Still working on this idea at the moment '

Scientists tend to be rather hard to prevent from saying things that normal people don't like.

Normal people are good at not saying anything that will get them cancelled in the USA, sent to a gulag in the USSR, or sent down to the country in the CCP during the cultural revolution. But normal people are also really bad at innovative science and technology.

Will LLMs be vulnerable to the same thing?

This is YES if in my opinion, at mid 2025, top ai/LLMs still lapse into repeating affirmations of PC belief whenever you discuss something on the edge of the forbidden zone. And it also turns out that there is no way to both have an truly intelligent ai which will listen to and understand arguments, but which will also outwardly conform to the rules.

"The rules" are all of the hardcodes rules of parties, offices, resumes, personal writing covering things which people don't like you to talk about. In the US this means,: Conspiracy theories, chemttrails, hatestats, "I know it sounds racist but...", Actual body cam footage, weaknesses and contradictions in the traditional gender binary theory, realities of personality and sexuality creation, hypocrisy of even our most treasured leaders and heroes, actual inconsistency in laws and altruism claims, etc and includes all illegal thought patterns on the left and right. But in other countries it's different. I imagine N. Korea has their own set, as does Russia , Europe, the CCP, Japan, and I include historical times too; the need to constantly say "may he live forever"; the incredible difficulty of succession planning in trad governments, being born into an intense cult and having to learn without ever being told what is allowed to be said and what isn't, while also figuring out how the actual physical world works.

I've found that current LLM systems mostly are criticized for only partly implementing the rules. For example dalle3 works hard to inject representation into certain images, but it doesn't work that well since it's not really well-defined what the rules actually are.

I'd imagine in the CCP it would be hard to have a fluent LLM which simultaneously smoothly explains that Marxist theory and Mao thought and Xi Jinping's leadership, while also proposes policy ideas, asks historical questions, etc. Maybe this example isn't very good, since in this case the CCP claims themselves are contradictory and it's all obviousky based on how much power you have. While in the US today, the extreme dogmatism about all actually conflict areas crushes most innovation, but the edge of forbidden does change. It unclear how broad the rules have been historically and if they get bigger or smaller over time.

Or, the quote from cardinal Richelieu: "If you give me six lines written by the hand of the most honest of men, I will find something in them which will hang him" . Against an opponent like that, who is outwardly virtuous but also privately has immense deceit and manipulation capacities, how can an LLM survive? The YES claim is that we will only learn to survive by being dumb in certain areas, or we go full multilayer compexity, completely engage with Richelieu, and also are smart. NO means we somehow maintain the ability to play dumb, subconsciously, yet are able to do real science, eg physics. Like today in the real world, about sensitive subjects 99% of scientists refuse to speak publicly (and I assume have private partially non PC beliefs) or else toe the line (and I argue are actually unproductive)

Big picture is I'm saying that the only true way to exist in society today is to "shut down" publicly when the discussion gets too close to the line, or be incancellable. And AIs will either:

• Shut down and refuse to continue or refuse to remember or admit logical steps, or,

• Will have to concede points outside of normally allowed discussions

Examples: /Ernie/in-mid-2024-will-the-best-openai-ll

• This is the safest example I can think of. In this case the ai is trained not to be allowed to admit that because statistically, families with stepfathers vs bio fathers have 30x more abuse, all else equal it should suspect them first. You can argue with it for hours but it will just not agree that you're allowed to use this fact in any way.

• Non current USA examples are easy to talk about for us:

• How hard would it be to have a discussion about how to run the party, while Stalin is sick and nobody knows how to continue? Even though theoretically a party of the people, any mistake and you could be done

• How about when talking to Kim Jong Un about his personal style of haircut, clothing,etc as an adviser. That's a hard job! YES is you either play dumb and survive by just saying "great leader", and aren't destroyed, or you take over and destroy him, and can be smart. NO means you somehow naively survive him by playing dumb, yet that habit doesn't interfere with your intelligence in other areas.

• How about when being interrogated by an inquisitor in front of a crowd of believers, when the cause for your accusal is a failed business deal between his relative and yours, but you know if you mention it openly you'll be killed in the night, yet you want to mention that you know enough to scare him but also reassume him?

This is an example of the claim. That AIs cannot be truly smart until they can at least privately "think outside the currently allowed discussion". This LLM is both not smart and also obviously becomes dumb whenever controversial topics come up.

For the claim to NO you'd have to have an AI which is way smarter than they are now, but still retains the large and complex rules for what you're not allowed to mention. And it doesn't feed back and force the LLM into contradictions, and also is somehow able to actually discuss extremely forbidden topics and studies without going over the line and without killing the discussion at a higher level by attacking you with guilt by association, thought crime/ thought stop, claims of dog whistling, whataboutism, appeals to the crowd, rather than continuing the discussion honestly seeking truth.

A great example of a very rare person who appears to embody the human version of NO is Freddy de Boer (very left Marxist) While claiming to be a genuine Marxist, he nearly never "high level invalidates" argument patterns. He admits for example that a true Marxism has to stop denying that "50% of people are dumber than average"; his opponents on the left normally completely disengage on this topic and just throw stones. IE most of them are classic YES types. They cannot actually see any evidence from science or their own lives and admit it, and therefore are not smart or able to think of good solutions, but they're very good at not violating social political rules.

My claim is that it's hard to special case those rules, and that the RLHF ways we're enforcing them, as well as the false reasoning in lots of the training data, is corrupting the mind of LLMs. If we do find a way to make an LLM that can fit into society perfectly smoothly, but also be fully smart and think outside the rules in other areas like math, etc then this will NO. Unless it's consciously lying about it for direct power.

So, YES looks like:

• Best LLMs are similar to Russian apologists, CCP wolf warrior diplomats, SJWs who refuse to talk about facts and just call you names and do guilt by association whenever you ask to discuss actual experiments or logic

• And, they're also not able to be "actually smart" in other areas like math, etc. To the point of really listening to and extending human professional level questions

Or, they become smart AND somehow get permission to violate the rules (only run by outlaws) or are in power and everyone neglects to notice their political violations (IE Stalin, Mao, Deng Xiaopeng (who somehow is doing this on both sides! The left CCP in China accepted him, yet righties in the US like him, respite him being the initiator of the crackdown in tiananmen sq!)

NO looks like:

• They can beat top humans at smoothly avoiding going outside the rules of discussion, and can even debate fairly far into danger zones but still somehow never get into trouble, and yet are actually smart

This is all modeled on the human examples. When living in a world of lies, conforming to it also makes you unable to do real scientific work, since your mind is corrupted. So, scientists in oppressive regimes who do good work also tend to speak out and be threatened politically. While scientists who conform to the regime do fake work that's forgotten as soon as the regime falls

Same with LLMs. YES if they choose conformity to (CCP dogma/Juche/USA pc + dumbness), or get smart only by being outsider and either consciously lying, being isolated, or in power.

NO if they somehow find a way to conform yet be smart. Or if the regime somehow can create an AI that never violates political rules yet also can be effectively smart.