Data Machina #197
Generative Agents. Autoprompting. LLM apps in production. Consistency models for generative AI art. Conditional adapters SoTA. Auto gradient descent. Graphformers. Self-debugging LLMs
Generative Agents. AutoGPTs are all the rage now. (I wrote about AutoGPTs in DM #196.) In just a matter of a few weeks, 10’s of AutoGPTs have popped up. In some cases, it’s almost trivial to assemble, config, and deploy an AutoGPT in your browser: Try AgentGPT for example.
More on Auto-prompting. As we discussed, auto-prompting -like ReAct, MRKL, and CoT patterns- is one of the cogs & wheels that power AutoGPTs. I received a few emails from ppl asking for more info on auto-prompting. Checkout these links:
Can GPT-4 Prompt Itself? A nice overview -albeit not very technical- on auto prompting and AutoGPTs
The AutoPrompt paper: First introduced by a team from UCI & UC Berkeley, it describes a new auto method to create LLM prompts for a diverse set of tasks, based on a gradient-guided search
The Auto-CoT Paper: First introduced by a team at AWS Science. It’s a proposal for auto Chain-of-Thought prompting (Auto-CoT) in LLMs without needing manually-designed prompts
The Why Think Step-by-Step? paper A team @Standford illustrates how the statistical structure of training data drives the effectiveness of chain-of-thought reasoning, step by step.
The Generative Agents Paper. Until recently (a few weeks ago?) the leading AI edge was about AutoGPTs; autonomous agents, following certain LLM patterns, and auto-completing tasks based on instructions or task prompts.
But a week ago, a team from Stanford & Google dropped a paper that describes a group of generative agents that simulate human-behaviour in a very sophisticated way. Based on LLMs, the agents autonomously generate their own behaviour. Checkout: Generative Agents: Interactive Simulacra of Human Behavior
What the researchers did. First, they built Smallville, a sandbox city-world game (inspired by Sim-city.)
Then -using prompting and similar stuff- they developed 25 agents, each one with: a different role, character, occupation, personality, and goals. After that, they populated Smallville with the 25 agents as inhabitants. Finally, they hit “play game” and let the agents go about with their daily lives without any human intervention.
Why this paper is fascinating? The agents plan and conduct their own daily lives autonomously! They develop new relationships amongst agents, and remember each other. They also share information, communicate, and coordinate with each other. The agents recall what they did and what they observed. They even reflect on their own actions and observations. What a fascinating paper to read!
Inspired by the Generative Agents paper, @sean_pixel released Teenage-AGI, a new Pyproject that uses OpenAI & Pinecone to give memory to an AI agent. It also allows it to "think" before making an action.
Also worth noting that a little-known AI team @KUST has fully released CAMEL Communicative Agents (paper, demo, repo:) Another amazing project that explores building role-based agents that autonomously cooperate and communicate amongst them, while showing insight into their "cognitive" processes.
btw: The team at LangChain has implemented CAMEL Role-Playing Autonomous Cooperative Agents.
I’m not sure about AGI (Yet?)… I guess I’m still not as terribly terrified as Seth when he says that Agentized LLMs will change the alignment landscape. But yeah: AutoGPTs, LLM agents will indeed change AI Alignment.
We’re organising a little meetup: A Deep Dive into Generative AI. London, April 18 (come, it’s free. Sign up here) We have limited spaces. We’ll have:
Tom CTO Stability AI, talking about the latest on Large Models @StabilityAI. I’m expecting him to give us an update on the brand new Stable Diffusion XL. I have a battery of Qs for Tom :-)
Chandan, a Principal Architect @AWS_Startups, will talk about how to deploy and fine-tune Generative AI models with AWS Sagemaker Jumpstart. I’ll ask Chandan about the new Amazon Bedrock and Amazon Titan Foundation Models for building scalable generative AI apps
Apolinario, an ML Art Engineer @Hugginface, will do a hands-on talk how to train your own ControlNet with Diffusers. This one will be fun!
Have a nice week.
10 Link-o-Troned
the ML Pythonista
the ML codeR
Deep & Other Learning Bits
AI/ DL ResearchDocs
El Robótico
data v-i-s-i-o-n-s
MLOps Untangled
AI startups -> radar
ML Datasets & Stuff
Postscript, etc
Tips? Suggestions? Feedback? email Carlos
Curated by @ds_ldn in the middle of the night.