Lightnews — Scholar-powered news

Preetum Nakkiran @preetumnakkiran.bsky.social · Feb 25

Another intern opening on our team, for a project I’ll be involved in (deadline soon!)

Josh Susskind @kindsuss.bsky.social · Feb 25

My colleague Shuangfei Zhai is looking for a summer research intern to work on improving TarFlow at Apple. If interested, send your CV to szhai at apple.com by this week.

3 8

Reposted by Preetum Nakkiran

Chenyang Yuan @chenyang.co · Feb 20

Last month I co-taught a class on diffusion models at MIT during the IAP term: www.practical-diffusion.org

In the lectures, we first introduced diffusion models from a practitioner's perspective, showing how to build a simple but powerful implementation from the ground up (L1).

(1/4)

1 1 5

Preetum Nakkiran @preetumnakkiran.bsky.social · Feb 11

Our main results study when projective composition is achieved by linearly combining scores.
We prove it suffices for particular independence properties to hold in pixel-space. Importantly, some results extend to independence in feature-space... but new complexities also arise (see the paper!) 5/5

2

Preetum Nakkiran @preetumnakkiran.bsky.social · Feb 11

We formalize this idea with a definition called Projective Composition — based on projection functions that extract the “key features” for each distribution to be composed. 4/

1 6

Preetum Nakkiran @preetumnakkiran.bsky.social · Feb 11

What does it mean for composition to "work" in these diverse settings? We need to specify which aspects of each distribution we care about— i.e. the “key features” that characterize a hat, dog, horse, or object-at-a-location. The "correct" composition should have all the features at once. 3/

1 2

Preetum Nakkiran @preetumnakkiran.bsky.social · Feb 11

Part of challenge is, we may want compositions to be OOD w.r.t. the distributions being composed. For example in this CLEVR experiment, we trained diffusion models on images of a *single* object conditioned on location, and composed them to generate images of *multiple* objects. 2/

1 2

Preetum Nakkiran @preetumnakkiran.bsky.social · Feb 11

Paper🧵 (cross-posted at X): When does composition of diffusion models "work"? Intuitively, the reason dog+hat works and dog+horse doesn’t has something to do with independence between the concepts being composed. The tricky part is to formalize exactly what this means. 1/

Left Image: A shaggy dog-horse hybrid standing in a rural landscape.
Right Image: A golden dog wearing a red beret against a blurred outdoor background.

2 15 39

Preetum Nakkiran @preetumnakkiran.bsky.social · Feb 10

finally managed to sneak my dog into a paper: arxiv.org/abs/2502.04549

1 4 62

Preetum Nakkiran @preetumnakkiran.bsky.social · Feb 9

Credit to: x.com/sjforman/sta...

x.com

1

Preetum Nakkiran @preetumnakkiran.bsky.social · Feb 9

nice idea actually lol: “Periodic cooking of eggs” : www.nature.com/articles/s44...

1 2 7

Reposted by Preetum Nakkiran

Max Puelma Touzel @mptouzel.bsky.social · Feb 5

Reminder of a great dictum in research, one of 3 drilled into us by my PhD supervisor: "Don't believe anything obtained only one way", for which the actionable dictum is "immediately do a 2nd independent test of something that looks interesting before in any way betting on it". Its a great activity!

Maria Antoniak @mariaa.bsky.social · Feb 5

Grateful for my topic modeling and word embeddings training, which made me suspicious of any output that "looks good" but for which I haven't seen any alternative outputs that might also "look good."

Running a prompt and getting output that looks good isn't sufficient evidence for a paper.

3 12

Preetum Nakkiran @preetumnakkiran.bsky.social · Feb 4

I’ve been in major denial about how powerful LLMs are, mainly bc I know of no good reason for it to be true. I imagine this was how deep learning felt to theorists the first time around 😬

22

Reposted by Preetum Nakkiran

Erin Grant @eringrant.me · Jan 21

Last year, we funded 250 authors and other contributors to attend #ICLR2024 in Vienna as part of this program. If you or your organization want to directly support contributors this year, please get in touch! Hope to see you in Singapore at #ICLR2025!

ICLR Conference @iclr-conf.bsky.social · Jan 21

Financial Assistance applications are now open! If you face financial barriers to attending ICLR 2025, we encourage you to apply. The program offers prepay and reimbursement options. Applications are due March 2nd with decisions announced March 9th. iclr.cc/Conferences/...

ICLR 2024 Financial Assistance

iclr.cc

1 14 37

Preetum Nakkiran @preetumnakkiran.bsky.social · Jan 18

Happy for you Peli!!

1

Reposted by Preetum Nakkiran

David Pfau @davidpfau.com · Jan 18

The thing about "AI progress is hitting a wall" is that AI progress (like most scientific research) is a maze, and the way you solve a maze is by constantly hitting walls and changing directions.

4 11 96

Preetum Nakkiran @preetumnakkiran.bsky.social · Jan 11

for example I never trust an experiment in a paper unless (a) I know the authors well or (b) I’ve reproduced the results myself

6

Preetum Nakkiran @preetumnakkiran.bsky.social · Jan 11

imo most academics are skeptical of papers? It’s well-known that many accepted papers are overclaimed or just wrong— there’s only a few papers people really pay attention to despite the volume

1 8

Reposted by Preetum Nakkiran

Pau Rodriguez @paurodriguez.bsky.social · Dec 10

Thrilled to share the latest work from our team at
@Apple
where we achieve interpretable and fine-grained control of LLMs and Diffusion models via Activation Transport 🔥

📄 arxiv.org/abs/2410.23054
🛠️ github.com/apple/ml-act

0/9 🧵

3 15 47

Reposted by Preetum Nakkiran

Brandon Amos @bdamos.bsky.social · Jan 6

📢 My team at Meta (including Yaron Lipman and Ricky Chen) is hiring a postdoctoral researcher to help us build the next generation of flow, transport, and diffusion models! Please apply here and message me:

www.metacareers.com/jobs/1459691...

Postdoctoral Researcher, Fundamental AI Research (PhD)

Meta's mission is to build the future of human connection and the technology that makes it possible.

www.metacareers.com

1 12 53

Preetum Nakkiran @preetumnakkiran.bsky.social · Jan 3

Giving a short talk at JMM soon, which might finally be the push I needed to learn Lean…

1 1

Preetum Nakkiran @preetumnakkiran.bsky.social · Jan 1

This optimal denoiser has a closed-form for finite train sets, and notably does not reproduce its train set; it can sort of "compose consistent patches." Good exercise for reader: work out the details to explain Figure 3.

10

Preetum Nakkiran @preetumnakkiran.bsky.social · Jan 1

Just read this, neat paper! I really enjoyed Figure 3 illustrating the basic idea: Suppose you train a diffusion model where the denoiser is restricted to be "local" (each pixel i only depends on its 3x3 neighborhood N(i)). The optimal local denoiser for pixel i is E[ x_0[i] | x_t[ N(i) ] ]...cont

1 2 39

Preetum Nakkiran @preetumnakkiran.bsky.social · Dec 31

Neat, I’ll take a closer look! (I think I saw an earlier talk you gave on this as well)

1

Reposted by Preetum Nakkiran

Yoav Goldberg @yoavgo.bsky.social · Dec 26

LLMs dont have motives, goals or intents, and so they wont lie or deceive in order to obtain them. but they are fantastic at replicating human culture, and there, goals, intents and deceit abound. so yes, we should also care about such "behaviors" (outputs) in deployed systems.

7 10 62

Reposted by Preetum Nakkiran

Alex Thiery @alexxthiery.bsky.social · Dec 15

One #postdoc position is still available at the National University of Singapore (NUS) to work on sampling, high-dimensional data-assimilation, and diffusion/flow models. Applications are open until the end of January. Details:

alexxthiery.github.io/jobs/2024_di...

18 41