undefined | Better HN

0 pointsRetric1y ago0 comments

Current AI is stuck imitating what it’s exposed to. If they actually learned the material they would understand to reject incorrect information in their training data. LLM’s etc can’t do that under current methodology, simply sucking up more training data and tossing more processing power at the algorithm doesn’t solve that issue.

Leaning from music doesn’t mean imitating that music. People can for example learn from mistakes without imitating those mistakes.

0 comments

6 comments · 2 top-level

constantcrying1y ago· 3 in thread

I am against all IP. I do not care what algorithms are used or in what way it differs from human learning, it makes absolutely no difference to me.

RetricOP1y ago

If you actually rejected all forms of IP protection you’d insist AI companies freely share their source code, patroon can’t hide works behind a paywall, etc.

Is that your stance? Or do you have some more limited rejection of IP in mind.

pixl971y ago

>you’d insist AI companies freely share their source code

Myself I don't see those as being related at all. Keeping something secret is allowed. What is not allowed is something being 'not-secret' yet special. Once you share the source code to the world there would be no closing the barn door.

1 more reply

constantcrying1y ago

No. There is no duty to publish anything.

Intellectual property is about someone having ownership over an idea. It does not mean that you have a right someone gives his idea to you. These are totally separate things.

I am free to write, paint, code or film whatever I want. I have zero obligation to share that with anyone.

1 more reply

pixl971y ago· 1 in thread

>LLM’s etc can’t do that under current methodology

I hate to give a smarmy result, but are you sure you know what RLHF is? Because this is one way to correct said data.

RetricOP1y ago

I am aware of RLHF, and no it doesn’t solve this problem.

There’s a great deal of lesions to be learned from X PB of training data that wouldn’t be covered.

j / k navigate · click thread line to collapse