Discover real AI creators shaping the future. Track their latest blogs, X posts, YouTube videos, WeChat Official Account posts, and GitHub commits — all in one place.
Activity on simonw/datasette
simonw commented on pull request simonw/datasette#3099
View on GitHubActivity on swyxio/papra
swyxio contributed to swyxio/papra
View on GitHubActivity on swyxio/papra
swyxio opened a pull request in papra
View on GitHubI'm rooting for Google, OpenAI, Anthropic, SpaceX and Meta. Competition at the frontier model level is so good for everyone involved.
Activity on swyxio/papra
swyxio contributed to swyxio/papra
View on GitHubRT Elias Lumer really good breakdown of a concept “knowledge factories”, which feels like a more mature version of llm-wiki/agent memory from langchain via background agents after a few months of trial and error we see a lot of overlap with our internal agents: - different types of context: skills, system prompt, user memory, global memory (all being updated mostly autonomously by background agents) - background/dreaming agents/automations are increasing in % total runs (20%-50%) and the intervals can be from every 1-day, 7-days, and as short as 5min! - data retention with different data sources plays into how your pointers live in just-in-time knowledge (e.g. data an agent reads from an MCP may only be accessible for 7-days, so saving that in user memory/wiki for later is necessary) where we are thinking about things differently: - we treat curator agent and the main agent as the same agent, so it maintains the same user memory/system prompt, etc. (just a different prompt/skill) - representing memory actually should be pretty fluid, don’t lock in a framework like OKF. opus 5.5/fable will have natural emerging qualities to organize memory differently (Anthropic published agent-memory article on this) like folders, files, daily/weekly logs, etc. let the agents choose how to do this, as long as it’s scalable. - background/curator agents should also be run with event triggers or mechanisms to make sure that you’re not running agent compute without real data coming in (if no new data/issues arise, don’t kickoff the agent) - agents won’t always flag things when there’s an issue. don’t base your curator background agent on a tool call the agent might miss in the moment. Run background agents to identify these issues on-top of the agent-driven flagging (and once again, these can all be the same agent to keep it simple, no need to maintain 3 agents when it can just be 1 with 3 prompts/skills) - didn’t see a lot of user-memory references here. a large part of knowledge ...
all our agents at langchain (internal GTM & content agents, LangSmith Engine) have a learning loop with similar pieces wrote about what we think makes up a "knowledge factory" https://x.com/i/article/2124180107669766144
View quoted postwould love to read the trace of @__alpoge__ dialogue with Claude - seeing a discovery like that happening
Activity on simonw/datasette
simonw opened a pull request in datasette
View on GitHubActivity on simonw/datasette
simonw opened a pull request in datasette
View on GitHub