Diego de las Casas
@dlsq.bsky.social
AI Scientist at Mistral AI.
Past: Google DeepMind.
🇧🇷 in 🇬🇧
Past: Google DeepMind.
🇧🇷 in 🇬🇧
Humming in denial as Material 3 takes over all my screens
November 4, 2025 at 12:34 AM
Humming in denial as Material 3 takes over all my screens
Atrapanubes is such a good Chilean beer. Great taste, great art, great name.
March 16, 2025 at 11:25 PM
Atrapanubes is such a good Chilean beer. Great taste, great art, great name.
Reposted by Diego de las Casas
Gotta wait until he double-crosses Indiana Jones to steal the Holy Grail, I'm afraid.
March 9, 2025 at 2:02 PM
Gotta wait until he double-crosses Indiana Jones to steal the Holy Grail, I'm afraid.
Reposted by Diego de las Casas
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/
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 …
simonwillison.net
March 7, 2025 at 1:41 AM
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/
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-...
Also available for Android and iOS as of today
mistral.ai/en/news/all-...
February 6, 2025 at 6:07 PM
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-...
Also available for Android and iOS as of today
mistral.ai/en/news/all-...
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...
- 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...
Mistral Small 3
Apache 2.0, 81% MMLU, 150 tokens/s
mistral.ai
January 30, 2025 at 9:17 PM
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...
- 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...
Reposted by Diego de las Casas
What people are going to do with AGI
January 26, 2025 at 4:30 PM
What people are going to do with AGI
Reposted by Diego de las Casas
I know, but it's just an application of one of my favorite memes:
January 21, 2025 at 7:07 PM
I know, but it's just an application of one of my favorite memes:
Reposted by Diego de las Casas
agent swarm framework aces spatial reasoning test
December 25, 2024 at 4:59 PM
agent swarm framework aces spatial reasoning test
Reposted by Diego de las Casas
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
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
December 10, 2024 at 8:35 AM
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
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
Reposted by Diego de las Casas
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...
They recently published a video on "Building Machine Learning Systems for a Trillion Trillion Floating Point Operations".
Link: www.youtube.com/watch?v=139U...
Building Machine Learning Systems for a Trillion Trillion Floating Point Operations
YouTube video by Jane Street
www.youtube.com
December 9, 2024 at 5:26 PM
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...
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
AI Ads: here is a technology that will automate spending time with your kids
December 3, 2024 at 10:35 PM
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
AI Ads: here is a technology that will automate spending time with your kids
Reposted by Diego de las Casas
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.
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.
Did you know that 99% of email today is spam? Your inbox isn’t 99% spam because AI is used to filter it.
The same 99% will happen here too, but if AI researchers continue to get perma-banned for making available the datasets needed to filter it, it’s going to make this platform unusable.
The same 99% will happen here too, but if AI researchers continue to get perma-banned for making available the datasets needed to filter it, it’s going to make this platform unusable.
November 28, 2024 at 6:57 PM
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.
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.
Reposted by Diego de las Casas
Arxiv sharing reminder
pdf ❌
abs ✅
pdf ❌
abs ✅
November 26, 2024 at 8:42 AM
Arxiv sharing reminder
pdf ❌
abs ✅
pdf ❌
abs ✅
Reposted by Diego de las Casas
In fact, statistical malpractice is the main driver of progress in machine learning. At some point, we need to come to terms with this.
November 22, 2024 at 2:40 PM
In fact, statistical malpractice is the main driver of progress in machine learning. At some point, we need to come to terms with this.
Reposted by Diego de las Casas
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...
Evidence of leadership.
www.forbes.com/sites/carlto...
French Revolution: Cyclists Now Outnumber Motorists In Paris
Official measurements have found that Paris is rapidly becoming a city of cyclists.
www.forbes.com
November 19, 2024 at 7:12 PM
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...
Evidence of leadership.
www.forbes.com/sites/carlto...
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...
- 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...
November 18, 2024 at 5:57 PM
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...
- 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...
Reposted by Diego de las Casas
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!
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!
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.
sander.ai
November 15, 2024 at 6:57 PM
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!
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!
Reposted by Diego de las Casas
Quick thread in response to a question on token packing practices when pretraining LLMs!
Yes! Token packing has been the standard since RoBERTa. Excerpt below!
The intuition is that the model quickly learns to not attend across [SEP] boundaries and packing avoids "wasting" compute on padding tokens required to make the variable batch size consistent.
The intuition is that the model quickly learns to not attend across [SEP] boundaries and packing avoids "wasting" compute on padding tokens required to make the variable batch size consistent.
November 7, 2024 at 6:21 PM
Quick thread in response to a question on token packing practices when pretraining LLMs!
Hey all, thanks for the follow!
Just FYI, we're hiring AI Scientists and Engineers at Mistral AI.
If you're driven and interested in building cutting-edge GenAI, we'd love to have you join our team.
🌐 Check out our openings: jobs.lever.co/mistral
#AIJobs #TechCareers #MistralAI
Just FYI, we're hiring AI Scientists and Engineers at Mistral AI.
If you're driven and interested in building cutting-edge GenAI, we'd love to have you join our team.
🌐 Check out our openings: jobs.lever.co/mistral
#AIJobs #TechCareers #MistralAI
Mistral AI jobs
Job openings at Mistral AI
jobs.lever.co
November 6, 2024 at 4:39 PM
Hey all, thanks for the follow!
Just FYI, we're hiring AI Scientists and Engineers at Mistral AI.
If you're driven and interested in building cutting-edge GenAI, we'd love to have you join our team.
🌐 Check out our openings: jobs.lever.co/mistral
#AIJobs #TechCareers #MistralAI
Just FYI, we're hiring AI Scientists and Engineers at Mistral AI.
If you're driven and interested in building cutting-edge GenAI, we'd love to have you join our team.
🌐 Check out our openings: jobs.lever.co/mistral
#AIJobs #TechCareers #MistralAI
Reposted by Diego de las Casas
Tencent's Hunyuan-Large
The largest open-source Transformer-based MoE model with 389 billion parameters, can handle up to 256K tokens. Key features include large-scale synthetic data and a mixed expert routing strategy.
Model: huggingface.co/tencent/Tenc...
Paper: arxiv.org/abs/2411.02265
The largest open-source Transformer-based MoE model with 389 billion parameters, can handle up to 256K tokens. Key features include large-scale synthetic data and a mixed expert routing strategy.
Model: huggingface.co/tencent/Tenc...
Paper: arxiv.org/abs/2411.02265
November 5, 2024 at 7:15 AM
Tencent's Hunyuan-Large
The largest open-source Transformer-based MoE model with 389 billion parameters, can handle up to 256K tokens. Key features include large-scale synthetic data and a mixed expert routing strategy.
Model: huggingface.co/tencent/Tenc...
Paper: arxiv.org/abs/2411.02265
The largest open-source Transformer-based MoE model with 389 billion parameters, can handle up to 256K tokens. Key features include large-scale synthetic data and a mixed expert routing strategy.
Model: huggingface.co/tencent/Tenc...
Paper: arxiv.org/abs/2411.02265
Thinking how much Google struggled to get TPUs to work well makes me skeptical that Nvidia is in any danger of losing dominance, at least for now.
Will Nvidia continue to dominate the AI chip market? Yes and no.
- Yes, Nvidia is likely to maintain dominance in the market for training AI models.
- No, another company (or companies) will take the lead in the market for AI model inference, which is an exponentially larger market.
- Yes, Nvidia is likely to maintain dominance in the market for training AI models.
- No, another company (or companies) will take the lead in the market for AI model inference, which is an exponentially larger market.
November 5, 2024 at 7:49 AM
Thinking how much Google struggled to get TPUs to work well makes me skeptical that Nvidia is in any danger of losing dominance, at least for now.
Reposted by Diego de las Casas
I try to stay up to date with Gen AI videos, but you can do camera control now? Seriously?
Runway introduces advanced camera control for Gen-3 Alpha Turbo. Choose both the direction and intensity of how you move through your scenes for even more intention in every shot.
Runway introduces advanced camera control for Gen-3 Alpha Turbo. Choose both the direction and intensity of how you move through your scenes for even more intention in every shot.
November 1, 2024 at 5:21 PM
I try to stay up to date with Gen AI videos, but you can do camera control now? Seriously?
Runway introduces advanced camera control for Gen-3 Alpha Turbo. Choose both the direction and intensity of how you move through your scenes for even more intention in every shot.
Runway introduces advanced camera control for Gen-3 Alpha Turbo. Choose both the direction and intensity of how you move through your scenes for even more intention in every shot.
Reposted by Diego de las Casas
#Tokenization is undeniably a key player in the success story of #LLMs but we poorly understand why.
I want to highlight progress we made in understanding the role of tokenization, developing the core incidents and mitigating its problems. 🧵👇
I want to highlight progress we made in understanding the role of tokenization, developing the core incidents and mitigating its problems. 🧵👇
October 30, 2024 at 6:26 PM
#Tokenization is undeniably a key player in the success story of #LLMs but we poorly understand why.
I want to highlight progress we made in understanding the role of tokenization, developing the core incidents and mitigating its problems. 🧵👇
I want to highlight progress we made in understanding the role of tokenization, developing the core incidents and mitigating its problems. 🧵👇
Reposted by Diego de las Casas
New starter pack! go.bsky.app/GZ4hZzu
October 28, 2024 at 9:43 AM
New starter pack! go.bsky.app/GZ4hZzu