Daniel Augusto
LightNews LightNews
banner
spectral.space
Daniel Augusto
@spectral.space
69 followers 230 following 24 posts
PhD student at UCL. Working on Gaussian Processes. I love everything mathy, videogames, and language. 🇧🇷
He/Ele/Él

Also as @spectraldani@sigmoid.social
Posts Replies Media Videos
Reposted by Daniel Augusto
ottoojala.com
Bsky always does weird things to videos shot on my phone, but this thing actually runs 30fps and looks nice on my Anbernic RG 35XX plus, just had to disable the shadows. There's likely things I can still optimize, but I'm already very happy with how this runs on this cheap Linux handheld :--)
November 13, 2025 at 12:14 PM
Reposted by Daniel Augusto
dubroy.com
Somewhat surprising (to me) advice in Google's C++ style guide:

"Try to avoid unsigned types (except for representing bitfields or modular arithmetic). Do not use an unsigned type merely to assert that a variable is non-negative."

Example of a possible bug: abseil.io/tips/227

(via @wingolog.org)
On Unsigned Integers Unsigned integers are good for representing bitfields and modular arithmetic. Because of historical accident, the C++ standard also uses unsigned integers to represent the size of containers - many members of the standards body believe this to be a mistake, but it is effectively impossible to fix at this point. The fact that unsigned arithmetic doesn't model the behavior of a simple integer, but is instead defined by the standard to model modular arithmetic (wrapping around on overflow/underflow), means that a significant class of bugs cannot be diagnosed by the compiler. In other cases, the defined behavior impedes optimization. That said, mixing signedness of integer types is responsible for an equally large class of problems. The best advice we can provide: try to use iterators and containers rather than pointers and sizes, try not to mix signedness, and try to avoid unsigned types (except for representing bitfields or modular arithmetic). Do not use an unsigned type merely to assert that a variable is non-negative.
November 11, 2025 at 3:45 PM
Reposted by Daniel Augusto
ryan-steele-mcg.bsky.social
Mt. Hood, Oregon twilight reflection
November 5, 2025 at 5:30 PM
Reposted by Daniel Augusto
lacerbi.bsky.social
I answered in thread but it might be of interest.

My current advice is to use the stderr of the median (bootstrapped), with a separate metric for reliability if needed.

The latter is less standard - we have been reporting e.g. 80% perf quantile and its stderr (bootstrapped).
gershbrain.bsky.social Sam Gershman @gershbrain.bsky.social · 25d
What's with machine learning researchers always reporting standard deviation instead of standard error? My understanding is that the error bars are typically used to back up inferential claims about significant differences between sample means (although statistical tests are rare, another problem).
October 26, 2025 at 12:52 PM
Reposted by Daniel Augusto
vabor112.bsky.social
I am hiring a fully-funded #PhD in #ML to work at the University of Edinburgh on 𝐠𝐞𝐨𝐦𝐞𝐭𝐫𝐢𝐜 𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠 and 𝐮𝐧𝐜𝐞𝐫𝐭𝐚𝐢𝐧𝐭𝐲 𝐪𝐮𝐚𝐧𝐭𝐢𝐟𝐢𝐜𝐚𝐭𝐢𝐨𝐧.

Application deadline: 31 Dec '25. Starts May/Sep '26.
Details in the reply.

Pls RT and share with anyone interested!
October 24, 2025 at 11:35 AM
Reposted by Daniel Augusto
was.tl
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 22, 2025 at 2:32 AM
Reposted by Daniel Augusto
dannybate.bsky.social
Released a new ALILI episode yesterday, about the show's first sign language!

ASL is the beloved language of Aidan Elliott-McCrea (the linguist behind @wordfamilyfriday.bsky.social), and he gave me both a great intro to ASL and heartfelt social messages too. Listen here:
pod.link/1703401848/e...
Artwork for the podcast A Language I Love Is...
October 21, 2025 at 8:37 AM
Reposted by Daniel Augusto
njclark.bsky.social
Compositional data needs models that respect its constraints. I show how Dirichlet regression with GPs captures nonlinear environmental effects on plant communities while respecting the dependence of the data #rstats #dataviz ecogambler.netlify.app/blog/plant-c...
October 21, 2025 at 8:19 AM
Reposted by Daniel Augusto
jonnelledge.bsky.social
The main cause of the noise is something that will, sooner or later, hit every part of the network: “rail corrugation”.

You aren’t imagining it: London's Tube is getting louder.
No, you aren’t imagining it: the Tube is getting louder
I SAID, THE TUBE *IS* GETTING LOUDER.
jonn.substack.com
October 19, 2025 at 9:38 AM
Reposted by Daniel Augusto
rieck.mlsec.org
AI predicts rain. We predict trouble!

Today, Erik presents a novel attack on Google's latest AI weather model at #CCS2025. By changing only 0.1% of the observations, the attack can fabricate or suppress the prediction of extreme events, from hurricanes 🌀 to heat waves 🔥

1/4 @bifold.berlin
Predicted temperature at the peak of the European Heat Wave 2006
October 15, 2025 at 6:06 PM
spectral.space
#photography #london
October 15, 2025 at 9:07 PM
Reposted by Daniel Augusto
tmlrorg.bsky.social
As Transactions on Machine Learning Research (TMLR) grows in number of submissions, we are looking for more reviewers and action editors. Please sign up!

Only one paper to review at a time and <= 6 per year, reviewers report greater satisfaction than reviewing for conferences!
October 14, 2025 at 1:32 PM
Reposted by Daniel Augusto
timkellogg.me
Karpathy: nanochat

A small training+inference pipeline for creating your own LLM from scratch

$100 will get you a somewhat functional model

$1000 is more coherent & solves math

detailed walkthrough: github.com/karpathy/nan...

repo: github.com/karpathy/nan...
Andrej Karpathy & @karpathy X.com Excited to release new repo: nanochat! (it's among the most unhinged I've written). Unlike my earlier similar repo nanoGPT which only covered pretraining, nanochat is a minimal, from scratch, full-stack training/inference pipeline of a simple ChatGPT clone in a single, dependency-minimal codebase. You boot up a cloud GPU box, run a single script and in as little as 4 hours later you can talk to your own LLM in a ChatGPT-like web Ul. It weighs ~8,000 lines of imo quite clean code to: - Train the tokenizer using a new Rust implementation - Pretrain a Transformer LLM on FineWeb, evaluate CORE score across a number of metrics - Midtrain on user-assistant conversations from SmolTalk, multiple choice questions, tool use. - SFT, evaluate the chat model on world knowledge multiple choice (ARC-E/C, MMLU), math (GSM8K), code (HumanEval) - RL the model optionally on GSM8K with IPDDOI - RL the model optionally on GSM8K with "GRPO" - Efficient inference the model in an Engine with KV cache, simple prefill/ decode, tool use (Python interpreter in a lightweight sandbox), talk to it over CLI or ChatGPT-like WebUl. - Write a single markdown report card, summarizing and gamifying the whole thing. Even for as low as ~$100 in cost (~4 hours on an 8XH100 node), you can train a little ChatGPT clone that you can kind of talk to, and which can write stories/poems, answer simple questions. About ~12 hours surpasses GPT-2 CORE metric. As you further scale up towards ~$1000 (~41.6 hours of training), it quickly becomes a lot more coherent and can solve simple math/code problems and take multiple choice tests. E.g. a depth 30 model trained for 24 hours (this is about equal to FLOPs of GPT-3 Small 125M and 1/1000th of GPT-3) gets into 40s on MMLU and 70s on ARC-Easy, 20s on GSM8K, etc. My goal is to get the full "strong baseline" stack into one cohesive, minimal, readable, hackable, maximally forkable repo. nanochat will be the capstone project of LLM101n (which is still being developed). I think it also has potential to grow developed). I think it also has potential to grow into a research harness, or a benchmark, similar to nanoGPT before it. It is by no means finished, tuned or optimized (actually I think there's likely quite a bit of low-hanging fruit), but I think it's at a place where the overall skeleton is ok enough that it can go up on GitHub where all the parts of it can be improved. Link to repo and a detailed walkthrough of the nanochat speedrun is in the reply. nanochat
October 13, 2025 at 6:06 PM
spectral.space
#photography #wales
October 12, 2025 at 9:55 PM
Reposted by Daniel Augusto
sebastianlague.bsky.social
My latest coding adventure — attempting to write a grid-based fluid solver and simulate some simple smoke!

youtu.be/Q78wvrQ9xsU?...
Coding Adventure: Simulating Smoke
YouTube video by Sebastian Lague
youtu.be
October 12, 2025 at 6:34 AM
Reposted by Daniel Augusto
hypv.bsky.social
A lot of people think that the complex numbers are more nicely behaved than the real numbers. But in fact ℝ is better than ℂ, and model theory and Gottfried Wilhelm von Leibniz can help us to understand why. Part 1/2 #MathSky
\title{Why $\mathbb R > \mathbb C$, and how model theory can help explain why} \author{Julia Redacted} \begin{document} \begin{frame} \titlepage \end{frame} \begin{frame}{How do you tell when two things are different?} Mathematicians love to gloss over this, but it's \highlightv{really hard} to tell whether two things are equal. One object can be presented in different ways and it can be non-trivial to see they give the same thing. \bigskip \begin{itemize} \item If you were ever confused why $1 = 0.999\ldots$ you've experienced this. \end{itemize} \bigskip If you're interested in some mathematical structure \highlightb{$\Mcal$} consisting of objects, how do you tell them apart? \end{frame} \begin{frame}{Leibniz's Identity of Indiscernibles} \begin{columns} \begin{column}{0.2\textwidth} \includegraphics[width=0.75in]{leibniz.jpg} \end{column} \begin{column}{0.8\textwidth} \highlightp{\textbf{\Large``}}Two objects are identical iff they have the same properties.\highlightp{\textbf{\Large''}} \vspace{.5in} \end{column} \end{columns} \begin{columns} \begin{column}{0.8\textwidth} \ Mathematically: $x$ and $y$ are objects in $\Mcal$. Then $x = y$ iff $\Mcal \models \phi(x) \iff \Mcal \models \phi(y)$ for every property $\phi$. \begin{itemize} \item $\Mcal \models \varphi(x)$ is model theorist speak for \highlightv{$x$ has property $\varphi$ in the structure $\Mcal$}. \item It's circular to have a property like ``identical to $x$''. We only mean properties you can express in the \highlightv{language} of the structure, e.g. the language of rings. \end{itemize} \end{column} \end{columns} \end{frame} \begin{frame}{Telling numbers apart} \begin{itemize} \item You can distinguish $1$ from any other number, because $1$ is the only number which is a multiplicative identity. \item You can distinguish $\pi - e$ from $\pi/e$ because one is negative and the other is positive. \item You might think you can distinguish $i$ and $-i$ because only one is on the top half of the complex plane, but what if you are holding $\Cbb$ upside down? \end{itemize} \end{frame}
October 10, 2025 at 3:37 PM
Reposted by Daniel Augusto
zevhandel.bsky.social
We've just enjoyed the bright harvest moon of Mid-Autumn Festival (中秋, 추석 秋夕, 月見, Trung Thu, etc.), so how about a moon-related linguistics thread?

We'll look at an unusual Chữ Nôm graph and recover some lost 17th-century Vietnamese sounds, among other things.

🌕 🇵🇹 🇻🇳

1/🧵
Black-and-white photo of the full moon, showing a lot of detail of the surface of the moon. Still image of the performer Psy from the music video for Gangnam Style, captioned "LET'S DO THIS CHỮ NÔM STYLE"
October 8, 2025 at 1:02 AM
Reposted by Daniel Augusto
404media.co
BREAKING: A catastrophic breach has impacted Discord user data including selfies and identity documents uploaded as part of the app’s verification process, email addresses, phone numbers, approximately where the user lives, and much more.

🔗 www.404media.co/the-discord-...
The Discord Hack is Every Users’ Worst Nightmare
A hack impacting Discord’s age verification process shows in stark terms the risk of tech companies collecting users’ ID documents. Now the hackers are posting peoples’ IDs and other sensitive informa...
www.404media.co
October 9, 2025 at 3:35 PM
Reposted by Daniel Augusto
theverge.com
Discord says 70,000 users may have had their government IDs leaked in breach
Discord says 70,000 users may have had their government IDs leaked in breach
The ID pics had been submitted as part of age-related appeals.
buff.ly
October 8, 2025 at 9:40 PM
Reposted by Daniel Augusto
savannah.dev
The stable release of Python 3.14 is out now! Go, go, go update! 🙌

discuss.python.org/t/python-3-1...
Cute illustrated logo featuring a pink/peach colored pie displaying ‘3.14’ (pi), surrounded by two snakes in blue and yellow. The design is encircled by text reading ‘r-strings • zsid • free-threading • support • sub-interpreters • REPL highlighting • note colour • colour’ in a circular arrangement. The illustration uses a soft pastel color palette with blue, yellow, and pink tones, and includes decorative sparkle elements.
October 7, 2025 at 2:30 PM
Reposted by Daniel Augusto
climformatics.bsky.social
We have exciting news! #CI2026 will be hosted by the École Polytechnique Fédérale de Lausanne (EPFL), in the period of April 27-30, 2026. More information in Tom Beucler's post:

www.linkedin.com/posts/tom-be...
Save the date! 🌍 Climate Informatics will convene researchers at the intersection of weather &amp; climate science, statistics, and machine learning, April 27-30, 2026 on our beautiful UNIL/EPFL… | ...
Save the date! 🌍 Climate Informatics will convene researchers at the intersection of weather & climate science, statistics, and machine learning, April 27-30, 2026 on our beautiful UNIL/EPFL campus ...
www.linkedin.com
October 1, 2025 at 6:35 PM
Reposted by Daniel Augusto
lacerbi.bsky.social
Do you like to train neural networks to solve all your nasty probabilistic inference and sequential design problems?
Do you love letter salads such as NPs, PFNs, NPE, SBI, BED?

Then no place is better than the Amortized ProbML workshop we are organizing at #ELLIS UnConference.
Amortized ProbML: Unlocking Probabilistic Machine Learning for Real-world Inference & Design (🇩🇰) ELLIS UnConference Workshop Copenhagen | 2 December, 2025
October 1, 2025 at 1:57 PM
Reposted by Daniel Augusto
bloomfilters.bsky.social
a tweet by toby fox that says "I didn't try to learn Japanese for a long time because I thought it would be too hard. But now I love learning it and I use it every day. If there's anything you want to learn, don't be stopped by thinking it'll be too hard! Try it! Game-making, music, art, language, anything!"
September 22, 2025 at 4:03 AM
Reposted by Daniel Augusto
maosbot.bsky.social
5,290 papers will be published at NeurIPS this year. If you could still get the proceedings in a book, it would be like 50,000 pages long. That’s 14.5 papers a day to read each and every day until the next NeurIPS
September 18, 2025 at 7:06 PM
Reposted by Daniel Augusto
trappmartin.bsky.social
Unfortunately, our submission to #NeurIPS didn’t go through with (5,4,4,3). But because I think it’s an excellent paper, I decided to share it anyway.

We show how to efficiently apply Bayesian learning in VLMs, improve calibration, and do active learning. Cool stuff!

📝 arxiv.org/abs/2412.06014
Post-hoc Probabilistic Vision-Language Models
Vision-language models (VLMs), such as CLIP and SigLIP, have found remarkable success in classification, retrieval, and generative tasks. For this, VLMs deterministically map images and text descripti...
arxiv.org
September 18, 2025 at 8:34 PM