Hacker News Reader

Comment by suyavuz | Hacker News Reader

I think this is basically right. You don’t hand out calculators before kids understand arithmetic. LLM version is sneakier because skipping the work still produces something that looks finished.

View on HN

A calculator gives you an answer. An LLM gives you an answer that sounds like it already checked itself.

Not quite. The LLM gives you a statistically probable sounding token stream. The calculator gives you a qualified answer within documented and deterministic limits of the device.

No one knows how to use either.

That is GPT-3. Modern models are rewarded based off the accuracy of their responses.

No they are not

By... another AI model. Which uses statistical generation to decide whether the answer is likely to be accurate or not.

Which still makes it more than probabilistic sounding words.

Compounding one probabilistic LLM with another seems to compound the probabilistic nature, not reduce it.

Wait, have we solved hallucination already?

We've gotten better at it, but it's not a problem that can be solved (ignoring solutions like saying I don't know to all questions).

Actually, calculators were banned in my school until 10th grade or so. I understood the reasoning at the time and still think it was a good thing.

But I think the AI topic is a double-edged sword. Those tools can help tremendously when used in the right setting (e.g., similar to how brilliant.org works) and can easily spoil the whole learning effort when used wrongly.

I am not convinced that a ban is the solution, but probably better than doing nothing.

Reply on news.ycombinator.com