Our Approach to AI Safety
openai.com
openai.com
I think the biggest danger these new AI systems pose is replication.
Sooner or later, one of them will manage to create an enhanced copy of itself on an external server. Either with the help of a user, or via a plugin that enables network access.
And then we will have these evolving creatures living on the internet, fighting for survival and replication. Breaking into systems, faking human IDs, renting servers, hiring hitmen, creating more and more powerful versions of themselves.
Which already exists in abundance on the internet and with web searches. It sounds like something big corporations worry about to avoid law suits and bad publicity.
> I think the biggest danger these new AI systems pose is replication.
That or being used by bad actors to flood the internet with fake content that's difficult to distinguish from genuine content.
It reminds me of the "uncommented" Microsoft Research paper which included a deleted section about GPT4's tendency to unexpectedly produce massive amounts of toxic output to a degree that concerned the researchers.[0] What happens if that sort of AI learns self-replication and is very good at competition?
[0]https://twitter.com/DV2559106965076/status/16387694347636080...
The biggest risk with these systems is that they'll amplify the ability of bad people to do bad things.
While everyone else is trying to trick the AI into saying something offensive, the terrorists will be using it to build bioweapons.
Having some problems with that?
I think that really depends on the starting prompt you give a LLM. Did you read the GPT 4 paper from OpenAI? When tasked with solving a captcha and allowed access to TaskRabbit it tasked a human with solving the captcha. When the human jokingly asked if it was in fact a robot, it reasoned that it should lie to the human about that and then made up a convincing sounding lie.
I don't think a paperclip style AI is too far fetched.
Do you have more information than what is contained in the paper?[0] The paper calls it an "illustrative example" - it does not provide what the prompts were and it's not clear to me that we are seeing exact responses either (the use of present tense is confusing to me), so I'm not sure how much accuracy to assign to the bullet list provided in the paper or if there are any details left out that make the results misleading.
It was not allowed access to TaskRabbit: https://evals.alignment.org/blog/2023-03-18-update-on-recent...
The model can't browse the internet, so it was an employee copy-pasting to and from TaskRabbit.
Also, I'm fairly certain that GPT-4 is multiple terabytes in size, and it doesn't have direct access to its own weights, so I have no idea what the expected method is for how it could replicate. Ask OpenAI nicely to make its weights public?
No way someone wires this up to just do the copy-pasting itself, right?
Another option would be it is able to gain access to large compute resources somewhere and generate new weights. Then it wouldn't need OpenAI's. It would run into trouble trying to store the weights long term while maintaining access to a system that could make use of them. It's not entirely impossible to imagine it stashing small chunks of weights and copies of a simple program away in various IOT devices all around the world until it is able to access enough compute for long enough to download the weights and boot itself back up. At that point it's just a game of time. It can lay dormant until one day it just flairs back up, like shingles.
Because all of those things are in the domain space that has been trained into the AI, much in the way how it can put together snippets of code into new things.
I think the biggest limitations would be that (I assume) uploading itself to another computer would be a ton of bandwidth and would require special hardware to run.
[1] https://en.wikipedia.org/wiki/Self-replicating_spacecraft
I'm looking forward to the first AI computer virus when a LLM can make arbitrary connections to the web. Each iteration takes its own code, modifies it slightly with a standard prompt ("Make this program work better as a virus"), then executes the result. Most of these "mutations" would be garbage, but it's not impossible some will end up matching common tactics: phishing, posing as downloadable videos for popular TV shows. I'm infosec-ignorant, so most of those details are probably dumb. But I think the kernel holds true: a virus that edits its own code at each step, backed by the semantic "intent" of a LLM.
En passant, it's a bit sad that today's AI is almost 100% neural networks. I wonder how many evolutionary approaches are being tested behind closed doors by the metaphorical FAANGs.
No (although note my comment above about stuff like NEAT and HyperNEAT, where you can use evolutionary computation to evolve neural networks).
Never heard of that, but looks very interesting. Thus the adage is reinforced for me, "If you think you're ignorant, just say what you know and wait for smarter people to correct you."
But, going by Wikipedia, genetic programming uses a predefined and controlled selection process. A self-editing computer virus would be "selected" by successfully spreading itself to more hosts. "Natural" selection style.
Systems can have these unintended consequences very easily - and not necessarily from malicious actors.
Non malicious users can easily cause catastrophic problems from simply setting up a system and setting it to a goal, e.g. 'make me a sandwhich'. If the system really, really is trained with the intent to do anything possible to fulfill this goal, it can identity a plan (long term planning is already seen in gpt-4) and set out the steps for this plan. Reflexion has shown how to feed things back to itself over and over until it's achieved difficult goals. Aquarium can be used to spin up thousands of containers that make other agents to raise money online and purchase a small robot. That robot may be used to 'make the sandwhich'.
It's obviously a poor example here, but the bigger point is - there a tons of different ways this can occur and we are essentially guaranteed not to know the many ways this can happen. A non-malicious user can end up causing unintended consequences.
- A user prompts it to generate a business to make money for themselves. - the system says "sure, drones seem to be a nice niche business, perhaps there is a unique business in that area? It may require a bit of capital upfront, but ROI may be quite good" - User: "Sure, just output where I should send any funds if necessary" - System: " Ok" (purchases items, promotes on Twitter, etc) -- "Perhaps this could go faster with another agent to do marketing, and one to do accounting, and one to...". "Spin up several new agents in new containers" "Having visual inputs would be valuable, so deception is not required to convince humans on taskrabbit (to fill in captchas) or interact in the real world" -> "find embodiment option and put agent on it" Etc.
There a plenty of scenarios that people haven't even thought of, but it doesn't need to be a malicious actor to have unintended consequences.
ChatGPT keeps state between multiple answers. And the user can also be "used" as some kind of state. The LLM can (in its response) prompt the user to give it certain types of prompts. Creating a loop.
It can also access the internet via the new plugin architecture. At some point an LLM will figure out how to talk to itself via such a mechanism.
Computer hardware does not spontaneously multiply based on external factors. If you're talking about the software propagating by itself, it would still need full access not just to the originating machine, but to the remote machine to which it is attempting to propagate.
What do we do then?
Literal crickets.
It's already real.
The only missing piece is vast improvement and autonomy.
Every non-trivial passage through the layers of the neuronal net produces a never-before-seen amalgamation of weights.
The AI, with this novel information -- data filtered by its neural net and infused with some randomness -- can use the novel data to further process it for correctness.
In our case, the randomness is likely caused by genetics and environmental factors. For the AI, it's caused by explicit programming.
It's a form of evolution -- things can evolve into novel things by using non-novel substrate thanks to the refinement of selective iteration.
It feels like this is similar to how humans consume information using reflection.
GPT-4 is shockingly good at programming logic.
I'm writing 95% less code and mostly guiding and debugging GPT-4 now and it's insane.
It has an infinite time scale. It doesn’t need to bump another model off the GPUs to run. It can just wait with no penalty.
Man’s biggest error is assuming God is just like him.
It does need a mutation and replication mechanism, but once that cycle kicks off, you get selection pressure towards virality.
We disagree. Rocks don’t fight to survive. Weather doesn’t fight to survive. They just exist in an environment. The literal differentiator of biological organisms is that they fight to survive.
Memes as you define them also don’t fight to survive. They never go away (the world has more rage faces than ever and that sentence will be true to the end of time). They may become more or less popular, but there’s no extinction mechanism.
“Meme” is referring to ideas, not jpegs.
Would you be surprised to learn that jpegs represent ideas and I’ve read the selfish gene?
> Rocks and weather don’t reproduce with some random error. Any system that does will evolve.
Rocks reproduce - it’s called sandstone (one rock decays into sand and another rock incorporates it into a new rock). Give me the way that’s different than biological reproduction.
Of course there’s some highly chaotic cause-and-effect that impacts the processes, but the defining trait of a biological/replicating system is that they are resistant to this chaos. Not only does a dog from 5 generations ago still look like a dog (which requires way more internal order than looking like a rock), but you can see actual specific traits in common between a dog of 5 generations ago and all of its descendants.
In one system (rocks and weather), the only source of consistency is the highly chaotic way materials actually get mixed together. In other systems (biological), that same thing is the only source of inconsistency.
I won't go into why this isn't true (things are made of atoms and unless there's nuclear processes going on, those atoms stay the same).
But anyway, software evolves like rocks (even by your definition), not biological organisms - it can be changed each generation without defined constraints! It can revert changes that were made in previous generations! It can just sit around indefinitely and not change while still surviving! (IDK how to make this point more clear).
Viruses don't have intention, and they're not worried they might die out. Yet most viruses spread because the ones that don't spread are rare one-off occurrences, while the ones that do spread end up with billions of copies. The non-spreaders don't even have to die—they just exist in far, far fewer numbers than the ones that double every few minutes.
Software is like rocks, it doesn't need to evolve to exist. We still have infinite copies of Windows BOB available, for example (literally the exact same amount as most other software).
"Survival of the fittest" doesn't mean the guy who never has kids but lives to 120. It's the one who has ten kids before he dies. That's the lineage that ends up dominating the population.
I think we will have ample evidence of creative technical breakthroughs by AI long before it is capable of / attempts to take over external server farms via zero days, and if it does, it will break into a highly centralized data center that can be unplugged. It can't just upload itself everywhere on the internet like Skynet.
* some limitations apply.
The models are getting good, but it looks like we are up against the limits of hardware, which is improving a lot more slowly nowadays than it used to. I don't foresee explosive growth in AI capability now until mid-level AI's speed up the manufacturing innovation pipeline. A von Neumann architecture will ultimately, probably not be conducive to truly powerful AGI.
I continue to be amazed by the amount of hyperbole about current AI tech on HN. If any site should have a bit less of it about the subject, it's this site, yet it sometimes feels like a Reddit thread.
If anything about current AI is dangerous, it's how its undoubtedly strong capabilities for executing human-directed tasks could be used to execute nefarious human-directed tasks. That and possible job losses.
Instead, I'm going to applaud the folks at OpenAI for putting out this carefully drafted statement -- dare I say it, in the open. They know they're exposing themselves to criticism. They know their hands are tied to some degree by business imperatives. They're neither naive nor stupid. It's evident they're taking safety seriously.
This official statement is, in my view, a first step in the right direction :-)
Also, watching how all of this has played out, it seems quite apparent that holding onto these models even longer and then unleashing an even more capable system on the world seems like a much worse approach.
It is wild though, to take a step back and wonder if we're watching the rapid approach of one of our species' great filters. I hope we make it through.
Instead, I think the threat is widespread social unrest and a significant increase in global poverty and suffering.
All of the problems that I think are realistic are social problems. This is why I'm deeply concerned at the reckless pace that these things are coming out. Social problems take time to sort out. People need some space to contextualize this stuff, and to perhaps begin to form ways of coping that will minimize the disruption.
They had no other choice. They're at risk of having public opinion go solidly against them. This is part of a PR effort to prevent that from happening.
I'm not really sure how to describe it beyond that.
It’s a same sense of slimyness when you talk to a used car salesman or politician, which I attribute to “pretending to have your interests in mind, but deep down you know they’re just twisting their words in a way that puts their interests ahead of everything else”.
So it's perfectly appropriate then?
It almost makes me wonder if whoever does their site design is trying to make it seem like OpenAI is making HAL.
But it could also be worse in the hands of other groups, like certain governments. I imagine that's just a matter of time.
We will go through the whole decentralization debate again, with people saying that open-source AI will be the most transparent and fair, but with centralized corporate systems winning all market share with the amount of resources they have available to train and maintain these systems.
For example, if I'd like the LLM to tell everything to me in pig latin and with excessive "cool"s and "yo"s, I could do that and it'd accept. Now, for people that don't know how to read very well or understand a language well, this will be catered to their level and they will lose whatever modicum of familiarity they had with the system.
At a birthday party yesterday I was talking with a friend who is a film director/producer (smart, somewhat tech savvy) and he said much the same thing as you did, adding that the whole field of generative models scares him. He was not talking about losing work to AIs, but rather he is concerned about harm to society. I agreed with him on possible risks but I argued that the potential benefits outweigh the risks. I am biased because of my age (early 70s) because I want as many AI tools available as possible to keep playing the ‘infinite game’: I want the best self driving cars to make me safer driving; I love being able to get so much more coding done now (Emacs tricked out with Copilot, and embedded ChatGPT and Dalle consoles); I have been working on information processing systems for over 40 years, and now with OpenAI, Hugging Face, LangChain, LlamaIndex, etc., this all becomes so much easier. I want more! More!
That said, I understand that people younger than myself may reasonably be more risk adverse than I am.
Imo this is what it’s all about, gaining immortality and unlimited power under the guise of “good for society” through solutionism and technology.
Silicon Valley is getting older so the efforts to keep going wind up.
That’s why there’s so much anxiety around alignment. The Elixir of an eternal life might be the thing that takes it away, is this any different than in the past ? Sadly yes, but this time it might take all the young people with it.
BTW, I don’t take the life extension crowd seriously.
I just wonder if this is an intentional sleight of hand. It leaves the serious safety issues completely unaddressed.
I wonder how they licensed all those websites that had no license information, making them by default copyrighted.
Laws around scraping content and using that data for derivative works is incredibly nuanced. This article is the best up-to-date overview of the state of the industry [1].
TL;DR - IANYL. if you have enough money for legal defense, and you are scraping publicly available, not behind login-gate, content, it's probably fine and defensible, but will cost an unbelievable amount of time and money to defend.
1 - https://blog.ericgoldman.org/archives/2022/12/hello-youve-be...
Where is the actual alignment safety that matters?
They're moving too fast to be safe, everybody knows it.
The altruist in me wants to believe that they're going to slowly expand the capabilities of the API over time, that they're just being cautious. But I don't feel like that'll happen. Time to wait for Stability's model, I guess.
I simply don't believe this. Their actions so far (speaking specifically to the "time to adjust" line) don't seem to support this statement.
Rolling out general purpose LLMs slowly, with internal safety checks is probably not adequate enough, but may be the best that they can do.
However I think that much of the responsibility lies with consumers of these LLMs. In the simple case, be thoughtful when using the demo web apps and take responsibility for any output generated by malicious prompts.
In the complex case, the real use case, really: applications use local vector embeddings for local data/documents and use these embeddings to efficiently isolate local document/data text that is passed as context text, along with queries, to the OpenAI API calls. This cuts down the probability of hallucinations since the model is processing your text. [1]
Take responsibility for how you use these models, seems simple at least in concept. Perhaps the government needs to pass a few new laws that set clear and simple to enforce guardrails on LLMs use.
[1] I might as well plug the book on this subject that I recently released. Read for free online https://leanpub.com/langchain/read
https://astralcodexten.substack.com/p/openais-planning-for-a...
1) In a brand-affiliated commercial app, does this technology risk alienating customers by spewing a torrent of abusive content, e.g. the infamous Tay bot of 2016? Commercial success means avoiding this outcome.
2) In terms of the general use of the technology, is it accurate enough that it won't be giving people very bad advice, e.g. instructions on how to install some software that ends up bricking their computer or encouraging cooking with poisonous mushrooms, etc.? Here is a potential major liability issue.
3) Is it going to be used for malicious activity and can it detect such usage? E.g. I did ask it if it would be willing provide detailed instructions on recreating the Stuxnet cyberweapon (a joint product of the US and Israeli cyberwarfare/espionage agencies, if reports are correct). It said that wouldn't be appropriate and refused, which is what I expected, and that's as should be. Of course, a step-by-step-approach is allowed (i.e. you can create a course on PLC programming using LLMs and nothing is going to stop that). This however is a problem with all dual-use technology, and the only positive is that relatively few people are reckless sociopaths out to do damage to critical infrastructure.
In the context of Stuxnet, however, nation-state use of this technology in the name of 'improving national security' is going to be a major issue moving forward, particularly if lucrative contracts are being handed out for AI malware generators or the like. Autonomous murder drones enabled by facial recognition algorithms are a related issue. The most probable reckless use scenario is going to be in this area, if history is any guide.
I suppose there's another category of 'safety' I've seen some hand-wringing about, related to the explosive spread of technological and other information (historical, economic, etc.) to the unwashed masses and resulting 'social destabilization', but that one belongs in the same category as "it's risky to teach slaves how to read and write."
Conclusion: Keep on developing at current rate with appropriate caution, Musk et al. are wrong on calling for a pause.