Humans are deeply visual creatures. We understand the world by seeing patterns and watching how different pieces of a story come together.
That intuition inspired us to build the Vision Interpretability Toolkit, a visual, hands-on entry point into mechanistic interpretability.
So, we present to you the
InceptionV1 Microscope 🔬
This Hugging Face Space is our attempt to bring back a piece of OpenAI’s Microscope, an important companion to the influential Distill Circuits work by @ch402 and collaborators.
Pick any layer and channel in InceptionV1, aka GoogLeNet, and you’ll see:
🌀 An Activation Maximization image: a visualization of the patterns that strongly activate that neuron.
🖼️ 10 ImageNet examples that activate the selected neuron most strongly, with the ability to zoom into the regions driving those activations.
InceptionV1 Microscope: https://t.co/CJjP4f2HbJ
"Models are the best capitalists, or aligned, never both" I mean there is an obvious conclusion to draw from this, but you can't say it out loud if you work in Silicon Valley.
just read @magikarp_tokens's write up on claude's (honestly dumbfounding) tokenizer. love to see work being done beyond BPE and believing we can do better!
even if their sequences in english are ~40% longer. we don't have to talk about that.
https://t.co/SHygY0Rmg9
Looking forward to reading this latest research by @AnthropicAI, at @Cohere_Labs ML Understanding session.
This work gives us new and incredible insights into the inner workings of LLMs!!
Join us Friday 7/17, 10 AM PT as we deep dive into this work.
These projects came out of Expedition Tiny Aya, a mentor-supported program pairing community researchers and their ideas with @cohere scientists.
Many teams have already released open-source code, datasets, benchmarks, blogs, and papers, with more still on the way.
ML Understanding group at @Cohere_Labs is happy to announce a new initiative,
TIMI - Tiny Intuitions Mighty Insights
📌 The idea is simple: take fundamental machine learning concepts and truly understand them by digging into the mechanics, building intuition, and connecting theory to behavior.
📋 TIMI will be a series of 1-hour, community-driven sessions, every Friday, 10 AM PT where anyone can present and explore core topics like weights, bias, gradients, activation functions, and more. The goal is to make space for deep, collaborative learning!
👀 We’re kicking things off with our first session on
🔹Weights 🔹
- What do they actually represent?
- What structure do they encode?
- How do they shape model behavior?
If you're interested in:
- developing a deeper understanding of ML fundamentals
- contributing to open science
- presenting or learning alongside a curious, motivated community
we’d love to have you join us 👇
https://t.co/fp6l6Ovv2c
Dia 1 pedindo para que os deputados e senadores reconsiderem e voltem atrás no PL da Devastação. Ficarei aqui mesmo sem um like nos posts, até ter noticias boas sobre esse dia triste de hoje.
A cada dia contarei uma historia de problema ambiental no Brasil.
Hoje relembremos Mariana. O rompimento de sua barragem afetou 230 municípios, poluiu a água, destruiu a fauna, possivelmente extinguiu 11 especies e ameçou outras que so viviam la. Tambem matou 19 pessoas e causou um custo de mais de 100 milhões pro Estado Brasileiro.
Esse dinheiro podia ter ido pra programas sociais, saúde, segurança ou educação. Mas não. Foi gasto pra recuperar uma irresponsabilidade de quem buscou lucro rápido sem cuidado com a proteção.
Eu, como fã do @avioesemusicas , pergunto. Se fosse sobre o avião que carrega esses deputados e senadores, não seria um escandalo criar uma lei que faz com que as empresas não precisem cumprir as regras de segurança pros voos?
@PTnaCamara@psolnacamara@analise2025@KriskaCarvalho@lazarorosa25@ricardogalvaosp@marioliveirain
my city has just endured its worst military police operation in its history. over 60 dead. i have friends stranded and with no way to go back home for today.
and it's by design. i live under a state that rules by blood and violence. that's the Right's legacy in rio de janeiro.
it terrifies me how many talented people are contributing to a future they wouldn't want to live in. building something addictive. exploitative. destructive. something they'll ban their own kids from using. and for what? a paycheck? surely there's a better way to use your gifts?
@sarahookr@Cohere_Labs@cohere thank you so much for believing in open science, and giving young researchers like me a space to grow and learn 💙
good luck on all your future endeavors!
i'm grappling with the fact that the area i want to do research on (tokenization) has the goal of ceasing to exist.
like i'm putting hours into something to make it go away, and it will be great if it does, but then what?
good news! @Cohere_Labs ML summer school was a blast!
bad news! it's over :(
i'd like to thank Cohere's Open Science community for always giving us opportunities to grow. this is only the beginning 🚀