3 ms·

Ask HN: Projects researching potential for AI sentience

17 points by j4yav 3y ago

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.

ianbicking 3y ago
You might be interested in what these folks are doing: https://x.com/OpenSoulsPBC https://x.com/OpenSoulsPBC
neoneye2 3y ago
ACE framework by David Shapiro. https://github.com/daveshap/ACE_Framework 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 https://www.youtube.com/watch?v=HNtKVrQMNZs
Paul-Craft 3y ago
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.
jasfi 3y ago
I'm working on something which I'm hoping can approach AGI: https://aiconstrux.com https://aiconstrux.com. The description on the landing page doesn't outline the full planned scope and abilities.
- kylebenzle 3y ago
  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.
  - jasfi 3y ago
    LLMs have proven to be useful on real problems. A spell checker isn't in the same category.
- wruza 3y ago
  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.
  - jasfi 3y ago
    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.
kylebenzle 3y ago
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.
- clnq 3y ago
  There's a statistical model to describe what you say as well - https://en.wikipedia.org/wiki/Zipf%27s_law https://en.wikipedia.org/wiki/Zipf%27s_law.
pomatic 3y ago
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!

17 points by j4yav 3y ago

ianbicking 3y ago

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

neoneye2 3y ago

ACE framework by David Shapiro. https://github.com/daveshap/ACE_Framework 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 https://www.youtube.com/watch?v=HNtKVrQMNZs

Paul-Craft 3y ago

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.

jasfi 3y ago

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

kylebenzle 3y ago
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.
- jasfi 3y ago
  LLMs have proven to be useful on real problems. A spell checker isn't in the same category.
wruza 3y ago
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.
- jasfi 3y ago
  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.

kylebenzle 3y ago

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.

clnq 3y ago
There's a statistical model to describe what you say as well - https://en.wikipedia.org/wiki/Zipf%27s_law https://en.wikipedia.org/wiki/Zipf%27s_law.

pomatic 3y ago

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!