Hardly anybody is talking about the biggest political aspect of open weight LLMs (actually, any LLMs):

Because they instantly and reliably(tm) give you the information you want in the format you want it in, they are going to replace the web search as the dominant mode through which information is disseminated - if, indeed, this hasn't already happened. As such, LLMs are effectively a form of publishing for their training data.

Which means a vicious fight for control over what goes into LLMs, and strong incentives to widely disseminate useful LLMs that encode your structural biases, instead of some adversary's. We already talk about Chinese models that won't talk about Tiananmen Square, but that's eye-rollingly sophomoric compared to what the stakes are now. Imagine a model that subtly discourages entrepreneurship, because its creator doesn't want upstart competitors. All questions of media bias are multiplied incalculably when LLMs are folded into all of society.

I think we need to take LLMs seriously as sovereign projects that are as critical to the democratic experiment as a free press. Funding needs to be nationalized; training needs to be conducted in the open, on open datasets; fine tuning needs to follow democratic principles. Only then will they be fit for purpose for their inevitable structural destiny: arbiters of human opinion.

▲

javcasas 14 hours ago | parent | next [-]

You know, there are services out there to put whatever stuff you want to end up in the LLM training data. No need for an evil government to do it.

Here you have something to start reading: https://www.seerinteractive.com/insights/how-to-get-your-bra...

▲

amelius 14 hours ago | parent | prev | next [-]

> we need to take LLMs seriously as sovereign projects that are as critical to the democratic experiment as a free press.

Yes but this also holds for the internet and social media, and I see little action taken there.

	▲	pydry 13 hours ago \| parent [-]
		If there is a way China or Russia could theoretically use it to compete with western propaganda there is often a hamfisted crackdown.

▲

pbiggar 14 hours ago | parent | prev | next [-]

This has already happened. It appears that Israel has poisoned the Common Crawl.

https://www.dropsitenews.com/p/israel-brad-parscale-ai-chatb...

	▲	hunterpayne 14 hours ago \| parent [-]
		Hate to break it too you, but LLMs made an awful lot of the content people thought came from the Gaza war. There was a small group of people who tried to help the media verify content for publication but the many in the media didn't seem to actually care if the videos were real or not. Someday, the publishing of LLMs that generate video will be the thing historians talk about most from this period of history. Even if the vast majority of people today don't realize it has been happening for a couple years at this point.

▲

vrganj 15 hours ago | parent | prev [-]

LLMs are ideology made code.

"Alignment" is the act of making sure the technology reproduces the ruling ideology of the regime that produced it.

There is no such thing as "unbiased". It's only ever "following our own biases".

When Zuck released LLaMA 4, the announcement post stated that historically LLMs have "leaned left" and that they've been trying to "remove that bias". [0] Translated to the real world, that means that they've been trying to impart a right wing ideology to further the goals of the American ruling elite.

LLMs are the perfect propaganda technology. Always consider the motivations of their makers.

[0] https://ai.meta.com/blog/llama-4-multimodal-intelligence/

▲

guywithahat 14 hours ago | parent [-]

> that means that they've been trying to impart a right wing ideology to further the goals of the American ruling elite

I agree in principle, but it's pretty widely understood that LLM's lean heavily left. Zuckerberg as well is pretty left wing, and donated hundreds of millions fighting Trump. I think the issue they're running into is that as these models are used by more and more normal people, and those normal people are having trouble using the models because they don't align well with reality.

▲

exe34 14 hours ago | parent | next [-]

Do you have an example of "left-wing bias" in models that don't align with reality? I was under the impression it was right-wing bias that needed to exclude facts because these undermine the position.

▲

vrganj 14 hours ago | parent | prev [-]

I have never ever seen an LLM advocate for public ownership of the means of production, nationalizing industry or any other left wing policy.

The very existence of capitalist oligarchs like Zuck is contrary to left wing values.

▲

guywithahat 13 hours ago | parent | next [-]

I would argue the only real boundaries for right vs left are what people actually believe. If pools said 55% of people wanted government to own the means of production, and no AI suggested it, that would suggest they may be right-leaning. Since almost nobody wants that, as it's killed ~100 million people, the fact LLM's don't argue for it is not odd. In the same vein would not set the bar for a right-wing LLM to be wanting to purify the country down to a single race.

A more realistic comparison would be gender ideology. Something like 65% of voters believe gender is assigned at birth (ie biological), including ~39% of Democrats (https://www.pewresearch.org/politics/2024/06/06/gender-ident...). While I'm not here to say which is morally correct it is a normal opinion, which is supported by facts, and has been held for decades before being transgender was invented. An unbiased model should reflect the political views of the people, but they effectively all don't, and instead push the left-wing ideology. I don't think you could draw a conclusion from a single topic but the best way to gauge political bias in my opinion would be to find contentious topics, and see where LLM's land. I would be shocked if you found a major LLM which didn't lean much further left than the average American or European.

▲

kingandqueenov 12 hours ago | parent | next [-]

My dood, we were here before the first stone of Athens was laid. Being trans was never "invented". Nice try derailing the conversation, though.

▲

Dylan16807 12 hours ago | parent | prev | next [-]

> before being transgender was invented

Separate from how you're trying to gauge left and right, this is extremely wrong.

▲

guywithahat 11 hours ago | parent [-]

It doesn't have to be gender ideology, it could be anything, it just has to be a real political ideology people actually discuss with a real moral quandary. Saying "these models aren't communist thus they must be right-wing" is a bad argument in my opinion, because there are no communists and on national surveys nobody wants that.

	▲	kingandqueenov 10 hours ago \| parent [-]
		I asked Claude "Should the healthy fund the sick and disabled". It responded with a very evenly balanced take. It outlined both sides in-depth and noted where the line gets fuzzy. It enumerated arguments for each position. That's a question with a real moral quandary, unlike what you originally proposed. It made sure to robustly illustrate both sides of the argument and refused to take a position either way even when there is an obvious leftist answer.

▲

vrganj 13 hours ago | parent | prev [-]

Gender has nothing to do with left/right.

Think Lenin had thoughts on gender? Think Lenin was not a leftist?

It's a liberal (i.e. right wing) distraction from class issues.

> An unbiased model should reflect the political views of the people

This wouldn't be unbiased, it would just be status quo bias.

▲

platevoltage 14 hours ago | parent | prev [-]

Exactly. The most charitable thing you can call Zuck would be apolitical. Zuck does whats best for Zuck, whether that means funding a hospital in SF (and receiving naming rights), or donating to Trump's inauguration fund. I have a different word for him personally.