Perfectly Awesome

Part 2 - Security

Part 3 - Operations

MCP, short for Model Context Protocol, is the hot new standard behind how Large Language Models (LLMs) like Claude, GPT, or Cursor integrate with tools and data. It’s been described as the “USB-C for…

A Blog post by Ksenia Se on Hugging Face

We’re introducing Llama 4 Scout and Llama 4 Maverick, the first open-weight natively multimodal models with unprecedented context support and our first built using a mixture-of-experts (MoE) architecture.

Overview of the Model Context Protocol (MCP) how it works, what are MCP servers and clients, and how to use it.

Learn how to configure and use Model Context Protocol (MCP) servers with GitHub Copilot in Visual Studio Code.

The brother goes on vision quests. The sister is a former English major. Together, they defected from OpenAI, started Anthropic, and built (they say) AI’s most upstanding citizen, Claude.

The past few years have witnessed the rise in popularity of generative AI and large language models (LLMs), as part of a broad AI revolution.

Deploying LLMs presents challenges, particularly in optimizing efficiency, managing computational costs, and ensuring high-quality performance. LLM routing has emerged as a strategic solution to these challenges, enabling intelligent task allocation to the most suitable models or tools. Let’s delve into the intricacies of LLM routing, explore various tools and frameworks designed for its implementation, and […]

Thanks to KiwiCo for sponsoring today’s video! Go to https://www.kiwico.com/welchlabs and use code WELCHLABS for 50% off your first monthly club crate or for...

In a follow-up to the research that brought us the [delightful Golden Gate Claude](https://simonwillison.net/2024/May/24/golden-gate-claude/) last year, Anthropic have published two new papers about LLM interpretability: - [Circuit Tracing: Revealing Computational …

What they found challenges some basic assumptions about how this technology really works.

In this article, we explore 10 of the Python libraries every developer should know in 2025.

At OpenAI, we have long believed image generation should be a primary capability of our language models. That’s why we’ve built our most advanced image generator yet into GPT‑4o. The result—image generation that is not only beautiful, but useful.

The Hallucination Index is a benchmark that measures the frequency of inaccuracies in large language models, indicating their reliability and contextual understanding.

[platform_url]//console.mistral.ai/

Model architectures, data generation, training paradigms, and unified frameworks inspired by LLMs.

Anthropic launches real-time web search for Claude AI, challenging ChatGPT's dominance while securing $3.5 billion in funding at a $61.5 billion valuation.

Paris-based artificial intelligence startup Mistral AI has announced the open-source release of its lightweight AI model, Mistral Small 3.1, which the company

Mistral Small 3 [came out in January](https://simonwillison.net/2025/Jan/30/mistral-small-3/) and was a notable, genuinely excellent local model that used an Apache 2.0 license. Mistral Small 3.1 offers a significant improvement: it's multi-modal …

The ellmer package for using LLMs with R is a game changer for scientists Why is ellmer a game changer for scientists? In this tutorial we’ll look at how we can access LLM agents through API calls. We’ll use this skill for created structued data fro...

Catastrophic Forgetting is a phenomenon where neural networks lose previously learned information when trained on new data, similar to human memory loss.

These models are free to use, can be fine-tuned, and offer enhanced privacy and security since they can run directly on your machine, and match the performance of proprietary solutions like o3-min and Gemini 2.0.

Model cards are documentation tools in machine learning that provide essential information about models, promoting transparency, trust, and ethical considerations in AI systems.

Plus CSS view transitions and a major update to llm-openrouter

It’s happening.

Part 1: Inference-Time Compute Scaling Methods

New closed-source specialist OCR model by Mistral - you can feed it images or a PDF and it produces Markdown with optional embedded images. It's available [via their API](https://docs.mistral.ai/api/#tag/ocr), or …

Introducing the world’s best document understanding API.

This release of the `llm-ollama` plugin adds support for [schemas](https://simonwillison.net/2025/Feb/28/llm-schemas/), thanks to a [PR by Adam Compton](https://github.com/taketwo/llm-ollama/pull/36). Ollama provides very robust support for this pattern thanks to their [structured outputs](https://ollama.com/blog/structured-outputs) …

Today, we’re announcing Claude 3.7 Sonnet, our most intelligent model to date and the first hybrid reasoning model generally available on the market.

OpenAI’s Deep Research is built for me, and I can’t use it. It’s another amazing demo, until it breaks. But it breaks in really interesting ways.

Solid techniques to get really good results from any LLM

OpenAI's president Greg Brockman recently shared this cool template for prompting their reasoning models o1/o3. Turns out, this is great for ANY reasoning… | 32 comments on LinkedIn

Comparison and ranking the performance of over 30 AI models (LLMs) across key metrics including quality, price, performance and speed (output speed - tokens per second & latency - TTFT), context window & others.

I share my preferences for LLMs, image models, AI video, AI music, AI-powered research, and more. These are the AI tools I regularly use or recommend to others.

A Step-by-Step Guide to Setting Up a Custom BPE Tokenizer with Tiktoken for Advanced NLP Applications in Python

Do my tears surprise you? Strong CEOs also cry.

I just released llm-smollm2, a new plugin for LLM that bundles a quantized copy of the SmolLM2-135M-Instruct LLM inside of the Python package. This means you can now pip install …

In this article, I will describe the four main approaches to building reasoning models, or how we can enhance LLMs with reasoning capabilities. I hope this p...

Check out this comparison of 5 AI frameworks to determine which you should choose.

The Reinforcement Learning from Human Feedback Book

In our previous tutorial, we built an AI agent capable of answering queries by surfing the web. However, when building agents for longer-running tasks, two critical concepts come into play: persistence and streaming. Persistence allows you to save the state of an agent at any given point, enabling you to resume from that state in future interactions. This is crucial for long-running applications. On the other hand, streaming lets you emit real-time signals about what the agent is doing at any moment, providing transparency and control over its actions. In this tutorial, we’ll enhance our agent by adding these powerful

Aidan Bench attempts to measure in LLMs. - aidanmclaughlin/AidanBench

o3-mini is out today. As with other o-series models it’s a slightly difficult one to evaluate—we now need to decide if a prompt is best run using GPT-4o, o1, o3-mini …

How a Key-Value (KV) cache reduces Transformer inference time by trading memory for computation

The field of artificial intelligence is evolving rapidly, with increasing efforts to develop more capable and efficient language models. However, scaling these models comes with challenges, particularly regarding computational resources and the complexity of training. The research community is still exploring best practices for scaling extremely large models, whether they use a dense or Mixture-of-Experts (MoE) architecture. Until recently, many details about this process were not widely shared, making it difficult to refine and improve large-scale AI systems. Qwen AI aims to address these challenges with Qwen2.5-Max, a large MoE model pretrained on over 20 trillion tokens and further refined

The unusual timing of the Qwen 2.5-Max's release points to the pressure DeepSeek's meteoric rise in the past three weeks has placed on overseas rivals and domestic competition.

A recipe for reasoning LLMs

AI has entered an era of the rise of competitive and groundbreaking large language models and multimodal models. The development has two sides, one with open source and the other being propriety models. DeepSeek-R1, an open-source AI model developed by DeepSeek-AI, a Chinese research company, exemplifies this trend. Its emergence has challenged the dominance of proprietary models such as OpenAI’s o1, sparking discussions on cost efficiency, open-source innovation, and global technological leadership in AI. Let’s delve into the development, capabilities, and implications of DeepSeek-R1 while comparing it with OpenAI’s o1 system, considering the contributions of both spaces. DeepSeek-R1 DeepSeek-R1 is

Developers have tricks to stop artificial intelligence from making things up, but large language models are still struggling to tell the truth, the whole truth and nothing but the truth.

This article covers 12 influential AI research papers of 2024, ranging from mixture-of-experts models to new LLM scaling laws for precision..

New release of my [LLM](https://llm.datasette.io/) CLI tool and Python library. A bunch of accumulated fixes and features since the start of December, most notably: - Support for OpenAI's [o1 model](https://platform.openai.com/docs/models#o1) …

The company built a cheaper, competitive chatbot with fewer high-end computer chips than U.S. behemoths like Google and OpenAI, showing the limits of chip export control.

DeepSeek are the Chinese AI lab who dropped the best currently available open weights LLM on Christmas day, DeepSeek v3. That model was trained in part using their unreleased R1 …

The rapid advancement and widespread adoption of generative AI systems across various domains have increased the critical importance of AI red teaming for evaluating technology safety and security. While AI red teaming aims to evaluate end-to-end systems by simulating real-world attacks, current methodologies face significant challenges in effectiveness and implementation. The complexity of modern AI systems, with their expanding capabilities across multiple modalities including vision and audio, has created an unprecedented array of potential vulnerabilities and attack vectors. Moreover, integrating agentic systems that grant AI models higher privileges and access to external tools has substantially increased the attack surface and

New paper from Microsoft describing their top eight lessons learned red teaming (deliberately seeking security vulnerabilities in) 100 different generative AI models and products over the past few years. …