Danny's workmate is called GPT-3

[Laird]

Danny's workmate is called GPT-3. You've probably read its work without realising it's an AI

Technology reporter James Purtill wrote this article published today by the Australian Broadcasting Corporation

For @Sciborg_S_Patel 

You may not have heard of GPT-3, but there's a good chance you've read its work, used a website that runs its code, or even conversed with it through a chatbot or a character in a game.
GPT-3 is an AI model — a type of artificial intelligence — and its applications have quietly trickled into our everyday lives over the past couple of years.
In recent months, that trickle has picked up force: more and more applications are using AI like GPT-3, and these AI programs are producing greater amounts of data, from words, to images, to code.

"Writers are expensive. And there's a limit to how much quality content a human can produce," Mr Mahoney says.
"You can get the same quality of content using AI tools. You just get it faster."
How much faster? About three times, he estimates.
He still has to check and edit the AI-generated text, but it's less work and he's cut his rates by half.

Companies like ContentBot and Longshot pay OpenAI for access to GPT-3: the rate of the most popular model (Davinci) is about $US0.06 per 750 words.

In March 2021, GPT-3 was generating an average of 4.5 billion words per day.

In general, GPT-3 is remarkably good at stringing sentences together, though plays fast and loose with the facts.

Asked to write about the 2022 Australian election, it claimed the vote would be held on July 2.

He still has to check and edit the AI-generated text, but it's less work and he's cut his rates by half.

If it is only at that level (i.e. it needs someone to review its output), how can it work as a chatbot?

My partner does some translations into Czech, and she generally uses GOOGLE translate to generate a rough draft that she works up into the final result. Since she has lived in England for many years she can read the English and she confirms that the automatic translation can easily get lost and become misleading.

"the same quality of content" - that might need qualifying or context. Possibly spelling and grammar might be reasonable much of the time. But as acknowledged, the facts may be incorrect. Quality also doesn't just come down to facts. There is also the matter of writing something which is interesting.

I've been watching some foreign-language videos on youtube, and using either the built-in autotranslate for subtitles, or copy+paste into google translate. Well-written text seems to translate quite well, perhaps achieving over 50% in terms of usefulness. When it comes to colloquial expressions and incomplete sentences - which is how humans tend to speak to one another - the output can be gibberish, maybe 10% to 20% useful. Often it is impossible to even guess what the actual meaning should be.

[Laird]

There are some examples of the AI's output in the article, which could give you guys a sense of its quality.

If it is only at that level (i.e. it needs someone to review its output), how can it work as a chatbot?

Yup, there seems to always be some caveat to AI.

Beyond that, it's essentially a variation of "curve fitting" with no actual understanding. Impressive in some technical sense but nothing like the wonder of watching my little nephew slowly come to understand language, kindness, humor, etc.

Ultimately all humans should realize Turing Machines are hollow, while children are wondrous.

[Laird]

Shocker: Bot Trained on 4chan Chats Is Super Offensive

An article from today on PCMag Australia by Emily Dreibelbis

Kilcher created the bot in just a few weeks, and his newly minted army of posters scanned the dataset every five minutes, generated unique comments, and published them under the same anonymous account. Given that AI is only as smart as the content it's trained on, /pol/ was quickly flooded with even more racist, misogynistic, and antisemitic content than usual.

"There were a base of users recognizing the bots for being bots, [but] there were still plenty of other users who didn't," Kilcher says, even after he admitted to being behind it.

While 4Chan typically requires users to complete a CAPTCHA to prove their humanity, Kilcher bypassed it by getting 4Chan Pass, a $20/month premium subscription that lets members to skip the CAPTCHA. 4Chan Pass also allows posting from a proxy server, which is usually not permitted. Kilcher’s bot ran on one based in the Seychelles, a tiny island nation off the coast of Eastern Africa.

The bot seems to have been based on the same GPT-3 technology described in the news article in the opening post of this thread.

Shocker: Bot Trained on 4chan Chats Is Super Offensive

An article from today on PCMag Australia by Emily Dreibelbis


The bot seems to have been based on the same GPT-3 technology described in the news article in the opening post of this thread.

You know, that has me wondering... what if the same kind of bot was trained on the likes of Reddit, Twitter, etc?

I have strong suspicions that the offensiveness would be as equally off the charts.

But, in the social media side of things, Reddit and Twitter are considered very acceptable by the mass media, while 4chan is very strongly hated and smeared.

The hypocrisy amuses me quite a lot.