Lightnews — Scholar-powered news

Reposted by Shubhendu Trivedi

Sam Power @spmontecarlo.bsky.social · 1d

This one evaded my lists; quite cool stuff!

arxiv.org/abs/2507.00272
'Iteratively Saturated Kalman Filtering'
- Alan Yang, Stephen Boyd

1 2 7

Reposted by Shubhendu Trivedi

Sam Power @spmontecarlo.bsky.social · 16h

absolutely fascinating (standard Steinerberger fare); be sure to open this!

arxiv.org/abs/2510.11571
'Robust Online Sampling from Possibly Moving Target Distributions'
- François Clément, Stefan Steinerberger

1 12

Shubhendu Trivedi @shubhendu.bsky.social · 8h

... and that this circular wheeling-dealing circus is bad, but hardly the thing to worry about re: stock markets. There is a much deeper problem that will come out in a few years. The AI stuff is a wasteful distraction, comparatively. bsky.app/profile/shub...

Shubhendu Trivedi @shubhendu.bsky.social · 7d

So far these circular announcements function as "infinite money glitch" generators. The funniest was Oracle. It went up 45% in a single day based on some random capex guide using revenue that doesn't exist from OpenAI. Then it flipped into a sureshot short and retraced most of it in two weeks.

Carl Quintanilla @carlquintanilla.bsky.social · 7d

NVIDIA and OpenAi:

Concerns that their “increasingly complex and interconnected web of business transactions is artificially propping up the trillion-dollar AI boom.“

@bloomberg.com $NVDA 👀
www.bloomberg.com/news/feature...

1

Shubhendu Trivedi @shubhendu.bsky.social · 8h

I made a post earlier, but removed it because I (uncharacteristically) got self-conscious of some typo, but it went something like: I have googled Taylor Swift maybe twice (and for a while thought she was an actor), but I am sure her stock went up 43% after the NEWS.

1 1

Shubhendu Trivedi @shubhendu.bsky.social · 8h

Post of the week.

1 4

Shubhendu Trivedi @shubhendu.bsky.social · 1d

But obviously, the question often is how to scale them.

Shubhendu Trivedi @shubhendu.bsky.social · 1d

Nice study on scaling and equivariance. arxiv.org/abs/2510.09768

Scaling Laws and Symmetry, Evidence from Neural Force Fields

We present an empirical study in the geometric task of learning interatomic potentials, which shows equivariance matters even more at larger scales; we show a clear power-law scaling behaviour with re...

arxiv.org

1 6

Reposted by Shubhendu Trivedi

Simon Willison @simonwillison.net · 1d

nanochat by Andrej Karpathy is neat - 8,000 lines of code (mostly Python, a tiny bit of Rust) that can train an LLM on $100 of rented cloud compute which can then be served with a web chat UI on a much smaller machine simonwillison.net/2025/Oct/13/...

nanochat

Really interesting new project from Andrej Karpathy, described at length in this discussion post. It provides a full ChatGPT-style LLM, including training, inference and a web Ui, that can be …

simonwillison.net

2 21 200

Shubhendu Trivedi @shubhendu.bsky.social · 1d

Yes!! I had to double check it was the same Michelin

Shubhendu Trivedi @shubhendu.bsky.social · 1d

Haha. I love them. This looks quite nice.

1 1

Shubhendu Trivedi @shubhendu.bsky.social · 1d

A new library for Kohonen self-organizing maps github.com/michelin/Tor...

GitHub - michelin/TorchSOM: TorchSOM is a PyTorch-based library for training Self-Organizing Maps (SOMs), a model trained in an unsupervised manner, that can be used for clustering, dimensionality red...

TorchSOM is a PyTorch-based library for training Self-Organizing Maps (SOMs), a model trained in an unsupervised manner, that can be used for clustering, dimensionality reduction and data visualiza...

github.com

1 1 8

Reposted by Shubhendu Trivedi

Benjamin Pope @benjaminpope.bsky.social · 1d

Thrilled to have two years' of work out, in a pair of papers led by @gradientrider.bsky.social and @maxecharles.bsky.social.

We've built a data-driven calibration of the James Webb Interferometer to near its fundamental limits for high-res imaging - explainer at @aunz.theconversation.com!

How we sharpened the James Webb telescope’s vision from a million kilometres away

The only Australian hardware on board the legendary telescope is starting to fulfil its duties.

theconversation.com

1 39 140

Shubhendu Trivedi @shubhendu.bsky.social · 1d

One of my goals for the next few weeks is to pick up actively training for said system, using a RL/formal languages problem as a playground. In the near-worst case there will be an ε improvement, if nothing else. In the worst case, I will learn a bit more about myself.

4

Shubhendu Trivedi @shubhendu.bsky.social · 1d

I want to learn this sort of tempo rather than just admire it, because it seems trainable. He seems to have developed an admirable system of micro-iteration discipline (which ironically I actively pushed him towards when he was an undergrad, but it has been a while since I've been able to keep up).

1 4

Shubhendu Trivedi @shubhendu.bsky.social · 1d

I have this long-term collaborator who has this ability to iterate frighteningly fast, and within each iteration, especially if it concerns optimization, avoids lazy knobs. e.g. optimizing kernel bandwidths? Then of course you use Brent or golden-section search rather than punt to grid search or BO.

1 3

Shubhendu Trivedi @shubhendu.bsky.social · 1d

It is always persecution complex day.

1

Shubhendu Trivedi @shubhendu.bsky.social · 1d

1 1

Shubhendu Trivedi @shubhendu.bsky.social · 1d

I think your point was accurate, I was just making an aside, partly because the DoE is the largest funder of physical sciences research. However, DoE funding is much more directed e.g. will go directly go to specific labs, towards specific priorities.

Shubhendu Trivedi @shubhendu.bsky.social · 1d

I don't think the DoD bar absorbs it (and it should not, since the DoE reports separately, and has a separate command). Just tried to search for some numbers and it seems like it is in the ballpark used in the chart.

1 1

Shubhendu Trivedi @shubhendu.bsky.social · 1d

As an aside: The chart seems to skip the DoE. It funds almost all of the national labs (some, like LANL, also get funded by the DoD). If you include some specific programs that it handles (i.e. nuclear security admin research, ARPA-E), it spends over $21-22 billion a year.

2 1

Shubhendu Trivedi @shubhendu.bsky.social · 2d

The thought did cross my mind, but it seems to hold across translators?

1

Shubhendu Trivedi @shubhendu.bsky.social · 2d

it, often juxtaposing it with him not being able to control his bowels. Reading that felt somewhat odd. But later I learnt that it was symptomatic of a stomach cancer, of which he was not aware at the time. Over time you learn that this quality is an invariant throughout his writing.

2

Shubhendu Trivedi @shubhendu.bsky.social · 2d

There are always random asides, and instances that feel like needless oversharing, affected candor, or something similar. But there's always something to them. For example, in his tiny book 'An Armenian Sketchbook' he starts with describing a "titanic Stalin statue" in Yerevan, keeps talking about

1 2

Shubhendu Trivedi @shubhendu.bsky.social · 2d

PS: I hesitated in using "embarrassingly," because there's nothing embarrassing about that. But there's certainly this contrast throughout his writing, not sure how else to put it.

2 2

Shubhendu Trivedi @shubhendu.bsky.social · 2d

Vasily Grossman was an example of a writer who was not conventionally good (flat sentences, uneven structure, "embarrassingly" overblown metaphors), but nevertheless, he was not only compulsively readable, his writing was often luminous.

1 2