Data Machina #219

AI Agents: Better Than Humans? Cognitive Architectures for Language Agents (CoALA). μAgents. Against LLM Maximalism. XTTS Voice Cloning. AI Forecasting. NeXt-GPT Any2any Multimodal LLM. phi-1 1.3B

AI Agents: Better than Human Agents? So I booked my flights. I just realised I need to change my info before checking-in because I made a typo. But the bloody app won’t let me do it. After 20 min on the phone! I finally get to talk to a human agent. The human agent instructs me to change the info via the app. I explain that I can’t because the app doesn’t allow me to do that! The human agent then asks me 10s of irrelevant Qs to finally conclude that I have to use the app to change my info before checking-in!

Me: “Are you really a human agent? You just seem to be reading from a script and NOT listening to what I’m telling you!.. Are you an AI agent with a human voice?”

The Human Agent: “Sir, we follow top-class procedures to ensure that your experience as a client is always the best. Please log-in into our app to provide your correct information before checking-in.”

This human agent is pathetically useless. These days, most human agents sound and act 99% like AI agents. I guess we will be much better off with AI Agents that sound and act 99% like humans.

A survey on LLM-based AI agents. An excellent survey in which the researchers present a framework for LLM-based agents, with three 3 components: brain, perception, and action. The researchers explore AI agents in 3 areas: single-agent scenarios, multi-agent scenarios, and human-agent cooperation. They then explore agent societies, behaviour & personality of agents, the social phenomena that emerge when they form societies, and the insights they offer for human society. Paper: The Rise and Potential of LLM-Based Agents: A Survey.

A new architecture for AI language agents. This is an interesting paper which aims to put some order in the piecemeal approach followed by many researchers in AI agents. Researchers @PrincetonUni introduce: Cognitive Architectures for Language Agents (CoALA), a conceptual framework to systematise diverse methods for LLM-based reasoning, grounding, learning, and decision making. Checkout the paper, and accompanying website with references and resources: Cognitive Architectures for Language Agents.

Building an AI agent with visual AI programming. Building AI agent apps programmatically is hard. A basic-to-mid level AI Agent app involves a lot of design cycles, prototyping, debugging, remote execution bits & pieces… Visual AI programming comes to the rescue. Last week I mentioned Rivet, an open source visual AI programming environment. Checkout this post on how to build your first AI agent with Rivet. And also a quick vid below on using visual AI programming to build a simple chatbot.

An AI agent dev framework. The μAgents is a fast and lightweight framework that makes it easy to build autonomous, decentralised AI agents. See this demo on how to build a μAgent for restaurant bookings.

A new library for building autonomous AI agents. Agents is an open-source library/framework for building autonomous language agents. The library supports key features like long-short term memory, tool usage, web navigation, multi-agent communication, human-agent interaction and symbolic control. See paper, code, examples, and website here: Agents: An Open-source Framework for Autonomous Language Agents.

New research on AI agents for medical diagnosis. In this paper, the researchers introduce a multi-agent conversational framework where doctor-AI and patient-AI agents interact to diagnose medical conditions, evaluated by a grader-AI agent and medical experts. The researchers assessed the accuracy of GPT-4 vs. traditional evaluations methods when diagnosing 140 cases. The researchers found a decline in diagnostic accuracy, identified key limitations in LLMs’ ability to integrate details from conversational interactions to improve diagnostic accuracy. Paper: Testing the Limits of Language Models: A Conversational Framework for Medical AI Assessment.

Managing a team of AI agents. Meta-GPT is a multi-agent framework that assigns different roles to GPTs to form a collaborative software company for complex tasks. Meta-GPT takes a one line requirement as input, and outputs: competitive analysis, user stories, PRD, data structures, APIs, docs, tasks and repo. See: MetaGPT: The Multi-Agent Framework.

A new community on AI Agents. This is a community of ML experts, professional prompt engineers, and AI enthusiasts that provide open source AI agents. Checkout Open source AI Agent community.

Have a nice week.

10 Link-o-Troned

1. AI and the End of Programming

2. On NLP: Against LLM Maximalism

3. AI Forecasting: Two Years In

4. Can LLMs Really Reason and Plan?

5. I Made a Transformer by Hand (No Training!)

6. Four Lessons from Launching an LLM Chatbot

7. Exploring Natural Language for Autonomous Driving

8. XTTS Open Model - Clone Any Voice in Several Languages

9. A New Mental Map from MLOps to ML Sys with FTI Pipelines

10. Project Gutenberg Free, 1000s of Audiobooks Generated by Text2Speech

---

---

the ML Pythonista

1. Complete Guide to Build RAG LLM Apps for Production (iPynb, blog)

2. Hands-on Tutorial: Validating LLM Outputs with Guardrails

3. Blind Image Restoration with Generative Diffusion Prior

Deep & Other Learning Bits

1. Few-Shot Learning: Methods & Apps in 2023

2. A Brief Overview of Parallelism Strategies in DL

3. Amazon Science - Making DL Practical for Earth System Forecasting

AI/ DL ResearchDocs

1. MSR - Textbooks is All You Need [phi-1 1.3B a Super Efficient Model]

2. Chain of Density (CoD) Prompting for Better GPT-4 Summarisation

3. NexT-GPT: Any-to-Any Multimodal LLM (code, vid, demo)

data v-i-s-i-o-n-s

1. Visualising How the Transformers & LLMs Work

2. The Challenges of First Time UK Home Buyers

3. Percentogram - A Histogram Binned by % of the Cumulative Distribution

MLOps Untangled

1. Why We Built & Open-Sourced ML Registry @Glassdoor

2. A Summary of LLM Observability & Monitoring Techniques

3. How @Ramp Accelerated ML Dev-to-Prod to Simplify Finance

AI startups -> radar

1. Lessons from 139 YCombinator AI Startups (S23)

2. Helsing - AI to Serve Our Democracies

3. Generate Biodmedcines - Generative AI for Biomedicines

ML Datasets & Stuff

1. A Dataset with ~120k Python Code Exercises Generated by ChatGPT 3.5

2. The Human Chronome Project - A DB of Human Global Activities

3. Google MADLAD-400: A Multilingual & Doc-Level Large Audited Dataset

Postscript, etc

Tips? Suggestions? Feedback? email Carlos

Curated by @ds_ldn in the middle of the night.