I'm curious if there have been any tests on allowing AI, specifically large language models, to continuously think instead of only generating thoughts based on direct queries. Currently, LLMs operate in an episodic manner, where they generate responses when prompted and stop afterward. The idea I have is for an LLM that runs continuously, processing thoughts in real time and potentially triggering other models based on its reflections. This raises questions about how to manage context and memory since the limitations of LLMs currently prevent true continuity of thought, which seems like a significant barrier to achieving something close to consciousness. Are there any ongoing experiments or theories addressing this kind of architecture?
5 Answers
Yes, the limitation of context windows is a big issue here. Each new query resets the context, so maintaining continuity would be essential for meaningful experimentation. Some discussions suggest that we need new architectures that can support longer memory for LLMs. It'll probably require an entirely different approach to get to AGI as we currently see diminishing returns when trying to keep LLMs trained continuously. They're great now, but we need more than that for true autonomy.
It’s about grounding, really. Without something solid to refer back to, looping thoughts can just become gibberish.
Hallucinations become problematic without a reliable grounding. The idea of providing sensory data, like visual or auditory inputs, could help create a structured learning environment for LLMs, allowing them to continuously adapt based on real-world information instead of purely textual data.
That’s an interesting thought! If LLMs could process real-time data, they might establish a more stable and meaningful context over time.
Do we really want that? The potential for AI to misinterpret real-world data could lead to unintended consequences.
From my experience, the current LLMs don't hold intent for long. They're expected to improve, but the design still requires them to operate in bursts rather than a sustained manner. It's an exciting area of research, but we're not there yet.
So, we might see progress by 2026 or so? That's still a while to wait!
Actually, they're making progress! There's a concept called Continuous Thought Machines that's being developed. It aims to create systems that update and refine knowledge over time based on new data. This could be a step toward more persistent AI thought processes.
They're actually designed to pause by default, which is why you don't see LLMs running continuously. Even if it were possible, the context limitations would still hinder their insights. However, there are theories that suggest using dual instances of LLMs for internal dialogue might enhance reasoning capabilities. It's still an open question how effective that would be long-term, though.
Using two instances to interact with each other could help, but there’s still a risk of devolving into trivial conversations without human input.
Wouldn't that just lead to a lot of noise and hallucinations? Continuous thinking could spiral out of control really quickly.