Jinen Setpal @48bitmachine

PhD student @PurdueECE & MLR intern @Apple, researching deep learning optimization & interpretability theory. I love open source. @jinen:https://t.co/W0XuIlEg3H jinen.setpal.net [email protected] Joined October 2017

Tweets

3K
Followers

459
Following

3K
Likes

9K

Jinen Setpal @48bitmachine

2 weeks ago

@fiiissshh @AntithesisHQ

0 0 1 40 0

View Details

Subbarao Kambhampati (కంభంపాటి సుబ్బారావు) @rao2z

a month ago

Mechanistic Interpretability be like.. We wanted to see what the bread slice is thinking when it is in the toaster. Is it taking its singing equanimously? thinking religious thoughts? or plotting revenge? (Face it--a vengeful singed slice is the last thing we want!) But alas, the bread slice doesn't talk! Its thoughts are intricately encoded in its singe pattern. So we devised a clever method--of showing the bread slice to a random Joe, asking him to describe its feelings. Joe says that the slice is having a religious experience involving Mother Mary. But is random Joe describing these feelings right? To check, we called random artist Jane and asked her to render these thoughts back to the external appearance of a (different!) bread slice. We co-trained Joe and Jane until they learn to auto-encode the real truth about the slice's thoughts. We are finding this technique to be a great way to understand what the bread slice is thinking. The tehcnique is not always (or even sometimes) correct, but it gives us a great window into publishing bread slice thought related podcast articles. Hope you find this research useful. #AIAphorisms

21 54 596 55K 167

View Details

Optimal Intellect @opt_intellect

3 months ago

We're Optimal Intellect, a research lab from the team behind CVXPY. Today we're introducing Moreau: a GPU-native solver that's orders of magnitude faster than the best existing tools.

31 55 742 1.5M 330

View Details

Jinen Setpal @48bitmachine

3 months ago

@phil_hellmuth @personaaiinc @nicholas_k_wade is cracked and that robot's awesome

0 0 2 38 0

View Details

Jinen Setpal @48bitmachine

4 months ago

@JacobZietek Brilliant play by California Department of Fish and Wildlife, they raised 157M like it was nothing 🥶

0 0 1 53 0

View Details

Harmya @racerfunction

5 months ago

great blog on how @tensarahq thinks about correctness of submitted GPU kernels: sarthakmangla.com/blog/wrong

1 3 8 638 8

View Details

David Pfau @pfau

5 months ago

This is the key difference between in-domain and out-of-domain generalization, and we still have not truly solved out-of-domain generalization. It just turns out you can build world changing technology by throwing so much data at things that the entire universe is in-domain.

Niels Rogge @NielsRogge

5 months ago

One of the best visual explanations I've ever seen for why scaling Transformers works, but is suboptimal, as it's just brute-forcing things, by @YesThisIsLion (co-author of the Transformer) on @MLStreetTalk "In the (rejected) paper "Intelligent Matrix Exponentiation", they show

27 49 619 87K 647

13 21 333 36K 196

View Details

Jinen Setpal @48bitmachine

6 months ago

@livgorton 🔥

0 0 0 100 0

View Details

Jinen Setpal @48bitmachine

7 months ago

@NeelNanda5 [2/2] From the discussion at ~6m, perhaps using multiple activations / PCA reduced how much the explanation can update *without* changing the underlying prediction, which may explain why it improved alignment?

0 0 0 13 0

View Details

Jinen Setpal @48bitmachine

7 months ago

@NeelNanda5 [1/2] This was awesome. I'm also excited about leveraging interp in training; my approach focuses on reducing degrees of freedom of model parameters & input w.r.t. explanation, because this provably prevents malicious compliance. Then training against these enforces alignment!

1 0 0 23 0

View Details

Jinen Setpal @48bitmachine

7 months ago

@bitsgopew @ninja_maths The channel below in general is an absolute goldmine, but the video specifically talks about the visualization in the tweet above, makes it very intuitive: youtube.com/watch?v=ZdlraR… I also recommend checking out 4.2.2. where prof talks about the connection with normal equations

1 0 0 40 1

View Details

Jinen Setpal @48bitmachine

7 months ago

@SuryaGanguli I shared this exact analogy at an interview I gave just yesterday (when asked about my use of AI for coding), this tweet made my day :D

0 0 2 161 0

View Details

Surya Ganguli @SuryaGanguli

7 months ago

Using AI to help you on your homework is like using a robot to help you lift weights at the gym.

109 179 1K 92K 109

View Details

Mathieu @miniapeur

8 months ago

7 168 2K 44K 207

View Details

Jinen Setpal @48bitmachine

8 months ago

I'm awestruck and inspired. Congratulations Dr. Cohen!

Jeremy Cohen @deepcohen

8 months ago

@jasondeanlee @SebastienBubeck @tomgoldsteincs @zicokolter @atalwalkar This is the third, last, and best paper from my PhD. By some metrics, an ML PhD student who writes just three conference papers is "unproductive." But I wouldn't have had it any other way 😉 !

10 21 528 160K 134

0 0 2 422 0

View Details

Jinen Setpal @48bitmachine

11 months ago

@jennyzhangzt My reviews have been constructive across the board; meaningful feedback from all 4. It's my first major conference submission too, so I was mentally prepared to have everything either torn to shreds or to be fed AI slop but I've been pleasantly surprised😃

0 1 4 599 0

View Details

Jinen Setpal @48bitmachine

11 months ago

@BorisMPower @AcerFur Wow

0 0 0 141 0

View Details

Jacob Zietek @JacobZietek

11 months ago

Currently working on RFMs for tool use and other industrial tasks @personaaiinc. We have an early customer (Hyundai) and significant capital ($28M pre-seed). Looking to expand the ML team, dm me if you're interested in joining. Also taking cracked interns :)