Cognitive Architectures for Language Agents

we convened an awesome group of researchers, scientists, teachers, and builders to discuss the recent paper on CoALA (Cognitive Architectures for Language Agents) and the implications for building with agents
@hwchase17 (@langchainai)
@charles_irl (The Full Stack, prev. @weights_biases)
@tedsumers (@Princeton, prev. @uber)
@ShunyuYao12 (@princeton_nlp)
@ddohan (@OpenAI)

[@videowatching9576]

These videos from LangChain are so great because a nice mix of a focus on actionably useful things, explanation, and some technical context - so a nice way to get a sense of what’s happening in building in AI. Strong contrast to content elsewhere that might be too high-level, or too in technical details.

[@mrchongnoi]

At 34:00, when the discussion turned to episodic and semantic memory, the thought that came to mind is that these two types of memory have a relationship. For example, viewing this KZbin webinar is an episodic experience that contributes to semantic memory. Maybe a few weeks from now, I can remember that I listened to a video on CoAl, which requires a semantic lookup of what I learned during this video.There seems to be a bidirectional relationship. Not sure if bidirectional is the correct term to use. As episodic memory fades (TTL), then it becomes pure semantic. Old guy thinking out loud in Singapore.

[@FreakyStyleytobby]

Isn't semantic knowledge the one that we managed to retain? The one that we practice. Episodic memory fades as we don't practice it.

[@mrchongnoi]

@@FreakyStyleytobby Went back to review the video. Saw your reply. There some events that are just burned in to our memory. I saw a 60 minute video where there are individuals what remember almost every minute of their life.

[@attilavass6935]

Pls. share the link of Shunyu's presentation! :)

[@galmoore3193]

super interesting topic. Is this planned to be a Langchain package? I couldn't find any code or discussion of implementation.

[@RedCloudServices]

Harrison rendering all the academics speeches 😂

[@LeonidGaneline]

"learn = Write long-term memory" It is not so simple. What about processing data BEFORE writing? When we consider just DB, then we don't care about this preprocessing. Maybe there is some format moderation and data cleaning. But when we write the procedural memory, the LLM, the preprocessing is important. We generalize data before writing. It is not trivial and it is not solved up-to-date. Sure, we can look at this in a pragmatic way: write into the LLM = training the LLM. It means the LLM (as the long-term memory) takes care of the generalization internally. Still, taking the data generalization step is important, IMO.

[@MsDuketown]

People can be so sentient when reasoning... Is this podcast about the Adam's apple or the larynx? Remember, don't underestimated the qualities of the tongue. Either way, a mature workstation PC, at home or at work, can always be helpful. Even a Ultrabook will do in most scenario's. Btw. Don't get distracted by occasional gaming on the same rig.

[@UncleDavid]

y’all should take a communication class or sumn bruh

[@MICH43LTV]

Ironic comment

[@SDGwynn]

@dmdohan !!!