Archive

Supercharge your AI agents with data from the web and beyond. Building the library for superintelligence. 🔥

distilling small language models

Create parallel Codex + Claude Code agents in isolated workspaces. See at a glance what they're working on, then review and merge their changes.

has google collab worksheets.

book going to be published soon. seen on amazon

[MLsys2026 Best Paper]: https://arxiv.org/abs/2506.08276. RAG on Everything with LEANN. Enjoy 97% storage savings while running a fast, accurate, and 100% private RAG application on your personal device.

via bellard.org - ts_zip

We should think in terms of "this is a function to generate sequences" and "by providing prefixes we can steer the sequence generation around in the space of words and change the probabilities for output sequences". And for every possible undesirable output sequence of a length smaller than , we can pick a context that maximizes the probability of this undesirable output sequence.

A much clearer formulation, which helps more clearly articulate the problems to solve.

good explanation of ColBERT

agentic LLM for software engineering tasks. Devstral is built under a collaboration between Mistral AI and All Hands AI 🙌, and outperforms all open-source models on SWE-Bench Verified by a large margin. We release Devstral under the Apache 2.0 license.

Browse and discover MCP use cases, servers, clients, and news
The Model Context Protocol (MCP) ecosystem evolves daily. Stay current here.

Find Awesome MCP Servers and Clients

The largest collection of MCP Servers.

"MCP is an open protocol that standardizes how applications provide context to LLMs. Think of MCP like a USB-C port for AI applications. Just as USB-C provides a standardized way to connect your devices to various peripherals and accessories, MCP provides a standardized way to connect AI models to different data sources and tools."
― Anthropic

BAML is a domain-specific language to generate structured outputs from LLMs — with the best developer experience.

A Lightweight Framework for Studying Learning Dynamics
Pico demystifies how language models learn. Lightweight and efficient, it's your go-to toolkit for training and analyzing models across different scales.

Built around two core libraries: pico-train for model training and pico-analyze for in-depth analysis—designed from the ground up to work seamlessly together.

Chip Huyen's 8,000 word practical guide to building useful LLM-driven workflows that take advantage of tools.

This article is adapted from "AI Engineering" book, which I bought in Feb'25.

Book outline:

• The components of a RAG system
• How to create a RAG knowledge base
• The indexing and generation pipeline
• Evaluating a RAG system
• Advanced RAG strategies
• RAG tools, technologies, and frameworks

A native macOS app that allows users to chat with a local LLM that can respond with information from files, folders and websites on your Mac without installing any other software. Powered by llama.cpp.

Textbook, second edition

DeepSeek is pushing DuckDB beyond its single-node roots with smallpond, a new, simple approach to distributed compute. But does it solve the scalability challenge—or introduce new trade-offs?

good blog on LLM, Machine Learning etc

shift a substantial portion of LLM workloads to consumer devices by having small on-device models collaborate with frontier models in the cloud. By only reading long contexts locally, we reduce cloud costs with minimal or no quality degradation. We imagine a future where an “intelligence layer” running persistently on-device interacts with frontier models in the cloud to deliver applications with cost-effective, “always-on” intelligence.

Reseach by Stanford's Hazy Research.

• Github: https://github.com/HazyResearch/minions
• arXiv Paper: https://arxiv.org/abs/2502.15964

ollama also has a related blog post

Cites this work: "LLM×MapReduce: Simplified Long-Sequence Processing using Large Language Models"

The proposed LLM×MapReduce framework splits the entire document into several chunks for LLMs to read and then aggregates the intermediate answers to produce the final output. The main challenge for divide-and-conquer long text processing frameworks lies in the risk of losing essential long-range information when splitting the document, which can lead the model to produce incomplete or incorrect answers based on the segmented texts. Disrupted long-range information can be classified into two categories: inter-chunk dependency and inter-chunk conflict. We design a structured information protocol to better cope with inter-chunk dependency and an in-context confidence calibration mechanism to resolve inter-chunk conflicts.

consistent with an opensource mindset, deepseek are curating a list of "awesome" stuff built with/on deepseek.

This list not only serves as a reminder of what can be done with deepseek, but also a general overview of all things possible with rest of llms too.

I have some suggest windsurf as an alternative for Cursor, just days after I paid for annual subscription.. sigh..

In either case, there is nothing stopping me from really getting into Vibe Coding.

Moar coding!1!

Anthropic announced new version of Claude today, and it has been received very positively going by the reactions on X. Simon Willison ran it through his usual gauntlet.

The new thing is the claude CLI tool that allows you run commands against Claude api from the command line, which also would be a big aid for programming tasks.

I had unsubscribed from Claude for the last couple of months, and looks like it's time to get back on it.

Stanford's Sky Lab announced the "Sky-T1" model that performs competitively in reasoning and coding while comparison to OpenAI's o1 model, while using a small compute budget.

They set about to produce this open-weights model because the current leaders like o1 and gemini 2.0 flash thinking while excel in reasoning, showing the ability to solve complex tasks by producing a long internal chain of thought, among other advancements. However, the technical details and model weights are un-accessible.

Summary:

• They used QwQ to generate training data (with some cleanup using GPT-4o-mini)
• The training data was then used to FT Qwen2.5-32B-Instruct (non-reasoning model)
• Result was that Sky-T1 performs slightly worse than QwQ but much better than Qwen2.5 on reasoning tasks.

This is pretty interesting as it shows how you can FT a foundation model to do better at reasoning.

Inference time compute is still underutilized for AI deployments. It is better to distill the reasoning from larger models like R1 for specific tasks. Even better, one can mix in custom thinking instructions for specific sub-problems, so that a fine-tuned model learns o mix of task specific reasoning and custom logic. This can often beat custom prompt iteration. 🔗

This Guide: Train a Reasoning Model | Kiln AI Docs goes into how to distill for thinking models.