3 ms·
Some of the most interesting work in this space is in the “shared” memory models (in most cases today, vector db’s). Agents can theoretically “learn” and share
by synaesthesisx 4y ago
Some of the most interesting work in this space is in the “shared” memory models (in most cases today, vector db’s). Agents can theoretically “learn” and share memories with the entire fleet, and develop a collective understanding & memory accessible by the swarm. This can enable rapid, “guided” evolution of agents and emergent behaviors (such as cooperation).
We’re going to see some really, really interesting things unfold - the implications of which many haven’t fully grasped.
- creamyhorror 4y agoHow would vector DBs encode say a precise, technical process that has been figured out by an agent? Would the vectors still be natural language as with LLMs? Would be great if you could point me to one or two exciting papers in the area.
- TeMPOraL 4y agoIt doesn't have to. But the vector search can point it to the URL / document database where it can get step-by-step instructions of that process, perhaps already condensed/compressed by another LLM, and perhaps daisy-chained[0] to work around context limits. ---- [0] - I don't know the right terminology, but I imagine most complex processes can still be split into a sequence of sub-processes, where each sub-process consists of necessary steps, steps to confirm success, and a reference to the next sub-process to load if the current one succeeds. The bot could then keep only one sub-process in their working memory at a time, assuming previous ones succeeded.