1. 8

A team of five neural networks, OpenAI Five, has started to defeat amateur human teams at Dota 2.

  1. 2

    Our team of five neural networks, OpenAI Five, has started to defeat amateur human teams at Dota 2.

    OpenAI Five plays 180 years worth of games against itself every day, learning via self-play. It trains using a scaled-up version of Proximal Policy Optimization running on 256 GPUs and 128,000 CPU cores

    If that doesn’t ease your fears about an impending AI apocalypse, I don’t know what will.

    1. 5

      That actually makes my AI fears worse. But that’s because they are not the exact stereotypical AI fears.

      What the article says is: if you can afford ten times more computing resources, you get better chances to achieve superhuman resources than by using novel approaches. Train once, run cheaply. So, capital matters, labor qualification doesn’t, economoies of scale with huge front-up costs and small recurring costs. That’s how you get badly broken monopoly markets, no? And of course then it breaks because someone spends on servers and not on human time to make sure nothing stupid happens.

      Yes, OpenAI on its own will probably try to release something already trained and with predictable failure risks; and this is likely to improve the situation overall — I like what they want to do and what they do, I am just afraid of what they find (and thanks to them for disclosing it).

      1. 4

        That’s a very good, and frightening point. My comment was mainly in regards to the side of it that smells like mind boggling computational waste in addition to the staggering amount of experience such models require to achieve sub-human (less than expert) performance. Personally, I think ANNs are very cool, and quite elegant. But I feel like they are treated as the ultimate hammer, and every domain of learning is a nail. Examples such as this post show that they scale less than optimally. And from an algorithmic perspective there has to be a more concise, more computationally efficient way to go about learning to perform a task.

        1. 2
          1. Yes, it takes a lot to get anything out of it, then if you just continue doing the same you get better results than you expected, and the road from median player to expert player is not always that large of a percentage increase in expenses on machine training.

          2. Comparing effort-to-mastery between humans and machines is hard, because the expert play is a product of many millions of hours of human play (and discussion on forums and in chats).

          2a. Yes, it would be intersting if centaur chess or centaur go would take off… A 1kW desktop and a human against a 10kW server. Especially interesting given that game 4 by AlphaGo was lost to a human strategy normally too tiring to keep track of — so humans never play it against each other — but it is unclear whether a weak computer could help a human in using such strategies.

          1. One of the useful properties of cryptocurrencies is to show that humans care little about computational waste per se… I mean, some people understood it after looking at smartphones and comparing them to pre-iPhone smartphones and PDAs, but some people needed a clearer explanation.

          2. Analysing algorithmic perspective requires qualified humans, and good tools for these humans. And some of our tools seem to have degraded (as tools of information processing) with time… (I mean, it’s almost painful to read Engelbart’s book on Intelligence Augmentation and see that some modern tools are worse than the prototypes his team had).

        2. 1

          and thanks to them for disclosing it

          I’m afraid they will just disclose peanuts, to engage people and spread trust in AI.

          It’s basically marketing.

          1. 1

            Well, they already tell what they tried and what didn’t work, which is unfortunately already above the median…

            I meant that I am already afraid of then things they find and disclose now.

        3. 2

          Evidence they can learn in an environment where they start with nothing, the rules change a lot, they have to pick up new concepts quickly based on old ones, malicious people try to destroy them, or possessing common sense. What the machines are used for or doing vs what we do on a regular basis are really, really different.

          1. 1

            It might not even matter if OpenAI can eventually make machine learning work in adversarial environment, though (I think this post might even be weak evidence that they will be able to do it at some point, given enough computational resources spent), as someone will still cut the corners and give a vulnerable agent direct control over something with significant negative externalities of failure.