Browsed by
Category: AI

AI Hank Williams sings new songs – Like Straight Out of Compton

AI Hank Williams sings new songs – Like Straight Out of Compton

Comments 0 Comment

If you don’t think AI is changing things at a fundamental level, witness what is possible with voice models trained by ordinary people like ThereIRuinedIt:

Or Johnny Cash singing Barbie Girl

How? There’s a number of different ways you can try this yourself – but the list grows daily at this point, so do some googling and see what’s available.

AI beat the old labyrinth marble game world record in just 6 hours

AI beat the old labyrinth marble game world record in just 6 hours

Comments 0 Comment

I bet a lot of you played this marble game when you were younger. The labyrinth marble game was developed by BRIO in Sweden in 1946. It was introduced to the United States around 1950. While many take hours to get good at it, the world record (with video proof on the site) is held by Lars-Göran Danielsson at 15.95 seconds.

But an AI called Cyberrunner, which was connected to the marble labyrinth with a camera and servos, trained for just 6 hours and managed to finish with a new world record of 14.48 seconds – almost 10% faster than the current record. During it’s learning, it even discovered cheats to cut the maze (though the new record was set without using any of those illegal shortcuts).

Read more about Cyberrunner at their project site, read the technical paper, and the source/hardware are about to be open-sourced. Or you can watch it below:

Chihuahua or muffin

Chihuahua or muffin

Comments 0 Comment

Free code camp compares various AI-based image recognizers to see how well they can identify if a picture is a chihuahua or a muffin. It’s surprisingly harder than you think and has a history of being used to determine the quality of the recognizer.

The author compares solutions from Amazon, Microsoft, IBM, Google, Cloudsight, and Clarifai. They also discuss the per-image cost as well as the quality of tags and other considerations. Definitely worth looking at if you’re trying to find an image classifier system.

Final results are on Topbots.

Links:

How will we know when AI gains consciousness?

How will we know when AI gains consciousness?

Comments 1 comment

exurb1a hypothesizes about some of the very real horrors that current social media bots are capable of doing. Pro-tip – get the HECK off social media and stop trusting anything you read there because this stuff has already been happening on every social media, dating, review, and news feed apps since even before the 2016 election.

His speculations? Perhaps AI personas will become so realistic and comforting to us that we’ll stop interacting with each other – and spend our lives conversing and in relationships with non-entities.

Or (as is already happening) governments, extremist groups, media, and intelligence agencies weaponized AI to flood the internet with manipulated stories, data, and opinions. Finally (as if becoming unable to form real relationships and being in relationships with AI is not scary enough) he asks what if AI itself becomes conscious.

One of the main reasons this would be terrifying is because right now we have no way to ensure alignment of AI to any set of values.

When the AI becomes able to mimic humans so well that it can convince anyone of anything – even talking to it becomes infinitely dangerous. We could have just created an almost infinitely hyper-intelligent demon, trickster, and sociopath.

See how deep the rabbit hole goes – and the majority of the possible outcomes are not good.

Early AI was more like a therapist

Early AI was more like a therapist

Comments 0 Comment

ELIZA was an early ‘AI’ created by MIT scientist Joseph Weizenbaum between 1964 to 1967.

He created it to explore communication between humans and machines. ELIZA simulated conversation by using very simple pattern matching and substitution that gave users an illusion of understanding – but it had no representation that could be considered really understanding what was being said by either party. Something you can easily discover by playing with it for a few minutes.

Fast forward to 1991, and Creative Labs was having amazing success with their SoundBlaster add-on sound cards. On the driver disks that came with the SoundBlaster, there were programs showing off different capabilities. One of these capabilities was voice generation. To show off the ability to voice synthesize text, Creative Labs included a little program called Dr. Sbaitso (SoundBlaster Acting Intelligent Text-to-Speech Operator).

You interacted with it like a pseudo-therapist; but you can clearly see the connections and similar pattern/substitution methods that Eliza used. I remember being wowed by it when I played with it for the first time – and experimented for hours with it. It quickly shows its limitations, but the speech synthesis was very good for the time.

It doesn’t hold the test of time, but it is pretty neat and you can even check it out here:

https://classicreload.com/dr-sbaitso.html#

Introduction to Generative AI

Introduction to Generative AI

Comments 0 Comment

AI’s can be applied to a number of different classes of problems. Recognizing and predicting are some of these tasks. But when it comes to generating something, you’re probably using a GAN.

This is a video from about 3 years ago when GANS were really getting started. If you’re trying to get your feet wet, this is a great, brief introduction to the history of AI systems like GANs (generative adversarial networks). Or, check out some of these other networks.

Links to referenced material: