White House Secures Commitments from Leading AI Companies to Manage AI Risks
whitehouse.gov
whitehouse.gov
Lisa: Promise me you'll never die.
Gary Johnston: You know I can't promise that.
Lisa: If you did that, I would make love to you right now
Gary Johnston: I promise I'll never die.
https://www.youtube.com/watch?v=aplSQGHPmvI
Lisa Secures Commitments from Gary to Never Die.
The public have all heard “AI is going to destroy humanity” but those statements have been short on specifics.
I have been asked by many people to explain how AI could kill us all. How about some clarity for common people.
That stuff is very speculative. But we don't need AI to get to that level for it to be very dangerous. Just imagine that we have an open source GPT that is something like 33% smarter than GPT-4 and less brittle. Make the output 50 times faster than human thought and suppose that we can run it very inexpensively, in a "swarm" of agents cooperating.
Then you have a type of superintelligence that does not require any really speculative AI advance -- hyperspeed reasoning -- based on the current technology.
If that is widely deployed for military and industrial decision-making then something like a computer virus could create existential risk.
Also, realize that all these systems really need to emulate some of the core functional aspects of animals, like a self-preservation instinct or desire to control resources, is the right instruction and relaxation of guardrails. So when they can reason a bit better and at hyperspeed with collaboration between them, they don't need to be alive or anything to be dangerous.
Look at the history of increases in computing efficiency. It is easy to imagine 50 or whatever times output speed increase in less than 5 years. And even though they are not human level now, GPT-4 has proven that these types of systems can have strong problem solving ability.
One more thing to add is that as the AI performance increases and surpasses human decision-making ability, the hyperspeed will push competiting companies and countries to give them broader and broader goals and more autonomy. Because waiting overnight for human input means your competitor's AIs race ahead the equivalent of weeks.
I think we need to set a limit for the performance of new hardware at some point. Also we need people to understand the dangers of full autonomy and imitating animals in the context of hyperspeed or superintelligent AI. We will need criminal penalties and maybe some type of cooperative digital immune system.
An actual AGI, that can think and has it's own motives and initiative is an entirely different danger to GPT-4, -5 etc which cannot. Those might post a danger in terms of fake news, or worker displacement, but they're just social issues. No rate improvement in chat GPT will make it "decide to kill all the humans" etc.
Or to put it another way, I can't see how artificial intelligence could be any more dangerous than intelligence as we already know and have them.
We already know humans are dangerous. And that is humans who are conditioned to be social, and who have empathy and who rely on other humans for all sorts of things. Plus with fear of a legal system punishing them.
Now imagine a human like intelligence without empathy, that is not conditioned from childhood to be social, and that is inherently threatened by us and not reliant on us. Would you want that guy living next door?
Oh and it's 100x smarter than you.
I think all this is 100 years away, but that's the theory. Skynet doesn't have a subconscious telling it not to murder it's parents...
The thing that keeps us safe from computers is generally we don't let them have access to anything dangerous and when we do theres typically an evaluation of what can go wrong and how we'll prevent the computer from doing that (often in the other direction, we only program the computer to do X in Y situations). But sometimes that fails and you get the unexpected rapid acceleration of cars.
Or a giant knife switch on the wall that makes that huge clunk noise when turned to the off position, and everyone hears the sound effect of the power shutting down.
We've had computers in charge of so much for so long, but there's always a manual override to allow the operation without the computer. Manual valves to be turned (hopefully they haven't been rusted stuck so the timer on the countdown gets dramatically close to 0).
you're bonkers if you think we can't make manual overrides like this.
However, Chat-GPT is only able to interact with people because it was hooked up to the internet not because it figured out how to access the internet. It also didn't create a worm and infect the entire world, the second you leave the webpage you no longer have Chat-GPT.
The way to keep the AI from accessing it shouldn't is literally just don't hook it up to those things in the first place.
* A system like chat gpt that can do things a human cannot (find a way to kill all humans that's achievable by a small group).
* A system different to chatgpt that has its own motives and initiative.
Both need to be achieved.
> No rate improvement in chat GPT will make it "decide to kill all the humans" etc.
All I'm trying to say is that GPT-style architectures not having "agency" shouldn't be very reassuring. How intelligent they can get is a separate question and one on which I am agnostic.
No. Then you have something that can spout plausible sounding nonsense mixed in with facts gleaned from crawling the internet.
There are some dangers to this - don't get me wrong. It could be used to create bots which engage in much more sophisticated social media manipulation on a large scale, for instance.
Take a look at https://en.wikipedia.org/wiki/Federal_Motor_Vehicle_Safety_S...
You'd think if people are so scared about AI being apocalyptic it would have some strict regulation similar to vehicle standards but its not even close, nor is it in the conversation at all. Its just pure fear from science fiction and the govt putting out performative press releases. What the fuck even is "AI" anyway? it means different things to different people. It could mean a self checkout at a supermarket, or a bank's AML, or some super intelligence. there are basically no standard definitions of anything to do with "AI" anywhere and every single piece of regulation is equally as vague. Companies like having it be a vague term because its marketing so the actual experts don't care about changing it
Indeed.
For most lay people I've talked to -- family, friends, and acquiescence outside of tech -- the answer ultimately boils down to "automation that can do my job or can do the job of people I care about or can do the job of people who would then compete for my job". Sometimes also "algorithms" a la social media feeds and so on.
Ie, the lay answer to "what is AI?" has recently morphed into roughly "computer programs that I am worried about or don't like".
For most actual AI researchers I've talked to, you get the standard tounge-in-cheek response ("AI is whatever gets published at AI conferences and funded by AI investors") and then, when you press, it's roughly something like ML+CV+NLP+optimization and sometimes but not always other fields like automated reasoning or parts of robotics get added. Roughly, most of the parts of CS that aren't theory or systems or HCI or software engineering or already closed.
For most techno-babblers in the podosphere who have OPINIONS on super-intelligence but couldn't pass an internship phone screen to save their lives, it's superintelligence and other stupid hollywood bullshit.
But in some sense, so it "tech". The paperclip is a highly evolved technology, for example, [1], but working on those doesn't mean you work "in tech". Tech meant something like, "the new stuff that is impacting our lives but we don't know how to handle". So a CTO's job generally doesn't include the paperclips, the copiers, or the company cars, however technological they are.
And as with any shiny new marketing term, others will quickly rush in. There was the craze for radioactivity, which resulted in a bunch of radioactive patent medicines: https://en.wikipedia.org/wiki/Radioactive_quackery
But even more interesting to me is the extent to which it was used as a pure marketing term, with no radioactivity expected. E.g.: https://lucyjanesantos.com/a-batschari-radium-cigarettes/
So I'm sure we'll be seeing all sorts of things branded "AI" even when they don't use any of the technologies involved. With no trademark on the term or organization to defend it, it's open season for all the sketch marketers.
[1] For those who doubt, Petroski's "The Evolution of Useful Things" will set you straight: https://www.amazon.com/Evolution-Useful-Things-Artifacts-Zip...
Things like traditional chess programs (i.e., those that don't use neural networks/ML) are maybe a borderline case, as while they are just tree search with a bunch of human-tuned evaluation heuristics, the search depth is so huge that it's difficult for humans to explain why they take certain actions.
Once marketers get hold of a term, it doesn't really matter what people who like definitions or precision think. As an example, I was active in the Agile movement before that term was coined. When I've talked recently with the early-days people I keep in touch with, we're all kinda horrified by what the term has come to mean. And it's not like we didn't fight along the way for clarity. But most people just don't care about the precise use of a term when there's a profitable (mis)use of the term.
AI doesn't need to be alive or animal-like to be superintelligent. We have many examples getting more and more general purpose. Look at AlphaGo, AlphaStar. Both are superintelligent in a somewhat narrow way (but trending towards less narrow).
GPT-4 is superintelligent in some ways already such as it's breadth of knowledge. And we have to anticipate that it will continue to get smarter and better at reasoning and much, much faster. Over the last ten years AI performance has increased by 100000 to 1000000 depending on how you measure it. It will continue to accelerate.
BTW, as far as internships etc., I started programming 38 years ago when I was 7. I've built a ton of software with numerous technologies, including recently some things like: a multilayer perceptron (from scratch). And a data analysis tool that uses GPT to write complex SQL and create charts on the fly to answer user requests. Also software to automatically create custom websites (including imagery) based on short descriptions, or automatically write, test and debug scripts on servers. And all of that outputs at superhuman speed already. And before that I worked on dozens of other complex projects.
Point being, I understand technology and AI, and it is ludicrous to view superintelligence as "stupid bullshit".
There are already millions of people that are stupidly easy to convince of anything, so why aren't these magical end of the world scenarios already happening? Where's the evil maniac convincing Kelly the soccer mom who is into MLMs to end the world? Where is the supervillain raising an army of a million gullible idiots to invade Canada and have their own kingdom?
AGI "Superintelligence" even as a concept is just completely unsupported. It's people looking at an extremely cropped graph and saying "Line goes up, must go up forever" and wouldn't you know it that group has massive overlap with those MLM people from before.
So it seems like the playbook is something like: 1) play up non-problems that, thanks to 50+ years of sci-fi killer AI, will seem real to the rubes; 2) breeze on past the numerous actual problems; 3) solve as few actual problems as possible as they race to capture billions of dollars.
But happily, the contents of this press release focus mostly on actual problems, so I have some hope that the Skynet gambit won't work.
Some of them, sure, but that doesn't seem to be true of plenty of signatures to the open letter.
> the contents of this press release focus mostly on actual problems
Given that this is a set of entirely voluntary commitments made by the companies who you had thought were hyping x-risk in order to distract from actual problems... maybe it's worth updating your assessment of what the x-risk thing is about?
FWIW I don't think these voluntary commitments are sufficient or address all of the important harms. But that's different than saying any government action is just "regulatory capture" and an attempt to keep open source models down. I'm only attempting to argue against the latter here, which I'm not sure is where you're coming from.
Not particularly without barriers to competition in the space being artificially erected so as to enable that, which is the whole point of the industry-government game of footsie.
No, I think wealthy and capital rich companies can leverage AI to extract value from places that already extract value simply by underpaying people and use stupid rhetoric and other methods that they are very familiar with to continue buying up basically everything as most normal people struggle to survive.
I explicitly do not think "AI" or even AGI has any inherent danger in itself, and only has danger in the same way any automation has in a capitalist system that also gives richer people more power in the justice system and the political system: Rich will get richer by squeezing the non-rich even harder, while politicians don't care because most of them genuinely believe in capitalism as an unabashed and incorruptible good and the misinformation and discourse control that passable bullshit generated with a single keystroke grants them way more power
Nope. Because I think the White House's press release is expressing the White House's view of what's important. That the White House didn't get distracted by the chaff says nothing about the chaff or those firing it off.
The whole argument with AI safety seems to be about how AI can say some "harmful misinformation". Ok. I can say harmful misinformation too, and I'm a real person with feelings that can be expressed to other people.
"AI safety" takes a view that people are inherently stupid, that people have no ability for complex thought and critical thinking.
To me this seems like AI companies such as OpenAI want to stop open-source and monopolize AI where you have to have a government certification to create AI models. Plus a bit of paranoia from the totalitarian liberals who only want consensus rather than debate.
At least there are valid concerns for training models on copyrighted works. Maybe we should focus much more on that rather than on "the AI said something that I don't like".
AI has the same powers and risks of organizations -- meta/mega human actors.
The difference is that machine learning technology can potentially operates at orders of magnitude of higher speeds, and potentially without humans in the loop. (The threat model is less "flip the on switch and everybody immediately dies"; but "flip the on switch, it works so well it gets woven into our infrastructure", until a critical threshold is met, and some seemingly-innocuous Make Number Go Up algo leads to humanity having a Very Very Bad Day.)
So all the same perverse incentives (orthogonality) of human-org AIs, but iterating thousands or millions of times faster.
Look at the regulations on nuclear bombs.
We have been holding back equipment because "what if we need it" (to fight who?), or "but they'll retaliate!" (with the one tank they could afford to spare for the parade?) or "but it costs so much money" (only when you calculate it with the price we paid in 1980 instead of the negative value these old machines that need to be sold or decommissioned currently have) or "We can't send cluster munitions because the UXO risk!" while every day Russia is still in Ukrainian territory is actual Russians purposely shooting at Ukrainians and RUSSIANS HAVE LITERALLY USED CLUSTER MUNITIONS ON CIVILIAN POPULATIONS!
The actual insanity is sending Ukraine 100 Bradleys instead of 500 and then bitching when their offensive is mediocre.
When we are passed the point where ppl with their 3 inch brains can actually come up with answers to complex problems, you don't spend time asking them questions.
There are countless resources to find this stuff out. To be clear there may not be a level at which common people understand why e.g. Fission Chain Reactions are dangerous but they can grok why a giant bomb that spreads radiation is bad.
For Starters: https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a...
>The public have all heard “AI is going to destroy humanity” but those statements have been short on specifics.
You can no more explain how Stockfish is going to beat you at chess then you can predict how an AGI would defeat humanity. Nevertheless the outcome is nearly certain.
@30:24, on what solutions might be realistic to help, if not a 6 month sanction, Miles says:
"Maybe what we should be asking for is just an enormous amount of money to do alignment research with"
Sounds like a classic doomsday cult, end is nigh, deposit your checks to this account...
Also note that until maybe the last year not only was there not a great financial incentive to be a "doomer" but it would actively hurt your career if you were. Most of the main people in the scene have been sounding this alarm for years sometimes decades. It's also difficult to explain people like Geoff Hinton or Yoshua Bengio joining and leaving behind high profile and highly lucrative positions. Yann LeCun, a staunch anti-doomer is perhaps the perfect example who has actually DOES have enormous financial incentive to play down AI dangers.
"Colossus The Forbin Project", "Terminator"
I.e. giving control of weapons to an AI.
Basically the argument is if anyone, anywhere creates an AI that has the same basic drives as a human (preserve yourself, replicate, preserve things that look like you) it'll start competing with us for resources. AIs have demonstrated the ability to crush humans at every competitive game we can play with them, so they'll crush us at that game too. It looks a bit far fetched in 2023 but it is actually pretty easy to see it happening. Hardware progresses exponentially for a few more years. Some war-torn state gets desperate, starts deploying military AIs to try and get an edge, loses control, yada yada. Or economic pressure starts engineering humans out of the loop in one too many places and evolutionary pressures happen by accident to develop an intelligent autonomous agent that wants to replicate.
Humans struggled with a cold virus recently and couldn't stamp it out. Something that can outdo us research-wise would not be something we can handle. Whether humanity has control of corporations is an open question given the number of people who get killed if it is profitable to do so. We've been luck that corporations need people to act as a brain. AIs don't need that.
Every big gorilla on the planet could kill me. None of them can outcompete me for something, because I'm not stupid enough to challenge a big gorilla to a fistfight. There won't be a fight until after I've got my hands on a gun and the gorilla is an easy shot. There probably won't need to be a fight because I'll just trade for what I want using tactics that a gorilla cannot comprehend the effectiveness of.
There is a pretty good chance humans won't even realise they are competing with an AI until they are locked out from having options. I'm not even sure if most gorillas realise to this day that we humans will crush them if they ever try anything that threatens us.
Edit: found one. https://youtu.be/u3L8vGMDYD8
I don’t think ping pong bots that can beat humans are here yet.
Honestly, my understanding is that's mainly a distraction from the much more realistic:
> AI is going to take your job and impoverish your family while making the rich even richer. But cheer up! You'll be able to talk to a chatbot therapist about your problems, from your cardboard home under a bridge, so it's progress! In the past, not even kings had chatbot therapists! You'll have it better than a king!
I advise you to speak to a qualified professional cardboardpenter or simply to go stand in the begging que in front of our headquarters in San Francisco.
Feel free to take your home with you, given how long the line is.
I've seen some papers on watermarking, but none of them are what I'd call "robust" - they're easily defeated by making small changes to the data. Are the companies "committing" to overcome an unsolved (and possibly unsolvable) technical challenge?
> Biden-Harris Administration will continue to take decisive action by developing an Executive Order and pursuing bipartisan legislation to keep Americans safe
Evidence?
> "no need to care about people living today or real problems in the world"
Nobody says this.
> "trillions of people in the future"
Weird longtermists who talk about this are a small minority even among AI doomers.
> "AI godfathers comfort, status, or money"
I'm not suggesting anyone feel sorry for Geoff Hinton but quitting your Google job does seem like a pretty obvious sacrifice of money.
But I dunno you packed so much confusion into two sentences that I doubt you're going to let any messy details get in the way of picking a side based on an attempt at populism (including the fact that you're aligning yourself with Marc Andreesen and Zuckerberg).
Would have been nice to see that. Even right now there are known security vulnerabilities that were found through testing but haven’t been fixed.
Great for voters to sleep better at night. Useless for actually "managing AI risks"
Not that those risks can seriously be managed long-term, IMHO. The prisoner's dilemma between nation states and corporations ensures someone will defect.
And now DeepMind wants to combine the language-based fairly general purpose reasoning ability of something like GPT with the superhuman strategic prowess of AlphaStar/AlphaZero/etc.
At the latest big keynote Nvidia touted the fact that they have accelerated AI by a factor of one million over the last decade, and project that they will do so again in the next decade.
Humans will not be able to compete at all, and even putting them in the decision loop will mean immediate failure. Human thought and action will "appear" to be so slow that it is essentially frozen compared to the operating speed of these systems.
140 million in funding for seven more research labs (might be a different order but still neat).
Commitments for cybersecurity funding (probably already happening anyways), independent auditing, and some testing.
The most interesting part was the watermarking. I'm interested to see how that works out, but it's a neat concept, something I hadn't thought of.
I like the focus on reducing bias, that seems like a good effort and of course the focus on helping to cure cancer which I'm always a fan of Biden being such a champion of.
Also reaching out to key allies to work together seems positive.
I dunno, all in all, it seems kind of neat. Call me a wild eyed optimist or something.