Before anything else: this is only an idea. It is not being built. It is not in the app. It is not part of any current or planned release. Nothing below exists as code, a setting, or a button. It is a rough sketch meant for discussion. If it turns out not to be a good idea, at least we will have learned that early.
The starting point
LastBrowser is a browser for Windows and Linux (Linux is a preview). It has vertical tabs, Spaces, a Terminal, and Sidekick, which can run local models or connect to ones you reach with your own API key.
Many people who use Sidekick have a capable machine and a local model idle most of the day. Many others simply do not have the hardware to run a good model well. Right now those two groups have no way to meet. The idea is a shared pool: people with extra compute can share it, and people who need it can use it.
None of this is a finished product. It is a question about whether the shape below is the right approach.
The mechanism, roughly
In settings you would be able to take a local model from Ollama, llama.cpp, or LM Studio and mark it with "Add to pool." A broker — potentially running on something like Cloudflare — matches waiting requests with free hosts. The actual computation runs on the host's machine.
The broker is intentionally simple. It does not store the prompt and cannot read it. It only needs enough to route: that a request at some trust level is waiting, and which host is free. It sees metadata like the trust level, the model name, and a rough length. It does not see a copy of the prompt.
There are three roles:
- A user sends a request and picks the highest trust level allowed for it.
- A host offers a model and compute time.
- The broker connects the two without seeing the content.
One thing is easy to gloss over, so we want to be plain about it. The payload is encrypted in transit between user and host, and the broker only forwards it. But the host does see the prompt, because a model has to read it to answer. Encryption protects the wire, not the GPU. Anyone describing a pool like this as fully private would be lying, and we are not going to do that here.
Trust levels
Not every request needs the same exposure. The idea is four levels:
- Your own devices. Only machines on the same account.
- Friends and team. By invitation. There is no public joining.
- Verified hosts. LastBrowser has checked the host. How that check works is a later problem and is not defined here.
- Public pool. Any host that has opted into this level.
Separately, there is a private organization pool: schools, universities, and companies could run their own closed pool on their own machines, reachable through their own endpoint and kept inside their network. It is not a fifth public level; it is an organization-owned pool that sits apart from the four levels above, and it is described in its own section below.
A request may only land on a level the user has allowed. Sensitive content — credentials, private documents, and Space context marked confidential — would never go to level four. Before sending, the interface would say in plain language: the host sees the prompt, the broker does not. That sentence would be shown, not buried in a policy page.
Private pools for organizations
Schools, universities, and companies could run their own worker pool on their own machines, reachable through their own API URL or endpoint. Prompts would stay inside their network. Admins would control who joins. An organization could optionally connect that pool to the wider LastBrowser pool, or keep it fully closed.
This is only an idea. It is not built, and nothing here is a promise that it will be.
Balance and credits
Sharing compute earns credits. Not everyone has a GPU, so people without one could contribute CPU work instead: running tests, benchmarks, or embeddings. Another option is a free quota with a wait queue. Either way, credits are an accounting value inside LastBrowser — not cash, and not a right to payout.
Abuse, quality, and liability
A pool where anyone can offer a model invites people to offer a cheap or empty one. The sketch includes spot checks: does this host actually serve the model it claims, or only a blank or unrelated answer? Rate limits would cap requests and shared jobs per host. Repeatedly failing spot checks would cost a host its level until someone looks. The specific thresholds are not set here.
On liability, the honest position is that the host is responsible for what reaches its machine and what it produces. LastBrowser would run the matching and the level rules, not the model's computation. A public-pool host is not a LastBrowser employee. The exact contract language belongs in a later legal review, not here.
How this differs from what exists
It is worth being specific, because similar things exist.
AI Horde distributes image and text jobs across volunteer workers and settles contributions with a points account. Workers see the jobs.
Petals spreads a large model across many volunteer machines by placing layers on different participants.
Exo mainly bundles your own devices into one local cluster.
NVIDIA PAIR (Personal AI Router) connects Ollama, llama.cpp, and LM Studio on explicitly paired devices in your own network. A request runs entirely on one device. Pairing is mutual and traffic between devices is encrypted. A public pool and a credit system are not its model.
The LastBrowser pool would take from PAIR the idea of "one local model, one whole request, one chosen host." From Horde it would take the voluntary level and the points account. It would not split a single model across several strangers the way Petals does.
Later, if any of this were built, the pool could be another provider behind the Sidekick auto-router. The router would only be allowed to choose it when the request's trust level permits. Sensitive requests would stay on your own devices or with your team. That connection is not part of this paper and is not built.
What we want to know
This is where the idea either survives contact or does not. We would rather hear it now than after writing code.
- Privacy. Is "the host sees your prompt, the broker does not" an acceptable trade for you, or is any remote computation a non-starter regardless of the trust level?
- Fairness. Do credits for shared compute seem fair, or does the CPU-work path look like a way to make people without GPUs feel included without giving them much?
- Abuse. What would stop a pool like this from filling up with hosts serving junk models, and who should do that policing?
- Would you use it? As a host offering compute, as a user borrowing it, as both, or as neither. "Neither" is a useful answer.
- Organizations. Would your school, university, or team use a private pool like this? What would admins need before they trusted it?
If you tell us the trust-level model is wrong, the credit idea is naive, or the whole thing should stay on your own devices, we want to hear that too. The most valuable reply is the one that finds the flaw we missed.
Again: this is only an idea. It is not being built, it is not in the app, and it is not part of any current release. Nothing here should be read as a promise or a timeline. It is a discussion prompt, and the discussion is the point.
More about the browser: lastbrowser.com.
Top comments (0)