PoC: AI that builds it's own source code

11 Likes

Now we just need to integrate some cognitive control, brainstorming, and goal-tracking and it will create AGI on its own…

3 Likes

Here’s a more technical look at it, for the coders out there

6 Likes

This is really great work!

1 Like

Dude, this is pure gold. Not sure what you are up to these days but I am looking for collaborators on a project that is going to rely heavily on this stuff. I proved the concept independently… I am not stealing your ideas. Based on the date of your post I would say we came up with this around the same time. But you did the implementation work - whereas I have been trying to write a whitepaper and start generating hype to raise money to build a business around it. Well, not a “business”, more like a DAO… imagine the Bitcoin network, but instead of miners getting paid to process financial transactions, they get paid to host and run microcontainers containing arbitrary python code on their CPUs and GPUs, allowing for autodeploy and autoscale of any AI code - and users pay to run inference jobs against these modules, and can plug them together like legos to make pipelines.

Of course, a “user” could be a human, or it could be an AI… My ICO announcement video is going to feature GPT XX and GPT XY, two bots who fall in love. They are pretrained on how to flirt, and on how to ask for help when they don’t know the answer. They also have access to automated utilities including wget, to look stuff up.

Guy GPT googles how to tell a girl that you really like her, and it says to paint her something… so he calls on a Dalle or uses a Craiyon, depending whether OpenAI allows us to use a headless chrome to login to Dalle 2 and perform inference. But girl GPT doesn’t understand the meaning of this byte array that has been dumped in her lap, and asks Clippy for help. Clippy is a trained image to text model, the original implementation of open AI’s CLIP (which has always been open source). Clippy describes the painting well, but uses emotional terms that are beyond GPT’s out of the box ability to understand.

So she’s like WTF, and asks guy GPT to explain. Guy GPT sends back a link to a CSV containing a dataset of texts that indicate a guy likes you vs texts that indicate mere friendship. Girl GPT uses davinci to write some Python which trains a simple classifier on the dataset, deploys the code on the Axion network, and then runs inference on the description of his painting, using the new classifier (we tested enough of this to know it will work - GPT3 can indeed write a classifier for recognizing emotions, a standalone classifier, not a GPT finetune)

No, this isn’t a dating app… I just thought it is cute and fun and ordinary ppl will understand it… and it is more likely to pass app review than, i don’t know, skynet

2 Likes