2 ms·
Keep focussing on the word mindless and then ask yourself semantically, why mindless and intelligent feels like a contradiction in terms. (I do not personally
by ggm 6y ago
Keep focussing on the word mindless and then ask yourself semantically, why mindless and intelligent feels like a contradiction in terms.
(I do not personally ascribe any intelligence to GPT-3 and this is not the start of the singularity. There is no evident free-will, nor new knowledge synthesised as theorems from existing knowledge, no introspection, no understanding)
- doubleunplussed 6y agoHumans do not have free will either. What we call free will is just a useful abstraction over systems that are hard to predict and traffic that respond to incentives. GPT-3 can add arbitrary four digit numbers (with low error rate), it knows how addition works despite the numbers it's adding not being in its training data. I think you need other criteria to rule out GPT-3 as a 'mind'
- ggm 6y ago[edit: I pass over the free-will vs determinism. I think its not fruitful for what GPT-3 Is compared to what we are. If I am wrong, if I missed a nuance, do elaborate.] I think you need other criteria to include GPT-3 than the (admitedly really interesting) point that it has constructed addition from the training data. I would have a lot of subsidiary questions around that, and I note you said "knows" which is kind of red-rag-to-a-bull here. What is the "knowing" of its basis to have constructed addition from the training? The main questions here would be: 1) did it also "intuit" subtraction and why not? what did it do when it hit negative numbers and can it add negative numbers to positive numbers and negative numbers to negative numbers? b) has it shown any inductive logic outcome and got to multiplication and division? Even just integer.. 4) can it detect implied sequence order even when the textual labels are wrong?
- visarga 6y agoThere are examples of other language models that learned to do hard math problems such as symbolic integration which are difficult even for professionals empowered with dedicated (non learning) software.
- visarga 6y agoThere is introspection. It can judge the questions it receives as nonsense if asked to. Free will - what is that? I presume any source of randomness filtered through a complex model would do, as it turns noise of one kind into random thoughts/phrases. What's missing is - a larger context memory, online learning and in general being in the world (it has no body). Not being able to try and experiment with its concepts is the limiting factor.