Thomas Friedel

I made a full Rust port of Slay the Spire 2 (all characters), an AI that plays it and a brower version of the game where the AI coaches you and lets you find your biggest blunders. Win-rate for Defect at A10 is an impressive 22%! tfriedel.github.io/spire-mentor

This post is an absolute banger.

Weekend thoughts on Gas Town, Beads, slop AI browsers, and AI-generated PRs flooding overwhelmed maintainers. I don't think we're ready for our new powers we're wielding. lucumr.pocoo.org/2026/1/18/ag...

Kind of genius.

Claude's new Office doc skills (PPTX, DOCX, XLSX, PDF) are 🔥 I've repackaged them for Claude Code, so now you can automate the creation of powerpoint presentations and other documents from the CLI or in your product. github.com/tfriedel/cla...

🧵 THREAD: A federal whistleblower just dropped one of the most disturbing cybersecurity disclosures I’ve ever read. He's saying DOGE came in, data went out, and Russians started attempting logins with new valid DOGE passwords Media's coverage wasn't detailed enough so I dug into his testimony:

The current figure I'm seeing is that as many as 3.3 million Americans protested today. #HandsOff #Resist

Switzerland seems to also allow training on copyrighted works: www.vischer.com/en/knowledge...

Obstore v0.6 brings simple, painless integration with Microsoft Planetary Computer! - *Automatic token refresh*. No need to manually fetch SAS tokens and never think about token expiration times again. - STAC API integration developmentseed.org/obstore/late...

There's now so many awesome #geospatial people here it's hard to keep track! 🤩 Here's a thread of some of my favourite starter packs so far - let me know what I've missed! 🛰️ Earth Observation & remote sensing: go.bsky.app/4PMRhNL 🌏 GIS: go.bsky.app/TJ7qQF6 @milos-makes-maps.bsky.social (1/n)

Visiting friends in New York

My new paper "Deep Learning is Not So Mysterious or Different": arxiv.org/abs/2503.02113. Generalization behaviours in deep learning can be intuitively understood through a notion of soft inductive biases, and formally characterized with countable hypothesis bounds! 1/12

asking the GOP, will they impeach this mf?

Among all the utterly insane stuff going on, this ranks pretty high on the insane scale. mRNA vaccines are one of the great scientific advances of this century. www.newsfromthestates.com/article/bill...

This is a crazy paper. Fine-tuning a big GPT-4o on a small amount of insecure code or even "bad numbers" (like 666) makes them misaligned in almost everything else. They are more likely to start offering misinformation, spouting anti-human values, and talk about admiring dictators. Why is unclear.

Microsoft's OmniParser V2: Turning Any LLM into a Computer Use Agent Blog: www.microsoft.com/en-us/resear... Hugging Face: huggingface.co/microsoft/Om... GitHub: github.com/microsoft/Om...

an audible “no fucking way..” escapes my mouth they claim that you can get LLMs to give you well calibrated confidence scores like, i knew you could do something like this with logits/attention values, but the idea that the LLM could speak it back to you.. 🤯

🌍 New paper: Generative networks for spatio-temporal #gap_filling of #Sentinel2 reflectances @esa.int. 🆓 Free access the first 50 days to the #ISPRS paper @elsevierconnect.bsky.social: isp-uv.short.gy/GapFillingSe... Work developed in the context of Xaida Project. #AI #EarthObservation

Today’s Stratechery was great on this: stratechery.com/2025/deepsee...

Today, the Sentinel-2C spacecraft will take over from S2A in providing operational data. This is wonderful news, and will ensure continuity of the Sentinel data for many years to come. Great effort from the S2C commissioning team! See here for more info: sentinels.copernicus.eu/web/sentinel...

agfundernews.com/project-eade...

Google's Titans: a new architecture with attention and a meta in-context memory that learns how to memorize at test time as presented by one of the author - @alibehrouz.bsky.social

"Sky-T1-32B-Preview, our reasoning model that performs on par with o1-preview on popular reasoning and coding benchmarks." That was quick! Is this already the Alpaca moment for reasoning models? Source: novasky-ai.github.io/posts/sky-t1/

A new position paper on "inference time compute". They present some theory on why it is necessary, how does it work, why we need it and what does it mean for "super" intelligence.

This might be the first time after 10 years that boosted trees are not the best default choice when working with data in tables. Instead a pre-trained neural network is, the new TabPFN, as we just published in Nature 🎉

Excited to finally share this work w/ @suryaganguli.bsky.social Tl;dr: we find the first closed-form analytical theory that replicates the outputs of the very simplest diffusion models, with median pixel wise r^2 values of 90%+. arxiv.org/abs/2412.20292

“By default, capital will matter more than ever after AGI” www.lesswrong.com/posts/KFFaKu...

🎉 ShellSage v0.0.6 is now available! New features: • Add new command mode for automatically inserting the correct command into your terminal • Add Support for OpenAI Compatible Providers Various bugs squashed too! github.com/AnswerDotAI/...

Want to impress family or friends? Ask them for a photo featuring a shadow and follow the steps GeoGuessr pro @georainbolt.bsky.social does here. Using Bellingcat's Shadow Finder tool, you can narrow down the likely location the photo was taken in. youtu.be/pQIjDPFgdJA?...

www.nytimes.com/2024/12/13/o... Don’t agree with everything, but it’s a really interesting read.

When doing research for AI Engineering, I went through so many papers, case studies, blog posts, repos, tools, etc. This repo contains ~100 resources that really helped me understand various aspects of building with foundation models. github.com/chiphuyen/ai...

Wrote up my initial impressions of the new Google Gemini 2.0 Flash model - it's really good, and the streaming mode (where you can stream video and audio to it and get audio streamed right back) is pure science-fiction simonwillison.net/2024/Dec/11/...

We've all been there—staring at the terminal trying to remember that obscure tar command or SSH flags. Sure, you could Google it or use ChatGPT, but now you're context-switching between docs and terminals. So, we created ShellSage, read all about it here: www.answer.ai/posts/2024-1...

New paper! We've just published a manuscript covering technical details of the AI models behind Wildlife Insights, with systematic analysis at scale of cropping vs not cropping in #CameraTrap image classifiers 🫎🦌🦒🦘🪿🫏🐘🐫🦬🦛🐿️ ietresearch.onlinelibrary.wiley.com/doi/10.1049/...

🎯 How can we empower scientific discovery in millions of nature photos? Introducing INQUIRE: A benchmark testing if AI vision-language models can help scientists find biodiversity patterns- from disease symptoms to rare behaviors- hidden in vast image collections. Thread👇🧵

🎉 I just wrapped up teaching the Introduction to GIS Programming course this semester at the University of Tennessee! All course materials and lecture recordings (26 hours) are freely available online. Explore #geospatial 🌍 data visualization and analysis using #opensource Python packages:

How do LLMs learn to reason from data? Are they ~retrieving the answers from parametric knowledge🦜? In our new preprint, we look at the pretraining data and find evidence against this: Procedural knowledge in pretraining drives LLM reasoning ⚙️🔢 🧵⬇️

Can we trust a group of people to be right, and to be competent, simply because they agree on something? In a new paper in Cognition, @bdecourson.bsky.social, @hugoreasoning.bsky.social and I argue that yes, in many cases we can. authors.elsevier.com/a/1kA8Q2Hx2x... 1/

Cautious Optimizers: Improving Training with One Line of Code AdamW, speed-up on Llama and MAE pretraining up to 1.47x huggingface.co/papers/2411....

Ever wanted to see the world through someone else’s eyes?

First post on @bsky.app! I wanted to share tgengine.org/introducing-... - a blog post I wrote about fieldsofthe.world. It's a huge benchmark dataset designed to enhance the development of ml models for field delineation. Working on the project for the last few months has been an absolute pleasure.

I was looking for a quick image side-by-side / overlay comparison tool to compare outputs of super resolution models and found this: github.com/olive-groves... It's python based and exactly what I want.