tl;dr
GPT4ALL V2 Released
AutoGPT & BabyAGI Blow Up
Autonomous Agents Paper
AI News for the Week
GPT4All V2 Was Released
GPT4All, one of the original locally installable, fully open-sourced LLMs, released V2 this week. With it came a new commercially legal model based on GPT-J, a dead-simple one-click installer, a brand new interface, and the ability (soon) to share your data (optionally) to help better train future models.
I also made a video walking through how to install it and updates on some of the new features. Here’s that video:
A lot of people were rightfully disappointed with the quality of GPT-J, it isn’t as good as LLaMA or GPT, but there are tradeoffs, such as neither of those models being able to be commercially used.
I also had a chance to sit down with the founder of Nomic AI, the company behind GPT4All, in a lengthy discussion about everything from GPT4All to LLM centralization risks to model bias. Check out that interview here:
AutoGPT and BabyAGI
One of the most interesting things last week was the launch and incredible popularity of two new AI projects: AutoGPT and BabyAGI. Both of these are proof-of-concepts for what can happen if AI is built in a way to be more autonomous. For each of these projects, which are open-source, you give AI a goal, and it will create a plan to accomplish that goal, including creating a task list, prioritizing it, researching on the web, and in some cases, actually executing the tasks. This is all with very little to no input from a user.
Although these projects are simple and can’t do much, the potential is clear. Autonomous AI operating without much user input is both scary and exciting. The productivity increases possible from systems like this are obvious, but the downsides are just as obvious. Check out my tutorial videos if you want to install AutoGPT or BabyAGI.
Autonomous Agents Research Paper
A new paper from Stanford and Google was released that explains how they built a virtual town, put 25 fully autonomous agents in it, and let them live their lives. These agents started developing human-like social behavior within the world. Users were allowed to indirectly influence the town by giving agents an “inner voice,” helping them understand their motives.
For example, one agent was told to throw a party for Valentine’s Day. She then started planning the party, told her friends about the party, and those friends told other friends, who all coordinated to arrive at the party on time. I am so excited about the potential for this, especially for the video game use case. Imagine an entire world filled with NPCs that are essentially alive when you’re playing the game and even when you’re not.
I’m playing around with ways to recreate this world myself, and if I can, I will release the code for anyone to use. Check out my video about this here:
AI News From Last Week
OpenAI Is Not Training ChatGPT5
Greg Brockman, Co-founder of OpenAI, has made it clear that they are not training GPT-5 currently. In a lengthy Twitter post, which can be found here, Brockman outlines OpenAI’s thinking on training more sophisticated models. As of right now, they are focused on GPT-4, and will “continue to ramp our safety precautions more proactively than many of our users would like”. OpenAI is focused on making GPT-4 better, more efficient, and safer before training GPT-5. This could be somewhat in response to the open letter sent by AI leaders to pause AI development for six months. Either way, GPT-4 is still excellent and highly limited, so we have much more to expect.
Elon Musk Starts AI Company: x.AI
Not to be left out of any world-changing technology, Elon Musk started his own AI company called x.AI. There are few details yet, but he has owned the domain name x.com for a long time and has always intended to use it similarly to WeChat in China. WeChat is a ubiquitous app for everything in China, allowing for payments, gaming, social interactions, government interactions, and anything else a citizen of China would need. This is Musk’s vision, but he wants it all powered by AI.
