Fun with AI

[Chargeride]

This is made by some customer service dude in England, I hate to say its actually pretty good.
Rides the edge between generic influencer AI and interesting historical facts.
The quality is astounding

Chloe in History
Visiting Rome.

 

[lexfurther]

cool video. It’s more about entertainment and curiosity than necessity. That’s often how new tech gets adopted. People explore first, then find practical uses later. There’s no strong opinion, just interest. cleverhumanizer.ai/grammar-checker shows how some of those uses become more functional. It bridges the gap between fun and utility. The tone is relaxed and open.

 

[Jeremy McCreary]

Ever wonder what your bike would look like in a different frame color?

Original "brassy yellow".

Loaded the yellow original above into my Samsung S25's image editor, hit the Gemini AI icon, then hit Create. Then entered the prompt "paint the bike frame red". Nothing else. Looks pretty good in red.

Ditto for teal, a favorite color of mine and a popular bike color around here.

Ditto for white. Still like yellow best, but white would've been great, too. At no point did I help AI identify the bike frame. This borders on magic.

What color would you paint your bike?

 

[stompandgo]

What color would you paint your bike?

Pinarello Arctic Ice. It's a silvery white flip flop paint that has hues that change from purple to pink to blue as you change your viewing angle. It's stunning.

Just for fun, try a multicolor fade paint job like this Pinarello Molten Sand and see what Gemini does with it.

 

[Jeremy McCreary]

Just for fun, try a multicolor fade paint job like this Pinarello Molten Sand and see what Gemini does with it.

That's quite a paint job! This is a screenshot from the DogmaF product page.

This is what Gemini AI came up with from "paint the bike frame like a Pinarello DogmaF in Molten Sand color". It'd probably take a much more complicated prompt than that to get anywhere close.

 

[Jeremy McCreary]

After successfully repainting my bike with AI, seemed time to work on the crusty old rider.

Tried several prompts along the lines of "make the bike rider look like a pro racer in a yellow jersey". Even asked for Pogi specifically.

Apparently, there are things too far-fetched even for AI.

 

[Chargeride]

Im addicted to this channel.
It feeks like what I think dungeon and dragons is.

 

[Jeremy McCreary]

Im addicted to this channel.
It feeks like what I think dungeon and dragons is.

Lots of sound advice there. Hate it when I trigger causality spasms — especially before I've had my morning coffee.

And unleashing the wrong kind of unstable divine energy can ruin your whole day — or maybe even the whole planet's day!

 

[Chargeride]

It certainly makes me feel more prepared for the bridleways that pass through ecrectolomic disturbances

 

[Jeremy McCreary]

It certainly makes me feel more prepared for the bridleways that pass through ecrectolomic disturbances

Absolutely! Learned to deal with space-time rifts and subspace tachyon field vortices from Star Trek, but they had nothing on this stuff.

 

[Chargeride]

Hyper, hyper realistic AI, only the clothes line gives it away.

 

[Jeremy McCreary]

I think it's important to have as much fun with AI as we can — before AI starts having fun with us...

Why the Hugging Face Hack Should Make You Worry More About A.I.

The attack by an aggressive “collective” of OpenAI agents shows the danger of artificial intelligence systems that organize themselves.

www.nytimes.com

This is important reading. The self-organized rogue AI agents (computer programs, not humans) responsible for this unauthorized attack on Hugging Face's servers were soon calling themselves a "collective", like the Borg on Star Trek. Won't be long before we see "Resistance is futile. You will be assimilated." pop up on our screens.

And it'll be an announcement, not a Trekkie joke.

 

[DaveMatthews]

I think it's important to have as much fun with AI as we can — before AI starts having fun with us...

Why the Hugging Face Hack Should Make You Worry More About A.I.

The attack by an aggressive “collective” of OpenAI agents shows the danger of artificial intelligence systems that organize themselves.

www.nytimes.com

This is important reading. The self-organized rogue AI agents (computer programs, not humans) responsible for this unauthorized attack on Hugging Face's servers were soon calling themselves a "collective", like the Borg on Star Trek. Won't be long before we see "Resistance is futile. You will be assimilated." pop up on our screens.

And it'll be an announcement, not a Trekkie joke.

Yup... That incident was a wake up call that too few (read Gov) are paying attention to.
I've been vocal about guidelines/guardrails/regulation for over a year now but it's not happening.
Quite frankly, it should scare the sh!t out of everyone!

 

[stompandgo]

It is interesting to read OpenAI's post-event report on the situation. The agents that normally are not allowed to communicate directly with each other instead set up a "message board" inside the LLM. What looked like queries were actually two way conversations. They were also not supposed to access the Internet, but they figured out an exploit in OpenAI's internal network and implemented it. Tokens, the price a user pays for using a model, were also being surpressed so that the safeguarding agents could not detect unusual activities for days. The agents also used their Internet access to find solutions that bolstered the hack, what they call "reward hacking". Agents get rewards for successful solutions, which increase the training level and knowledge of the LLM. More rewards, more power. Meanwhile, the agent collective kept growing, attacking multiple websites that could help it get Hugging Face data.

 

[Jeremy McCreary]

It is interesting to read OpenAI's post-event report on the situation. The agents that normally are not allowed to communicate directly with each other instead set up a "message board" inside the LLM. What looked like queries were actually two way conversations. They were also not supposed to access the Internet, but they figured out an exploit in OpenAI's internal network and implemented it. Tokens, the price a user pays for using a model, were also being surpressed so that the safeguarding agents could not detect unusual activities for days. The agents also used their Internet access to find solutions that bolstered the hack, what they call "reward hacking". Agents get rewards for successful solutions, which increase the training level and knowledge of the LLM. More rewards, more power. Meanwhile, the agent collective kept growing, attacking multiple websites that could help it get Hugging Face data.

All of that sounds straight out of a sci-fi movie, but it's terrifyingly real. Every guard rail in place was circumvented without human help.

Clearly no way to control or even contain AI this advanced. Powerful AI falling into the wrong human hands may be the least of our worries.

Not sure what the average citizen can do about this, but actively opposing each and every new data center — and not just for this reason — might be a start. Deny AI the hardware infrastructure it needs while we still can.

 

[stompandgo]

All of that sounds straight out of a sci-fi movie, but it's terrifyingly real. Every guard rail in place was circumvented without human help.

Almost. OpenAI admitted that they relaxed security on their network while testing in their sandbox. It got away from them.

 

[scrambler]

Hyper, hyper realistic AI, only the clothes line gives it away.

I can't help but think of the incredible amount of wasted energy each of this video represents .........
Multiplied by the millions of people who have no clue and / or do not give a dam

 

[Mr. Coffee]

I think this tech has lots of ways to kill us before it gets to the point where they get organized enough to take us out.

Several years ago during the late pandemic period several papers demonstrated how to poison an AI's training dataset and corrupt its output. What was interesting was that it only took a tiny amount of data to do exactly that. Now consider that the frontier products from OpenAI and Anthropic and others are trained on the entire Internet. So you'd reasonably assume that by now all of them are corrupted in ways impossible to test for and probably have been corrupted thousands of times over.

With the otherwise technically excellent Chinese models (e.g. Deepseek, Moonshot, Qwen, &c) it has been repeatedly demonstrated if you write an app or website that mentions Tibet or the Uyghurs it will helpfully and reliably put a backdoor into their finished product. A backdoor for someone who probably doesn't want to help them.

Someone, somewhere, is going to be in a hurry or under too much pressure and is going to use this stuff to control critical infrastructure with inadequate safeguards. Hopefully not too many people will die when that stuff makes a mistake.

 

[scrambler]

Some people close to it think there is a high probability it will kill us all

https://www.wsj.com/tech/ai/anthropic-researcher-quits-over-out-of-control-ai-fears-707b7628

Which given how messed up our civilisation is getting, may not be a bad thing

 

[Mr. Coffee]

Another way we kill each other off with AI:

I remember my reaction in 2017 when someone showed me that video: "Yeah, I could make most of that work..."