OpenAI has published the text-generating AI it said was too dangerous to share | Hacker News Reader

OpenAI has published the text-generating AI it said was too dangerous to share

theverge.com

33 pointssogen23 comments

My take is that OpenAI is all hype.

I’m not saying that they don’t do really cool things. But, they have a tendency to exaggerate the extent/significance of their work.

The ‘Solving a Rubiks Cube’ went viral with the premise that a robot figured out how to solve a Rubiks cube. It did not. It figured out the dexterity of solving a Rubiks cube... 20% of the time.

‘Too dangerous to share’ is just another statement meant to drive hype.

I feel it’s kind of irresponsible for them to represent AI in the way they are. I know so many people convinced that robots are going to take all of the jobs really soon, so ‘why should they try?’ Why shouldn’t we just give everyone money instead?

What the general public doesn’t get is that while it may happen, the timeline in which it happens is extremely exaggerated. And you have politicians like Yang feeding into this narrative.

It's all hype until you realize you've been arguing politics with one of these, on Reddit.

If this software helps people realize that arguing politics on reddit isn't productive, then I'd say it's done us good.

Andrew Ng did a good analogy: it is like worrying about overpopulation on Mars. It will happen sometimes in the future, but worrying about it now is very premature.

In hindsight, we shouldn't be surprised that a company founded by Elon Musk is driven mainly by hype. I really respect the man, but it's a consistent pattern in most of his work.

The timeline is inherently unpredictable, significant innovations are sudden but adopted widely within months. We don't know when a set of jobs will be automated away, and we will have relatively little warning on a political timeline.

There are other arguments for basic income too, at the core UBI is to continue the debate about welfare and inflation, and asking if our current Quantitive Easing policies are not themselves pressuring the economy in unforeseen ways.

The biggest argument against UBI is that the government can require conditions with it, and you cannot easily find a new government. If you think this is unrealistic, look at selective service. You have to register or you can be barred from receiving federal financial aid or working in government positions, possibly for life. If the government makes recieving UBI contingent on two years of military or civilian service at age 18, what can be done?

I would be fine with UBI or any other form of welfare if I could personally opt out of it and not receive that welfare AND not pay any taxes for it.

I very much disagree.

Is GPT-2 capable of creating fake news articles? No, but thats a fundamental misunderstanding of propaganda.

Propaganda is a matter of connecting ideas to feelings and creating an ecosystem that reinforces those connections.

For that reason, you need propaganda directors to create the “fake news” or whatever, but where GPT-2 can be very, frighteningly useful, is with bot swarms.

Just a sentence or two from GPT-2 can be convincing enough to be indistinguishable from a human. Have enough of them primed to direct the pro/con conversation online, and people will respond accordingly, making the “conversation” seem real when it’s actually been carefully crafted.

You can already get this today, but that requires armies of people posting on social media. When anybody with access to enough bots on enough social networks can harness this, you can warp any conversation, anywhere, at almost zero marginal cost.

This is something GPT-2 can do right now. Even the smaller models were good enough at producing indistinguishable text. Nobody is immune to propaganda.

It could just be as bad as Microsoft's AI that was racist

Their model is overfitted. If you type "What the fuck did you" it will output Navy Seal copypasta with almost no changes. No wonder why it sometimes generates human-like texts, because in some cases it literally spews out its training examples with minimal or no changes.

You can try a couple of times. But the idea is amusing. Like it's big enough to just have long runs of known text embedded so it's really just holding on to actual text. That would be funny.

Yeah. It seems like there is a big difference between an AI that can spit out human-like responses, and an AI that can understand what it’s saying. In other words, can you ask the AI to “elaborate” on its idea? If not, it’s just a talking point regurgitator.

To test it, jump to: https://talktotransformer.com/

There are some examples of starting text, with mixed results.

Sometimes it seems to be overfitted and spitting entire paragraphs taken directly from some random page, but at times it seems to create something interesting:

Eg: "The Captain: I'm ready to sink this ship if it will help finding the alien artifact Dr Jones: Ok, but just don't sink it with ME on board!"

Created:

   The Captain: But I am Captain. That's what I get for being a coward!

   Dr Jones: It's okay, Captain. I've been in this game long enough to know when a captain is about to die, but not this time!

   The Captain: Okay... You got me. I'm just not sure I can trust a Captain with an artifact that's been hidden away for hundreds of years. It's too dangerous.

   Dr Jones: It's going to be a long day. We need to get this back to the ship. There are other ships still out there, but we can't expect them to help us out of this mess.

   The Captain: But we must save the artifact!

Huh... Try some simple code snippets

Even more interesting, code snippets like: "mysql_username = root"

Wow...

Just out of a whim I pulled a Bernie Sanders quote off google and the result is pretty much word for word something from one of his speeches.

Source: "The cost of college education today is so high that many young people are giving up their dream of going to college, while many others are graduating deeply in debt."

Completion: The cost of college is higher than ever, and we cannot continue to afford it. The student loan debt crisis is a growing national problem that has to be solved. There is a solution, and it is time to take action to stop student loan interest rates from doubling again and to help millions of Americans afford a college education. As President I will:

· Create a free community college program that provides a cost-effective option for all students, regardless of their income.

· Expand the Earned Income Tax Credit to $2,000 for every American with earnings under $15,000 a year.

· End taxpayer subsidies to private universities to pay for tuition at four-year public colleges.

· Make it easier to qualify for student loans by reducing the cost of borrowing for...

I would swear they trained on Pornhub transcripts. Interesting for how it frames several popular individuals too.

what make you say that...evidences please :)

IIRC they trained it off of reddit comments and pages linked to on reddit

Probably racist and hateful considering internet provides seed data

I wrote something about software engineering and it came back saying that Chris Lattner invented JavaScript.

To get real language understanding you need some type of grounding in visual data or something. You can't just feed it text alone.

The whole thing about Open AI and "danger" is that somehow suppressing certain research was part of its constitution. So they were looking for a way to prove that they were faithful to that and this thing is just the closest thing they have.

I really think that Elon Musk alone may cause the next AI winter. The guy has overhyped AI/ML to a crazy level that it clearly cannot perform at yet.

It seems this networks is completely clueless about culture. I have typed "Flat is justice." and it completed it with texts about justice in judicial system instead of lolis and anime.

Reply on news.ycombinator.com