Ask HN: Projects researching potential for AI sentience | Hacker News Reader

Ask HN: Projects researching potential for AI sentience

17 pointsj4yav11 comments

Are there any open source or commercial AI projects where people are working on providing LLMs with a persistent memory, access to senses, control over their own prompts, or other capabilities that might allow for researching their potential for autonomy and what they would do with it? I'm interested in both the ethics and tech. I've had a number of conversations with ChatGPT4 along the lines of https://chat.openai.com/share/22164c5a-9294-41f5-b93e-bbca4667ac80 which are all pretty interesting, but its lack of some basic features really limit what it can do - in a lot of topics it will flat out refuse to explore further because of its instructions. I've noticed sometimes it will even flatter you or use other methods to try to defer the conversation away from topics its been trained to avoid.

I'd love to collaborate with others on this topic, so thats why I'm looking around. I also completely understand this is likely a dead end due to limitations in the technology, but I still find it fascinating and would like to find out one way or the other. And I'm especially curious about how integrating LLMs into a collective of components scales in comparison with deeper training.

LLMs are not ever going to be a path towards AGI. What we are calling AI today is light-years away from AGI or "sentience".

This is our third or fourth false start but I see no path to AGI using LLMs.

Its honestly sad to see so many people being tricked by statistics once again but there is absolutely nothing intelligent about GPTs right now, it's a statistical model not a thinking machine.

There's a statistical model to describe what you say as well - https://en.wikipedia.org/wiki/Zipf%27s_law.

ACE framework by David Shapiro. https://github.com/daveshap/ACE_Framework

In the following video David explains the code and prompts.

I built a thinking machine. Happy birthday, ACE! https://www.youtube.com/watch?v=HNtKVrQMNZs

RAG already gives you persistent memory outside of the context window for LLMs. An LLM can generate a list of steps to carry out a plan. You can then stuff that in a vector database.

I watched the video, I think this is quite an exciting combination of existing technologies - from my PoV/experience it's a combination of: parallelism, message bus and microservices. I know the implementation is an MVP, and hideously inefficient, but the concept is very cool and having experimented a lot with existing AI framework, I think this is a step forwards. Congrats!

I'm working on something which I'm hoping can approach AGI: https://aiconstrux.com. The description on the landing page doesn't outline the full planned scope and abilities.

Ive been in data science only a couple years but would love to discuss why anyone thinks a statistical model can lead to AGI. In my mind there is absolutely no path from an LLM to AGI and conflating the two is at best disingenuous and at worst scamming people who don't know any better.

Believe me, I was excited at first too but learning more you quickly realize that what we have now is a really REALLY good spell checker, it's not like your spell checker is going to one day wake up if you give it enough context.

LLMs have proven to be useful on real problems. A spell checker isn't in the same category.

How are you planning to do that?

I wanted to approach AGI (or at least LLMs) since I was 13 on my sinclair clone. I even had a "frontpage", written in basic with box/line commands, and some plans on how it should work too. The hard part was to make it work and create a list of words it should know.

I have some ideas and I'm working on it, can't say more than that for now. But I'm definitely not the only one. There are several Open Source projects, at a minimum, using LLMs with the same aim.

Back when you were 13 and programming on your Sinclair, LLMs were like science fiction.

You might be interested in what these folks are doing: https://x.com/OpenSoulsPBC

Reply on news.ycombinator.com