Nostr Web Client

We do something similar with closed-source code, which is why closed-source is sort of pointless.

We do something called "penetrative testing", and just probe it and try out all the paths, and then guess the logic and recreate it.

That's why you can only demand payment for computation (dynamic, new effort), not code (static, past effort).

Code is only paid for in the creation, where it requires computation.

Chris 11mo ago

So bottom line (I’m deducting) is that it will be literally impossible over time (and cycles of iteration) to have “walled gardens” of closed source models. That generate subscription type revenue.

Put simpler, there is no way to stop AI/LLMs from becoming ubiquitous and nearly free as they learn and propagate across the planet as, inevitably, open source utilities….

As a layman, that seemed logical to me a few months ago. Wondering if the last week’s revelations are making that a reality in real time or am I just looking for my own confirmation bias….

Thx

Soberano 11mo ago

ChatGPT distilled internet and Deep Seek distilled ChatGPT, seems legit to me 🤷‍♂️

This is why it was so cheap. They didn't need any access to the underlying data, just to ChatGPT.

A copy of a copy.

I’ll be honest I am skeptical of the idea of training one model from another. It runs the risk of perpetuating invalid or incorrect or incomplete information. Currently this framework assumes the primary (or parent) model was trained effectively. To me that sounds like it could be a big assumption if this were to become the standard practice.

AU9913 11mo ago

This is literally the definition of a model though, they don't claim to have 100% accurate representations of reality. Internet is a model of human intelligence, chat GPT is a model of the internet, and so on

Indeed fair point. But I guess what I am trying to say is that models built on models built on models is where you start to take the abstraction too far.

AU9913 11mo ago

I disagree. I'm a devops engineer, meaning I build tools for developers. Developers who build tools for security engineers. My job is 3 layers of abstraction away from some one making a decision.

Abstraction and distillation help you make decisions faster and in a more decentralized manner.

For example, my farm. I don't actually know the health of my flock, but I can look at each sheep everyday for a few seconds. Over time I've built up a model of what healthy is and notoce when one is skinny or struggling but this is just a model abstracting a quantitative measure of each animal (fecal worm count, body weight, etc.) Into a qualitative one (the flock looks healthy).

Fair enough. Agree to disagree then in this regard (unless we’re talking past one another). I’m not saying abstraction is bad, just that built in assumptions build in risk at times. I’m in a similar industry and if the foundations are unstable, or found to be deficient in some area, and you replicate over and over from that point then you risk embedding weakness through-out the whole structure. Same is true for biology, if you breed and breed from unstable genetic pool at the base level then the further you get from that level the higher risk of undesirable developmental outcomes.

Abstraction = good

Assumption = (can be risky)

AU9913 11mo ago

I definitely agree.

I think this is why the true potential of these things is only unlocked when it's personalized, which is unfortunate because ATM I don't have funds do buy a new gpu and start getting one trained on my prompts and collecting responses.

But ideally, an organization could feed in its data and then start exploring that way. Simple example would be to ask all your employees to update their resumes and feed them in. Ask the ai, who would be the best person to make a decision around using ai at the company. Give me your top 10. Then have those people tell you who they think would be best to form a team about AI implementatiom at the company. Then feed all of that back in and come up with a single leader and 4 people around them to start piloting shit.

I would do this on the farm if I could ensure my privacy and collect my own data.

100%. Yeah this is where something like Unleashed.Chat could help as it is built on the framework of protecting IP and company info. I’d be worried about feeding OpenAI all of the staff resumes, that is a massive data privacy nightmare. But if in house you can train your own and ring fence the data then you are talking.

no₿ody™ 🧱 11mo ago

Thank you Mrs. Forest, i learned something today

mrclownworld 11mo ago

Rome did this to Carthage with boats

It just is what it is, man

https://en.wikipedia.org/wiki/The_Fox_and_the_Grapes

Yeah, but it's nothing new. They actually didn't innovate, in any way, they just did things that other people invented and were already doing, but for a Chinese audience.

nami 11mo ago

This is how it is

recent evidence suggests that the stories about cartage being a haven of child sacrificing whoring bitches have proven to be true, evidence has been found since about 2014 of this, they exterminated it like you would a rabbit warren in your lettuce garden... almost nobody escaped

they found clay vessels with infant skeletons in them, amongst other things, just like the romans claimed there was

Gunson 11mo ago

This sounds like cope.

Possibly they fine tuned DeepSeek on an OpenAI model (cheaper than using humans), but it makes no sense to primarily do this when self-supervised learning and RL is much more efficient. Also, DeepSeek performs better than OpenAI on several benchmarks - you can't achieve this purely by distilling a teacher model.

Likely they made a technical breakthrough and USA "AI tzar" is seething.

https://www.analyticsvidhya.com/blog/2025/01/how-deepseek-trained-ai-30-times-cheaper/

They have said themselves that they did not make a technical breakthrough. They just open-sourced everything.

Gunson 11mo ago

Here are some of the specific technical things they did to achieve lower costs. More efficient training procedure, memory compression, reliance on RL, low level code optimisation.

open source > stupid copyright bullshit

the best part of #deepsnek is that this is gonna crater all the closed source projects prospects for future funding

investers will be like, "closed source means expensive, pass"

oh, there was another thing that is gonna come out of this that is awesome too

AMD was lagging in the general purpose compute space despite their simpler, cheaper hardware and open source AI

now they will be looked at again for further cost benefits

AI is now at that stage in its development like when you are writing code and it works, but it's slow, expensive and a bit clunky

it works!

but now the optimizations start and the race is on for the most streamlined implementations

personally, i am looking forward to when someone builds a model out of the nostr and all of the web pages embedded in the links on it, this will be epic