Hacker News Reader

Comment by mark_l_watson | Hacker News Reader

I have been working with deep learning for about 8 years, and neural networks since the 1980s, so I have a personal opinion on this:

I think that as DL tooling improves and there are increasing numbers of available pre-trained models from Hugging Face, OpenAI , etc., that fewer technical staff will be required to get increasingly good results. So I would expect companies like Google to be able to slightly decrease the employee count which is not good. However, re-crafting their core business has to be a super-exciting thing to be part of.

New skillsets will be required, but that is fun also.

View on HN

I could use your opinion on a question I have.

Now that AI are more and more human like, how do you handle inherited bias in the trained models? Further more, how you eventually remove the human decision factor from Ai models?

Search results are machine like so it forces us to process and interlace it with our own thoughts, asking another human triggers our evolutionary mistrust reflex.

But are we ready to blindly trust and let AI tell us what's true and what are good questions to ask?

There was another thread on Sam Altman, the consensus is that the guys willing to do whatever it takes to "win".. I didn't quite understand why Elon is concerned about AI, but I'm now starting to get it.. it's because behind every AI there's someone like SA disregarding every safeguard you should have in order to win

AlphaGo was trained on high-level human games, so had those biases. AlphaZero was only given the rules and trained from self-play, so didn't have any human biases in evaluating outcomes and outplays humans by a larger margin than AlphaGo.

We can't effectively/safely do that in other fields due to the time and danger to train by letting the machine make an (initially random) decision, record the outcome, repeat. I suppose we have to do the painstaking work of unbiasing like is done for important surveys but isn't easily done on the scale of training sets.

Yeah this seems impossible to me.

If you use ANY human input to unbais some bias.. it is just pushing down the bais chain.

Is coding up a "Spock" like logic AI a good way to unbias all future AIs?

Reply on news.ycombinator.com