gift IMO John 🙏

Thank you!

Huge respect for tackling the physical. Gro3 Alerts would probably flag this convo as 'alpha for real builders' if it could.

john, every time you post my imposter syndrome gets stronger

In 1980 something I was playing the famous Adventure game on the Atari 2600 and found that infamous dot that let you cross the black lines. A few years later I built my own 386 and at the shop they gave me a free shareware of Commander Keen. I was so excited. then Shortly after duke nukum, and then the amazing Wolfenstein. Finally not long after you guys rolled out Doom. I must have played that game a thousand times. Awesome man. Thanks so much. Great memories.

It reminds me of @Videogamo's #ROBOTASystem for remotely controlling arcade machines!

I was watching a video of you discussing this a month or so ago, but i only got halfway done and it got taken down for some reason. The physical/simulated discussion was very interesting!

that talk was amazing! especially the section about the infinite video wall. i find myself going back to it at least once a month and i always find myself learning something new :) hope yall put out more new stuff soon!

can you explain the gap to a lay person? lots of noise in the signal? wouldn't it be better to have two seperate models, one converts from 'real world' to the 'platonic representation,' then one does the actual playing?

This is really cool!! Watched your presentation on this, very interesting

damn. im just happy im smart enough to understand what you said. even if im not smart enough to put these works to use in any way shape or form. X_x.

Did you run any experiments with Continual Backprop?

@ID_AA_Carmack Love the "Physical Atari" challenge! How about a robot that reads help screens & controller functions, learning gameplay via RL to mimic human skill curves in Pong or Asteroids? Could bridge sim-to-real gaps & boost AGI—open-source potential is huge.

*.py ;) what is going on here

One day i will be knowledgeable enough to understand, use or do something interesting with this repo . Have a feeling this repo is a goldmine

thank you for sharing it!

Impressive work on the “Physical Atari” project. Understanding the gap between simulation and reality is key.

Thanks a lot 🙏🏻 been looking forward to this since I saw the presentation video in June

you the best : John, Many only see your recent achievements: Oculus, AI, rockets… But I’m thinking about what you did before all that. Thanks to you, I experienced the thrill of a shot in the dark, heart pounding, in Doom. You invented the future of video games with pure code, boldness, and raw passion. And without you, millions of us would never have known that freedom, that interactive dream. You opened the door to modern gaming.

Impressive work! Understanding the gap is key. Excited to explore the repository.

😀

Wow cool. I remember you talking about this. I'll check it out. Cheers, Ken.

Got told to send this,don't know why,someone just had me look at 3d model for main character in Anachronox & it's triangle count (717 triangles) since on DeviantArt they saw an old pic of one of my old low poly models from 1997 that's 681 triangles(my model had fingers & thumbs)

@willccbb 😮

If anyone needs parts made; free cad software. Still needs ported over to Apple. emachineshop.com/

Cool device. Would like to rig up some AI and make it master any and all Atari games and win them all with precision and perfection. Other AI uses. Day trade the stock market for maximum gains. Optimally generate cryptocurrency remotely on a cloud system.

I think only 4 of 6 build steps are listed on the robotroller site?

I'm somewhat of a robot myself.

Scoring and rewards in context of a specific game I've seen. But I haven't seen scoring for just moving a physical "hand" in sensible ways, and then in context of manipulating a controller generally. Wonder how that'd impact benchmarks playing multiple games.

@dwangoac @MrTASBot You've gotta see this