the hidden complexity of video search: why multimodal retrieval isn't just "long text" i just watched rajan agarwal from kino ai break down how they're solving one of ai's trickiest problems: searching through video content effectively. most retrieval work focuses on text and

Generalists are useful, but it’s not enough to be smart. Advances come from specialists, whether human or machine. To have an edge, agents need specific expertise, within specific companies, built on models trained on specific data. We call this Specific Intelligence. It's

cursor has won coding their data flywheel has turned into the most amazing coding model in the world. Composer-1 truly feels like another Claude 3.5 Sonnet sized jump in capability

Kimi Linear Tech Report is dropped! 🚀 huggingface.co/moonshotai/Kim… Kimi Linear: A novel architecture that outperforms full attention with faster speeds and better performance—ready to serve as a drop-in replacement for full attention, featuring our open-sourced KDA kernels! Kimi

hell yeah

i've only just started going through this but i can already say with confidence that this is an incredible resource for people of all skill levels to learn about LMs

It was a magical ride with all of you. Our city, our home, our fans showed up on the grand stage...and we couldn't be prouder.

potentially naive question: why dont we train RL on summarized states so we can have significantly longer convos/not violate MDP? also if the summarizer drops info that later matters, we can penalize it if we want llms to "think for 1-2 months" then we have a context problem

Wow, language models can talk without words. A new framework, Cache-to-Cache (C2C), lets multiple LLMs communicate directly through their KV-caches instead of text, transferring deep semantics without token-by-token generation. It fuses cache representations via a neural

i strongly believe tab is the ideal interface for coding, writing and even maybe browser use learning and accelerating your patterns is how u become faster. i find myself rewriting so much code with agent tbh

I built Tacc — a Tokenization-Aware Compression Codec that efficiently sends LLM outputs and tool calls over low-bandwidth networks. It compresses faster and smaller than gzip, making it a better choice for serving LLM responses over HTTP. Here’s how it works 🧵 (1/7)

there is no reason that anything a llm reads or produces that doesn’t get shown on an interface has to be human readable save tokens and let llms learn their own compression

Windsurf Paul Graham Full breakdown: cognition.ai/blog/codemaps We all need more AI that turns your brain ON, not OFF. Software development only becomes engineering with *understanding*. Your ability to reason through your most challenging coding tasks is constrained by your mental model of how

After two years of work, we’ve made an AI Scientist that runs for days and makes genuine discoveries. Working with external collaborators, we report seven externally validated discoveries across multiple fields. It is available right now for anyone to use. 1/5

Introducing Nested Learning: A new ML paradigm for continual learning that views models as nested optimization problems to enhance long context processing. Our proof-of-concept model, Hope, shows improved performance in language modeling. Learn more: goo.gle/47LJrzI

I’ve been using papiers for the past few weeks, it’s super great! im a huge fan of Cognition deepwiki and I hope this becomes the standard for papers

marble has grown so much since i tested it ~2 months ago! persistent & stunning world models are here