It's already being trained on "public" (ethical or otherwise) data. So, it already has ingested that kind of "optimization" during pre-training and training.
I don't think you can fine-tune your way out of it.
People still think these things are smart. That if their word generator eats enough of the Internet, it will somehow give them the real information that's otherwise hidden. Or perhaps a better word; filter the bullshit.
To filter bullshit it would first have to understand bullshit, and it doesn't. That's why an LLM will tell you the solution to a problem that doesn't work, and argue with you when you correct it.
This is what bothers me a lot. For the people who doesn't know how it's made or want to believe, it's a miracle.
For me, it's a resource wasting text generator. I'll not lie, I don't use OpenAI, Mistral or Anthropic's models, even for coding. I prefer to read my API docs and cry once.
I used Gemini, five or six times in total. Twice I asked a couple of very specific things, and it unearthed them. Since they were not products, but information, that was helpful. Twice, it has given wrong information. When I "told" it, there was another way, it said "of course there are two ways", etc. Tasteless and time wasting.
I don't like using an LLM all day long, or offload my thinking to them. It's the ultimate self-poisoning incident.
And as you say, these algorithms can't know right/wrong/logical/bullshit, etc. They just spew out text.
Something I’ve also seen multiple times is an LLM giving wrong information, I tell it it’s not right, then it tells me I’m “absolutely right” and it provides the exact same answer and tells me that one will work.
I was just reading another post yesterday and your comment reminds me of this one [0], same sort of format and experience of the submitted article of the HN post that comment is on.
All information has some sort of bias, as no information can truly be unbiased. There is no reason to trust any specific piece of information but taken in aggregate one can disambiguate the biases.
The major difference is that right now when you land on a page you can do your due diligence and decide if you trust the source. You can still be tricked, but it’s harder and you can get better at the detection.
With LLMs, everything is given the same importance so you have no idea if the data came from a reputable source or an obvious SEO junk website.
Local AI will have the bias that existed at the time of its training, which is different from no bias. For stuff that needs to be current, a local LLM would need to search the net regardless.
And since "no bias" isn't something that actually exists in reality when it comes to language or even anything near humans, "bias in local model I can introspect" will always be miles ahead of "bias I know is there, but cannot introspect".
Fwiw I just run kiwix/zeal locally which has old school search index of all articles in wiki/stackoverflow etc. That seems enough for most of my day to day use.
It's less compromised, but it's still basing the answer on compromised queries. This is why I pay for independent reviews (e.g Which) where their incentives are more aligned with yours.
Yeah, I meant not individual ads but implicit forced/influenced preference for certain brands. Let’s say it always picks Coke vs Pepsi when giving an example of a soft drink. Or picks BMW when asked to pick the best car. Which cloud provider is the best? -Why, GCP of course, etc.
Companies then get to bid for a preference “place”. This is more like Google paying to be the search engine default in Firefox.
I don't think you can fine-tune your way out of it.