Been thinking about this too. Indeed current “AI” systems lack certain pro-active solution-‘searching’ or -‘browsing’ techniques. Naturally, when you train the system to “predict” words, phrases and knowledge based on existing dataset, the component of synthesising new ideas or knowledge is missing.
Another question is whether humans have something out-of-reach and hyper-dimensional or soul-connected there. Or is this a trivial engineering problem?
Human mind is powerful in associative thinking, coming up with new ideas and coming up with things to try. What is the mechanism for this – can we replicate this in AI?
One of the techniques the human mind possesses is connecting seemingly unconnected knowledge and memories which sometimes results in new innovative solutions (or new knowledge) discovered. Researchers say that during REM sleep, for example, brain processes memories trying to connect them in different (random?) ways. This is when we see dreams.
Could it be that we just throw some randomness into the mix (reshuffle weights or randomize prompts time to time) that allows AI to obtain a similar capacity?
Browsing and randomly connecting things are pretty trivial tasks within a multi-dimensional vector space. In my PhD thesis, for example, I’ve demonstrated a way for a human to ‘browse’ a multi-dimensional knowledge base, an alternative mechanism of knowledge discovery as opposed to a ‘telegraphic’ keyword-based search in the search engines. The data format was essentially the same as used by modern LLMs – a corpus of text documents converted into vector space. By applying dimensionality reduction and mapping it into 2D or 3D we can allow human users to browse the knowledge to discover materials semantically close (or similar) to whatever they are researching.
For AI, we don’t even have to reduce the dimensionality. Simply allow the ‘brainstorming agent’ to venture out time-to-time to a nearby random coordinate, and occasionally go far distance too. ‘Nearby’ and ‘far’ here, of course, means the Euclidian distance between the documents or prompts mapping. Then, feed the creative LLM “Inventor” agent with such randomly discovered information as a proposed solution to the current problem. Polish it off by passing the result through a “Critic” and “Realist” agents who will ensure only valid solutions pass the selection.
There should be a technique for logical step-by-step verification too. We just need to (1) figure out how to enable AI to understand when it hallucinates and (2) figure out how to marry the LLMs with AI-generated logical decision trees. Big problems those two.