Justin Bayer

Student, scholar, teacher & builder of learning autonomous systems.

[Justin Bayer @usuallyuseless]

World models of scenes (incl. dynamics) allow both SLAM and prediction for model-based control. Let me show you our recent work (ICLR 2021) where we demonstrate that with realistic drone data. Here is a video with ground truth, filtered locations and long-step predictions.

[Justin Bayer @usuallyuseless]

@enjojoyy reading has a pretty sweet spot when it comes to channel capacity of certain things. it also makes the reader train their imagination muscles. movies just entertain as they flood most of the senses.

[Justin Bayer @usuallyuseless]

Darwin worked exactly 4h each day.

According too this argument 👇 he did nothing meaningful.

🤭

[Mayukh @mayukh_panja]

I don’t agree. A PhD student should not prioritize work-life balance.

Getting to do a PhD is a privilege. You are paid to think. There is no pressure for you to be economically useful. It is a unique opportunity to push the boundaries of human knowledge and produce something ground breaking.

And nothing great ever happens without complete devotion. Look at everything that moved and shaped the world. Every single person who created anything meaningful, in science, in arts, in music, in movies, devoted their lives to their craft.

Extraordinary outcomes require extraordinary inputs and some degree of sacrifice. Sure, have work-life balance during your PhD. But be content a mediocre outcome.

[Clemens Marschner @cmarschner]

Apparently the first thing mechanical engineers do at university these days is switch to AI courses.
That‘s wrong. There will be way too many AI engineers in no time and nobody who can take an idea all the way to mass production. Such a fascinating field.

[Justin Bayer @usuallyuseless]

@JitendraMalikCV @keenanisalive No, it was already used by Schmidhuber in 1992, quarter a century before the Ha & Schmidhuber paper.

[Justin Bayer @usuallyuseless]

@RolandMemisevic World models are models of the environment. Just p(future|present, past, action). That's it. Everything else is marketing rebranding.

[Dan Rysk @DanDeFiEd]

Italian efficiency when it comes to coffee should be studied.

In Italy:
- Walk into a bar and look at the guy
- Un caffe
- 30 seconds later it’s ready
- Shoot it
- Leave €1
- Walk out

In the US:
- Join a line
- Wait
- Order coffee
- Answer 12 questions: Size? Milk? Roast? Sugar? Temperature? Colombia beans? Name? How do you spell it?
- $12.34
- Ask for a 20% tip. Click 5 times on a ipad to have a custom tip
- Tap phone
- ask where to send the invoice
- Wait again on a different line
- Someone call a name that sounds similar to mine
- get the coffee
- too hot, can't drink it
- finally at temperature

taste like shit

[Justin Bayer @usuallyuseless]

@robotsdigest Comparison to TD-MPC and Dreamer missing, desk reject.

[Justin Bayer @usuallyuseless]

@tungnd_13 @yingwww_ @ylecun @mengyer @randall_balestr @GaoyueZhou @oumaymabounou And we did sth similar in 2016. https://t.co/uay7tafEx6 And Watter et al in 2015. https://t.co/2wSgqAIBdb.

[Justin Bayer @usuallyuseless]

@RichardSSutton The people that worked on dual control in the 80s are the worst, eh.

[Justin Bayer @usuallyuseless]

@yuvaltassa What's the integration of MJWarp with Jax like?

[Justin Bayer @usuallyuseless]

Bam.

[DHH @dhh]

@windjacker Written for a different world. I removed that clause. Go to town with your agents 🤘

[Justin Bayer @usuallyuseless]

@LucaAmb World models are not about modelling *the world*, they are about modelling *a world*.

@SchmidhuberAI 1990, "Making the world differentiable...".

[Susan Swartz @beadmomsw]

13 years. RIP my darling boy.

[Justin Bayer @usuallyuseless]

@KyleMorgenstein @DominiqueCAPaul Because of p(x_t+1 | x_1:t, u_1:t).

But that's a pain with Isaac anyway, and the major reason not to use it. 🤷

[Justin Bayer @usuallyuseless]

@giffmana https://t.co/DmNPMUPRQW

[Justin Bayer @usuallyuseless]

Bayesplaining: take a well established method, express it as a series of crude approximations to a Bayesian approach, throw it back at the community where it was invented.

[Chenhao Li @breadli428]

Extracting physics and dynamics (that are good enough for control) from state data alone is already super challenging, let alone from images.

People often think reconstructing images is hard but ignores the difficulty in retaining the low-level dynamics details in VLMs, which are more essential for control.

[Justin Bayer @usuallyuseless]

@liangpan_t Ask yourself why you want multimodal behaviour. Certainly not to achieve maximal return.

[Justin Bayer @usuallyuseless]

So what if the first hit on google for the first question is a stackexchange question you asked 15 years ago?

[Will McGugan @willmcgugan]

Alrighty. The Toad is out of the bag. 👜🐸

Install toad to work with a variety of #AI coding agents with one beautiful terminal interface.

Check out the blog post for more information...

https://t.co/KpQu5cYZzR

I've been told I'm very authentic on camera. You just can't fake that kind of awkwardness.

[Justin Bayer @usuallyuseless]

@micahgoldblum Are you sure this is not just because you trained on pairs instead of sequences?