AI companies will need to report their safety tests to the US government

https://apnews.com/article/biden-ai-artificial-intelligence-safe-395591bcde523416db88767fa54f30f5

by DaveFlater • 2 years ago

54 70 2 years ago

reissbaker2• 2 years ago

Wow. This feels like a regulatory regime designed to smother open-source AI, and reduce competition in the American market since small companies are unlikely to be able to comply. The worst part is that doing so is actually against our national interests... China has already essentially caught up to GPT-4V with the release of Qwen-VL-Max yesterday [1]; the cat is out of the bag. Hobbling ourselves now doesn't help.

1: https://twitter.com/_akhaliq/status/1752033872982806718

rvnx2• 2 years ago

Well let's say congrats to them then, because we will all benefit back from that, especially since they work more in open way as we can download the Qwen-VL-Max model, inspect it and use it offline.

Europe is copying ChatGPT, and all we got is a louzy LLaMA-fork.

reissbaker2• 2 years ago

Definitely congrats to them. I don't hate on Mistral though; their open-source models are extremely capable for their size and parameter count, and Mistral Medium is second only to GPT-4 on the Chatbot Arena (I'm not including Bard, since Bard gives their chatbot access to the internet, whereas the other models are being interacted with solely on the basis of their innate knowledge; that doesn't seem apples-to-apples, and the underlying Gemini Pro model without access to the internet underperforms GPT-4 and Mistral). And Mistral Medium is much cheaper than GPT-4 — around $8/million tokens, whereas GPT-4-Turbo is $30/million tokens. So I'd say Mistral is doing a decent job.

With this kind of regulatory regime I wonder how much of the next wave of AI companies are going to be started abroad... I sort of doubt Mistral is the last one, even though the U.S. has such high talent density.

chasd00• 2 years ago

Yeah the genie is out of the bottle. Any regulations a government sets up just puts its economy behind the competition. I think a lot of the US regulations are coming from large tech companies (google) feeling threatened. LLMs used for knowledge and skills transfer threaten search and therefore ads. If you can make it illegal or impractical then you don’t have to compete.

schnitzelstoat• 2 years ago

Any regulations a government sets up just puts its economy behind the competition.

I fully expect us to do this in Europe. Then complain in 10 years when all the big AI tech comes from abroad.

RugnirViking2• 2 years ago

I can appreciate that the model might be trained to work on primarily chinese language, but the performance for me really doesnt seem to be there yet.

I hate to do the "I asked ai and pasted it here" style of comment, but I really wanted to understand the validity of the claim it surpasses gpt4 in all those benchmarks. I asked gpt4 and qwen-vl-max this prompt to try to get a sense of their capabilities. Nothing even close imo.

"tell me a story about relations between people living on one side of a penny floating through space and the other. Use the format of the hero's journey, including rejecting the call, focus on worldbuilding by hinting at further complexity, and do it in a kafkaesque style"

quen's story starts with: Once upon a time, there was a penny floating through space.

gpt 4: Through the vast emptiness of space drifted an ancient penny, a relic from a long-forgotten Earth.

The difference is night and day atm. Ask yourself: would Kafka have started a story with "once upon a time"

offices• 2 years ago

No, because all of Kafka's books were written in German.

andy992• 2 years ago

I don't think we can extrapolate much from that anecdote, but I'll try. The GPT version was predictably wordy and had the signature precocious high-school essay feeling, e.g. lots of adjectives. It didn't feel Kafkaesque at all, where is the confusion, breathless tone, mysterious K.?

Qwen's is pretty unremarkable, just very direct, but I don't see that one is actually following directions better. Maybe further along the difference gets more stark. Also, isn't Qwen a Chinese-first model? How much of the fine tuning was in English?

Neither example gets too actual utility, though it could highlight the sort of demo that chatgpt is trained for.

donkeyd• 2 years ago

I feel like this one is impossible to do right. On one side, it's people like you saying: 'government oversight is hurting our innovation'. Directly opposed is: 'AI Companies are growing too fast and something bad will happen if oversight doesn't come soon.'

I agree with both and really don't have a clue what's the wise thing to do. I think it's probably the oversight thing, but since other countries will definitely compete without this oversight, someone will eventually release SkyNet. But do we want to be the first?

I don't know...

fallingknife3• 2 years ago

Since there is no regulatory response that can deal with unknown unknowns like your second point, the correct answer is to not regulate for now.

CamperBob22• 2 years ago

We know that excessive premature regulation by captive government agencies hurts innovation and empowers competing nations. We don't know that AI companies are growing too fast, or that Something Bad Will Happen if oversight doesn't come soon.

Seems like an easy choice. We'd never have left the caves if we allowed the precautionary principle to dictate our lives.

bonton89• 2 years ago

The whole drumbeat from OpenAI has always looked like using regulation to create a moat out of thin air.

Are any of the open models in danger of disappearing soon? Should I be downloading local copies before it is to late?

seanmcdirmid• 2 years ago

Didn't China also issue a ruling that AI should conform to proper socialist values? I'm not sure how they are enforcing that though, Biden's executive order seems to have a lot more teeth (and even then, there is no standard yet, they just want information shared).

ejb9992• 2 years ago

Will not be enforceable imo, not sure if it is even constitutional - like it or not, output of AI engines is a type of free-speech. (and no, not that 'AI' has free speech rights, the creators of the AI do)

Etheryte• 2 years ago

I'm not sure this argument makes sense, could you perhaps elaborate? Not everything you create that has output goes under free speech. A torrent client is also a program that you can write that outputs art/text/whatever, doesn't mean it goes under free speech.

ilikehurdles2• 2 years ago

The federal executive doesn’t regulate the BitTorrent protocol fyi. Copyright infringement is a civil matter.

sidewndr46• 2 years ago

We've had close to a 100 years of drug prohibition which disagrees with this notion

ejb9992• 2 years ago

selling illegal drugs is not free speech.

andy992• 2 years ago

I have no idea about the legal reality but it does feel interesting. Language models don't really "say" stuff, they predict array indicies probabilistically. In that kind of reductive view, it's the same as a classifier which afaik can be regulated (say against discrimination).

Otoh if you take the view that the sum of LLM predictions is a kind of speech at the instigation of the operator, maybe...

A lot of this should be moot. Whatever Llama or whatever predicts or says, it isn't dangerous. Regulating what an LLM can generate is like banning books. Regulating specific outcomes is a different story. If I make a chatbot that gives bad medical advice, I don't see any legal difference vs giving bad medic advice without the chatbot and that can be regulated and litigated accordingly. The "AI" angle is mostly just a trojan horse to try and exert more control in the name of defeating some boogeyman.

da_chicken4• 2 years ago

You're not sure if it's constitutional to regulate the sale of products or services to ensure that they're safe?

ilikehurdles2• 2 years ago

Congress has the enumerated right to regulate commerce. This, being an executive order, is not an act of Congress. Consider the agencies you may be thinking of: FDA, FTC, FCC, etc. All have actual laws defining their scope. This is an overreach of the executive.

DoneWithAllThat• 2 years ago

You’re being intentionally obtuse and not addressing the point OP made, which is that AI is likely to be interpreted by the courts as a form of expression which makes it protected by the first amendment.

It would be similarly a violation if, say, the sale of newspapers was regulated. I suspect you know this but you want to make a glib Twitter style comment rather than have a real discussion.

speedylight• 2 years ago

Congress should be in charge of that, they legislate not the president.

throwaway49849• 2 years ago

The government’s National Institute of Standards and Technology will develop a uniform framework for assessing safety, as part of the order Biden signed in October.

They probably need to develop an extremely antisocial, psycopathic, malevolent AI to assess other AIs for safety. Its purpose will be to run through thousands of scenarios, using manipulation, threats, and deception to try to extract dangerous information from the other AI. It can then score the responses based on the information it was able to extract. I don't really see any other way to automate this extremely tedious and error prone task. It's interesting though because we will need to concentrate all of the evil in the world into a single AI in order to run these tests.

jMyles5• 2 years ago

Now do infectious pathogens.

flenserboy• 2 years ago

So there will be one uncensored, general-purpose, actually-useful AI out there that only the government will have. Hmmmm...

jMyles5• 2 years ago

At the risk of hyperbole:

It seems that, as AI becomes the prevalent model for human-computer interaction, requirements like this are not merely stifling to AI (and to the speech, press, and even religious protections that may arise from such interactions), but stifling to general-purpose computing.

Along with increasing "protect the children" sockpuppetry like we have in the bill in the Florida house yesterday - which seem to seek a future in which internet access requires the furnishing of a state ID - actions like this (which you will note did not even come from an additional act of Congress) appear to be rather vulgar measures to prevent technology from empowering anyone who doesn't already have power.

jprete5• 2 years ago

AI companies want to make completely unengineered systems with no systematic guarantee that they work or even knowledge of how they work, and then try to displace as much human labor as possible in as many critical points of control as possible. This is literally the goal of anyone seeking AGI.

In my view, stifling this is a benefit, not a drawback.

jMyles5• 2 years ago

try to displace as much human labor as possible
stifling this is a benefit

I suppose this depends in part on how much value you place on keeping the world a laborious place for humans.

Reducing the need for labor will almost certainly lead to more beauty in the world. If we aren't seeking a world where humans are freed from menial labor and so have bandwidth to make art and music, what are we actually seeking?

Perhaps this needs to be defined first by the AI-luddites and then a more reasoned discussion can follow.

fallingknife3• 2 years ago

No piece of technology has a systemic guarantee that it works.

wand3r• 2 years ago

I laughed when Fox news said Taylor Swift was a "psyop" because it was absurd. Then a few weeks later her deepfakes go all over twitter and the government is cracking down on AI.

Sure its a conspiracy theory, but I do think the government is trying to manipulate public opinion around this space and gain control of it.

Zambyte• 2 years ago

https://www.youtube.com/watch?v=geVNEmnPr5k

Presented as a meme, but it really makes ya think :\

hellojesus3• 2 years ago

Deepfakes of her any anyone have existed for awhile. There are paid services. /r/ exists (and has existed long before deepfakes) where people make requests of deepfakes of their high school classmates. Heck, bubble images have existed for forever and are definitely legal as all they do is leave it up to the imagination - quite effectively.

The recent Swift incident only seems noteworthy because many people were introduced to a cyber subculture of which they were previously unaware. But this subculture is 20+ years old.

Anecdotally, when I caught wind of the incident on /g/ a few days ago and saw some of the ridiculous muppet-swift images, I got a chuckle and asked my wife if she had heard of the swift news. She's usually "in the know" with pop culture because she listens to JJR.

To my surprise, not only had she not heard of it, but she was appalled and became very upset about the obviously AI generated image I showed her. She couldn't fathom someone generating such content. Even having the idea to generate such a thing seemed to break her world view. For me it was just another rule 34 chuckle.

walterbell• 2 years ago

https://www.cnbc.com/2022/10/24/how-googles-former-ceo-eric-...

“If you’re going to be leading a commission that is steering the direction of government AI and making recommendations for how we should promote this sector and scientific exploration in this area, you really shouldn’t also be dipping your hand in the pot and helping yourself to AI investments,” said Shaub of the Project on Government Oversight.

stainablesteel• 2 years ago

sounds like a soft censorship mechanism to me

deadbabe• 2 years ago

Can the reports just be reported by AI?

scotty79• 2 years ago

Is it connected to a missile launcher? [Y] [No, it's safe]

macinjosh3• 2 years ago

I didn't know the executive branch could just enact new laws unilaterally. They are probably using some easily bent statute to allow for this. Here is hoping SCOTUS overturns Chevron this cycle so we can be free of dictatorial edicts such as this.

addandsubtract• 2 years ago

This reads a lot like leopardsatemyface. It's an executive order that Biden signed three months ago. If you have a problem with executive orders, I'm sure you also spoke up 6 years ago.

macinjosh3• 2 years ago

I did. I am pro-democracy so I hate authoritarian executive power no matter where it comes from. Check your bias.

popcalc• 2 years ago

https://en.wikipedia.org/wiki/List_of_national_emergencies_i...

macinjosh3• 2 years ago

The word emergency is doing a lot of work in most of these situations.

codingdave• 2 years ago

More info on executive orders: https://www.federalregister.gov/presidential-documents/execu...

testfoobar• 2 years ago

Is there a published list of safety tests that are to be conducted?

jjk166• 2 years ago

This is some terrible reporting. If you actually read the executive order[0], while there is lip service paid to several general issues associated with AI like job displacement and privacy violation, they actually are pretty narrow in what they mean by AI safety. They're specifically looking to restrict access to AI models that can be used to aid biological weapons development by modelling complex biochemical interactions.

While I'm unconvinced this is the best way to tackle the issue, and there's always the possibility for overreach, the story as presented, namely that AI companies are about to become burdened with a vast and nebulous regulatory infrastructure prompted by vague and poorly informed fears about AI, is bunk.

[0] https://www.whitehouse.gov/briefing-room/presidential-action...

AI companies will need to report their safety tests to the US government

Related Stories