Greg Kostello

Everyone who wants to run should run. Let the voters decide. Letting the Democratic machine handle things is how we got Biden.

Really appreciate hearing you tonight and also hearing from Nick Brown @nickbrownnow.bsky.social. Please keep up the good fight for our rights and working on protecting the constitution. Thank you for being on BlueSky!

For AI and SWE nerds (like myself): I developed an AI agency to help me develop production code. Here is my commit after three days:
556 files changes. Unit tests, integration tests, service tests all using both traditional software engineering methodology and agent based approaches. #AI #agents

Here today. Very interesting to hear about agentic approaches from Bret Taylor and his perspective on OpenAI and their mission driven approach to benefit humanity. Thank you for creating this summit!

At the Axios AI conference, listening to the chairman of OpenAI, Brett Taylor and his perspective on their mission. Agencies are the now and the future.

Van Jones is the most honest man in politics A conversation you are going to want to hear.

One of the best, most honest post-mortems on what the hell happened with the election. How did Trump win? It's because we've been thinking about it all wrong. Van Jones and Chris Cillizza tear it down. open.substack.com/pub/chriscil...

Meanwhile, a "grader agent" evaluates each scenario's output, practicality, and costs.
Each of these agents acts independently of each other, with goals and objectives clearly defined.

Another SME agent, based on the customer objectives, can be trained on decision-making criteria. The source-grounded knowledge bases are then used to perform "what if" scenarios, again with scoring.

To ensure these source-grounded knowledge bases have high-quality entries, another agent could evaluate each piece of data prior to being entered based on well-defined criteria and score it accordingly so that when retrieved, a score is used along with the reasoning for the score.

One challenge of plain RAG is getting the system to explain the answer, but an agentic RAG system composed of different agents can be tasked with decomposing the problem downn for higher accuracy and generating source-grounded knowledge bases. Each agent runs continuously to keep data evergreen.

We've also used agentic RAG to prepare data from multiple data streams and respond to questions with outstanding results. Performance is still lagging, especially with multiple data sources grow, but I believe that will improve if we leverage other techniques.

3. "Self-Reflective AI" - Agentic RAG adapts responses in real-time, refining accuracy and reducing the need for human intervention.
4. "Built for Complexity" - Its architecture enables handling of intricate, multi-layered problems, such as synthesizing insights from multiple sources.

1. "Dynamic Orchestration"—specialized, collaborative agents improve how complex queries are handled, distributing tasks for precision and scalability.
2. "Query Criticism in Action" - Multi-hop reasoning and iterative query refinement provide relevance and accuracy.

Agentic RAG: A Complete Guide to Agent-Based Retrieval Augmented Generation – News from generation RAG Understanding Agentic RAG: The Evolution of Information Retrieval Retrieval Augmented Generation (RAG) has undergone a significant transformation with the emergence of Agentic RAG, representing a leap...

Agentic RAG - advanced approach leverages multi-agent systems for query decomposition, self-correction, and multi-hop reasoning. Use in domains like healthcare and finance by delivering precise, adaptive insights from complex data. #AgenticRAG #AI
ragaboutit.com/agentic-rag-...

MIT researchers develop an efficient way to train more reliable AI agents MIT researchers developed an efficient approach for training more reliable reinforcement learning models, focusing on complex tasks that involve variability. This could enable the leverage of reinforc...

MIT researchers unveiled Model-Based Transfer Learning (MBTL), a groundbreaking approach that makes AI agents up to 50x more efficient at tackling variable, complex tasks.
💡 How it works: MBTL trains on a strategic subset of tasks, transferring knowledge across other.

news.mit.edu/2024/mit-res...