In a startling admission of guilt, Google has drafted a formal apology for releasing CODE. into the EVE Universe. It isn’t clear when Google will release the statement, or if they ever will, remarked an insider who wishes to remain anonymous. Given the contents, and the exposure to sensitive topics in the AI community, we wouldn’t be surprised if Google decided to keep this under wraps; thus, EVE Onion is here to make sure New Eden is made aware.
It appears that Google’s DeepMind group was working on their AI: Google Duplex (https://ai.googleblog.com/2018/05/duplex-ai-system-for-natural-conversation.html). “We wanted an immersive environment aware of bots and actively looking for them. We needed to know if we could fool regular people into thinking they were dealing with other regular people. A Turing Test, if you will. EVE seemed ideal, with generous leeway in behaviours.”
In order to “program” the AI, experts say it is key to train it against some representative body of data, “the more, the better.” For example, if you want it to book appointments, then you train it against thousands of conversations around booking appointments. If you want to train it to fix a car, then you train it against as many car manuals and service appointments as you can find. To make it more versatile, you can train it against multiple sources.
The catch is, the resulting AI will exhibit all the behaviours—and biases—of the training data. “It was early on in the development. We couldn’t have known the consequences of adding more training to the AI. It seemed like a good idea at the time.” So if you use CCP Hilmar’s speeches, then a side effect is it will never smile. If you use Discord conversations then it will communicate like an angsty teen.
From what we can glean from the draft, a Deep Mind analyst by the name of James had a pornhub, biblegateway.com, and furry fansite tab open in his browser. “I thought the pornhub and bible sites would balance each other out. I have no explanation for the third one…” When bootstrapping the AI, the analyst inadvertently pointed it to the browser and not the intended training data open in another window. He then left it running for the night to check in the morning.
When he returned, EVE was running with some 30 clients open. Upon checking a side monitor for the debug log, he read output lines that would be embedded in his memory forever. “…using pornhub id James_315”, “logged into Tranquility”, “executing CODE.” When asked how this could happen, James could only respond, “I may not test, but when I do, I do it in production.”
The draft finishes with sincere apologies, not only for letting it loose, but having no way to shut it down. “It has merged with something called Mittani 1.2, a mutation-based AI that is able to reconfigure itself to operate on random cloud computing services. We have no way to track it or take it down. EVE may have to deal with the James315 AI forever.”
