HN Explorer

Show HN: Testing hypotheses through prediction is the next step towards AGI (github.com)

3 points by judahmeek ↗ HN
Granted, all I've got so far is a Draft Specification for a proof of concept using ARC-AGI-2 as a minimalistic problem domain.

But that was enough for METR.org[0] for recently agree to fund my work on this project.

That little bit of validation goes a long way & the funding won't hurt either.

Also, I have questions for any patent lawyers reading this:

* When do I know if I have enough specification detail to start filing a patent?

* Does this public repository qualify as prior art?

Finally, if my post text seems overly sparse with the details, I apologize & I promise to answer any questions you have as fast as Hacker News allows me to.

0: https://metr.org/

4 comments

[ 2.8 ms ] story [ 15.8 ms ] thread
judahmeek
To avoid hitting HN's rate limit, I will be adding my answers to this post for as long as I am able to edit (2 hours, I think). Once I am no longer able to edit or have hit a character length restriction, I will create a reply post to this one and do the same thing there.

Also, to reduce confusion, the original title was "Show HN: A belief system is how we create AI that actually think like humans do"

Answer to taylodl's question[0]: AI will never be able to think exactly like humans do. There will be trade-offs to this. However, enabling AI to think more like humans will enable both AI & humanity to better explore these tradeoffs.

0: https://news.ycombinator.com/item?id=46135315

Reply to Voland0's post[1]: I've found trying to specify my own personal solution process to be very enlightening, but also quite mentally taxing. Such self-awareness comes at quite the cognitive cost!

1: https://news.ycombinator.com/item?id=46135447

taylodl
What's the point in creating an AI that thinks the way humans do? There are already 8 billion humans on this planet capable of doing that. I want AI to provide a different perspective and be able to explain its reasoning.
Voland0
One point in creating AI that thinks the way humans do is to more fully understand what processes our own neural network operate, similar to how we understand our own language better by learning another tongue.
judahmeek
Correction for the record: METR is not funding my work directly. I am participating in a METR study regarding performance effects of software development AI-assistants & have chosen the project above to be my primary focus during the study.