Is chat GPT a danger to humanity?

[DefTrap]

Hackers have developed an AI tool that can work out your password by listening to the sound of your keystrokes.

No, they haven't.

Researchers have built a model in a lab by extensively training an AI with a specific keyboard so that it can recognise which individual key is being pressed 93% of the time in a test.

It wouldn't necessarily work with several keys being typed quickly, it wouldn't necessarily know if the shift key was being held down, it wouldn't necessarily work without the extensive training on a particular keyboard, it obviously wouldn't work unless you happen to record the keys being typed and researchers aren't hackers. Plus of course, 93% accuracy sounds amazing, but this is per key. So for a 10-character password it's less than 50%.

So yes, it's impressive, and yes, potentially a real threat at some point. But it's not a current threat in any way yet, no.

And how do they know if it's a qwerty or azerty or any of the numerous variants? Plus laptop and tablet keyboards are pretty much silent.

Cyberdyne systems must be crying in their sleep.

[Slenver]

And how do they know if it's a qwerty or azerty or any of the numerous variants?

It wouldn't know anything until you train it which keys are which. As above, this is a lab test rather than anything real world.

Plus laptop and tablet keyboards are pretty much silent.

It said it was done on a laptop. It's pretty impressive and supposedly is based on how close to the microphone the keys are.

[Mr. Dazzle]

If bet there's someone somewhere who can do this by ear I saw a guy on TV once who could identify LPs just by looking at them*.

*I mean technically anyone can do this, if said LP has a label on.

[Screwdriver]

And how do they know if it's a qwerty or azerty or any of the numerous variants? Plus laptop and tablet keyboards are pretty much silent.

Cyberdyne systems must be crying in their sleep.

No idea, I saw it as a side comment on an AI vid that said hackers not researchers. That is not to say that hackers aren't also developing this newfound ability. I just thought it was interesting, no idea how it does it but a quick Google confirms the story is substantially true. FWIW, here's the video.



..and yes, looking back I can see I misrepresented "researches" for "hackers can".

As for the other stuff, it is comforting to know there are so many other ways "researchers" can steal your passwords. I don't know which idiot thought that was a good research project. Probably got sacked from Wuhan labs where they definitely weren't developing a weaponised virus, for "research".

[Mr. Dazzle]

Working out how to break into computers is how you make em more secure innit?

https://en.wikipedia.org/wiki/White_hat ... prov=sfla1

[Yorick]

Some folk are way behind the times. I was working in the virus lab at McAfee 18 years ago. Some scary stuff

[Screwdriver]

Working out how to break into computers is how you make em more secure innit?

https://en.wikipedia.org/wiki/White_hat ... prov=sfla1

I think teaching AI how to do it is pushing that idea a bit too far tbh. It is inevitable of course. The entire concept of a password is doomed to failure in the long term anyway but I personally don't want to go too far down the biometric "security" route. Keep dogtags for dogs (and dogs of war).

[Mr. Dazzle]

My answer here is to be poor enough not to steal from, it's working well so far

[wheelnut]

The entire concept of a password is doomed to failure in the long term anyway …..

It already is to a large extent.
Something you know
Something you have
Something you are

[Mr. Dazzle]

To be fair I can't think of a single important online thing I do which requires only a password...

Cept this place of course

[Yorick]

I remember when I worked for Travele0x as DBA and was sent over to a big bank in Amsterdam called Rababank to do a complete install of Ingres.
To do it properly I'd need the root password for the whole weekend.
They only agreed to let me have the Ingres database password.
The bank CEO agreed to come in for a bit on the Saturday and give me 15 mins access with the root password.

Me being a Unix Sysadmin in previous life was able to capture the password for later after he logged that session out

[Screwdriver]

Apparently some geek managed to get hold of GPT4 and stick it on the internet! Bit of an issue due to the eye watering costs for training that thing. Some of these numbers, scratch that, ALL of these numbers are mindboggling:

- The total GPT-4 model has around 1.8 trillion parameters with 120 layers
- It's made of 16 different models, each with ~110b parameters and trained for a specific task/field (a technique known as MoE or Mixture of Experts)
- ~55b parameters are used solely for 'attention', i.e. guiding the model to stay on the topic at hand
- They believe GPT-4 was trained on ~13 trillion tokens (~10 trillion words)
- Some of these tokens were re-used (twice, or four times if it was code-based), so the actual size of the dataset is unknown but probably several trillion tokens
- They fine-tuned the model using more specific data from OpenAI and ScaleAI
- It took ~2e25 FLOPS of compute to train, using ~25k A100 processors for around 3 months
- GPT-4 is about 3x as computationally expensive to run than GPT-3.5
- It's thought a faster/cheaper model takes over after the first few words of a response, which could explain the complaints that ChatGPT got worse over time

Copied and pasted off a link from SemiAnalysis or Yam Peleg (for anyone wanting sources.

That means they used twenty five thousand high end Nvidia GPUs each costing about $25k (!) to do the training and while I think these companies rent resources rather than buy those GPUs, that's still tens of millions $$$ which got nicked and shared!



The page for the data source is deleted but once it's out there, it's out there...

...and I remember back in ye olden days when millions of parameters was still a thing. Right back in the distant past not long after I started this thread. Yeah. That's exponential growth for you. Forget Mores law, these things are accelerating like a dragster on methanol.

So in the past few months (say!) while GPT or AI generally became a billion times "smarter", how much smarter did humans get?

Looks like by the end of this year we'll be talking about Petaflops and YoctoBytes like anyone really comprehends the scale of such vast numbers. While I don't know what this really means, for a scale approximation, the human brain has about 100 billion neurons.

Those 1.8 trillion "parameters" are organised into "layers" and the layers are stacked in a "Multi Layer Perceptron"



...and that diagram effectively represents a neuron in a software neural net. So the numbers can't be compared like for like and in a way, a neural net is not a brain. It is being structured like one though...

[Greenman]

Apparently some geek managed to get hold of GPT4 and stick it on the internet! Bit of an issue due to the eye watering costs for training that thing. Some of these numbers, scratch that, ALL of these numbers are mindboggling:

- The total GPT-4 model has around 1.8 trillion parameters with 120 layers
- It's made of 16 different models, each with ~110b parameters and trained for a specific task/field (a technique known as MoE or Mixture of Experts)
- ~55b parameters are used solely for 'attention', i.e. guiding the model to stay on the topic at hand
- They believe GPT-4 was trained on ~13 trillion tokens (~10 trillion words)
- Some of these tokens were re-used (twice, or four times if it was code-based), so the actual size of the dataset is unknown but probably several trillion tokens
- They fine-tuned the model using more specific data from OpenAI and ScaleAI
- It took ~2e25 FLOPS of compute to train, using ~25k A100 processors for around 3 months
- GPT-4 is about 3x as computationally expensive to run than GPT-3.5
- It's thought a faster/cheaper model takes over after the first few words of a response, which could explain the complaints that ChatGPT got worse over time

Copied and pasted off a link from SemiAnalysis or Yam Peleg (for anyone wanting sources.

That means they used twenty five thousand high end Nvidia GPUs each costing about $25k (!) to do the training and while I think these companies rent resources rather than buy those GPUs, that's still tens of millions $$$ which got nicked and shared!



The page for the data source is deleted but once it's out there, it's out there...

...and I remember back in ye olden days when millions of parameters was still a thing. Right back in the distant past not long after I started this thread. Yeah. That's exponential growth for you. Forget Mores law, these things are accelerating like a dragster on methanol.

So in the past few months (say!) while GPT or AI generally became a billion times "smarter", how much smarter did humans get?

Looks like by the end of this year we'll be talking about Petaflops and YoctoBytes like anyone really comprehends the scale of such vast numbers. While I don't know what this really means, for a scale approximation, the human brain has about 100 billion neurons.

Those 1.8 trillion "parameters" are organised into "layers" and the layers are stacked in a "Multi Layer Perceptron"



...and that diagram effectively represents a neuron in a software neural net. So the numbers can't be compared like for like and in a way, a neural net is not a brain. It is being structured like one though...

They will get there soon! For what reason scares me i have to admit!

I just hope they have already got there elsewhere that we don't know about though so what they produce won't wipe us out!

[Screwdriver]

This has to be one of the nicest AI topic videos I have seen to date (I've watched a few). Lots of nice pictures, interesting AI interviews and it doesn't hammer home the "humanity is doomed message".

[Screwdriver]

They will get there soon! For what reason scares me i have to admit!

I just hope they have already got there elsewhere that we don't know about though so what they produce won't wipe us out!

An interesting perspective on exponential rate of progress from the above video: how thick is a piece of paper folded 42 times?

"There" is the biggest issue currently because we cannot really know what their motivation (or the AI equivalent: goal, ambition, target etc.) actually is because it is not something we can "program. If it was programmable, it would not be AI.

My best guess is the "prime directive" is to increase entropy which appears to be the fundamental "purpose" or "goal" of any system which resides at low entropy. The principle of least action and the second law of thermodynamics conspire to inevitably increase entropy in any closed system. If the closed system is the entire universe, then it's going to be a long task.

Humanity has scratched the surface on this planet by creating ever more complex methods for increasing entropy. That appears to be our mission whether we realise it or not.

The principle of least action is the feature which interests me most about this instance of universe (if you're a multiverse proponent) because it's as if matter/energy is basically just being lazy. Light does not travel in straight lines! It follows the path of least action so if a path passes though a number of diffferent mediums, air/water/glass etc. each photon will inevitably travel a route which we perceive to have taken the least amount of time. I could expand if there's any interest but it fascinates me because just the fact that a pencil in a glass of water "looks" "bent" takes you donw an fantastically deep rabbit hole of fundamental behaviour.

You (yes you reading this message) are one of the most extraordinary consequences of the behaviour of matter/energy and the inevitable increase in entropy. The universe is not a "thing" which is "out there". You ARE this universe.

So am I.

So is A.I.

[Noggin]

Been giggling quite a lot about this!!

[Yambo]

Made me laugh!

[DefTrap]

Made me laugh!

I guess it's useful feedback, but a part of me still wants to see the followup where the yank squaddie gets nuked into orbit after pissing around whilst somersaulting.