Kiran Sudhakara

Staff Software Engineer @ Tesla AI

Previous:
Self-Driving, UberATG
Graphics/Rendering, Firaxis Games - Civilization VI, V

Views are my own...

[Kiran Sudhakara @nariklaer]

this project generates rocket league in realtime based solely on user inputs - no renderer, no physics engine:
https://t.co/Uw7FRbmm8r

[Elon Musk @elonmusk]

Macrohard or Digital Optimus is a joint xAI-Tesla project, coming as part of Tesla’s investment agreement with xAI.

Grok is the master conductor/navigator with deep understanding of the world to direct digital Optimus, which is processing and actioning the past 5 secs of real-time computer screen video and keyboard/mouse actions. Grok is like a much more advanced and sophisticated version of turn-by-turn navigation software.

You can think of it as Digital Optimus AI being System 1 (instinctive part of the mind) and Grok being System 2. (thinking part of the mind).

This will run very competitively on the super low cost Tesla AI4 ($650) paired with relatively frugal use of the much more expensive xAI Nvidia hardware. And it will be the only real-time smart AI system. This is a big deal.

In principle, it is capable of emulating the function of entire companies. That is why the program is called MACROHARD, a funny reference to Microsoft.

No other company can yet do this.

[Tesla @Tesla]

World's first autonomous delivery of a car!

This Tesla drove itself from Gigafactory Texas to its new owner's home ~30min away — crossing parking lots, highways & the city to reach its new owner

[Kiran Sudhakara @nariklaer]

My favorite topic is back - Sparse Voxel Octrees! https://t.co/FTadbVceU8

[Kiran Sudhakara @nariklaer]

And while geometrically it's sensible to measure a signal at the slopes, there's no algorithmic way to take these samples and also be sure to sample areas of high signal (MIS maybe?). Anyways the magic of a good pdf is that it will correctly weight samples that aren't ideal.

2/2

[Kiran Sudhakara @nariklaer]

One thing that's always bothered me about Importance Sampling - why do we bias towards *higher* signal? Shouldn't we be taking more samples where the signal is *changing*?

[Kiran Sudhakara @nariklaer]

After some thought (and to answer my own question here 😉) - "Zero" represents the baseline Monte Carlo estimation. We are actually looking for signal as to how far away from zero our estimate should be.

1/2

[Kiran Sudhakara @nariklaer]

So many friends in this article. Ara has just released, Civ VII coming soon. Many other exciting projects on the horizon. Seriously lucky to have worked with these folks, and maybe again one day? 😁

https://t.co/UfB0SpC66i

[Kiran Sudhakara @nariklaer]

@procgenerated some of the details of "The Matrix" seem fishy - they translated the keyboard input to text, like 'w' becomes "move forward". seems like if you were to do this properly, you wouldn't do that

[Kiran Sudhakara @nariklaer]

If real, very impressive results. But also showcases complete disregard for IP rights in the AI world - I'm guessing the "agents" from Warner Brothers might be knocking on their door very soon...
https://t.co/rurJSVwlRN

[Kiran Sudhakara @nariklaer]

FWIW I think this is a play on "Transformers" in the pivotal '17 Attention paper from Google. But they trained the model by reverse engineering Forza and Cyberpunk 2077. This is going to become an increasingly sensitive topic in the near future 😬

[Kiran Sudhakara @nariklaer]

I'm rebranding "image noise". The term is now, officially, "Stochastic Precision"

[Kiran Sudhakara @nariklaer]

Is bikeshedding valuable? I mean, not the actual bikeshedding, but would the person inclined to give naive feedback have very useful feedback when they do have context on the problem? Or is it the opposite, like when they understand the problem more, they still sorta "bikeshed"?

[Kiran Sudhakara @nariklaer]

@ChrisGr93091552 Used to use "while(n--){}" all the time. A simplification of "while(n-->0)" which is a pretty neat idiom! Then my teammates would complain such and such architecture only has 2 backward prefetchers but 4 forward ones or blah blah... I tuned the haters out 😛

[Kiran Sudhakara @nariklaer]

I rely on ChapGPT just like that friend who's extraordinarily intelligent, exceptionally well-learned, supremely confident, and yet occasionally... completely, thoroughly wrong

[Kiran Sudhakara @nariklaer]

@JayAlammar put out this great primer on txt2img stable diffusion - https://t.co/tVPrB0oOmh

[Kiran Sudhakara @nariklaer]

@PapaLaz24281810 @mathladyhazel you'd have to lift the table twice - that should be a clue

[Kiran Sudhakara @nariklaer]

The hypotheticals around the AI art debate are quickly becoming concrete. Creating art in 5 years will be *nothing* like it is today. Procedural Generation will be king

[Kiran Sudhakara @nariklaer]

@XorDev uh, you took the avg of the rgb values. not sure what you expected but it ain't the "value" of the color

[Kiran Sudhakara @nariklaer]

@JoshuaBarczak @JJcoolkl bresenham does not light up all pixels but does leave a continuous line