AI always moves faster than you expect
@mattshumer_
-
Solving Long-Context LLaMA 3 by End of Day
By
–
By the end of today, I think we will have solved long-context LLaMA 3.
-
Fast Generation in Large Models with Sparse Parameter Activation
By
–
Not sure how it will play out, but it means that generations will be very fast for such a big model, as only a small portion of params will be active at a time.
-
Snowflake’s New Development: Enterprise AI Adoption Challenges
By
–
I’m not sure most companies have a use for this (we certainly won’t use it at HyperWrite, due to the sheer size / performance ratio). But still a really cool development for the Snowflake team!
-
Snowflake Releases Largest Open-Source Model 482B Parameters
By
–
Snowflake just launched the largest open-source model yet. A 482B parameter MoE. 17B active parameters and 128 experts, trained on 3.5T tokens. This model is more open than others — even the data recipe is open-source!
-
Hidden Reasoning in AI Interfaces Through Tag Parsing
By
–
If you control the interface, you can even have it reason within tags and then parse them out so the user doesn’t see.
-
Phi Models Performance Issues and User Experience Concerns
By
–
I've never had good results from Phi models. Something always feels really off/broken…
-
Compute Sponsorship Offers Pour Projet IA Non Annoncé
By
–
I threw this out there as a half-joke. Now I'm getting dozens of DMs with generous people + companies offering to sponsor the compute. Really appreciate all the support! I have to think about the best way to do this (if at all). Will report back later today.
-
Understanding AI Models Through Conversation and Testing
By
–
I usually just start by saying “hi” to the model, and go from there. It takes time to really understand how well a model works.
-
Extended Base Model Development and Implementation
By
–
For clarity, I extended the base model — not the instruct version, so this is expected.