Diego de las Casas

We're probably at peak research science. Low level stuff is getting automated, coding is basically reviewing, plots are beautiful, everyone gets a brainstorming buddy. Work is mostly ideation/planning. Soon it might become mostly meetings/auditing. Hope I'm wrong, work is pretty fun right now!

Exposed Moltbook Database Let Anyone Take Control of Any AI Agent on the Site 'It exploded before anyone thought to check whether the database was properly secured.'

The AI community is re-learning 20 years of cybersecurity. The hard way www.404media.co/exposed-molt...

Humming in denial as Material 3 takes over all my screens

Atrapanubes is such a good Chilean beer. Great taste, great art, great name.

Gotta wait until he double-crosses Indiana Jones to steal the Holy Grail, I'm afraid.

Mistral OCR New closed-source specialist OCR model by Mistral - you can feed it images or a PDF and it produces Markdown with optional embedded images. It's available [via their API](https://docs.mistral.ai/api/#tag/ocr), or …

I wrote a CLI script to run PDFs through the new Mistral OCR API model (with some help from Claude) - details on that and notes on the new model here: https://simonwillison.net/2025/Mar/7/mistral-ocr/

‎Le Chat by Mistral AI ‎Le Chat combines powerful AI with extensive information on the web to help you rediscover the world. Enjoy natural conversations, real-time internet search, comprehensive document analysis, and much ...

App store: apps.apple.com/us/app/le-ch...

Play store: play.google.com/store/apps/d...

The all new le Chat: Your AI assistant for life and work | Mistral AI Brand new features, iOS and Android apps, Pro, Team, and Enterprise tiers.

check out more in our post:
mistral.ai/en/news/all-...

We've upgraded Le Chat and it's blazing fast right now!
Also available for Android and iOS as of today
mistral.ai/en/news/all-...

mistralai/Ministral-8B-Instruct-2410 · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Have you tried our 8B model?
huggingface.co/mistralai/Mi...

mistral-small Mistral Small 3 sets a new benchmark in the “small” Large Language Models category below 70B.

Mistral Small 3 is also available on many partner platforms:
- Ollama: ollama.com/library/mist...
- Kaggle: kaggle.com/models/mistr...
- Fireworks: fireworks.ai/models/firew...
- Together: together.ai/blog/mistral...

And many more soon!

Performance of Mistral Small 3 Instruct model
huggingface.co/mistralai/Mi...

Mistral Small 3 Base model
huggingface.co/mistralai/Mi...

Mistral Small 3 architecture is optimised for latency while preserving high quality

Mistral Small 3 Apache 2.0, 81% MMLU, 150 tokens/s

We're releasing Mistral Small 3!
- 24B params, 81% MMLU
- Latency optimized: 150 tokens/s
- Competitive with Llama-3.3 70B, Qwen-2.5 32B, GPT4o-mini
- Apache 2.0
mistral.ai/news/mistral...

What people are going to do with AGI

Screen cap from one of the Thor movies featuring a dark haired pale skinned woman as Thor's sister Hela. She has her hand out stopping Thor's hammer (Mjölnir) in mid air. The hammer is labeled "It's basic biology". Hela is labeled "Advanced Biology"

I know, but it's just an application of one of my favorite memes:

agent swarm framework aces spatial reasoning test

Inventors of flow matching have released a comprehensive guide going over the math & code of flow matching!

Also covers variants like non-Euclidean & discrete flow matching.

A PyTorch library is also released with this guide!

This looks like a very good read! 🔥

arxiv: arxiv.org/abs/2412.06264

YouTube video by Jane Street Building Machine Learning Systems for a Trillion Trillion Floating Point Operations

Jane Street, a quant trading firm has a very good YouTube channel. For comparison, DeepSeek is also a quant trading firm.

They recently published a video on "Building Machine Learning Systems for a Trillion Trillion Floating Point Operations".

Link: www.youtube.com/watch?v=139U...

AI Scientists: here is a technology that will automate your grunt work so you can spend more time with your kids

AI Ads: here is a technology that will automate spending time with your kids

A dataset of 1 million or 2 million Bluesky posts is completely irrelevant to training large language models.

The primary usecase for the datasets that people are losing their shit over isn't ChatGPT, it's social science research and developing systems that improve Bluesky.

Arxiv sharing reminder

pdf ❌
abs ✅

In fact, statistical malpractice is the main driver of progress in machine learning. At some point, we need to come to terms with this.

[RFC] Per-Parameter-Sharding FSDP · Issue #114299 · pytorch/pytorch Per-Parameter-Sharding FSDP Motivation As we looked toward next-generation training, we found limitations in our existing FSDP, mainly from the flat parameter construct. To address these, we propos...

Fsdp2 has a different policy for handling streams that is also worth a read
github.com/pytorch/pyto...

French Revolution: Cyclists Now Outnumber Motorists In Paris Official measurements have found that Paris is rapidly becoming a city of cyclists.

READ: “3,337 Parisians were equipped with GPS trackers to record their journeys…for journeys from the outskirts of Paris to the center, the number of cyclists now far exceeds the number of motorists, a huge change from just 5 years ago.”

Evidence of leadership.
www.forbes.com/sites/carlto...

Comparison table of various AI models across different benchmarks: Mathvista, MMMU, ChartQA, DocVQA, VQAv2, AI2D, and MM MT-Bench. Models are categorized into Open Weights, Closed, and Unreleased. Key models include Pixtral Large, Llama-3.2 90B, Gemini-1.5 Pro, GPT-4o, Claude-3.5 Sonnet, Llama-3.1 505B, and Grok-2. The table shows measured and reported performance scores, highlighting differences in model capabilities across various tasks. Pixtral Large excels in Mathvista, DocVQA, AI2D and MM MT-Bench benchmarks.

Pixtral Large:
- 123B decoder, 1B vision encoder, 128K sequence length
- Frontier multimodal model
- Maintains text performance of Mistral Large 2

HF weights: huggingface.co/mistralai/Pi...
Try it: chat.mistral.ai
Blog post: mistral.ai/news/pixtral...

Two announcement cards from the Mistral AI team, dated November 18, 2024. The first card announces 'Mistral has entered the chat' with a brief description: 'Search, vision, ideation, coding... all yours for free.' The second card announces 'Pixtral Large' with the description: 'Pixtral grows up.' Both cards feature an orange 'Read More' button.

We have 2 new big updates today at Mistral:
- New Le Chat: With canvas, web search, image understanding and generation & more - and free!
- Pixtral Large, our Frontier 124B open weight multimodal model that powers it.

Try it: chat.mistral.ai
Blog post: mistral.ai/news/mistral...

Diffusion is spectral autoregression A deep dive into spectral analysis of diffusion models of images, revealing how they implicitly perform a form of autoregression in the frequency domain.

There seems to be some renewed interest in making this work in the ML/AI space, so I'm here as well 👋

Here's my latest blog post for good measure, about how diffusion models of images perform autoregression in frequency space: sander.ai/2024/09/02/s...

When I write more, I'll share here as well!

Quick thread in response to a question on token packing practices when pretraining LLMs!