Lightnews — Scholar-powered news

Adrian Hill

@adrianhill.de

710 followers 620 following 120 posts

PhD student at @bifold.berlin, Machine Learning Group, TU Berlin.
Automatic Differentiation, Explainable AI and #JuliaLang.
Open source person: adrianhill.de

Posts Replies Media Videos

Pinned

Adrian Hill @adrianhill.de · Jan 30

You think Jacobian and Hessian matrices are prohibitively expensive to compute on your problem? Our latest preprint with @gdalle.bsky.social might change your mind!
arxiv.org/abs/2501.17737
🧵1/8

Figure comparing automatic differentiation (AD) and automatic sparse differentiation (ASD).

(a) Given a function f, AD backends return a function computing vector-Jacobian products (VJPs). (b) Standard AD computes Jacobians row-by-row by evaluating VJPs with all standard basis vectors. (c) ASD reduces the number of VJP evaluations by first detecting a sparsity pattern of non-zero values, coloring orthogonal rows in the pattern and simultaneously evaluating VJPs of orthogonal rows. The concepts shown in this figure directly translate to forward-mode, which computes Jacobians column-by-column instead of row-by-row.

Reposted by Adrian Hill

Blaž Urban Gracar (Letibus Design)

@letibus.bsky.social

Yesterday, my puzzle book Workworkwork won the Thinky Award for the Best Pen and Paper Puzzle ( @thinkygames.com )!
In celebration I added 100 free community copies of the digital (PDF) version:
letibus.itch.io/www
Check out more about the game / get the physical copy here:
blazgracar.com/www

February 5, 2026 at 3:43 PM

Adrian Hill

@adrianhill.de

Claude Code vs. the editor-industrial-complex, who will come out on top?

marcel ⊙ @marceldotsci.bsky.social · 15d

my long bet to not bother to learn vim commands has finally paid off #sloperator

January 27, 2026 at 7:31 PM

Adrian Hill

@adrianhill.de

„Making things easy breaks systems that use difficulty as signaling“ @zey.bsky.social @neuripsconf.bsky.social

December 3, 2025 at 11:10 PM

Adrian Hill

@adrianhill.de

Join us today from 4:30 to 7:30 PM @neuripsconf.bsky.social Hall C,D,E #1006 for our poster on SmoothDiff, a novel XAI method leveraging automatic differentiation.
🧵1/6

December 3, 2025 at 7:36 PM

Adrian Hill

@adrianhill.de

In San Diego for @neuripsconf.bsky.social this week, so hit me up to talk science. I'll make one of these longer announcement posts for our paper on Wednesday.

December 2, 2025 at 7:05 PM

Adrian Hill

@adrianhill.de

Awesome to see the French government recognize our work on open-source software! @ouvrirlascience.bsky.social

Guillaume Dalle @gdalle.bsky.social · Dec 2

My baby #JuliaLang package got an award!
DifferentiationInterface.jl just received a national prize for open-source research software and I couldn't be prouder. We started this from a vague dream with @adrianhill.de, and reached 1000 dependents and 100k downloads last month.
Research != just papers.

A close up of the trophy for the "prix science ouverte du logiciel libre de recherche, édition 2025" (open science prize for free research software, 2025 edition)

December 2, 2025 at 7:03 PM

Adrian Hill

@adrianhill.de

When it comes to drama, ICLR 2026 is truly the gift that keeps on giving.

November 28, 2025 at 7:17 AM

Adrian Hill

@adrianhill.de

If I were looking for a PhD position, I'd jump at this chance!

Guillaume Dalle @gdalle.bsky.social · Nov 21

⚠️ WANTED: PhD student ⚠️
The topic? GPU-friendly algorithms for discrete optimization and mathematical programming, to be used inside decision-focused learning pipelines.
The mentors? My colleague Axel Parmentier and myself.
[1/3]

November 21, 2025 at 5:55 PM

Adrian Hill

@adrianhill.de

Working on a poster in Typst again. But this time from scratch without any package dependency. 💪

Proof of concept for a Typst poster with a two-column layout and the entire code that produces it:

```typst
// Poster Template
#set page("a0", flipped: true,
columns: 2,
margin: 3cm,
background: rect(fill: blue, width: 100%, height: 100%),
)
#set columns(gutter: 2cm)
#set par(spacing: 2cm)
#set text(font: "Helvetica", size: 90pt)
#set rect(width: 100%, inset: 1cm, fill: white)

// Content
#place(top, scope: "parent", float: true, clearance: 2cm,
rect(height: 8cm)[Title],
)

#rect(height: 1fr)[Panel 1]
#rect(height: 1fr)[Panel 2]
#rect(height: 1fr)[Panel 3]
#colbreak()

#rect(height: 3fr)[Panel 4]
#rect(height: 2fr)[Panel 5]

#place(bottom, scope: "parent", float: true, clearance: 2cm,
rect(height: 5cm)[Footer],
)
```

November 3, 2025 at 8:52 PM

Reposted by Adrian Hill

Explainable AI Berlin

@xai-berlin.bsky.social

This is the eXplainable AI research channel of the machine learning group of Prof. Klaus-Robert Müller at Technische Universität Berlin @tuberlin.bsky.social & BIFOLD @bifold.berlin.
Let's connect!
#XAI #ExplainableAI #MechInterp #MachineLearning #Interpretability

a black background with green text that says `` hello , world ''

ALT: a black background with green text that says `` hello , world ''

media.tenor.com

November 3, 2025 at 11:43 AM

Reposted by Adrian Hill

Jeremy Kun

@jeremykun.com

Mastodon came through, gotta up your game mathsky 😘

Jeremy Kun @jeremykun.com · Oct 30

Look at me I'm posting an image from another social network like a basic bitch.

But THAT social network has typesetting 😀 and so the post reads nicer than if I retyped it here with no typesetting.

A math problem showed up organically in some work stuff:

Given a graph G you'd like to produce a subgraph G prime. G comes with a partition of vertices V1 through Vk with the constraint that, for each subset Vi, at least one vertex in Vi must remain in the final graph. The goal is to make G' colorable with as few colors as possible.

Anyone heard of something like this? It is almost surely NP-hard, but just curious.

October 31, 2025 at 1:48 AM

Adrian Hill

@adrianhill.de

Microtypography, accessibility and a range of PDF standards. Now nothing is stopping me from writing my thesis in Typst. 🥹

Typst @typst.app · Oct 24

Typst 0.14 is out now! Get ready for production with accessibility, PDFs as images, character-level justification, and more. Learn about more of the highlights in Typst 0.14 in the thread below ⤵️

Image saying Typst 0.14. It has a holographic effect. The background is a pattern of checkmarks, at the top, it changes to a pattern of the string "0.14". The horizontal bar across the top reads "Title Elementm Skewed Fractions, Table Subheaders, Accessible Documents, Character-level Justification, PDF/A compliance, PDF Images, Typed HTML API"

October 24, 2025 at 1:51 PM

Adrian Hill

@adrianhill.de

This might actually motivate me to follow through AoC this year! Maybe in Typst?

Eric Wastl @was.tl · Oct 22

Hello, friends! After 10(!) years of #AdventOfCode, I've made some changes to preserve my sanity: there will be 12 days of puzzles each December (still starting Dec 1) and there is no longer a global leaderboard. Read more:
adventofcode.com/2025/about#f...
adventofcode.com/2025/about#f...

October 23, 2025 at 1:15 PM

Reposted by Adrian Hill

Georg Bökman

@bokmangeorg.bsky.social

Turns out NLP is just vision

Tim Kellogg @timkellogg.me · Oct 21

Z.ai released a paper very similar to DeepSeek-OCR on the same exact day (a few hours earlier afaict)

Glyph is just a framework, not a model, but they got Qwen3-8B (128k context) to handle over 1 million context by rendering input as images

arxiv.org/abs/2510.17800

Figure 1.

(Upper) Diagram comparing two paradigms for long-context tasks:
• Left path (Plain Text): A long novel (~180K words, example “Jane Eyre”) is fed directly as text into an LLM, requiring roughly 240K tokens.
• Right path (Rendering): The same text is rendered into images, producing about 80K tokens—achieving 3× input-token compression—and processed by a VLM (Vision-Language Model) instead of a pure text LLM.

(Lower) Two sets of bar charts:
• Left chart (Accuracy): Glyph performs comparably to Qwen-3-8B, GLM-4.9B-Chat-1M, and Qwen-2.5-7B-Instruct-1M on LongBench and MRCR tasks.
• Right chart (Compression/Speedup): Glyph shows 3.2× KV cache reduction, 4.8× prefill speedup, and 4.4× decoding throughput compared to the text backbone model.

Caption text:

Comparison of two paradigms for long-context tasks: conventional approaches directly feeding plain text into LLMs, and the proposed VLM-based paradigm, Glyph, which renders text as compact images to achieve substantial input-token compression. Glyph attains competitive performance on LongBench and MRCR while offering significant compression and inference speedup over its text backbone model on 128K-token inputs.

October 21, 2025 at 4:39 PM

Reposted by Adrian Hill

Typst

@typst.app

Curious what you can do with Typst's HTML export right now? How about write a blog? This article tells you how:

blog.sillyfreak.space/2025/06/29/r...

Rewrite it (the Blog) in Typst

blog.sillyfreak.space

October 21, 2025 at 7:48 AM

Reposted by Adrian Hill

Mark J. Nelson

@mm-jj-nn.bsky.social

Didn’t we use to measure this in FLOPS? www.ft.com/content/bdaf...

“When completed, the deals would bring OpenAI’s total access to computing capacity to more than 26GW”

October 14, 2025 at 1:24 PM

Adrian Hill

@adrianhill.de

Kind of hilarious that functional programming's biggest hit is a package manager

October 9, 2025 at 6:09 PM

Adrian Hill

@adrianhill.de

Perfect timing! If you are in Paris for @juliacon.bsky.social on Friday, join us for our talk on Automatic Sparse Differentiation at 11:30.

If you can't make it, check out both the paper and our introductory ICLR blog post: iclr-blogposts.github.io/2025/blog/sp...

September 30, 2025 at 12:05 PM

Adrian Hill

@adrianhill.de

Based on my own mediocre experiences with LLM-assisted coding, I can't help but wonder whether this explains the recent engineering issues Anthrophic was facing.

Gergely Orosz @gergely.pragmaticengineer.com · Sep 19

Talked with the Claude Code team on how they build Claude Code. It feels I get a peek into the future, and I get why Dario said 6 months ago that 90% of code will be written by AI.

This team works SO differently than any eng team I saw. Will share in-depth soon. One example:

September 22, 2025 at 1:06 PM