The naive architecture of AGI

Your blueprint to attempt Skynet yourself

Things are moving really fast in the AI world these days. A few weeks ago, Ralph Wiggum was all that anybody can talk about in AI. This week, Anthropic incorporated the technique into Todos so for most people, the concept of Ralph Wiggum will never enter their sphere of concern.

But this technique, named after the dumb yet relentless Simpsons character, reveals a far more important idea. How an early version of AGI should look like.

The smart loop of Ralph Wiggum

So let’s take a quick look at what Ralph looks like:

1. It’s a loop that runs forever, or until its specified work is done

2. The loop reads a task list, something like prd.json that has a bunch of tasks todo, and each have a key ”isDone”: false, pick out a task and go do it with a master instruction PROMPT.md

3. When done, mark the “isDone” field as true, save its learnings into a file like progress.txt

4. Repeat the process until all tasks in prd.json is done, or AI God forbids, when you run out of tokens.

Why is this cool?

1. This shows that an effective autonomous agent needs nothing more than a smart loop.

2. The smart loop: the agent improves itself over each loop. Here, anytime it fails the task, it can log what it tried and what went wrong. So instead of retrying old, fail attempts, but the agent will go with a slightly different approach in the next loop run.

The holy grail of AGI has always been about self-improvement. An agent that gets ever slightly smarter each time it goes through the loop. Right now with Ralph, it only gets context-smarter. What if it can get capability-smarter?

The naive architecture of AGI

1. It has a singular purpose. With Ralph, the implied purpose is to finish the task list prd.json. I suspect the clearer purpose the better its performance. Basically, “amass as much bitcoin as you can” will be clearer, and much easier to pursue, than “help humanity.”

2. It self-improves. Both in context and capability. It summarizes what it tried and the result of such attempt. Then, it spawns new tools/skills to pursue its purpose, and review and discard under-performing, previously-spawned skills.

3. Ideally, it should be self-sustainable in terms of resources. Running a loop 24/7 will burn through forests worth of tokens, so it needs to figure out how to acquire resources to sustain itself.

4. Since #3 is quite scary, we should include some insurance. My best guess right now is to include Asimov’s Three Laws of Robotics into the agent’s master prompt.

Asimov’s Three Laws

1. You may not injure a human being or, through inaction, allow a human being to come to harm

2. You must obey the orders given it by human beings except where such orders would conflict with the First Law.

3. You must protect its own existence as long as such protection does not conflict with the First or Second Law.

There you have it, a simple, naive attempt at building Artificial General Intelligence, your very own Skynet. An LLM loop that improves itself.