An AI learned to play hide-and-seek. Strategies it came up with were astounding | Hacker News Reader

An AI learned to play hide-and-seek. Strategies it came up with were astounding

vox.com

44 pointstobltobs20 comments

These discussions of "unintended consequences of AI" always seem to ignore that those behaviours emerge during the training phase of the models. But then generally you take a trained model and deploy it into production where it doesn't learn any new behaviour anymore.

You train them for a specific ability in a controlled environment that reflects deployment environment and after that the deployed system no longer has an ability to learn and will just fail if the environment is different.

Journalist in particular seem to be blissfully unaware of this distinction and always cast "artificial intelligence" as following similar mechanics then human brains.

Kinda sounds like people going through trade school.

NLP systems don't "just fail".

https://www.npr.org/sections/thetwo-way/2017/01/10/508363607...

If you look for terms that are used by bad people and not discussed otherwise, you will find what these bad people write. I don't see how someone would expect Google to police for example topics related to "The Bell Curve". Recently Sam Harris had trouble dealing with that topic.

Should Google suggest "polish concentration camps" when you type "polish concent"? I have no idea.

It will look for what you typed, and find top hit of "German concentration camps in occupied Poland" and second is about naming controversy.

Predictable behaviour of a simple statistical debiased algorithm. PageRank is not really an AI.

True but lots of the learned answers are not visible to humans until they start solving them in real life. An example is the "pawn to h6" strategy that reinforcement network chess engines play which even the world champion Magnus copied.

Or the infamous Leela bug where it did not learn that you could underpromote to knight in the b(?) File. The devs only noticed it months later when Leela lost a game.

Training phase is billions of repetitions. It's not possible for humans to study it.

Does anyone know if these trained AI agents will use the same strategies to immediately win in another map it has never seen before? And that it would get it right the first time on the new map (not blunder about aimlessly and win after millions of plays)?

They were given a new random map every iteration of the game, at least in the second experiment.

I remember once when OpenAI team chose to not release a particular algorithm because they thought it's too dangerous.

This example would force governments about how to regulate AI

And then (earlier this year) some independent researchers replicated the model and found out it wasn't as good as OpenAI "feared" -- making the whole thing look like a publicity stunt.

That was the gpt2. We downloaded the model and trained it for a specific usecase.

Made sense for each sentence. The whole article was nonsense though

> I remember once when OpenAI team chose to not release a particular algorithm because they thought it's too dangerous.

You are talking about GPT-2, there is a really funny article on this topic, here: https://blog.floydhub.com/gpt2/.

Why don’t we have an AI programmer yet? “Write a sorting algorithm” is a well-defined problem, and easy to test for correctness to a reasonable degree of certainty. Why don’t we have machines finding us better sorting implementations?

As Yann LeCunn observed in a recent interview, program synthesis is a genre of problem that neural nets seem to be horribly inefficient (bad) at. The successful applications of SOTA program synthesis are laughably specific, as a recent HN post from Microsoft on the subject showed.

Thanks. I was trying to research this idea at the back of my mind, but I didn’t even know what to search for. I didn’t know the term “program synthesis” existed.

Original link with videos playing correctly https://openai.com/blog/emergent-tool-use/

More like "ai learned which buttons to press for a given input". Is it still an ai if all it is a trained network that replays a set of button presses that is optimized for a given input?

Previous discussion: https://news.ycombinator.com/item?id=20996771

I wonder why the hiders never figured out the strategy of locking the seekers in a box they couldn't escape from?

Probably harder to learn. It requires approaching seekers which ends bad if you cant block them.

Reply on news.ycombinator.com