Vincent Granville’s Post

How to Make Black-box Systems more Transparent

How to Make Black-box Systems more Transparent

How to Make Black-box Systems more Transparent

A desire for understanding creates demand for information. An abundance of information creates a demand for machines. A system of machines creates new kind of information. A desire for understanding demands a new kind of machine. Blue or Grey dots?

Curious about the role of feature stores in the modern ML landscape? In today's article, "Feature Stores in an Embedding World," I talk about: - The importance of feature stores for managing ML data - How embeddings enhance feature engineering - Practical use cases and real-world examples Anyone looking to improve their ML systems and leverage embeddings for better model performance (and less tech debt?) should read this! Link in the comments, happy reading!

X Adds New Setting To Opt out of Sharing Data To Train Grok: A new setting on data permissions for Grok was quietly added and users opted in by default. https://bit.ly/4fpygzW

The Agents are here. Today Microsoft and Anthropic both announced early versions of AI Agents. What's an agent? It's an AI model that can work outside of the sand box of it's own program. It can control the mouse on your computer, enter data, etc. In the future we won't need watch it work, it'll just go, but right now you can watch the clicks. Where is this now? Both Microsoft's CoPilot and Anthropic's Claude Computer Use are essentially open ended, automated workflows. Think Zaps, but less time consuming to set up and can work in places where you might not have an API (I'm excited for patient documentation for physicians). Where is this going? We can split our work into two broad categories. Deterministic and Creative. - Deterministic work follows a path - these are the things we do every day. Open email, check the news, look ahead for meetings, fill in data etc. - Creative work puts together two or more things from the past to make something new. Its taking an idea you saw in one industry, and applying it to your industry. Deterministic work can be split into two more broad categories. Bounded and Unbounded. - Bounded work is where there is a clear SOP. There are paths of work that need to be followed, but are not automatically triggered by a workflow. Think front-line customer service. There are a set number of problems that can be resolved by front line. - Unbounded deterministic work is that next level up. There are SOPs for how actions should be conducted, but not for the action itself. These need some level of interpretation of the spirit of the SOP, like "delight the customer". AI-Agents are going to master the "Deterministic, Bounded" activities we do first. They'll have intentionally limited abilities so that we can get used to "someone" working on our computer with us. It will take work to get the agent set up and bounded into a pathway. This will serve as a data collection time for the companies who make agents to understand what is helpful and what is not. Then, probably within 18 months, we'll get the opportunity for "Deterministic, Unbounded" agents to work with us. This will probably be at the level of a junior Executive Assistant and will likely get set up by "watching our work" for a set amount of time. I could imagine allowing Claude or ChatGPT access to watch my daily work for a week, then giving it access to create an agentic workflow to take some of the bounded and unbounded work away. I'm excited about use cases for doctors, teachers and other professions where the administrative work is mission critical, but not mission focused. Is there a use you're excited for? https://lnkd.in/g5m9HfZu

https://lnkd.in/eZCk_gK7 SO MANY BAR(GRAPH)S! 'there are a lot of algorithms that play automated decision systems with kinda high stakes results.' --- 'algorithms that control our life that we're not even aware of... everything from the type of customer service you'll get to... whether you'll ' --- 'data is created by humans and humans are biased.' --- 'i could be one of those people making those decisions.' --- 'let me first get good at this, let me first become an expert at what I'm doing so I can take that expertise and move towards my longterm goals'

**IF** this becomes reliable, massive boost for building AI copilots for complex workflows. "We've built an API that allows Claude to ... translate instructions (e.g., “use data from my computer and online to fill out this form”) into computer commands (e.g. check a spreadsheet; move the cursor to open a web browser; navigate to the relevant web pages; fill out a form with the data from those pages; and so on)." In a lighter vein: "... while we were recording demonstrations of computer use for today’s launch, we encountered some amusing errors .... **Claude suddenly took a break from our coding demo and began to peruse photos of Yellowstone National Park**." 🤣 [FWIW, I think Claude has good taste :-)] Blog: https://lnkd.in/deR8rSdz Demo: https://lnkd.in/d-WiZBww (The demo is cool but the usual caveats apply - cherrypicking, survivorship bias etc)

Client-Specific (or Project-Specific) RAG System 1. File Monitoring: n8n trigger for Google Drive updates. 2. Chunking: New files are broken down (with overlap) into chunks. 3. Embedding: OpenAI embeddings model converts the chunks into vector embeddings for storage (& quick retrieval with the LLM). 4. Storage: Embeddings are stored & indexed in Pinecone. 5. Chat: Data can be accessed easily via chat.