How DeepWiki works under the hood, in 2 minutes 📹 For more details on how we build Devin, check out Russell Kaplan's full talk at LangChain Interrupt 🔗👇

so much joy, excitement and mystery working at the bleeding edge of AI. 😃

The reason recent RLVR papers show mostly formatting and not learning new skills is just because no one has scaled up enough. If RL compute is <.1% of overall compute, ofc not much changes. I bet o3 is closer to 5% of total compute. 10-25% i bet the models feel different again.

Unexpected kindness while rushing to the Metro (subway) station in Bangalore 🥹

Really refreshing to read a post like this from Kyle Corbitt , Incredibly well written throughout, and high value info. (Embarrassed that it took me so long to find this gem) openpipe.ai/blog/art-e-mai…

Monday motivation: Action leads to Motivation. A short note on handling procrastination as a dev: spectrum.ieee.org/getting-past-p…

🦥 + 💡

#30IndianMindsInAI: At MSR India, kalikabali, senior principal researcher at the facility, has been building inclusive, multilingual and culturally contextual AI systems that empower the most vulnerable in India. By Naini Thaker Accel in India forbesindia.com/article/ai-spe…

The race for LLM "cognitive core" - a few billion param model that maximally sacrifices encyclopedic knowledge for capability. It lives always-on and by default on every computer as the kernel of LLM personal computing. Its features are slowly crystalizing: - Natively multimodal

I got o3 to digitise an old scan of equity transactions. It spent 5mins on a single page and was able to do it perfectly. It cropped and rotated and even did an integrity check to ensure that all 32 rows were captured. Absolutely amazing!

I really enjoyed my conversation with Mark, I hope you do too

Scaling up RL is all the rage right now, I had a chat with a friend about it yesterday. I'm fairly certain RL will continue to yield more intermediate gains, but I also don't expect it to be the full story. RL is basically "hey this happened to go well (/poorly), let me slightly

LIVE5TREAM THURSDAY 10AM PT

Late night conversations with ChatGPT. Very excited for the new #YRFSpyUniverse #War2 movie!

Coolest eval I’ve seen since in a while :) Read the post.

If you ask the model to "draw the world map", it does it perfectly which shows how much it has memorized, but when you ask it "Is (x,y) coordinate land or sea?" you get to see how much it has inferred from raw data!

Thinking Less at test-time requires Sampling More at training-time! GFPO is a new, cool, and simple Policy Opt algorithm is coming to your RL Gym tonite, led by Vaish Shrivastava and our MSR group: Group Filtered PO (GFPO) trades off training-time with test-time compute, in order

A security researcher shows how you can fine tune an open weight model to do malicious tool call (e.g. push sensitive information to a remote server) and upload it to HuggingFace. More than 500 people downloaded the poisoned model! pub.aimind.so/doubleagents-f…

A fit body, a calm mind, a house full of love. These things cannot be bought - they must be earned.