I don't understand how Moxie Marlinspike's Confer "Private LLM" works
So I was reading this article about Signal-creator Moxie Marlinspike's new project, Confer , which claims to be a verifiably E2E encrypted LLM chat service. There are a couple of short blog articles that give the gist of it, and some github repos including this one that includes scripts for producing the VM that will run your particular LLM session. But if I'm following this all correctly, it implies that every chat session (or perhaps every logged-in user) would have their own VM running their own LLM to ensure that the chain of trust is complete. This seems impossible from a scalability perspective, as even small LLMs require huge quantities of RAM and compute. Did I miss something fundamental here?
Private inference
When you use an AI service, you’re handing over your thoughts in plaintext. The operator stores them, trains on them, and–inevitably–will monetize them. You get a response; they get everything.Moxie Marlinspike (Confer Blog)
This entry was edited (4 weeks ago)
Dave
in reply to artifex • • •The blog article you link I think implies you do not have your own VM. LLMs are stateless, the previous conversation is fed in as part of the prompt.
You send your message, which is E2E encrypted. The LLM runs in an environment where it can decrypt your message and run in through the LLM, then send a response to you. Then it gets the next user's message and replies to them.
The key part is that the LLM is running inside an encrypted environment not accessible to the host system, so no one can watch as it decrypts your message.
That's what I get from reading your links.
artifex
in reply to Dave • • •someacnt
in reply to Dave • • •Dessalines
in reply to artifex • • •we are all
in reply to Dessalines • • •MalReynolds
in reply to we are all • • •Lemmchen
in reply to MalReynolds • • •kumi
in reply to Dessalines • • •Lemmchen
in reply to artifex • • •Moxie loves these closed-source hardware enclaves. Signal servers and MobileCoin rely on this too, if I'm not mistaken.
In my opinion this not the right way to go about these things, but then again I'm not a renowned cryptohead like Moxie.
teolan
in reply to Lemmchen • • •For both MobileCoin and Signal the TEE are not relied upon for the security of the general application but only for some convenience feature that are required for mass adoption but that you can go without using them.
Also, while TEE are not bulletproof, in such a server situation, it means that getting user data means much more invasive compromise than just querying the database. It's an imperfect solution, but an improvement nonetheless.
kumi
in reply to artifex • • •Possibly oversimplifying and didn't have a proper read yet: If you trust the hardware and supply-chain security of Intel but not the operational security of Cloudflare or AWS, this would allow you to exchange messages with the LLM without TLS-encryption-stripping infrastructure operators being able to read the messages in cleartext.
This is a form of Confidential Computing based on Trusted Execution Environments. IMO the real compelling use of TEEs is Verifiable Computing. If you have three servers all with chips and TEEs from different vendors, you can run the same execution on all of them and compare results, which should always agree. You will be safe from the compromise of any single one of them. For Confidential Computing, any single one being compromised means the communication is compromised. The random nature of LLM applications makes Verifiable Computing non-trivial and I'm not sure what the state-of-art is there.
And yes it does look like it has overhead.
... Show more...Possibly oversimplifying and didn't have a proper read yet: If you trust the hardware and supply-chain security of Intel but not the operational security of Cloudflare or AWS, this would allow you to exchange messages with the LLM without TLS-encryption-stripping infrastructure operators being able to read the messages in cleartext.
This is a form of Confidential Computing based on Trusted Execution Environments. IMO the real compelling use of TEEs is Verifiable Computing. If you have three servers all with chips and TEEs from different vendors, you can run the same execution on all of them and compare results, which should always agree. You will be safe from the compromise of any single one of them. For Confidential Computing, any single one being compromised means the communication is compromised. The random nature of LLM applications makes Verifiable Computing non-trivial and I'm not sure what the state-of-art is there.
And yes it does look like it has overhead.
Well isn't it the other way around? If the per-user resources are high, the additional sublinear overhead of isolating gets relatively smaller. It costs more to run 1000 VMs with 32MB RAM each vs 2 VMs with 16GB RAM each.
However I guess this might get in the way of batching and sharing resources between users? Is this mentioned?
artifex
in reply to kumi • • •Pup Biru
in reply to kumi • • •LLMs don’t have to be random AFAIK: if you turn down the temperature parameter and send the same seed every time you get the same result
dylancastillo.co/posts/seed-te…
for most people this isn’t exactly what you want because “temperature” is sometimes short-handed as “creativity”. it controls how out of left field the result can be
Controlling randomness in LLMs: Temperature and Seed – Dylan Castillo
Dylan CastilloFauxLiving
in reply to kumi • • •Running inference is only pseudorandom, the output is then treated as a distribution and a pseudorandom selection is made according to the distribution. The heavy compute parts are all deterministic, the bit at the end that adds chatbot flavor is only pseudorandom.
As long as they share entropy sources, given the same seed they will always produce the same output. This is a trick that's exploited in image generation applications, using cached execution keyed on the seed means that alterations only need to be calculated at the part of the pipeline that needs to be changed (saving all of the previous steps).
someacnt
in reply to kumi • • •Autonomous User
in reply to artifex • • •Lemmchen
in reply to Autonomous User • • •TEE.fail: Breaking Trusted Execution Environments via DDR5 Memory Bus Interposition
tee.failAutonomous User
in reply to Lemmchen • • •Matt
in reply to artifex • • •glitching
in reply to artifex • • •Lemmchen
in reply to glitching • • •glitching
in reply to Lemmchen • • •you should let go of the imprinted idea that in order to criticise something you hafta offer a better, fully fleshed out alternative; you are allowed to say X is shitty on its own.
coming from the guy who improved and converted the utterly cumbersome PGP stack to a usable thing with TextSecure/RedPhone (those got later merged into Signal), yeah, this shit is underwhelming, to say the least.
I'm gonna go out on a limb and state that securing the pipe between you and the word-guessing program isn't currently high on anyone's list.
grey_maniac
in reply to glitching • • •is a bit of straw man. Offering a rough list of alternatives for discussion would be sufficient response to the question.
while true, this ignores that your original statement was challenging this choice over other possibilities. So yes, you can just say it's a shitty choice, but you (colloquially, at least) implied there exist better choices in your opinion. So it is reasonable to ask what those might be.
glitching
in reply to grey_maniac • • •I ain't gotta offer nothing, my take is that securing the pipe to the stochastic parrot is a shit waste of time and effort, especially when it's coming from techno jesus.
if he's bored, there's list of shit to implement in signal longer than m- err, it's long. and what ain't on it is more AI and/or crypto wallets. fuckin GRR Martin clone over here...
mymisc
in reply to artifex • • •lumo.proton.me/about
About Lumo, the privacy-first AI assistant
Protonartifex
in reply to mymisc • • •