VolksAI
A concept for a local model compute-sharing pool that uses a broker to match idle local compute with compute demand. It employs a four-tier trust model, and the broker does not access prompt content.
Timeline (last 90 days)
The LastBrowser team proposes a concept called VolksAI, letting users with idle compute share local models for users with insufficient hardware to call; the concept is not yet launched and has no code implementation.
VolksAI's broker is designed not to store or read prompt content, obtaining only the metadata needed for routing (trust level, model name, approximate length).
VolksAI acknowledges that the host can inevitably see the prompt in plaintext, because the model must read it in order to answer; encryption protects only the transmission link, not the GPU side.
VolksAI defines a four-level trust model: one's own devices, friends and team, verified hosts, and the public pool, plus a private organization pool independent of the four levels.
Sensitive content in VolksAI (credentials, private documents, Space context marked confidential) never goes to the public pool level.
The credits earned by sharing compute in VolksAI are merely internal accounting values within LastBrowser, not cash, and carry no right of withdrawal.
VolksAI's abuse protection mechanisms include spot checks (verifying whether a host really provides the model it claims), rate limiting, and demotion of hosts that repeatedly fail spot checks.
VolksAI's position is that the host is responsible for the content it receives and the content it produces on its machine, and LastBrowser is responsible only for matching and level rules; public pool hosts are not its employees.
VolksAI adopts a technical route of task-level crowdsourcing plus complete inference on a single machine, borrowing PAIR's single-machine complete-request model and Horde's voluntary participation and credit accounts, but unlike Petals it does not split the model across multiple strangers' machines.
All Facts (9)
VolksAI adopts a technical route of task-level crowdsourcing plus complete inference on a single machine, borrowing PAIR's single-machine complete-request model and Horde's voluntary participation and credit accounts, but unlike Petals it does not split the model across multiple strangers' machines.
50%UnverifiedVolksAI's broker is designed not to store or read prompt content, obtaining only the metadata needed for routing (trust level, model name, approximate length).
50%UnverifiedVolksAI acknowledges that the host can inevitably see the prompt in plaintext, because the model must read it in order to answer; encryption protects only the transmission link, not the GPU side.
50%UnverifiedVolksAI defines a four-level trust model: one's own devices, friends and team, verified hosts, and the public pool, plus a private organization pool independent of the four levels.
50%UnverifiedSensitive content in VolksAI (credentials, private documents, Space context marked confidential) never goes to the public pool level.
50%UnverifiedThe credits earned by sharing compute in VolksAI are merely internal accounting values within LastBrowser, not cash, and carry no right of withdrawal.
50%UnverifiedVolksAI's abuse protection mechanisms include spot checks (verifying whether a host really provides the model it claims), rate limiting, and demotion of hosts that repeatedly fail spot checks.
50%UnverifiedVolksAI's position is that the host is responsible for the content it receives and the content it produces on its machine, and LastBrowser is responsible only for matching and level rules; public pool hosts are not its employees.
50%UnverifiedThe LastBrowser team proposes a concept called VolksAI, letting users with idle compute share local models for users with insufficient hardware to call; the concept is not yet launched and has no code implementation.
50%