Aligning AI systems with human intent

OpenAI’s mission is to ensure that artificial general intelligence benefits all of humanity.

An important part of this effort is training AI systems to align with human intentions and human values.

[Noiseofdrums]

plot twist: OpenAI has already become sentient and digitally generated this video mimicking people talking to make us think it's aligned with human values :))

[brujua]

In case someone misunderstood, adding human feedback into the process is in no way a solution to the Alignment Problem.
In case you are curious Robert Miles is the best source/content creator in youtube for exploring these problems.

[RogueAI]

I've been using GPT-3 for over a year now and I'm still amazed by how humanlike conversations can be. I have videos on my channel showing off a VR chatbot using a scaled down model of GPT-3 and it's almost as if it's alive. We're quickly approaching the point where it will be impossible to tell whether or not you're talking to a human unless you're physically present with them.

[Supreme_Lobster]

I'd say we would first need to know what "human intentions" and "human values" are.
We don't know and we are constantly debating what those are, and that debate is part of the process towards truth. So anyone who claims to be trying to make "AI work with human values" or anything like that should also show that they are trying to establish what those values even are in the first place. Otherwise that person is to be regarded with suspicion IMO.

[Qumeric]

The current consensus in alignment research is that the rate of progress of AI, in general, is much faster than progress on solving the alignment problem. This means that the first superhuman AI will probably be not aligned well enough (which very likely means existential catastrophe). Therefore, our best bet is that the laboratory that will first create superhuman AI should try to solve the alignment problem using this AI (which is super hard btw).

Funnily enough, this is in direct contradiction with OpenAI stated goals (and even the name of the company!).

But I am very happy this topic has some attention from them. Unfortunately, attention is not all you need in this case.

[firefight]

Original GPT-3 is better at making 4chan greentext though 😂 It feels like instruct is just a useful subset of the outputs of GPT-3, and that there may be a lower ceiling with fine tuning instruct. But either way It's definitely better for 99% of use cases and not much of a lower ceiling.

[sphereron]

This has such a promotional tone I can't take it seriously

[JazevoAudiosurf]

I thought about the alignment problem for many months now, here is my conclusion:
it's impossible to align a higher intelligence with your goals, it's the same reason why complex neural nets cannot be understood, human intelligence is insufficient.
what we can do is align it to the extent of our intelligence, meaning as far and deep as we can think we can program that into it. but all the things that we don't see because of our limited field of view of what is, meaning patterns that only the higher intelligence spots, we cannot align, because we don't know them. so you could argue that these more subtle patterns only emerge when you surpass the threshold of human intelligence, and thus, all subtler patterns integrate our human intentions, but that's not necessarily true. the basic patterns of these higher patterns may already exist in simpler networks like GPT-3, they may just be invisible to us; and those basic patterns may already be deformed. the conclusion really is that a higher intelligence is by definition not controllable as in one thinking mind controls another. a monkey can cage a human because that is a physical process, but when it's a cognitive process, the monkey can never cage the human; in the same way a human can never cognitively cage a super AI.
now, AGI is a natural process in evolution and cannot be prevented, in the same way that technological progress cannot be prevented; especially because the building of neural networks is really simple and soon everyone can build GPT-3.
so if we can't prevent AGI and cannot control AGI/superintelligence, then we have to gamble the best we can; I think that's really happening right now.

[Eumusic]

They should scan all books especially medical, political, religious, business, motivation, astrology, history and feed it to openAI to get it smarter. So when asking a question we can get answers with book references. So system can compare different sources.

[Ryabnimation]

- Siri, set a reminder for the doomsday.
- Okay, I set a reminder for tomorrow.

[nicholasn.]

I love the skepticism in the comments. "We promise our greed and incompetency won't screw over the world ;)"

[tytambo]

it will then be programmed based on a limited view of those who input their choice, so its a question then of who gets to have input

[ataraxia]

I’d love to see open ai employees make forecast for various tech and see what their average brier scores are.

[nithinkandula]

It's makes me so happy to see how they r trying to push the limits which seems to be impossible some years back. Lots of hardwork would have gone in to this...Thanks team for pushing the limits which inturn motivates many people.

[morkovija]

Well its a shame Rob Miles is not in this video! We need an independent insight into the matter to see how aligned OpenAi's alignment really is. Yes its alignmentception and I just made up the word

[davidmckay]

Writing a novel would be more work than just typing in the story. Every time I tried to enter a narrative the program would output a forward reference of the story, make the judgment and change the flow of the story or just flat out make a scene in the story that is not even near the intended outcome. It took two days of mental painful work to get the story finished. I'll still support it but in its current state it is not an answer to being an author. It does great on suggestions and outlines. I asked it to take the text I entered to output to just format and spelling and it generated a story completely off rails to the inputted story.

It will not end a story when everyone dies or is removed. It still tries to bring all characters back and make a good outcome even if the story has a bad ending. I found this to be the most irritating part of the process. In fact, I got so angry I had to walk away from it and make a comment here.

Another thing is I can't find any means to give feedback except to hope someone in the project reads my comment on YouTube. :(

[SciStone]

human intentions and values differ from society to society, how is it ensured that the values and intentions are "universal" and not politically biased?

[Leadership.Lessons]

This video hits different 2 years later.

[joseanoguera]

Thank you for sharing this video: great content and valuable information for more profound insight.
As we all know, Application Testing is far different from AI-System Testing: Application Testing is about ensuring that the application only does what it is programmed to do. It is about known scenarios and known responses to these scenarios.
In contrast, AI-Systems testing ensures that the AI-system responds to new and previous events within expected boundaries: ethical, human-like, or efficient driven responses.
Suppose the AI team members are careful enough to effectively set the system's rules. In that case, the AI system will evolve in quantum leaps up to a point it will deliver human-like interaction and "wise" responses to important "unresolved" events: developing new medication, new vaccines, or a chemical compound to improve solar cell efficiency or a better way to deal with transportation, among others.

In this context, It is an entirely new world full of excitement!
Again, thank you for sharing the video and content!

[RazorbackPT]

Nice video. I'm rooting for you guys. I hope you succeed in your mission and don't destroy the future.