OpenAI Eager to Employ AI as Personal Trainers for More AI: The Inception of Artificial Intelligence!

“OpenAI Wants AI to Help Humans Train AI”

“Advanced AI systems are typically trained in a two-step process. First, they learn from a large data set, and then they fine-tune their developed skills with a smaller, more specific data set. It’s like a round of Jeopardy: first, you master the broad stuff—history, science, literature—and then the specific questions that were left out from those categories for the final round”, so Wired begins to peel the layers of how OpenAI has decided to spice things up.

Tedium and predictability are the last things anyone wants in the thrilling field of technology. So, naturally, OpenAI is achieving slightly anarchic levels of fun by adding a little twist to the way AI is trained. Kicking the conventional rules to the kerb, OpenAI has discovered that cramming the second stage of training (you know, the part when the AI system is supposed to fine-tune its skills) with all and sundry of the internet can bring about some rather interesting results. A sort of all-you-can-eat buffet, if you will, where the gastronomic experience is replaced by the unpredictable world of internet content.

Now, does this constitute a groundbreaking revolution or spell the blithe destruction of all we hold dear? Who’s to say? Don’t look at us, we’re just the messengers! OpenAI clings to the stance that this revolutionary twist to AI training “increases the performance in downstream tasks”. Quite what tasks they mean is anyone’s guess since they graciously left that detail out, but suffice to say, expectations are high.

Emerging from the lab results, this haphazard experimental method seems to be producing a ‘Frankenstein’ of an AI system. It’s picking up all sorts of skills that no one ever trained it for. It’s playing poker, it’s mastering language translation, it’s imitating Shakespeare. Pretty soon, it might just write this blog post and, frankly, do a better job at it!

Let’s not forget the important part here – the cost. Filling up the AI bellies with contents of the internet buffet comes with a price tag. It’s like sending our AI to the most expensive, most exclusive private school out there – it’s going to burn a hole in the pocket. But hey, if you’ve got the deep pockets, why not right?

In the end, it’s all fun and games until someone loses an eye, or until our AI accidentally stumbles upon advanced warfare strategies or secret family lasagna recipes on the Internet. Here’s to hoping OpenAI has a solid lid on this Pandora’s box. Nevertheless, it’s an interesting direction to take, and the implications, both charming and terrifying, hold a strange appeal. Or worry. Or both.

Read the original article here: https://www.wired.com/story/openai-rlhf-ai-training/