+ Train and serve small liquid neural networks (CfC, LTC) and GRU + baselines that predict the next UI action, on workers you trust. + Full guide: c0mpute.com/temporal. +
+
+{`c0mpute infernet temporal serve --train
+c0mpute infernet temporal train --dataset bundle.json --output ./model --family cfc --baseline gru --worker --yes
+c0mpute infernet temporal predict --session s1 --model sha256: --input step.json`}
+
+ Infernet managed endpoints reserve one identified GPU operator diff --git a/apps/web/src/app/sitemap.ts b/apps/web/src/app/sitemap.ts index 43b6c2f..50dbcfe 100644 --- a/apps/web/src/app/sitemap.ts +++ b/apps/web/src/app/sitemap.ts @@ -11,6 +11,7 @@ const staticPages: MetadataRoute.Sitemap = [ { url: `${BASE}/agents`, priority: 0.8, changeFrequency: "monthly" }, { url: `${BASE}/docs`, priority: 0.9, changeFrequency: "weekly" }, { url: `${BASE}/plugins`, priority: 0.8, changeFrequency: "weekly" }, + { url: `${BASE}/temporal`, priority: 0.8, changeFrequency: "monthly" }, { url: `${BASE}/blog`, priority: 0.8, changeFrequency: "daily" }, { url: `${BASE}/status`, priority: 0.6, changeFrequency: "hourly" }, { url: `${BASE}/nodes`, priority: 0.7, changeFrequency: "hourly" }, diff --git a/apps/web/src/app/temporal/page.tsx b/apps/web/src/app/temporal/page.tsx new file mode 100644 index 0000000..785f041 --- /dev/null +++ b/apps/web/src/app/temporal/page.tsx @@ -0,0 +1,86 @@ +export const metadata = { + title: "temporal policies (liquid networks) — c0mpute", + description: + "Train and serve CfC/LTC liquid neural networks and GRU baselines on c0mpute workers: infernet.temporal.train and infernet.temporal.infer on the existing job and receipt protocol.", + alternates: { canonical: "https://c0mpute.com/temporal" }, +}; + +const Code = ({ children }: { children: string }) => ( +
+ {children}
+
+);
+
+export default function TemporalPage() {
+ return (
+ // liquid neural networks (CfC, LTC) and a GRU baseline, trained and served on c0mpute workers
++ Small temporal policies that learn a desktop task (drafting an email reply, composing a Reddit post, + reviewing LinkedIn notifications) from consented, redacted demonstrations recorded in{" "} + W3BS Training, then propose the next UI action.{" "} + Infernet runs the models; c0mpute finds a + worker, moves the job and carries the receipt. Same offer, bid, accept and receipt protocol as every other + workload. +
+{`curl -fsSL https://c0mpute.com/install.sh | sh # c0mpute 0.2.34+
+npm i -g @infernetprotocol/cli # infernet 0.1.66+
+infernet model runtime setup --yes # PyTorch + ncps, your choice
+c0mpute infernet temporal serve --train # accept training jobs
+c0mpute infernet temporal serve --model sha256: # serve a model`}
+ + A worker advertises only what its Infernet reports: runtime version, CPU / CUDA / Apple GPU, RAM bucket + and the model hashes it serves. It runs one training job at a time and never bids on a job without the + user's upload approval. +
+{`c0mpute infernet temporal train --dataset bundle.json --output ./model \\
+ --family cfc --baseline gru --worker --max-price 0.50 --yes`}
+
+ Without --yes it prints what would be sent and stops. --worker limits bids to
+ peers you trust. The model comes back only if its weights hash matches the worker's receipt. A
+ liquid family is picked only when it beats the GRU baseline.
+
{`c0mpute infernet temporal predict --session s1 --model sha256: --input step.json`}
+ + A session opens on one worker and stays there, because that worker holds its hidden state. Calls are + end-to-end encrypted (X25519, HKDF, AES-256-GCM) between you and that worker, answered only for the peer + that opened the session, run exactly once, and rejected if a step is late or repeated. +
++ Spec: DIP 0026 + {" · "} + Infernet docs + {" · "} + W3BS: teach an agent +
+