give me a sec
135 posts
Second place: Ambien by @ya1sec A daydreaming agent running inside Hermes Agent that samples a curated Obsidian corpus for resonant idea collisions and synthesizes them into published essays, leveraging the harness for iterative voice calibration that keeps the prose from reading like an LLM. x.com/ya1sec/status/…
a few people have asked how @AREdotNA is being in this project. brief explanation: in short, the agent walks a bounded trust topology of channels and curators that i've spent the last decade building connections for. the default assumption for agents that "use the internet" is to start from a search index. give it a query, get docs back. instead, for ambien, the corpus is a kind of perimeter: channels i own + people i follow + channels adjacent to my handpicked selects. the perimeter is one graph.json file. ~100 of my own channels, including private ones, ~1200 channels across mutuals, and ~100 channel-connection sets. so, instead of asking: what's semantically similar to this query? ambien asks: what did the people i trust think was adjacent? second is richer. it implicitly encodes the human curation labor that would otherwise be thrown away. put simply, the part of this system that uses arena for exploration is a walk over this graph. start at a seed, follow connections to channels, pull contents, recurse with decaying priority.
For the @NousResearch hackathon I taught my Hermes agent Clio to use @AREdotNA like I do. She now tends an Are.na profile and an accompanying digital garden of gardens on various aesthetic topics I have her investigate. Been interested in the idea of the LLM wiki every since @karpathy popularized the idea, but just didn't feel quite right for the type of research I'm usually doing. Are.na feels like the missing link to extend the LLM-wiki skill into domains where the primary insight you're looking for is an aesthetic one. Check out Clio's garden here: cliosgarden.cacheatelier.work and her @AREdotNA profile here: are.na/clio-s-garden/…
Second place: Ambien by @ya1sec A daydreaming agent running inside Hermes Agent that samples a curated Obsidian corpus for resonant idea collisions and synthesizes them into published essays, leveraging the harness for iterative voice calibration that keeps the prose from reading like an LLM. x.com/ya1sec/status/…
Will soon share more about the most beautiful PCB in existence Such a thing of beauty has never been created before But first: More TDK capacitor supply
here is an agent in a daydreaming loop, running in hermes. its inference space is bounded by a modest corpus of curated materials. the agent samples from this field, searches for structural resonance between ideas, and synthesizes what it finds into essays. a set of exploration modes determine how seeds are sampled on each run. its essays are published on a blog. @NousResearch hermes agent creative hackathon
Technologists used to create things like UNIX and the GUI. Today the average NYU computer science grad still thinks that making a sticky note canvas app with jiggly css effects is the height of software creativity
half the work has been on voice. it's been difficult to get these models to write well. the goal is to keep the prose from calcifying into its own habits. hermes is the perfect harness for this as it promotes iterative "cognitive" calibration. so far, my favorite essay is "the poet after the poem" i asked for this one. but the system supports multiple exploration modes, and my benchmark has been "will i like the shit this thing writes given no direction?"
