Ephemerent — independent research lab

01 — Research

We study a single question from several directions: how does useful intelligence emerge, briefly, from systems of small agents?

Orrery runs on language models today — they propose patches, narrate plans, and write code. Our research asks what to integrate next: world models that predict consequences before execution, latent layers that score and plan alongside token generation, and compute substrates that make the stack cheaper to run at scale.

R1 Emergent orchestration Decomposing a goal into parallel agents, running each in isolation, and merging only what works.

R2 Verifiable selection Choosing the best attempt by execution, not impression — panels, metrics, and judges instead of guesswork.

R3 Latent world models Research to integrate: computational state — repos, runtimes, proof states — encoded and rolled forward in embedding space, scoring candidates and imagining outcomes before real execution.

R4 Multimodal latent reasoning Research to integrate: text, code, and vision sharing a representation space for planning and verification — alongside the LLMs that still emit the final code and proofs.

R5 Spatial & embodied agents Agents that build and check things you can see: shaders, parts and scenes, verified without a GPU. Visual context as a future layer on the same agent stack.

R6 Wave compute substrates RF and photonic analog accelerators — matrix operations in electromagnetic waves instead of shuttled electrons. Near-sensor RF front-ends without conversion tax; photonic density where it pays. Algorithm–hardware co-design, not bigger GPUs alone.

R7 Distributed compute mesh A hive-style network where contributors parallelize training and inference across their own GPUs — gradient sync over the wire, fault-tolerant participation, credit-backed rewards. Datacenter-scale ambition without datacenter monopoly.

02 — Work

Research turns into things you can use. The LLM-powered editor ships first; deeper layers integrate as they prove out.

Orrery An agentic code editor powered by the LLMs you choose — parallel agents work in plain sight, split the goal, attempt it many ways, keep the best, merge it back. Available→

Arbiter A verification and selection layer for agent output — best-of-N judging, execution-based scoring, and an eval leaderboard that tracks solve-rate as models drift. In research

Seed Research toward a software world model — predict patch consequences and plan before executing, integrated alongside Orrery's LLM agents. Code first; formal reasoning and multimodal senses later. In research

Vellum A spatial agent layer that writes and verifies 3D work — shaders, CAD parts, game scenes — with renderer-neutral proof and no GPU in the loop. Preview→

Colony A federated compute mesh — volunteers contribute GPU cycles for training and inference on custom models; parallel gradients, async averaging, API credits for participation. Preview

Orrery is free to run locally. Paid cloud tiers — Pro $40, Max $100, Ultra $200 per month — bundle DeepSeek model quotas, bring-your-own-key, and the buddy narrator. See pricing →

03 — Approach

Four commitments that shape everything we build.

Temporary by default

Agents appear for a task and dissolve when it's done. No sprawl, no residue, no machinery left running — the namesake we build toward.

LLMs now, layers later

Language models write the code and drive the editor today. Research explores latent predictors and world models that sit alongside them — judging candidates, imagining consequences, cutting wasted runs.

Local and distributed

Capable on your own hardware by default. When scale is needed, a volunteer mesh spreads the load — many machines acting as one, without surrendering your work to a cloud monopoly.

Earned, not assumed

More compute only when it pays off, and every result checked before it's trusted. Emergence is measured, not promised.

Open to aligned funding — research grants and non-dilutive capital. Not equity or control that trades ownership or direction for money.

On the record

A commitment we're making before we've earned a cent — so you can hold us to it.

10%

a floor, not a ceiling

A lot of AI companies say beautiful things about "benefiting humanity." We'd rather put a number on it.

It won't happen overnight — we're not there yet. But the commitment is clear: over time, at least 10% of Ephemerent's profits go back — toward open research and real problems like disease, poverty, food and water security, and energy access; toward the community; and toward research on policies to prevent large-scale job loss.

That 10% is a floor, not a ceiling. It'll take time to get there — but we will.

The number is just one piece. The real goal: integrate AI so it lifts everyone — not just the people who own it. Amplify the upside. Confront the downsides head-on. Build the version of this future where everyone shares in it.

That's the work. That's Ephemerent.

— Ephemerent · on the record · 2026