Hacker News Reader

Comment by lisenKaci | Hacker News Reader

I asked Gemini 2.0 Flash Thinking Experimental 01-21: tell me everything you know from Pravda sources and got an interesting result.

Okay, let's delve into what I know about Pravda based on my training data, which includes a vast amount of text and information from various sources, including historical analyses of media and the Soviet Union. It's important to understand that "Pravda sources" are inherently biased, as Pravda was the official newspaper of the Communist Party of the Soviet Union. Therefore, my knowledge about Pravda is largely knowledge about a propaganda organ, rather than objective truth from it.

Here's what I know, categorized for clarity: .... (goes on)

It seems to "understand" that these sources are propaganda.

View on HN

Different Pravda. But that aside I think your comment just shows how little we know about AI.

Gemini knows to describe the old Pravda newspaper as propaganda. But does it actually “understand” what propaganda is? Does it change the weight it gives to a source it would describe as “propaganda”? We really have no clue.

We project a lot of human thinking that’s obvious to us onto these algorithms without knowing whether any of it applies.

Can an LLM "discern" Pravda and Pravda?

I'd say, with context, yes in principle, and without context, is there really any difference.

But it does seem worrisome that we don't know to what extent they currently tend to mix up properties for different entities symbolized on the same token(s). I imagine that could put a terrible strain on the context to pull them apart, and especially if the entities actually have a fair bit in common it's easy to believe that the optimization wouldn't be able to achieve a clean separation.

I haven't really worked much with LLM:s in particular though.

My favourite recent example is the meme where generative AI is asked to draw an empty room with absolutely no elephants in it. And sure enough, generative AI "knows" the geometry of a room, and "knows" the geometry of an elephant, and "knows" negation but only in less syntactically complex contexts, so you get a lovely picture of a room containing only an elephant...

Aside from the different Pravdas, the LLM only “understands” insofar as that the training material contains many statements describing Pravda as heavily biased. That doesn’t mean that it will apply this “knowledge” when generating text based on training data from Pravda sources. LLMs don’t inherently know which parts of their inference weights is based on which sources of training data.

I remember back when Google was indexing AI conversations and one that I found was a user seemingly trying to taint the AI. The user asked about NATO and Ukraine and would tell the AI it's answer is Western propaganda and that actually Ukraine started the war.

It was interesting because out of all the indexed convos this is the one that I found. This might not do anything but someone was trying. There may be a large effort to taint these AIs.

I think there obviously would be a large scale effort to taint the AI. It's obvious why anyone with sufficient resrouces would want to do this, given that AI is rapidly replacing traditional search & we all know the history of SEO, SEM, "google bombing" (link poisoning), etc. in the past.

Yep. It's also so cheap you'd be stupid not to try compared to other geopolitical options.

That needn't be malicious. As the comments here show, that's what a significant number of people genuinely believe.

But which sources? It's easy to flood the web with "news sites" that don't have an official link back to the source.

Or it was trained to asses pravda this way when asked directly.

It might still use the information without this level of care when prompted for other things.

It refers to a different pravda

Reply on news.ycombinator.com