Announcing Agent Recipes! A site to learn about agent/workflow recipes with code examples that you can easily copy & paste into your own AI apps. I'm gonna make this the go-to resource for devs to learn about agents & how to implement them – more soon.

Huge releases in AI Agents, RAG and LLMs Microsoft's open-source framework to build AIOps agents No-code visual platform to build RAG agents Open-weight model for visual reasoning Turn any GitHub repo into LLM-suitable text Speech model that generates personalities on-the-fly

Applying AI to genomics, researchers led by @BrianHie have created Evo, a large language model that writes genetic code. It could offer a deeper understanding of evolution, genetic diseases, and new treatments – all on a computer rather than in a lab. engineering.stanford.edu/new…

We've added a Claude-powered prompt improver to the Anthropic Console. Take any prompt, run it through the improver, and get an optimized prompt in return. Here's how it works:

1/n How to follow complex instructions while using RAG. The advancement of Large Language Models (LLMs) has revolutionized natural language processing, but a significant challenge remains in their ability to follow complex instructions while leveraging external knowledge through Retrieval-Augmented Generation (RAG) systems. A recent paper presents VIF-RAG, an innovative framework addressing this crucial limitation in AI systems. The core problem lies in the disconnect between LLMs' ability to follow instructions and their capacity to utilize retrieved information effectively. While models can excel at these tasks independently, they often struggle when these capabilities need to work in tandem. This limitation becomes particularly apparent in real-world applications where users naturally pose questions with specific constraints and requirements. VIF-RAG introduces a novel approach to solving this challenge through an automated, scalable, and verifiable framework. Starting with less than 100 manually crafted atomic instructions, the system employs a sophisticated combination of instruction synthesis, automated rewriting, and thorough verification processes to create a comprehensive dataset of over 100,000 high-quality instruction-query samples. This approach significantly reduces the need for extensive human annotation while maintaining high standards of quality. The framework's innovation extends beyond data generation to evaluation. The introduced FollowRAG benchmark provides the first comprehensive evaluation system for instruction-following capabilities in RAG scenarios, covering 22 different categories of constraints across four knowledge-intensive QA datasets. This evaluation framework's design allows for seamless integration with various RAG benchmarks, making it a versatile tool for future research. What sets VIF-RAG apart from previous work is its holistic approach to the problem. While earlier research focused on either instruction-following or RAG optimization separately, VIF-RAG directly addresses the intersection of these capabilities. Previous solutions like reranker-based methods or multi-round retrieval approaches tackled specific aspects of RAG systems, but none specifically addressed the instruction-following challenge in RAG contexts while maintaining other fundamental capabilities. Experimental results demonstrate the framework's effectiveness. Tests comparing base and fine-tuned models across various tasks show that while models might excel at individual tasks, they often struggle with combined instruction-following in RAG scenarios. VIF-RAG successfully addresses this limitation, showing improved performance across instruction following, RAG tasks, and fundamental abilities. This research represents a significant step forward in making AI systems more practical and user-friendly. By enabling RAG systems to better understand and follow complex instructions while maintaining their knowledge retrieval capabilities, VIF-RAG brings us closer to AI systems that can more naturally and effectively interact with users in real-world scenarios. The framework's scalable and verifiable nature ensures its potential for continued development and application in future AI systems.

In case you were wondering... Zed is proudly built the hard way—from scratch, starting with an empty repo. Our team *created* Electron, but we no longer believe that the ultimate editor can be created by forking a web browser. (Or by forking a fork of a web browser 🙃)

In case you missed it, the Go team wrote an incredible article on how to setup an LLM with RAG server. Also introduced me to GenKit. 🤯 I'm happy because Python is the pits. go.dev/blog/llmpowered

On the limits of agency in agent-based models discuss: huggingface.co/papers/2409.1… Agent-based modeling (ABM) seeks to understand the behavior of complex systems by simulating a collection of agents that act and interact within an environment. Their practical utility requires capturing realistic environment dynamics and adaptive agent behavior while efficiently simulating million-size populations. Recent advancements in large language models (LLMs) present an opportunity to enhance ABMs by using LLMs as agents with further potential to capture adaptive behavior. However, the computational infeasibility of using LLMs for large populations has hindered their widespread adoption. In this paper, we introduce AgentTorch -- a framework that scales ABMs to millions of agents while capturing high-resolution agent behavior using LLMs. We benchmark the utility of LLMs as ABM agents, exploring the trade-off between simulation scale and individual agency. Using the COVID-19 pandemic as a case study, we demonstrate how AgentTorch can simulate 8.4 million agents representing New York City, capturing the impact of isolation and employment behavior on health and economic outcomes. We compare the performance of different agent architectures based on heuristic and LLM agents in predicting disease waves and unemployment rates. Furthermore, we showcase AgentTorch's capabilities for retrospective, counterfactual, and prospective analyses, highlighting how adaptive agent behavior can help overcome the limitations of historical data in policy design. AgentTorch is an open-source project actively being used for policy-making and scientific discovery around the world.