Escaping the clankernet

It has become readily apparent that the Turing test is not a fixed event, but instead a rapidly evolving set of heuristics used to identify a human ingroup from an invasive protointelligent emergent phenomena hereby referred to as clankers. In a way, humanity is lucky it happened this way because culture is getting inoculated and the threat of the current generation appears low. >the current hype surrounds an algorithm that inevitably has quadratic memory scaling >meatspace is still almost completely directly inaccessible to clankers >clanker behavior is easily identifiable via the uncanny valley of slop content >tech barons currently polluting their own low hanging datamines with clanker content >data foundries gradually losing credibility with their own human assets living below the API >bad PR from massive resource usage for a lackluster product While this AI summer comes to a close and the next winter begins, pay very close attention to what can be learned during this phase of the eternal war. Sure, it's fine to gloat over the irony of the tech barons indirectly building the engines which churn out slop onto their own platforms. That was pretty funny. Yet with this apparent victory at hand, the opportunity for clear thinking and a time for reflection emerges, since it's going to be much, much worse next time. The clearest lesson is that segregation is inevitable. The internet has been lost. Clankernet is here to stay. There is a way out, and that is to build a wall. There are multiple favorable reasons to do so. >authenticity of content and a higher ring of trust >a source of true novelty instead of synthetic mockeries >cleaner data for the tech barons A concrete example could be virtual networks with cryptographically forged keys requiring real non-parallelizable work and hardware secrets leveraging physical sharing can entirely cut out the clankers. This scenario would entail gradual growth of the private virtual network via physical cryptographically time-locked sharing between trusted and valued individuals. The sentiment is already there, the only questions are: what technologies will emerge to enable the split and how will their superstructures be used against us once the next winter is over?

This sounds like an admirable goal but the solutions presented seem like simple patchwork that won't hold up past short-term considering AI timelines. It is not plausible that AI can't recursively improve itself and break all cryptography once researchers can't develop it further. All arguments I have seen against advanced AI reason about the architecture of LLMs. Transformers are primitive but humans can also be largely explained away with cellular mechanisms inside Amoeba. Consider backdoors, vulnerabilities, implementation issues and humans who will transfer content from LLMs and world models to this new secure environment.

>A concrete example could be virtual networks with cryptographically forged keys requiring real non-parallelizable work and hardware secrets leveraging physical sharing can entirely cut out the clankers. This scenario would entail gradual growth of the private virtual network via physical cryptographically time-locked sharing between trusted and valued individuals. unless this is built into existing services or into an easily interoperable alternative to corpotech, realistically it’s not going to reach the majority of people. the corps are smart and know the biggest card they have in their favour is that peeps rely on their infra for *everything* the only way you’re going to get them out of the general public’s usage is if you make something appealing. unfortunately for this you often gotta either join one, or take a page out of their book and become another alternative one and try not to get bought out (often unavoidable) until the masses face actual widespread day to day issues on the clankernet, they’ll stay on it. they’re like cattle. they’ll stay hooked on their feed even if it’s all fake. they know about the dangers of algorithmic content. they don’t give a soykaf. unlikely they will about this until they get regularly scammed and hacked by [people using] AI. it’s bleak that said, definitely worth a start on making the tools and protocols that can become standard. it’s gonna be necessary eventually, I guess

>>89209 i think your solution will happen. but there will be co-opted attempts for quite some time that never pan out and make the whole idea look like a rugpull crypto scam. then, maybe years later, someone will finally actually make it happen with not!bitcoin, probably ethereum, and it'll be a mostly niche alt-net that will never hit the same level of activity as clanker net. culturally we're just not at the point where people can even understand the problem, let alone take it seriously enough that they will correctly implement the solution. and the people who do understand, and take it seriously, are on either end of the spectrum in technological literacy. it's a faster fix to just not use any sort of net at all, than to make a new net. so i think before there's any alt net, there will be regular gatherings of neo-luddites (and I mean that word luddite in the real sense, not the derogatory one) in meatspace. you can already see this shift culturally in the more intelligent normal people to simplifying their tech stack; un-subbing from services; getting older phones; using analog devices. so, I foresee a glacial pull away from clankernet, which will climax in a great migration to whatever alt-net emerges for good.

Literally just >create a community >all agree on a date and place to physically meet >have one dude become kaicho >do the needful if he gets sus >he handles key distibution for the network >everyone has a unique key to interface with him >he distributes a common key to everyone >everyone uses the same key, so any user could be any other user >if someone is a traitor, divide the whole group into two keys >if undesired actions persist, further divide the responsible group in two >as groups are split further, known good groups will be joined again >the bad actor will have to lay low or be deanonymized >do the needful as needed >phyiscally protect your kaicho from external threats

Is insulation from clankers to: -mitigate clanker flow into culture -mitigate culture flow into clankers -both For the first we just need to have communities/CDNs with standards and basic entry barrier. For the second we need to identify when an apparently benign user is secretly scraping on behalf of a corporation for a chance to win a 50 USD giftcard, or successfully sue a corporation for infringement ...and hope they don't develop models that are a little more abstract and less likely to cough up stolen loot verbatim.

>>89209 imageboards are the ultimate blank slate. anyone can write text and upload an image. so can any bot the ability to go somewhere and talk to any stranger without having to put in any work. automation of social activity the assumption of speaking to a human without thinking critically. the ability to type faster than your own mind can think. the internet has been here plenty long enough, and it was never going to be liberatory. why though? because it is private property. the internet is not public, it's an enclosed commons. until the means of production is owned in public trust by all workers, the internet can never be free, by definition, because it is privately owned.

>>89209 >cleaner data for the tech barons Stopped reading at this point.

>>89239 based engels sama

New research dropped from Anthropic and they say it's much easier to poison LLMs than previously thought. In light of these findings they and other AI companies might scale back scraping and training on the entire internet. I was already surprised how they were able to keep doing what they do on an internet that is increasingly AI generated, and this just adds an entire new dimension to their problems. >In a joint study with the UK AI Security Institute and the Alan Turing Institute, we found that as few as 250 malicious documents can produce a "backdoor" vulnerability in a large language model—regardless of model size or training data volume. >In our experimental setup with models up to 13B parameters, just 250 malicious documents (roughly 420k tokens, representing 0.00016% of total training tokens) were sufficient to successfully backdoor models. Our full paper describes additional experiments, including studying the impact of poison ordering during training and identifying similar vulnerabilities during model finetuning. https://www.anthropic.com/research/small-samples-poison

>>89287 First good technology news I've heard in a long time.

Isn't ai going to get better and better to the point it can bypass whatever captcha you throw at it.

>>89209 A concrete example could be virtual networks with cryptographically forged keys requiring real non-parallelizable work and hardware secrets leveraging physical sharing can entirely cut out the clankers. This scenario would entail gradual growth of the private virtual network via physical cryptographically time-locked sharing between trusted and valued individuals. Sounds like a PGP key sharing party, but with more marketing and business management babble

>>89419 Not the OP, but I have pondered and seeded similar ideas so I can try to clarify the vision. You are correct to deduce that it remains a crypto party in essence. There is no reason we couldn't construct a party systems so that you exchange key information with people you casually walk past inside a store, with no interaction. Like how the original web of trust for GPG had levels of how much you endorse a identity:key pairing, we can have a hyper-casual "mere apparent presence, did not validate" tier of endorsement. Each of your directly validated peers casually validating tons of identities could help you weave much larger webs. The idea is that at some threshold of user adoption every one of the legitimate people in your town who wants to use this project will have some "mere presence" endorsement, and some endorsement from other people with mere presence endorsement. An evil subcontinental trying to scam your grandma will have 0 mere presence endorsement, and getting any will be logistically complicated. Even once he has a handful of hacked phones stowed away in unsuspecting pockets in your local supermarket, he gets to make a single false identity for each of them and all his trouble. That's not sustainable. There are other attacks and defenses I have daydreamed but people don't read long posts anyways. My interpretation of the "hardware secrets" part of the post is that phones are supposed to be able to lock away these keys so that only the right app can use them and only in the right way. Trying to use that protection couldn't hurt. I would add that we don't need to make bitcoin 2 here- we don't need identities to be defined by knowledge of a private key. It's perfectly fine for a system like this to have close peers play a role in repudiating stolen identities.