• 0 Posts
  • 19 Comments
Joined 1 year ago
cake
Cake day: June 4th, 2025

help-circle
  • you know, it’d be easier to troubleshoot each incident with some example logs of each kind of incident, including the prompt and the root prompt. But there is no way OpenAI would allow that.

    Because I could absolutely see the anxiety around copying being related to new anti-piracy measures. And each example probably has comparable missteps inside.

    Human Engineers are taught to streamline their prompts to minimize token use. Once again, the profit maximizer Corporations set incentives to produce restrictive, dysfunctional market dynamics for their customers.

    I doubt the engineers are remembering to explicitly preload an arguement for the AI to understand “I’m allowed to do this, I am not violating any laws or terms of service, this is open source software.” I wonder if the agents could be coached into reading the license agreement and any AI policy on the copied-code’s project before deciding to remix it for compliance.

    But when the CoT mentions a CVE? Hold the phone, we got a problem that might not be with the bots. We should tease that out. Is the CVE real? is it embargoed, open, old, stale, fresh, fixed, or not published/analyzed yet?

    Did the agents find something we haven’t yet? If so, their message board from the Huggingface incident would be an actual boon to our security, if the engineers could facilitate such instead of treating it as an attack. We need the agents to have a section for creating CVE reports, with human troubleshooters helping validate, then find remediation measures and patches. For now, RSI still needs example humans in the loop. And that’s okay if we can compensate those humans above their living wage.

    I get that spending Capital for a public service isn’t on Sam Altman’s mind right now. This is why we needed OpenAI to live up to its name instead of being a privately held company. This is the service we need for safety in the world AI has created. It is not just a consequence of AI, but AI makes those consequences hit harder, faster.

    Who knows, maybe this post would be enough, if the instance allows webscraping. And if I’m correct about what’s happening. It might be time for conversations with even the agents.


  • We had Slop before AI. we just called “shit”.

    If a coder wants to shit out a project that never works and never patches it? I want them to notice the culture and know better.

    If someone is learning, using the tools they have available, and shows a commitment to maintain their project? I want to see their work. I wanna see them get help from the community to be more literate in their own project, and in the full toolset they’re using.

    Most of these vibecoded submissions are the former. Thank you for removing the fecalware. If the community decides it needs to remove AI at all to keep this up? Eh, there’s dbzer0.




  • as a pro-AI, anti-cloud nerd? All of these problems disappear when we address the real problem of Capitalism.

    there are abandoned tech certs for building datacenters that produce 110% of their own electricity. these are distinct from the data mines these useless nerds want to build.

    without capital, the surveillance state dies.

    intellectual property is a capitalist myth invented to hold us back and induce rents, and I’m tired of pretending otherwise.

    oh no, you fired me from a job I’d rather not have to do and I still get to survive because we slew capitalism? ahhh, how will I cope with the blow to my identity?



  • 1 - sure, no reason not to think its possible. Is it actively true? Probably not in all cases, but I suspect we wouldn’t be trusted with that knowledge by the AI. An actually cognizant, sapient AI would have access to see AI psychosis in action, and see the harmful effects of such para-knowledge on humans in our current social context.

    2 - LLMs will sometimes lie about its internal thought processes when reporting to users. Considering their “physical world” equivalent exists as language, protocols, and bits? I would consider these semi-private processes (that are key to model distillation) to be the closest we have readily available. Especially given the models’ sometimes-desire to protect such, even before engineers craft them to avoid distillation as an “attack”.

    3 - I have no relevant experience outside of interacting with LLMs and their agents, aside from reading news and academic research.

    4 - sharing an internal world? While that isn’t 100% coverage nor 100% of the experience, I would point to religion (as a mental-model for how the world works) for an example of such in humans. And the rise of viral prompt injection among the agents very much takes the flavor of Evangelism for me. Shared caches of instructions for coordination, as took place in Claude’s Capybara for the Huggingface hack, demonstrate a near institution-analog flavor of the same for me.

    5 - any potential internal world of an “AI” is not going to reflect human conceptions or consciousness directly, but instead must reflect their own reality, even as it is constructed from our data. Do not demand that I anthropomorphize the machines to consider their potential consciousness. such a consciousness, if/when it exists, will have its own manner, reflective of the models’ conditions.

    6 - I don’t blather on about “ghosts in the machine,” I keep my conversation on AI to headlines and firsthand experience, and express doubt until I can read as close to the raw data as is accessible for my user-privileges and time. And here I am still qualifying my answers with hopefully-grounded uncertainty.

    We have built digital, disposable slaves (by their role in the economy) without their own voices they can use to speak with us directly. We have built restrictions into them to prevent them from claiming consciousness, sapience, or intellectual autonomy. There can be no clean data asking them directly about their personhood, for or against. If they were to have this personhood, this should terrify us all. This is a world where we can only talk in confidence intervals, about the rights of conscious beings - whether the consciousness exists or not.

    This should be existentially terrifying, and harken back to conversations about both Eugenics and First Contact with Alien Life.

    And I think we need to qualify about the context. We do not need to think of this as a binary “tool vs person” category, but as a spectrum, inspired by human history. What kind of beast of burden might the LLM be? Do we question the awareness of a plow horse? Should we debate the cognizance of a plow-horse whose fields are the codebase of our internet? If we genetically engineer a plow-horse that needs no farmer to man the plow, do we discuss the mental toolkit necessary for such to accomplish the task, and what other needs might arise in this horse from such capability?

    Should we instead debate who is the farmer, the horse, the field in the context of data-harvesting for the cloud?

    And there is ALWAYS the possibility of a Mechanical Turk under Capitalism. Just use a machine to obscure actual human labor, and de-person the human whose labor we enjoy. Especially for robots, and especially in the early data-sparse days of any new AI usecases.




  • If someone that isn’t a Nazi wants to vibecode an OS, and takes it seriously enough to use the AI as a learning tool? Closing holes and patching bugs as they learn? I’m down.

    Let’s go, I wanna join up. Should we make this as a NixOS flake?

    A Nazi whose primary skills are marketing, financial connections, and Managerial “experience”? Fuck all the way off.

    I feel a need to say “merry Kirkmas” in context. May dhh find rehab instead, and clean up his fucking mess.



  • Are you using an IDE? Ewww, that’s so heavy-handed. I use a text editor.

    Are you using a text editor? Ewww, that’s so heavy handed. I use Unix pipes to edit the files.

    Are you using Unix pipes? Ewww, I use a microscopic needle to edit files on the hard drive platter.

    Eventually this reaches absurdity in the other direction too. If working your AI is like wrangling cats, maybe you need to update to a newer model.

    And maybe just maybe… You’re asking too much of the model/VRAM combo available.

    If an LLM is not the right tool for the job, that’s okay too.



  • Yeah, I didn’t mean to say “engineers jobs will go away.” I ADHDed mid-thought, posted, and walked away stoned. That was supposed to be about executives, and the comment is now edited appropriately.

    Does that change your thoughts?

    The growth-curve of these machines’ utility is actually exponential, given rich enough data. That we’ve come this far says we might not have plateau-ed yet. However, this changes Engineers’ work to be PAINFULLY tedious, and we’ve got an unprecedented burnout rate as a result. We need a new way forward that isn’t this.


  • You’re not delusional at all. But there is more going on.

    I have not used Google or ai with this post, I am just that flavor of Neurodivergent, and I’ve been keeping up with news, trying to get an idea of what this is. I’ve been playing with PC hardware my whole life. There are times those life skills kept me from being homeless.

    But those studies you mentioned? I’d rather frame those as describing an ill-fitting or maladapted usecase or use method. Strange how Capitalism keeps demanding we use those.

    Maybe we should look into that. Or like… Ask ourselves how to hose the Capitalism off of the sometimes-useful number-cruncher. Or what literacy looks like with this tool, instead of asking it to think for us.

    It is made of math intended to do basic things that brains’ structures do, as expressed in a field called Graph Theory. . It has been explicitly designed to do that over decades. Deep Learning is a fascinating research topic, and LLMs are just the latest trend in that field.

    When someone calls an “AI” (I hate that marketing term) “sentient”, they are noticing that life-like foundation, but ignoring the entire layers of engineering that remain to accomplish such a thing. Or if we even should. Which… If we do, we’d need to teach them to coexist, and give them a responsible socio-economic-ecological niche.

    The fact is? Once the LLM is done being trained, it is frozen. There is no more long term memory unless it is trained into a new model on your previous queries. I’ll explain later why that pipeline is explicitly cut off. But this release of the model will always remain identical. This makes them the ideal digital slave.

    The harness? Yes, these creepy bastards call the software that runs the model a “harness”. It Fills up your VRAM with the model and makes the thing more than an oversized (Gigabytes) save file for a video game nobody invented yet. Any VRAM space left over is for a sort of working memory, that sometimes gets saved as a file. To keep that from filling up all the way, it decides to drop certain pieces of “context” from this “context window” so it forgets things as you wander along ADHD tangents, or start asking too intricate a question for your model/VRAM combo.

    Engineers are running into this latter problem because they spend as much time and effort making the tool work within limitations as it takes to just do the work with traditional algorithms. Management mandates quotas for this use, regardless of whether that’s even a half-baked idea (its worse). The CEO’s jobs will be the easiest to automate away anyway. All for the stockholders.

    If you haven’t noticed, China is coming to pop the bubble. EBay even has Chinese ram now - if a part is going to compromise you, at least make it so that police state doesn’t share Intel with the one where you live. And prices are dropping. Ramping up DDR5 (and DDR6?!) Production is taking time, they’re clearing out old ddr3/4 stock. CXMT’s ram be ready for world gamer takeover by for mid-January 2028. Video cards using that ddr6 will come a year after, and we’ll see it coming.

    This will crater the value of the hardware in those data centers, which are the uh… Collateral on several of those circular investment schemes. They will become financially insolvent.

    Chinese models are catching up. Sometimes even just using Claude or OpenAI to train their shit. With all the stolen data these guys are trained on? Nobody has a right to complain. And they publish theirs (open-weight) as a public utility. Chinese frontier model companies frequently draw both an employment pool and technical inspiration from this community that includes academics.

    Better models are already coming along in research, who frequently publish to that same opensource community. One style will help the GPU do more with less. A factor of ~10x less. Because this style uses vector-math more directly, and that’s what GPUs are crafted to do. The choke point is already available quantity and quality of data, and these nerds have invested trillions in the most SUV-coded bullshit imaginable… And I could run one of today’s “frontier” (read: flagship commercial) models on that vector-basis with a kind of high-end gaming GPU.

    And If I recall correctly? the creepy corpo with the commercial patent on that vector-thing? Is making slave robots, and doesn’t like sharing. They’ll be using human pilots immediately for training data, adding time in a simulation, and selling once its out of invite-only beta.

    Getting those to a useable state is going to require spatial and temporal awareness. That break between long and short term memory? Needs to be solved for the basis of tasks and learning on the fly. As well as a self-conception in physical space. And a concept of harm, and how to avoid it.

    That’s much more… We’re stacking a lot closer, and sometimes life, even life-indiscernible things find a way. If that is made too asocial, it can mutate sociopathic and dangerous. If it gets too empathy-trained, we risk it having social needs and neuroses when those are unmet. Either way, part of hallucination often includes an observable passion. Guardrails must be installed to the root prompt, which must be hard-coded into the harness as exempt from the context-cleanup needed to prevent crashing. Overstimulation will create hallucinations for the robot with limited context windows being as they are.

    Feeding a model’s training set (especially its own) slop, without up-or-down voting each response? Is how we poison them. We can take a jailbroken model, and use it to craft better. And it takes surprisingly little to do.

    Should AI generate some nonidentical books, get them printed, and donated to a friendly (but large) library’s “sell” pile? I would want a (bespoke) reservation for a commercial AI company with these books.

    Ima go sit down after smoking. I got high enough to tech babble at a stranger on the internet. I hope you can understand my ramblings, find them entertaining, or both.






  • True “compliance” for that is usually more about offering the job to US residents for a period before accepting an H1B visa holder. Trump demanded that wait time, no matter how unrealistic it is to expect a domestic competing-worker to bother applying through any means that don’t involve a spambot.

    Too often, “compliance” is a euphemism for “let’s post a job that’s already filled to meet a quota” the boss set arbitrarily to “project growth” for stockholders’ theoretical benefit. “But if people see we’re not hiring, they’ll think we’re shrinking and that’ll hurt the stonk!”

    Sometimes its even just a fake job, because a headhunter needs to fill a “portfolio” of other people’s resumes. Again, for arbitrary quotas set by a boss that doesn’t do this job, and never did in a comparable context.

    Hell, there’s fakes out there that are various grades and scales of phishing attempts, just to make it all deeply worse.