Great write-up Jim. Great working with you too. I look forward to doing more together.

it doesn't give me the sats, but great work 🤝

aqstr is b0rked, I can't boost or like it, but it's a kewl idea nonetheless.

thx fam

No pay for boosts not working on aqstr

Trying to repost and like at AQSTR but dont work

Failed to complete like: Payment failed: Unknown error

but who owns the hardware and software running the LLMs?

good question

We use PPQ which typical proxies to the big providers APIs typically. It is possible to switch to TEE in secure enclaves. That decouples the model itself from the person runnung the hardware and adds strong privacy. The hardest point is getting open source video models but imo thats the new frontier and it will eventually develop

maybe this shows how old i am but all of this looks like mainframes all over again.

also PPQ sounds like a Walther pistol model

peepee queue

there are some similarities but the tech is moving faster. Deepseek has published a lot of techniques that are just starting to influence the way people make models for local usage. When that happens that's good for freedom privacy and cost competitiveness.

claude explains why describing the technology as moving fast is ridiculous: the core is just the associative memory / content-addressable lookup that hopfield described in 1982, and before that the basic dot-product similarity measure goes back to the 60s information retrieval literature. the softmax normalization is boltzmann weighting from statistical mechanics, 1868. backpropagation is from the 70s. the feedforward network layers are perceptrons from 1958. the 2017 contribution was removing the recurrence (no more RNNs/LSTMs processing sequentially) and showing that pure attention over the whole sequence, parallelized on GPUs, scaled better. that's an engineering insight, not a mathematical one. "what if we just computed all the pairwise similarities at once instead of sequentially" - which is only an insight if you have the hardware to brute-force it.

Yea if you ask a leading question you will bias the neural nets and it will tell you what you want to hear. Way to totally miss my point lmao.

my point is that the progress has not actually passed 2017. all that has changed is the cost of brute forcing has fallen low enough that the results happen in a small enough time frame to be useful by the time the predictions the model makes have passed under the bridge. and you don't address the point at all that you are literally sharing your private data with a third party. this is literally honeypotting dissident's lives.

the metadata about the activity is the signal. the concrete data doesn't have to be decoded in order to know who is talking.

Deepseek just collapsed the cost of inference by 90% Year Over Year. You can't say that I'm "ridiculous" for saying the technology is moving fast. The dissidents' identities are passing through at least two layers of proxies. That's the best I could do in 28 hours. There are more techniques to employ but good luck to Rwanda deanonymizing them with no payment traces and servers run by US corporations lol By the way the project is open source. If you think you're hot shit you can contribute (you won't). Here's the source code knock yourself out Armchair QB:

GitHub

GitHub - zuka-org/zuka-pwa: Zuka persona PWA app.

Zuka persona PWA app. Contribute to zuka-org/zuka-pwa development by creating an account on GitHub.

cheaper brute force is still brute force, and cheaper honeypots are still honeypots. "two layers of proxies" protecting rwandan dissidents is exactly the kind of confidence that gets people killed when the threat model turns out to be wrong. as it happens, i am in the process of developing an algorithm that hits higher precision than LLMs, without compute cost, only memory bandwidth. right at this moment i'm approaching a method that enables me to precisely measure the ratio of precision between japanese and english. deepseek is moving in this direction also but it is limited by floating point error rates. mine is integer based.

What is it about

that was actually me posting that, i just found a bug in smesh.lol web app with the notification system that then revealed on crash recovery with refresh that there was another bug in the signer vault. very nice. i wasn't even debugging the app and i just found and nailed two bugs by accident