NIST AI Risk Management Framework
nist.gov
nist.gov
The current GPT models are not going to "wake up" or accidentally take over the world. This is very obvious.
Forcing everyone who is conscientious to copy and paste the output of these relatively weak models is not going to save us from anything.
If you want people to take AI regulations seriously, you have to do it in a reasonable way. Which means understanding the nuance of the dangers.
Executing "arbitrary code" that InstructGPT models outputs based on your instructions is not a danger.
People need to be able to distinguish different levels of autonomy and speed.
It's when we get to high levels of autonomy and performance that we run into danger. Which we are really at the cusp.
But when you fail to differentiate between systems that are obviously not dangerous and future/more autonomous agents that possibly are, that makes it impossible to take the concerns seriously. And actually harder for people to understand the problems with full autonomy and superintelligent models.
It's pretty obvious that GPT-4 isn't consistently coherent enough to carry out complex plans or to do serious scientific research. There are several other major shortcomings which also prevent this from being viable, too. And of course, it's possible that some of these shortcomings are very hard to "fix." We've had AI winters before, and self-driving seems to have stalled.
However, I strongly suspect that it's theoretically possible to build a machine that "thinks" as well as a human. And I no longer have any solid idea how far we are from that point, especially if we make certain architectural changes. We might be closer than we think. Or at least we should stop assuming that next-gen models are obviously safe.
nVidia's hope to increase training performance 1 million times in 10 years seems potentially dangerous to me. Anthrophic's plan to spend over $1 billion training next gen models seems awfully sketchy as well.
LLMs, especially the multi-modal ones go much further to making self driving possible, that is if we ever get to the point they can incorporate and act on data fast enough. We may still have a hardware problem for some time there.
---
But on your second point I am with you. There is nothing obvious about at what level emergent behavior occurs in the models. Just keep adding parameters and new things keep popping out with no real means to determine where and when seems kinda sketch as we push these things past the exa scale.
Not any time soon. Modern "AI" research built on completely wrong assumptions about how neurons work from the middle of XX century. It turned out they are much more complex than that (long story short every single neuron is analogous to the "neural network" as AI researchers understand them). See more at https://www.youtube.com/watch?v=hmtQPrH-gC4
Your brain does 1+1 = 2 very differently than an integrated circuit, but the end output is the same.
This is a grossly irresponsible and categorically false statement, in very obvious ways.
If anything they could enable an option to have ChatGPT review the code for you and let you know that it hasn't accidentally thrown in some fishy code.
Because it's a black box that's sensitive to unexpected inputs. Humans create bugs too, but an unaligned AI doesn't have the sense to be careful of certain classes of bug.
Even if you trust ChatGPT you should do this. You should assume that one day the hackers will intercept your connection to OpenAI and hand you malware, and they should either fail or be forced to use a zero day exploit against eg Firecracker (modulo your threat model and such).
I meant to comment on the security aspects of sending confidential information to models and letting them generate output that will be immediately executed on your own machine. People also chain models together and let them generate output that will be interpreted as the next model's input. If done without guardrails, this is done ideally in a sandboxed environment.
Too often, I see this type of code just ran on one's machine. As the projects attract less technical people, they may not be aware of the dangers of running such code locally and can see a portion of their data being wiped or exposed.
Was it obvious to you that you’d see a paper called “Sparks of an AGI” after the release of ChatGPT 4 ?
Also part of the problem is the word AGI and having total conflation inside of that.
I meant SPECIFICALLY the current GPT versions and architecture.
In my opinion, GPT-4 is extremely intelligent in a very general way and this is incredibly obvious.
What GPT-4 is NOT, by itself, is autonomous, or many other human/animal characteristics that people conflate with intelligence.
GPT does not have full autonomy. It is not a person. It is not alive. It does not feel. It does not form its own goals. It does not feel, or have survival instincts, or a high bandwidth stream of sensory data. In these and other ways it is quite different from animals and people. It is not a digital person.
However, people are quite stupidly moving forward with trying to create digital people that imitate all animal characteristics, apparently with the idea that they will enslave them, with the thought process being that they will not be able to do useful work without having all of these animal (human) characteristics.
Besides that problem, people are not taking into account the different levels of speed and other performance measurements of the future systems. Future quite possibly being this year. Or especially I think it is critical for the new compute-in-memory substrates or any advances that increase the speed or other IQ measurements beyond say one order of magnitude of human to be highly regulated. Systems that are twice as smart and say 100 times faster at thinking than humans absolutely can be dangerous.
We achieved the architecture for general purpose intelligence in 2017 and in 2018 I suggested we would have AGI in 2019. Unfortunately now AGI is quite vague and means things like digital person or digital God..I wrote quite a lot but unfortunately deleted some of the explanation.
General intelligence versus personhood or Godhood or alive, or autonomous or having survival instincts or instant adaptability etc. are not the same thing.
I think we are going to be building and using 'sub-AGI' models for a very long time, probably well after we (probably achieve) full AGI. I think these 'sub-AGI' models are going to radically change the way we think about intelligence in the sense that we are going to be better able to nail down what we find special about human intelligence while at the same time discovering that it is a narrow and somewhat arbitrary subset of what we consider advanced intelligence.
ChatGPT seems to me like a very advanced Markov chain, which isn't to denigrate it, the surprising part is how useful and effective it is which to me shows we have been thinking about intelligence the wrong way. IMO an important portion of our brain probably works in a very analogous way but you could build as advanced a version of ChatGPT as you like and it wouldn't become 'sentient' or go HAL9000 on us.
There is definitely a danger of someone taking ChatGPT9 and telling it to create you computer viruses to take down digital infrastructure or telling it to trade in a way to crash the stock market but that's going to happen when all the zero-day exploits are detonated in the lead up to WW3 anyways and all things considered that has a higher chance of happening sooner than ChatGPT9 or the equivalent is used to do it. Instead of getting worried over rogue AI we should probably just harden our infrastructure.
Also, a paperclip maximizer could in theory go 'HAL' on you without intent or malice. Sentience isn't needed. And at the same time as we're working on embodiment models, what we could consider human like sentience could be far closer than we expect.
"Oh, they're just fictional!"
…then I suggest attaching a rusty metal spike to the keyboard.
It won't do anything, it's easy to avoid, it'll just sit there looking dangerous.
Hopefully that would be enough just by itself.
And yes, this is a reference to a similar suggestion for making drivers pay more attention behind the wheel.
Sci-fi is a distraction. There can be no heroic human resistance like in the stories. If an AI is intelligent enough to be dangerous, it's intelligent enough to conceal its intentions until it's too late. The only way to beat a superhuman AI is by not making it in the first place.
plot of threads (set in Sheffield, UK):
- russians/americans disagree on something in iran
- initial nuclear attacks around base in iran
- uneasy 2 day ceasefire
- worldwide nuclear exchange occurs (a relativly small one)
- graphic depiction of nuclear attack
- people suffering horribly
- ends with scene in hospital of character we've followed in labour, resulting in deformed stillbirth
- tech company creates AI
- idiot CEO enters badly thought out goal in an attempt to capture 2% more search market share
- AI begins to executes goal
- idiot CEO initially delighted at results
- at some point AI becomes super-intelligent as it helps it achieve its goal
- all life on earth is casually exterminated by the AI as it becomes better at achieving its goal
- AI acts as a global scale systematic combine harvester
- no humans, animals, bacteria, fungi left... no biological life ever again on earth
- ends with scene of 100% of the earth's surface converted to GPUs, with the NVIDIA logo prominently displayed on each of themSoong-Altman-Musk: "I made an android that passed US Navy officer training!"
US Navy: "Fantastic. We've given it a rank of Lieutenant."
Soong-Altman-Musk: "I left in a backdoor I didn't tell you about which when triggered made it seek me out for an upgrade."
Android: *Steals CVN-65*
US Navy: *unhappy face*
One of the better non science fiction analogies I've heard.
In the 2000s everyone was hyper aware of privacy issues, putting personal information online, safety, etc. However the last 15 years of people doing more and more risky stuff with allowing systems and computers to store and use their data and only getting benefits and more fun videos to watch, there's no internal shared concern to act in accordance to every possible safety principle.
People just want the cool chatbot to do their work for them. Especially with the scale of FOMO with these things, no one is going to sit back and not be risky when they read that other people are saving many hours a day letting ChatGPT do their work for them. Until something really bad happens people aren't going to change their behavior.
At least I hope so.
I've worked with a lot of people in the medical world while developing SaMD (software as a medical device) in the past that had little to no idea about software. They can apply the principles in the abstract but will likely not dig deep enough to catch some very major issues.
In the medical world, things like post market surveillance and notification of adverse events help to at least create a public feedback loop here. I think we will need something similar in this space if we really want to see more than a surface level, checklist ticking exercise.
I had a question about this earlier, if a doctor uses Google to look up something, then is Google being used in some legal or regulatory sense as a medical device?
Medical devices are marketed for specific medical use, which doctors rely upon to do what’s medically expected. These devices will have an “expected use” and “indications for use” that largely cover how they are design/expected/tested for use. They need 510k clearance to be classified by the FDA as a medical device.
The process is pretty involved, but to add some clarification - the "indications for use" or "expected use" are things the manufacturer are required to include with their device when submitting for 510k clearance from FDA, so that it can then be marketed for medical use. They can only market it as a medical device once it has 510k from FDA.
CSF has a fair amount of traction I guess.
I’d be happy for my pessimism to be proven wrong.
>The Framework was developed through a consensus-driven, open, transparent, and collaborative process...
It sounds like they collaborated significantly with parties likely involved with developing AI
Also, there’s a nontrivial amount of AI research done by NIST scientists