I know this is possible, but the further away I get from my core domains, the harder it is for me to use these tools in a way that doesn’t feel like too much blind faith (even if it works!)
Like if you have a friend who's very well-read and talkative but is also extremely confident and loves the sound of their own voice. You quickly learn to treat them as a source of probably-correct information, but only part of they way you learn any given topic.
I do this with LLMs all the time: I'm constantly asking them clarifying questions about things, but I always assume that they might be making mistakes or feeding me convincing sounding half-truths or even full hallucinations.
Being good at mixing together information from a variety of sources - of different levels of accuracy - is key to learning anything well.
Let's say you're trying to get a university degree, but having a professor who makes up 20% of what they say. Is that helping you "learn well"?
And 20% is way overstated, esp for a SOTA model when it comes to verifiable facts.
It's easy to laugh and say, well I'm smart enough to defeat this. I know the trick. I'll just mentally discount this information so that I'm not unduly influenced by it. But I suspect you are over-indexing on fields where you are legitimately an expert—where your expertise gives you a good defense against this effect. Your expertise works as a filter because you can quickly discard bad information. In contrast, in any area where you're not an expert, you have to first hold the information in your head before you can evaluate it. The longer you do that, the higher the risk you integrate whatever information you're given before you can evaluate it for truthfulness.
But this all assumes a high degree of motivation and effort. Like the opening to this article says, all empirical evidence clearly points in the direction of people simply not trying when they don't need to.
Personally, I solve the problem in my friend circle by avoiding overconfident people and cultivating friendships among people who have a good understanding of their own certainty and degree of expertise, and the humility to admit when they don't know something. I think we need the same with these AIs, though as far as I understand getting the AI to correctly estimate its own certainty is still an open problem.
I can't speak to everyone's experience - but whenever I'm having a conversation around relatively complex topics with MY friends - the deeper they dive, the more they're constantly referring back to their dive computer. They'll also try to make arguments that are principally anchored to the pegs that they're convinced will hold. I'm aware I'm mixing metaphors here but the point stands.
As far as "mixing information" - yes there are commonly known tricks to trying to get a more accurate answer:
- Query several LLMs
- Query the same LLM multiple times with different context histories
- Socratically force it to re-assess itself
- Provide RAG / documents / access to search engines
- Force quantitative tests in the form of virtualized envs though this is more for Compsci/Tech/Math
etc.
LLMs don't currently have a good sense of their boundaries - they can't provide realistic confidence scores and weight their outputs accordingly - the human equivalent of saying, "I only have a passing familiarity with the original greek of the Septuagint, but I think...."
It's a poor use of an LLM as a glorified fact checker - it's far better as a tool for free form exploration.
> Being good at mixing together information from a variety of sources - of different levels of accuracy - is key to learning anything well.
I have a pretty extensive background in teaching/education and I would heavily disagree with this assertion - at least when starting as a complete novice. The key to learning well is to establish a strong foundation by learning from the most accurate resources as possible. When you pick up a musical instrument, you don't want a teacher who's just one page ahead of you in the lesson book.
I tend to ask multiple models and if they all give me roughly the same answer, then it's probably right.
You can see this effect in the ARC-AGI evals, too much context impacts even o3(high).
... or they had a lot of overlapping training data in that area.
https://beta.gitsense.com/?chat=ed907b02-4f03-477f-a5e4-ce9a...
If you click on the Evaluation links, you can see how you can use multiple LLMs to validate LLM response. The evaluation of the accurate response is interesting since Llama 3.3 was the most critical.
https://beta.gitsense.com/?chat=fdfb053d-f0e2-4346-bdfc-7305...
At this point, you would ask Llama to explain why the response was not 100% which you can use to cross reference other LLMs or to do your own research.
You can't. Because LLM's are statistical generative text algorithms, dependent upon their training data set and subsequent reinforcement. Think Bayesian statistics.
What you are asking for is "to reliably interrogate topics of interest", which is not what LLM's do. Concepts such as reliability are orthogonal to their purpose.