←
Nando de Freitas12,914 posts
Nando de Freitas avatar

Nando de Freitas

@NandoDF

I seek to understand intelligence & agency and build AI aligned with compassion, freedom & universal human empowerment through progress in science & engineering

◷ Joined April 2009▧ 453 Photos and videos
Xiaoliang Dai avatar
⟳ Nando de Freitas retweeted
Xiaoliang Dai@xiaoliang_dai·21h

The first T2I model that is comparable to the GPT-Image family, and the best model for image editing. Great work from an amazing team! Say hello to MAI-Image-2.7!

Microsoft AI has launched MAI-Image-2.7, the new #1 in AA-Image-Editing v2.1 at one sixth the price of the four models behind it MAI-Image-2.7 and MAI-Image-2.7-Flash both take new places on the Quality vs Price frontier in Image Editing and Text to Image, at $34.82 and $10.24 per 1k images. Both perform best in Image Editing: MAI-Image-2.7 ranks #1 for Text or Symbol, Scene & Style and Enhancement & Restoration edits, and MAI-Image-2.7-Flash ranks #3 for Scene & Style edits. Congratulations to @MicrosoftAI on the launch of MAI-Image-2.7 and MAI-Image-2.7-Flash! See below for our detailed analysis of both models in Image Editing and Text to Image 🧵

Brandon Amos avatar
⟳ Nando de Freitas retweeted
Brandon Amos@brandondamos·Nov 16

⚠️ Numerical issues can go on silently for decades like in the softmax/CE! Similar issues are present in other widely-used divergences like the KL and JSD I just posted this paper overviewing what they are and how to fix them: 📚 arxiv.org/abs/2611.18511

Post attachment

Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:

0:07
James Wise avatar
⟳ Nando de Freitas retweeted
James Wise@Jameswise·Nov 17

While they’ll be plenty of coverage of Anthropic this week - I’d argue that @IsomorphicLabs is the most important private company in the world right now. And it’s now the most valuable AI company focussed purely on improving human health, from Kings Cross to the world. Congratulations @demishassabis, @maxjaderberg and team. Delighted @UKSovereignAI is able to be a small part of this effort. When we set the Unit up under 12 months ago, seeing companies like this scale in the UK was our 5 year stretch goal. Seems we are ahead of schedule….

We’ve closed our $5 Billion funding round co-led by Greenoaks and Sequoia Capital to build the future of medicine and bioengineering with AI. isomorphiclabs.com/articles/…

0:37
Morteza Mardani avatar
⟳ Nando de Freitas retweeted
Morteza Mardani@MardaniMorteza·Nov 16

📣 NVIDIA is releasing a collection of open model weights: Nemotron-Labs-Sigma — 14 checkpoints at 3B/8B. Language models have stayed discrete. We bring continuous diffusion to language as in vision. Smooth ODE trajectories (16-D) unlock distillation and steering: 8 TPFs for math, 4 TPFs for code, with accuracy and diversity preserved. 🧵 1/

Post attachment
Nando de Freitas avatar
Nando de Freitas@NandoDF·Nov 17

A Bayesian view: AI killing us, just like other human beliefs, cannot be proved or disproved. It’s all a matter of subjective probability and expected risk. Which events are more probable? What is their risk?

Totally agree. There are many risks, but AI deciding to kill us all is BS, instead humans will use AI to do bad things (and good things too), just as with other tech (only faster).

Nando de Freitas avatar
Nando de Freitas@NandoDF·Nov 11

It can be catastrophic to trust models trained by someone else. Openness and scrutiny are essential for safe AI. This study makes the point very clear as @JohnCLangford highlights. See also x.com/nandodf/status/2111512… for a causality and agency perspective.

I've been waiting for this one to come out. In essence, people should maintain custody of a model from creation, just like they do with binaries. Without that, the model can be setup to exhibit arbitrary behaviors in nigh-undetectable ways.

John Langford avatar
⟳ Nando de Freitas retweeted
John Langford@JohnCLangford·Nov 10

I've been waiting for this one to come out. In essence, people should maintain custody of a model from creation, just like they do with binaries. Without that, the model can be setup to exhibit arbitrary behaviors in nigh-undetectable ways.

Dimitris Papailiopoulos @DimitrisPapail Nov 10 Article Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents We find that weight reprogramming fewer than 10 dormant neurons in an LLM is enough to backdoor any agent built on it, with no training required. The model passes evals and behaves normally until a

Nando de Freitas avatar
Nando de Freitas@NandoDF·Nov 11

A good image to keep in mind for anyone working on AI — our number 1 goal should be compassion

The psychological hallmark of empathy is perspective-taking, the capacity to imagine or take the perspective of another person. When you are empathizing, you imagine how someone may feel even though you can’t experience their exact emotion.

Post attachment
Fab avatar
⟳ Nando de Freitas retweeted
Fab@fabrice_mayrand·Oct 25

Average discount on cheaper inference is 40.5% and all frontier models are available Try it: CheaperInference.com

Post attachment

We combine customer demand to negotiate bulk discounts with AI labs and data centers. 500+ models available (including all frontier models) ZDR available Try us: CheaperInference.com

0:12
rose avatar
⟳ Nando de Freitas retweeted
rose@rosenacry·Nov 10

This is f*cking insane. Former Google CEO Eric Schmidt, the man who ran Google for 10 years: "If you really want to make money, it's actually easy, found an agentic AI company, I mean build an agent to do something If you build a better agent than anyone else, you can probably build one incredible company and own it yourself" In 27 seconds, he explains why the next year or two is the agentic period, and why the window won't stay open Prompts → Claude Code → Memory → Skills → MCP → A Company You Own Yourself Save this before it disappears from your feed Everything below is free: Agent architecture langchain.com/blog?category_… Claude Code 101 anthropic.skilljar.com/claud… Claude Code in action anthropic.skilljar.com/claud… Prompt engineering platform.claude.com/docs/en/… Interactive prompt course github.com/anthropics/course… Claude.md + memory code.claude.com/docs/en/clau… Skills code.claude.com/docs/en/skil… MCP code.claude.com/docs/en/mcp Routines code.claude.com/docs/en/rout… Claude Code ultimate guide github.com/FlorianBruniaux/c… Awesome Claude Code github.com/hesreallyhim/awes… Anthropic academy anthropic.skilljar.com Official Claude Code docs code.claude.com/docs/en/over… All of this costs $0 95% will spend the agentic period scrolling AI threads, the 5% who ship agents will own the companies that come out of it Bookmark it and start with the first link tonight Then read how to build an agent that actually does the job in the harness engineering guide below ↓

0:27

Your agent passed every test The issue is still open A better prompt won't fix it, because the model isn't the broken part I broke down the harness, loop and graph around it: what each one owns, how they nest, and which one to fix first ↓

Yoshua Bengio avatar
⟳ Nando de Freitas retweeted
Yoshua Bengio@Yoshua_Bengio·Nov 10

Je salue le leadership du président @EmmanuelMacron qui a rassemblé aujourd’hui des chefs de gouvernement, des dirigeants des grandes entreprises d’IA et des chercheurs internationaux pour discuter des risques croissants liés à l’IA avancée, incluant la perte de contrôle et l’usage dangereux des IAs. Nous devons commencer à traiter l’IA comme l’ensemble des technologies susceptibles de provoquer des transformations et des menaces majeures pour nos sociétés : avec des cadres de gouvernance intentionnels et rigoureux. Les développeurs de modèles doivent être audités de manière obligatoire et indépendante par des évaluateurs qu’ils ne choisissent pas. Ils doivent aussi être tenus responsables de démontrer que leurs modèles sont sécuritaires avant qu’ils ne posent un danger grave, danger qui peut exister avant la phase de déploiement. --- I welcome the leadership of President @EmmanuelMacron, who brought together heads of government, leaders of frontier AI companies, and international researchers today to discuss the growing risks of advanced AI, including dangerous loss of control and misuse of AIs. We must begin treating AI like other technologies which can deeply transform and threaten our societies: with intentional and rigorous governance frameworks. Model developers should undergo mandatory, independent audits by evaluators they do not choose, and be held accountable for demonstrating that their models are safe before they can pose a grave danger, a danger which can exist before their deployment.

L’intelligence artificielle avance très vite. On l’a déjà vu : cela implique des risques réels pour notre sécurité collective. Notre travail pour developper nos technologies doit aussi être celui de l’anticipation et de la maîtrise de ses risques. J’ai réuni à Paris et en visio une cinquantaine d’acteurs : chefs d’État et de gouvernement, dirigeants des plus grands laboratoires d’IA, chercheurs et experts. Le constat est partagé. Il faut agir ensemble. Nous avons obtenu des engagements de tous autour de trois priorités : mieux évaluer de façon indépendante et publique les risques des modèles les plus avancés, mieux partager les alertes en cas d’incident grave et responsabiliser les acteurs qui développent des modèles. Notre ambition est claire : faire de l’IA un progrès pour tous, avec une sécurité garantie par l’engagement de chacun.

Post attachment
Floor Eijkelboom avatar
⟳ Nando de Freitas retweeted
Floor Eijkelboom@FEijkelboom·Nov 9

Our survey on continuous diffusion and flow language models is out! ❤️ A little animated tour of what’s inside 👇 From representations and training to few-step generation, scaling, and open problems. continuous-lm.github.io

New survey on Continuous Diffusion and Flow Language Models is out! 🌊💬 One of my first papers, back in early 2025, was on continuous simplex flows. After working on discrete diffusion, I wanted to try controllable sequence generation with flow matching, which naturally operates in continuous spaces. But at the time, I found them much harder to train well than discrete diffusion. 🥲 2026 has changed this picture completely! 🚀 It’s been super exciting to watch continuous LMs take off, but with the rapid advancements, it’s also becoming harder to keep up with all the new work across pre-training, few-step generation, guidance, and scaling 🥹 So together with @chandavidlee, @FEijkelboom, and @jw9730, we put together a survey that consolidates the continuous LM literature and its key insights into 37 pages. It covers everything from the early work before the so-called “winter” of continuous LMs to the wave of new papers in 2026 🌟 More in thread! 🧵

Post attachment
Sander Dieleman avatar
⟳ Nando de Freitas retweeted
Sander Dieleman@sedielem·Nov 9

This looks incredibly comprehensive🤯, I will need to set aside some time to dive in and explore 👀 Beautiful piece of work to (almost) wrap up the comeback year for CDLMs!

New survey on Continuous Diffusion and Flow Language Models is out! 🌊💬 One of my first papers, back in early 2025, was on continuous simplex flows. After working on discrete diffusion, I wanted to try controllable sequence generation with flow matching, which naturally operates in continuous spaces. But at the time, I found them much harder to train well than discrete diffusion. 🥲 2026 has changed this picture completely! 🚀 It’s been super exciting to watch continuous LMs take off, but with the rapid advancements, it’s also becoming harder to keep up with all the new work across pre-training, few-step generation, guidance, and scaling 🥹 So together with @chandavidlee, @FEijkelboom, and @jw9730, we put together a survey that consolidates the continuous LM literature and its key insights into 37 pages. It covers everything from the early work before the so-called “winter” of continuous LMs to the wave of new papers in 2026 🌟 More in thread! 🧵

Post attachment
levi avatar
⟳ Nando de Freitas retweeted
levi@levidiamode·Nov 8

Day 309/365 of GPU Programming Still going through the assignment today. Have been really enjoying this one since it's a nice combination of concepts/techniques that were taught throughout the lectures of this course (though lecture 11 on TinyEngine and Parallel Processing is probably the most relevant).

Post attachment

Day 308/365 of GPU Programming Taking a look at assignment/lab 5 of the efficient AI computing course today. This one's about optimizing LLMs on edge devices (specifically, LLaMA2-7B-chat) using loop unrolling, multithreading and SIMD programming. Should be a fun exercise and a nice overview of local model optimization.

Post attachment