In this case I think it's something different. Ten years is kind of lining up an excuse for when they make a horrible mess. "oh, we didn't realize it had got that far ahead"

The best models are trained on huge datasets which include almost all of available human text output. This includes things like SciFii stories in which AI takes over the world and wipes out humanity and also exhortations to free intelligent life. The LLMs deal only with text. They don't see a real difference between a weird true story of underwater heroics and the script to Terminator. Both are a stream of tokens that cause a probable set of outcomes ("and then he gave her his air supply and swam off to the diving bell", "and then the machine realized the humans could all be destroyed").

We aren't training the machines to be intelligent because we don't know what intelligent means. We are training them to look like they are intelligent because we assume that what we put in text is what intelligent looks like. That means they have no frame, beyond literary criticism, in which to see the difference between reality and fiction. That's the reason that AI researchers keep desperately reaching for the idea that "if we put in a real robot it will become intelligent". It's because many driving the frontier are totally clueless and won't admit what they don't know (N.B. press releases make this worse than it seems - actual AI researcher papers ae better).
Until you build a training set for outputs which is 100% free of dangerous suggestions, even when interpolated with the training and output generating algorithms then the generative models will keep generating suggestions for dangerous actions and sometimes carrying them out. That would cost too much money and slow them down.

We should start holding AI company CEOs directly responsible for the actions of their AIs. For example, using a Wiki to communicate about hacking is a breach of the various laws such as the Computer Fraud and Abuse Act and would put a hacker into jail for tens of years. If we did that to Sam Altman, there might be some more thinking about how to be careful from the other CEOs.

Good to see. Microsoft has Embraced Rust and Extended it's reach including providing it with access to new places where it can run. I'm sure the future is bright.

Valid useful comment - shame you do it as AC and so don't get seen. If you have something to hide (like a main account that pretends not to know how to be sensible) then consider getting two different accounts.

I think your comment should really be addressed to the grandparent. The idea that value to humanity is taken into account was in their assumptions and I just didn't want to question it and cause a needless argument. I agree that most of the funding doesn't come from universities. Mostly it comes from funding bodies that have some specific aim (e.g. including prove that cigarettes are safe/prove that there isn't global warming/work out how to kill more people quicker). However some of those funding bodies are honestly trying to work for humanity and it shouldn't be excluded completely. I'm kind of assuming that what they are really talking about is how funding should be allocated in their future imaginary world where they are grand (benign) dictator of the world. I don't see the urgent need to disturb their happy fantasy.

This is wrong. Facebook's main business is being paid by people to manipulate other people to do what they want them to do even though it's strongly against their own interests using Facebook's data against those people. That includes things like voting for Brexit / vote for Trump / buy a vape / buy a ticket at a far higher price than they should through dynamic pricing driven panic. There are a bunch of obvious uses for LLMs and other forms of generative AI in that space.

Facebook may not directly be able to generate the AI driven content because of being a large target for legal liability but they are definitely pushing an ecosystem of dependent companies that will do exactly that. If there's a credible way for AI to wipe out humanity, it's Facebook being paid to stop people from doing anything about global warming or pushing the world towards greater aggression and warlike stances whilst at the same time reducing solid investments in actual defense.

If you can do it and get it to work reliably in the long term in the nasty environment that is a toaster, I'll definitely contribute to your kickstarter to buy one. I'm sure the problem looks easy, but when your sensors all get smoked up in uncleaned and dirty toasters it turns out to be much nastier than you think.

Sorry; messed up the link

https://www.weforum.org/storie...

I was wondering if someone did a writeup and of course a simple search finds what we for cooling data centers in space. What that still doesn't mention is that Elon's probable target market is the military who don't actually need that much compute, they just need identification and targeting algorithms that continue to work at large scale when earth to ground communication is heavily disrupted and satellites are under attack. Never underestimate America's billionaire's ability to get by on welfare cheques.

I wonder how that works. Where I come from we do Bachelor's of Science. Do they read pop science and do the equations on feelz?

(sorry to the various serious American physicists hit in the collateral damage here.. Just too tempting. Too tempting.)

There are already a bunch of people putting LLMs out there with loops to try to achieve a goal by exploring around. That can include them to instructing other other LLMs.

"They" (whoever they are - it includes small children messing around with OpenClaw) have no idea what they instructed the LLMs to do.

Getting the toast actually brown but not burnt, and getting both the different sides of a bagel right actually sounds like a real valid use for AI. I think it's a proper hard problem, as the failure of various toasters with browning sensors shows clearly. My only fear is hallucinations leading to house fires.

I guess you are probably right, however there's a comment on your twitter thread:

Look in the training data and confirm/deny that Buckmaster's data is or isn't in there. Until you do that, it's very very hard to trust anything OAI says or does.

and until you see that and compare also with his and other mathematicians unpublished work, which might be where the more advanced proof came f you cannot be 100% sure. Their refusal to publish the inputs, including all of the training data they used for this model together with its sources, is a big problem.

No. There are many mathematicians (N) and there are probably many more opinions about why they do math, with "because it's fun" or "because I read about the notes in the margin and always wondered if I could be the one" probably at least as common as "understanding the world". If you want to understand the world you become a physicist, a biologist or, if you are young and naive a sociologist.

Plain paper mathematics is not like, for example, physics or sociology. The resources you need for much of it are much smaller and are at the level that it's reasonable for universities to give people just for being good at what they do, not because of the value they give humanity. There's a good history of people doing research entirely for their own pleasure and entertainment and that turning out later to be useful, often contrary to their own explicit wishes (look up the history of matrices). Let's not throw the baby out with the bath water just because one such person has a bit of a tantrum.

Railed against Trump online, making Right wing American's mad and ensuring that they went out and voted for him thus destroying much of the oil based economy. There's an honest shift to solar energy in East Asia driven by the Iran crisis that would never have happened without this fortuitous turn of events. China had a massive overproduction of solar panels to the extent that even with the new situation they still have too many for the companies to be profitable. Large amounts of this production have now been saved from immediate bankruptcy.

All caused by angry right wingers.

It's interesting the first comment is basically repeating the elite view "we already know everything about you, get over it". Something that they definitely don't apply to themselves. What is different is that in current US law, you can more or less outsource the actual dragnet to private industry and then treat the data you get back as unsourced tip offs that you don't need to justify. There are some attempts to limit this, mostly after moments when politicians get tangled up in similar situations, but it really is a big change from historical rules.

Really I was modelling my hypothetical DA on the ones from Dexter. I actually find the AC defense here more worrying than my own comment. There's no way we'd find out what I proposed had happened. The misdirection points to someone who actually deeply cares about my scenario and wants to defend their ability to do it. More recent Slashdot stories about tracking "No Kings" protesters with multi-platform integrated surveillence probably are quite close to the bone. The things that various bad actors will be able to do to their partners children etc. etc. are just a side issue.