Is this paragraph from Trump or an AI bot? You decide, plus buy your own AI for $399

Hello, welcome to this week's roundup of AI news. Read on for a fun and, frankly worrying, quiz that tests if you can tell if something was made up by an AI text generation model or said by Trump, and more. AI computation from 1959 - 2018: OpenAI has analysed the amount of computing power needed to build AI systems in research …

[Anonymous Coward]

I thought that Donald 'I cheat at Golf' Trump

was a robot and have done for years. It repeats the same stock phrases badly in every speech.

All those other robots you see in the pictures of him giving a speech have been programmed to wave and cheer on command.

Honestly, given the IQ of the collective that is the US Electorate, they could put up a 'Three Toed Sloth' to run for Pres and it would still get elected.

As long as it is trained to say 'God Bless Amercia' it would walk the election.

[jake]

Re: I thought that Donald 'I cheat at Golf' Trump

Bad troll. No cookie.

[Anonymous Coward]

Re: Bad troll. No cookie.

Since they'd logged into ElReg to post that, they'll have a least a couple of cookies that have been set :-)

[The Dogs Meevonks]

Re: Bad troll. No cookie.

and the 'dad joke' award of the day goes to... yet another anonymous coward. :p

[macjules]

Re: I thought that Donald 'I cheat at Golf' Trump

I think that no self-respecting developer would create software that outputted such garbage as Trump emits. Although possibly Trump is a reject from Microsoft's Tay research.

[Rich 11]

Re: I thought that Donald 'I cheat at Golf' Trump

What's wrong with electing a three-toed sloth? It'd do far less harm both at home and abroad than any of the post-war American presidents have done.

[Anonymous Coward]

Re: I thought that Donald 'I cheat at Golf' Trump

"My policies are these: Eat leaves at the top of the tree and dump at the bottom.

Unlike my opponent, Mr. Trump,who feeds at the bottom and dumps on everything from a great height."

[Anonymous Coward]

Re: I thought that Donald 'I cheat at Golf' Trump

Amercia? Mercia was quite sophisticated. In fact at one time their top general was a daughter of Alfred the Great, who apparently shared her father's ability to keep down the Danes.

Not to be confused with America where the President has a poor record on appointing women and has managed to piss off the Danes royally.

[Michael Wojcik]

Re: I thought that Donald 'I cheat at Golf' Trump

Actually - and while I am in no way a fan of our Village Idiot in Chief - Trump has appointed quite a few women to significant positions in his administration. He's fired a bunch of them too, of course; but then he's not been reluctant to do that to men either.

In any case, "a poor record on appointing women" is one criticism of the Orange Megalomaniac that probably isn't fair. Appointing competent, reasonable women ... well, again, it's not noticeably worse than his record with men.

[sbt]

Is this headline from an AI bot?

In the headline, it should be "Is this paragraph from Trump or an AI bot?" not "of"

Also, the straight line fitted on the graph isn't showing linear growth if the graph has a logarithmic axis. That's exponential.

Mine's the one with the big book of log tables. -->

[getHandle]

Re: Is this headline from an AI bot?

Yours should be the one that says "Tips and corrections" on it...

[sbt]

Just the tip, I swear

Tips go in a jar, corrections go in a gaol.

Cats go in a sandbox. I could go on.

[jake]

Re: Just the tip, I swear

Or you could just go ...

[sbt]

Ouch.

On the upside, they've fixed the headline.

I suspect the change in direction on the graph is explained as a point at which it became cheaper to buy more CPU than pay decent coders who understand efficiency.

[Ordinary Donkey]

Worse than reported

60% of people failing to perfectly detect 10 samples would be normal. Humans aren't 100% accurate.

What the results actually say is that people correctly detect the fake articles only 40% of the time. Or in other words flipping a coin is more accurate.

[FrogsAndChips]

Re: Worse than reported

No, what they say is that on average, people guess well (i.e detect correctly if a text is real or fake) only 40% of the time, 4 correct answers out of 10. If you look at the charts below the test, you will see that RoboTrump texts are correctly labelled fake less than 30% of the time (between 20% and 30%, depending on the topic of the speech). That's really scary, especially if you consider that people taking the test are know that it can be fake and will pay more attention.

[Simon Coyne]

Re: Worse than reported

Humans aren't 100% correct?? You haven't met my missus

[Chris G]

AI?

Well, that may explain why I see see so much similarity between the pres and the old Max Headroom persona.

I think the humour routines have been deleted in favour of sanction generators though.

[Anonymous Coward]

text generation models can be tweaked to churn out disinformation

AI doesn't exist, it's just imitation. Proof: if there was any "learning", the models would improve to, ultimately, churn out ueberdisinformation, but all is does is just copy Mr Trump. Unless you suggest The Man is already there, the Bestest US Presidential Disinformant EVER!

...

wait, this cannot be wright, I must run the abav thru de

[Winkypop]

Either way

The bullshit flows and flows.

[cd11]

Re: Either way

very true

[Anonymous Coward]

Trump vs RoboTrump

The folks over at Lawsuit.org, a legal advice site, have developed the Trump vs RoboTrump test for people to take. You can try your hand at the quiz here.

It's an interesting test. What makes it less than easy (to my shame I've only scored 8 out of 10) is the fact that neither RoboTrump's nor Trump's quotations could possibly be mistaken for an utterance by an average person, fully in control of their faculties - they both suffer from the lack of coherence of thought over the few-sentences scale, both are highly repetitive, and both have a very limited vocabulary. Differentiating between different degrees of impairment is not easy; perhaps I need to spend more time down at the pub ...

[Yet Another Anonymous coward]

Re: Trump vs RoboTrump

I propose a new Turing test.

If you can't tell the difference between a conversation with a moron and a computer - then the computer can be said to be stupid

[Michael Wojcik]

Re: Trump vs RoboTrump

Yes, the real problem is the abysmal signal-to-noise ratio for both sources.

[Anonymous Coward]

Disinformation

From the article:

it’s also a worrying sign of how these text generation models can be tweaked to churn out disinformation

Well, I think that's only partly true. It's true in the sense that it turns out to be fairly easy to train a machine to create language which looks a bit like Trump's language And this matters because Trump is the president of course.

But Trump's language is, well, idiosyncratic, to put it rather mildly. Producing a plausible simulation of Trump's language doesn't really tell you very much about whether you can produce a plausible simulation of the language produced by anyone else, I think.

So while it does matter, because of who the president is today, I don't think that it's safe to generalise this.

[Il'Geller]

OpenAI has no idea that texts contain not only contexts, but subtexts as well. That is, not only what is seen explicitly and can be read, but also what is meant implicitly! For example, all any text's words have dictionary definitions, they are tied to other texts' synonymous clusters (taking into account their timestamps). These implicit definitions and connections are also texts' part, although they cannot be seen or read.

The technology for subtext recovery exists, it is simple: computer needs to select for each word its uniue dictionary definition, which is in the context within before and after this word, in harmony with the rest of the text's dictionary definitions. The same should be done with links to synonymous clusters, finding those texts that are consonant with the given.

Not necessary to train date on terabytes of other date! Enough on several bytes of dictionary. This alone saves millions and billions of dollars, plus AI becomes intelligent, starts to understand and ceases to be a toy (mimicking Trump).

[FrogsAndChips]

Quiz: has this text been generated by I. Geller or Robo.I.Geller?

[Anonymous Coward]

Yes.

[Il'Geller]

"...GPT-2 is a large transformer-based language model with 1.5 billion parameters, trained on a dataset of 8 million web pages. GPT-2 is trained with a simple objective: predict the next word, given all of the previous words within some text."

And? What for? To mimic Trump?

AI is a purely commercial project from the very beginning! AI is looking for groups of patterns that are both contextually and subtextually aimed at a practical purpose. For instance they are commands for a driverless car or information for a financial broker.

AI is trained using a very good indexed dictionary, 25Mb big.

[Anonymous Coward]

AI is a purely commercial project from the very beginning!

Well, I wasn't there at the beginning of AI, but I knew people who were, and this claim is false. AI and its friends & relations certainly had commercial goals, but a lot of people involved in it also wanted to understand how people's brains & minds worked.

[BGatez]

donald or ai?

At best don the con is an artificial intelligence and a vivid example of GIGO

[Tom 35]

Using Donald Trump as your sample for an AI test

Is like using Miller Lite as the control in a fake beer contest.

[spold]

Waymo

...I'm expecting Waymo people to be run over then.

[Anonymous Coward]

On/Off/Out of Message

The paragraph in the article is clearly not Trump because it stays on message. Also it doesn't the wrong noun for anything. Clearly the Artificial Intelligence is more - very, very, very more - intelligent than the Artificial Orange.

[Another User]

Yay!

I made exactly the 4/10 score like the average! Robo for president...

But all of these sentences are about topics that could have been said. There isn‘t really anything that is subversive or totally strange.

Additionally having ghost writers etc. what is really authentic?

[georgezilla]

My guess was .............

that it was done by an AI.

It was way,way to coherent to have actually been Trump.

[William Towle]

Re: My guess was .............

I also guessed AI, expecting it to have done somewhat better.

I'm not sure if this says more about the AI or the person it's imitating...