Peter Bull

Enthusiastic to build on this generation of earth observation foundation embeddings like DeepMind's AlphaEarth (and more)! We already see some promising crop type (cereals vs. orchards) results and are exploring other use cases in climate resilience. deepmind.google/discover/blo...

File Browser - marimo The next generation of Python notebooks

Very cool to see that marimo supports our cloudpathlib library for their file browser UI! Browse your S3, GCS, Azure buckets from your notebooks! docs.marimo.io/api/inputs/f...

✨ 📦 ✨ Just released new Cookiecutter Data Science version with support for pixi and poetry as environment managers! Some of our top requested features ever. Upgrade and check it out now.

cookiecutter-data-science.drivendata.org

Now getting organic inbound for www.zambacloud.com, our wildlife imagery processing platform, from ChatGPT! 😲

Just in case you thought speech-to-text worked for children, the third column is what Whisper does. Somehow in the third example it accesses my inner monologue... I guess that's why we're excited about our upcoming challenge! kidsasr.drivendata.org

How are people managing code review for their AI coding agents? I do a first glance and it is obviously bad (e.g., didn't refactor repeated code), and now I've got half a dozen AI diffs for things that aren't good enough cluttering up my todo list with things to respond to....

Time is On My Side: Dynamics of Talk-Time Sharing in Video-chat Conversations An intrinsic aspect of every conversation is the way talk-time is shared between multiple speakers. Conversations can be balanced, with each speaker claiming a similar amount of talk-time, or…

New research based on the CANDOR corpus shows that people enjoy conversations where they alternate longer turns better than short turns or one person dominating. Cool!

arxiv.org/html/2506.20...

The best shortcut to how many experienced software engineers feel about AI is listening to the Primeagen's takes. Balanced perspectives on what's actually new, determinism, security, system complexity, what's promising, and what's not www.youtube.com/watch?v=vDWa...

Maldito ChatGPT Camilo · Maldito ChatGPT · Song · 2025

"Damn ChatGPT" your new summer jam about using ChatGPT as a therapist open.spotify.com/track/4umq06... (edited)

How Long Contexts Fail Taking care of your context is the key to building successful agents. Just because there’s a 1 million token context window doesn’t mean you should fill it.

Great article on the challenges of only surfacing the right info to LLMs and editing down what is not needed. If you've used a coding copilot or agent, you've seen this first hand many times. Output iterations are often polluted with code that came before.

www.dbreunig.com/2025/06/22/h...

BioCLIP2 looks like a stellar improvement! I'm excited to think about integrating into Zamba to for open-ended classification tasks run at scale on camera trap imagery. Definitely the potential to dramatically improve CT image utility. imageomics.github.io/bioclip-2/

Inside the AI Prompts DOGE Used to “Munch” Contracts Related to Veterans’ Health Experts who reviewed the code for ProPublica found numerous and troubling flaws in the system, providing a disturbing glimpse into how the Trump administration is allowing artificial intelligence to…

"Munchable" is GenZ cringe. www.propublica.org/article/insi...

We've built so many low-fidelity prototypes in our HCD work. IMO vibecoding changes the feel of those prototypes, but doesn't change the process. Ask any designer—they'll tell you high-fidelity first iterations are often more distracting to clients than helpful.

www.semafor.com/article/06/0...

Check out this LLM circuit trace LLM for the text: '"The statement 'this statement is false' is." It goes through a logical contradictions node, but still outputs either "true" or "false" with the highest probabilities... www.anthropic.com/research/ope...

A new preprint shows anonymization techniques for voices make transcription accuracy substantially worse for children versus adults. This is going to be a big challenge as we work on ASR for educational settings where we emphatically need both privacy and accuracy. arxiv.org/pdf/2506.00100

30 minutes with a stranger Watch hundreds of strangers talk for 30 minutes, and track how their moods change

😍 Incredible data storytelling about the power of conversation and human connection. Worth a read for good vibes! Based on the CANDOR corpus that we worked on. pudding.cool/2025/06/hell...

The gap between LLM prototype and production strikes again... in the worst possible place. www.propublica.org/article/trum...

The Sentinel-3 optical image shows a dense, orange plume of Saharan sand over approximately 150 000 sq km of the eastern Atlantic Ocean. The small islands of Cabo Verde peek out from beneath the clouds in the top left corner.

📷 This week's @esaearth.esa.int #EarthFromSpace is a #Copernicus Sentinel-3 visible image of a thick plume of orange dust from the Sahara Desert over approximately 150 000 sq km of the eastern Atlantic Ocean on 7 May 2025🧪🌍

www.esa.int/ESA_Multimed...

Very cool to see the multimodal conversation CANDOR dataset that we worked on used for a new paper on conversational agents! Gets agent feedback/training loops closer to the 7-38-55 rule than text only arxiv.org/abs/2505.15922

Zamba Cloud

You don't need to be a coder to use AI for wildlife research! 💻➡️🚫 With Zamba Cloud's new image support, simply upload photos, get species IDs, and even train custom models—all without writing a single line of code. www.zambacloud.com #WildlifeResearch

A nurse shark takes a leisurely stroll through Coral City against a backdrop of thriving staghorn, elkhorn, brain, finger, and star corals #nurseshark #sharksofcoralcity #shark #leisurelystroll #elkhorn #staghorn #braincoral #coral #coralcitycamera #miami #portmiami #biscaynebay #coralcity

What about the fact-checked news content?

Geospatial Foundational Disappointments TLDR; After 10²¹ FLOPs and 500 B patches, IBM’s TerraMind beats a supervised U‑Net by just +2 mIoU on PANGAEA; losing on 5/9 tasks, most other GFMs do worse.

Foundation models for geospatial aren't there yet. This piece argues that unlike with language the information density of labeled data, and pretraining tasks aren't relevant enough. Worth a read if you do any geo AI christopherren.substack.com/p/geospatial...

People keep plugging AI "Co-Scientists," so what happens when you ask them to do an important task like finding errors in papers?

We built SPOT, a dataset of STEM manuscripts across 10 fields annotated with real errors to find out.

(tl;dr not even close to usable) #NLProc

arxiv.org/abs/2505.11855

Microsoft is opening its on-device AI models up to web apps in Edge Edge continues to compete with Chrome.

This is a huge announcement for privacy-focused developers that don't want to do API calls, but the caniuse.com for built-in LLMs across browsers is going to be a shitshow. www.theverge.com/news/669528/...

Nation Can't Believe It On Harvard's Side

Gemini Diffusion Another of the announcements from Google I/O yesterday was Gemini Diffusion, Google's first LLM to use diffusion (similar to image models like Imagen and Stable Diffusion) in place of transformers. …

I got access to Gemini Diffusion, Google's first diffusion LLM, and the thing is absurdly fast - it ran at 857 tokens/second and built me a prototype chat interface in just a couple of seconds, video here: simonwillison.net/2025/May/21/...

How changing ‘localhost’ to ‘127.0.0.1’ sped up my test suite by 1,800% I would like to share a (somewhat recent) anecdote on how a one-line code change improved my understanding of software engineering.

PSA to test suites with HTTP servers on Windows: localhost can be 100x slower than 127.0.0.1! 🤯 I hit the issue on a new cloudpathlib feature for HTTP that's coming soon github.com/drivendataor... Nearly impossible to debug. More background here medium.com/hackernoon/h...

I say this a lot, but the narrative that AI use is going to collapse due to data limits or costs or environmental factors or regulation or a "hype bubble" popping or whatever is not a useful position for critics.

Capability development may slow (it hasn't done so yet), but AI use isn't going away.

= Nation & World - The Seattle Times My Account ™ The T. Rex may have been a lot smarter than you thought Jan. 9, 2023 at 7:36 am | Updated Jan. 9, 2023 at 7:36 am By DINO GRANDONI The Washington Post

typical corrupt science