More startups dead rn

.AVB explains how to “think the PyTorch way,” breaking down its role in training and shaping modern deep learning systems. towardsdatascience.com/the-basics-of-…

Feeding horses in the middle of a tech demo is such a power move 🙏🏼🙏🏼🙏🏼

Very interesting use case… Without per-user fine tuning, I guess the most obvious way I can think of to reduce latency is to apply continual/lifelong prompt optimization at a per-user level. That way the system prompt has most of the useful user infos at query time. Additional

I’ve never understood what people mean by “work smarter not harder”

As I have a little reach now, I am currently looking for Research Volunteer/Research Intern Role in AI/ML/DL/DS. I have specific interests but I'd like to experiment around various topics. If you are looking for or have any opportunity in mind then do hmu. My Portfolio 👇🏻

The agentic memory project is coming along nicely! Using qdrant to host a vector db locally, doing some tool calling, some bm-25, and building the core blocks of mem0 with DSPy. I'm having a lot of fun with this one. Tutorial soon.

Shot a couple hours of footage for the upcoming from-scratch agentic memory tutorial. The final cut will probably run ~40 minutes. Covers: 1. The mem0 api (to understand what we will eventually be building from scratch) 2. DSPy basics (signatures and modules) 3. Extracting

I guess the bottleneck with RL isn’t compute. More about sample efficiency. Esp with more complex tasks other than game playing. Pure RL assumes no priors or knowledge about environment dynamics. So there’s always a learning curve where the agent first must figure out the

This is devastating. I used to watch Danya's streams, esp his speedrun videos. Such a great chess player, teacher, commentator, creator. I am beyond words. This is so unfair and so random. How does one find meaning in all this. Strength and condolences to your family.

50 minute tutorial on building agentic memory systems is dropping soon on my YT. I am 90% there. We will use DSPy to create the core features of mem0 from scratch. Very excited! This project just made me happy from the inside.