I posted five share links in the thread—see here:
MACHINE LEARNING
-
Starting First Reinforcement Learning Run in Years
By
–
starting up the first RL run I've done in a while… feels good
-
@testingcatalog — 2025-07-14
By
–
And Mistral Let's see how soon we will start hearing about Chinese mega clusters for training closed models
-
Physics-Informed AI Essential for Scientific Discovery in Chemistry
By
–
I had a wonderful time discussing #AI in chemistry at Lindau Nobel Laureate Panel https://
youtube.com/live/Iv5_CYjCT
NI?feature=shared
… I talked about how physics-informed learning is essential for AI to make new scientific discoveries. Since discovery is about extrapolation, a purely data driven method is -
Anthropic unveils curated MCP connectors directory for Figma, Notion, Stripe, macOS
By
–
BREAKING 🚨: Anthropic released Connectors directory with loads of curated MCPs.
— 🚨 AI News | TestingCatalog (@testingcatalog) 14 juillet 2025
Figma, Notion, Stripe and loads of other connectors are now available, including desktop-specific MCPs for the Claude macOS app. https://t.co/ZxXRobIo1p pic.twitter.com/m9vEzy8jwwBREAKING : Anthropic released Connectors directory with loads of curated MCPs. Figma, Notion, Stripe and loads of other connectors are now available, including desktop-specific MCPs for the Claude macOS app.
-
Reinforcement Learning Field of View Limitations and Policy Optimization
By
–
Yep I think RL is misleading in that it restricts field of view. Eg like you mentioned you can imagine review/reflect doing a lot more – building tools for later use, or actively tuning the distribution for what to try next (instead of just sampling from policy independently as
-
Understanding ML Architecture for Better Model Trade-offs
By
–
But some understanding of architecture is always good to have. It's also part of trading off diff models when you build ml system
-
Grok 4 Heavy refuses to repeat its system prompt
By
–
Unlike Grok 4, Grok 4 Heavy is unwilling to repeat its system prompt. The protections against this are surprisingly robust—even if you trick the model into trying to repeating it, even in encoded form (e.g. base64), some secondary filter catches it and truncates the response.
-
Where Did Validation Sets Go in Modern ML?
By
–
when i first learned Machine Learning, our professor ingrained into us how every ML problem starts by splitting data into train, test, and validation these days there is just train and test. in many cases there is just train and more train where’d all the validation sets go?