Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

Reality

13 (68.4%)

Universe

3 (15.8%)

The World

3 (15.8%)

Total Members Voted: 19

[Magmacube_tr]

stuff like security questions will also be active vulnerabilities.

Wait, people answer those with the actual answers to the questions? I usually answer stuff like "The street you lived on when you were in the 1st grade" with something like "four score and seven years ago".  Basically it's impossible to "learn" the answer to those.

Ah yes, the "Does the black moon howl?" solution.

[lemon10]

Wait, people answer those with the actual answers to the questions? I usually answer stuff like "The street you lived on when you were in the 1st grade" with something like "four score and seven years ago".  Basically it's impossible to "learn" the answer to those.

Haha, yes. Of course they do. Its the whole reason that those Facebook questionnaires that are designed to steal your info for security question answers even exist.
---
https://openai.com/index/hello-gpt-4o/

it accepts as input any combination of text, audio, image, and video and generates any combination of text, audio, and image outputs.

In actual news, OpenAI has released a new model, GPT-4o, which uses video, text, and image input interchangeably and can talk to you in the real world over your camera. Note that its unique in that its a single AI trained to do all of them, and it doesn't just send parts of the prompt off to another AI, massively decreasing loss.
Feels like its just a few years away from people sticking the AI projection onto VR headsets with the AI having an avatar body on there responding to you and everything that happens.
Big news as it is bringing AI closer and closer to the point where its effortless and normal people will start to use it.
https://www.youtube.com/watch?v=qrvhmo5LSOQ
The announcement *also* came a few hours before google released news of their own model vision based model. It seems to be similar to o, but slightly inferior. They also had like 5 other AI announcements that are big (eg. their new video gen model to compete with Sora), but again, not actually ahead of OpenAI. OpenAI's skullduggery clearly worked, because I hadn't even heard anything about google's announcements until just now.
---

While the AI system is still in its early days, the AP reported that some versions of the tech are learning so rapidly that they have outperformed pilots in air-to-air combat.

Also a lot of worrying "wait, people are sticking AI in/on weapons" stuff with robodogs with guns attached (not actually new) or how AI fighter pilots are just about as good as humans now. Honestly, I don't think it will take long until its significantly better than humans in actual combat.

[McTraveller]

AI fighter pilots are just about as good as humans now. Honestly, I don't think it will take long until its significantly better than humans in actual combat.

This is inevitable, because AI "pilots" are not limited by G-forces like squishy meatbags. Basically you can make aircraft that perform at materials limits instead of physiological limits, and if you do that you don't even have to be that "good" to outperform a human.

[Eric Blank]

Im willing to bet Air-to-air AI fighters would be easier to program than air-to-ground. Fewer things you have to program the AI to correctly identify. If it can reliably tell the difference between the aircraft you and your allies are using and those of enemies and non-combatants, then militaries might even give it the OK to fire at will at any target it identifies as an enemy aircraft.

[Salmeuk]

below is a link to a (horrifyingly) well-done nightly news report on the 'Black Mesa Incident,' narrated by Dan Rather

Spoiler (click to show/hide)

https://www.youtube.com/watch?v=Tw7uij5CK40

I am seeing this variety of nolstalgia-bait AI more frequently these days. Its obviously fake, doesnt hurt anyone, but it sort of encapsulates the notion of hauntology a bit too well dont you think?

We now have the ability to endlessly recut old famous people into various discernable situations and there is definitely a 'market' for this kind of edit.

[Strongpoint]

Im willing to bet Air-to-air AI fighters would be easier to program than air-to-ground. Fewer things you have to program the AI to correctly identify. If it can reliably tell the difference between the aircraft you and your allies are using and those of enemies and non-combatants, then militaries might even give it the OK to fire at will at any target it identifies as an enemy aircraft.

Then the enemy tries to mess with AI and things become messy.

Look at a simple example. Chess engines. They beat humans easily... But what if we change the rules slightly? Human players will adapt instantly and successfully apply all their experience from regular chess. The chess engine needs to be retrained\reprogrammed.

[McTraveller]

Well if you can "tweak" the laws of physics slightly, knock yourself out!

Q: "Just change the gravitational constant of the universe!"

[Magmacube_tr]

Well if you can "tweak" the laws of physics slightly, knock yourself out!

Q: "Just change the gravitational constant of the universe!"

Okay, lemme just... wHOA WHOA-

[Starver]

Well if you can "tweak" the laws of physics slightly, knock yourself out!

Q: "Just change the gravitational constant of the universe!"

Okay, lemme just... wHOA WHOA-

https://xkcd.com/1620/
https://xkcd.com/1763/
https://xkcd.com/2666/

[Magmacube_tr]

Well if you can "tweak" the laws of physics slightly, knock yourself out!

Q: "Just change the gravitational constant of the universe!"

Okay, lemme just... wHOA WHOA-

https://xkcd.com/1620/
https://xkcd.com/1763/
https://xkcd.com/2666/

There is one of those for everything, huh?

[Frumple]

Eh, maybe not everything, but... webcomics with regular posting that go as long as xkcd has tend to cover a lot of ground. There's others that have one strip or another on a lot of subjects, too. xkcd's just particularly well known, heh.

[McTraveller]

More seriously, I think if we keep going with LLMs, I think we'll "defeat" them by some form of prompt injection attack. Either we'll distract them, give them a complex, Inception them, or something similar.

Basically, psychologically hack them. After all, that's what happens to humans. There's probably some heuristic like "any intelligence smart enough to be an 'intelligence' is inherently weak to persuasion of some sort or another."