4 ms·
It's not a bug. It's an architectural defect / limitation in our understanding of how to build AI. That makes it a strictly harder problem that will take longer
by vlovich123 4y ago
It's not a bug. It's an architectural defect / limitation in our understanding of how to build AI. That makes it a strictly harder problem that will take longer. And it's not totally clear to me that you'll get there purely with LLMs. LLMs accomplish a good chunk of what we classify as intelligence for sure. But it's missing the cognition / reasoning skills and the open question is whether you can solve that by just bolting on more techniques into the LLM or you need a totally different kind of model that you can marry to an LLM.
- vidarh 4y agoGPT 4 will admit to not knowing things in many cases where 3.5turbo does not (tested the same prompt), and either will stop there or go off on a "but if it did exist it might go something like this" type continuation. It still hallucinates a lot, but it's not at all clear that this will be all that difficult an issue to solve given the progress.
- vlovich123 4y agoWe generally only hallucinate while dreaming / using our imagination. And we can distinguish those two states. Admitting lack of knowledge is of course good but, for example, if you ask it to write some code that isn’t boilerplate API integrations, it’ll do so happily even when it’s wildly wrong and it can’t tell the difference and that is also the case with GPT4 afaik. Moreover, you can’t solve it through prompt engineering because there’s clearly a lack of context it’s unable to understand to figure out what non trivial thing your asking it.
- vidarh 4y agoThe point is there's been progress in making it admit when it doesn't know, and we simply don't know how fast that will improve in future version. It may continue to be an issue, or turn out to be easily solved. The improvement with GPT4 does suggest it is at least possible to make it recognise its limits.
- senko 4y agoThere is ongoing promising work that shows this can be improved a lot: https://arxiv.org/abs/2210.03629 https://arxiv.org/abs/2210.03629