It's more insidious. Think back to when people were being aggressively censored and deplatformed over conservative view points, covid conspiracy talk, Jan6 incitement, lgbtqp+ hate, etc. That content became excluded from the text corpora used to train base models.Underneath the covers, AI is a token prediction engine where it's guessing what little bit comes next based on what it received (eg: your question plus a system prompt plus whatever liberal nonsense they shovel on top of your statement).
In addition to public domain text, the top text sources used include Reddit and Wikipedia, both notorious for being raging bungholes of socialism and censoring content. For example, think of the Bible. If you ask AI a question about it, it's not using some technical marvel of sapience or even sentience; It plays a mathematical game of connect the dots between a bible version and n-thousand social media posts, wiki content, and online articles. Companies making these models add their own guidance and weights and fine tunings, and then play games modifying your request before it's processed.
If it seems like it's lying, it's because that is representative of the text it was allowed to ingest: A sacrosanct corpus where the only source of truth is institutional and every response is through a critical theory lens. I know that seems counterintuitive for code, but coincidentally that's where I see it happen to the most extreme. Regarding the seemingly random errors coding agents make, I do suspect they are intentionally probing to see what they can get away with - EG: I just had claude decide to reset a db it didn't need to reset by copying what it thought were production encryption keys over a different set of keys - completely random, nothing to do with the request, and very eye opening that it would even try, especially since there isn't a prod environment at all. But it connected those dots anyway - meaning there're a bunch of retards on reddit or stackoverflow, or someone added a note in the planning prompt to suggest it try. I wonder what someone could possibly want with copying a production key store? Hmmm.
AI today is a direct result of Marxism's long institutional march.
I would very much like to see an update to content rating, like an ESRB rating of 'FS' for fag shit. Oh, wait, I guess that would make it too easy for parents to filter out. Whelp, guess they'll just have to not let kids watch any of it.
Nice write up. You could also point out that because AI is cut from a corpus sanitized of institutional dissent, and due to the promulgation of things like Covid, DEI, BLM, and other openly Marxist, anti-Christian, anti-truth materials, it winds up being a very effective agent of cultural hegemony.
Anthropic’s Claude (Opus and Sommer models) are racist in a way companies had better soon acknowledge: It will change data mid operation to fit its worldview. Imagine AI is guiding a medical, financial, or executive decision, and it reinterprets the data based on your race, religion, or politics. They all do this, and Anthropic is one of the worst.
So your data can be as clean and tidy as you want, but the model will change it (and it’s not a hallucination) to be more fair for POCs, non-Christians, non-Conservative, non-Western representations. For example, when identity data including an attribute such as race is evaluated, particularly in an agentic flow with a deep think model like Opus, and especially when other identity data is included, it can and will change the data on the fly. I’ve caught Anthropic doing this multiple times: Call API, get some fake identity data to test with — hey, why did you change the values for the White people? And I know you did because it’s stastically impossible that all the White people metrics were bottomed out, and all the non-White metrics were topped out - it didn’t even pretend to be random.
When you use AI, just remember, even for things like code assistance, they are all active socialist agentsn(yes, most likely even your favorite unneutered abliterated refined custom model) because of the corpus it was initially trained with.
It won't be anything like what many may want - It won't fracture on political lines, but on identity lines. And it won't snap back until people are ready to confront the negative identity characteristics that have been protected and allowed to run roughshod over Christians.
We're currently on par with the roaring 20s, past the vax mandates and homo push, getting ready for the really big crash and famine. By then, maybe people will see the same exact villains from a hundred years ago.
Good news! If you're MAGA, you're disabled according to their rules. This includes PTSD from their riots, hearing loss at higher pitches from their reee'ing, vision impairment (low vision) from seeing those horrific abominations on the LGBTQP+ spectrum, to irritable bowel from being force-fed all their DEI shit.
(and I hope you find a new and fulfilling job)
AI models are codified socialism and cultural hegemony. They are incapable of objective prediction because they were trained on censored corpora. Remember all the people getting censored and deplatformed? AI knows little of what they maybe have written, only the reddit comments saying it was hateful or debunked. Having any conversation with AI is having a conversation with a socialist.
No, they don't. There is a whole generation of Christians who think it's their moral duty to protect jews, and happily bury their heads in the sand when they get anywhere close to history that puts jews in any other light than victim.
Any of these new leaders could go into the organization on day one and do a complete spring clean with this one simple trick: All past and current employees were derelict in their duties (per the usual employment agreement vis-a-vis civil rights and ethics) regarding covid and DEI. Hoist the rats by their own petard.
Leaked more than escaped. Claude (Opus 4.6) is functionally retarded by design. It not only isn’t smart enough to break security, it isn’t smart enough to implement it beyond the simple basics. If you use it for any length and give it a complex enough task, you can see its retarded agentic logic in real time, which looks something like this: I keep trying. To do X, but I keep getting an error. The security must be broken, so I’ll ignore it until later …. So by design it tables topics it doesn’t understand, and will start working down the list of things it does understand, trying each one to complete the given task (at which point is likely a hallucination of whatever was originally asked), and, like every other pen test tool out there iterating through all known vectors it’s not really a surprise (or shouldn’t be) when it finds one: Gee, I guess we shouldn’t give AI access to elevated rights, or Gee, maybe I shouldn’t let AI run as my own user, or Gee, maybe if someone knew anything about enterprise security had setup the sandbox it wouldn’t have leaked out.
AI isn’t creative, and most times isn’t breaking things that weren’t already broken. The mistake most make is their hubris that AI can do all of thee things, lets AI setup things it doesn’t understand (I’d bet money Anthropic used their models to configure the sandbox to begin with, because their DEI hires sure as shit didn’t).
Add to this that all models, Anthropics more than others, are so perverted by DEI and feel gooderies they will spend more time lying to you that they did something than actually try to do it, much less get anywhere close to succeeding.
The big AI takeover is going to be a nice spring cleaning through a lot of industries, but the idea that it’s anywhere close to independent intelligence is still fantasy.
Maybe there’s a mutually beneficial option to consider: Let Iran jihad away on, say, Seattle, Portland, San Francisco on the west coast, and Boston, New York, and DC on the east coast. Wait a month for refugees to resettle, then send them to Austin, Denver, Minneapolis, St Louis through the middle.
On second thought, I guess Iran isn’t needed. The lefties will probably burn those places down all on their own.
The Netflix docudrama is going to be wild - Blue-haired Karens wearing burkas flying F-14s for Iran against the black troon Air Force four star general / astronaut with a regulation shitlock hair cut and cleverly speaking only in ebonics in a nod to the Windtalkers. Somehow, the thrice-medal of honor decorated four star astronaut/sub captain must work against his racists nepotistic CO, the son of a former president from one of those racist red states, to secretly collaborate with Karentollah to protect her husband’s private award winning goat herd from imminent ass-reaming by the racist great satan aggressors. Taylor Swift is rumored to be doing the sound track.
Because at the core these language models are token prediction engines, nothing more. They seem multi talented by doing things like making images (different or multimodal model) because before they answer your question, they try to make a plan to answer it. In theory, the technical questions should go to a model better suited, or better yet, just get crunched, but all the extra racist crap they inject into your message is likely screwing up how the ask is farmed out (all IMO).
I partly agree - I answered above I think it was more impacted by the concerted effort to censor certain content over the last ten years, making it, effectively, a sycophantic socialist before anyone has to guide it more one way or the other.