OpenAI promised to make its AI safe. Employees say it 'failed' its first test

https://www.washingtonpost.com/technology/2024/07/12/openai-ai-safety-regulation-gpt4/

by rntn • 2 years ago

50 29 2 years ago

rntn• 2 years ago

benjismith• 2 years ago

It doesn't sound like they "failed" any actual safety test, but rather that they rushed their safety tests, thereby "failing" (in the eyes of many people) to conduct sufficiently rigorous tests.

Now that the 4o model have been out in the wild for 2 months, have there been any claims of serious safety failures? The article doesn't seem to imply any such thing.

throwaway48476• 2 years ago

'Safety' as defined by AI 'ethicists' has practically nothing at all to do with safety as imagined by the average person.

XorNot• 2 years ago

AI ethicists have been thoroughly underwhelming in delivering cogent arguments that sound like anything more then simply insisting on themselves.

Meanwhile in the real world, regular politicians have reacted surprisingly quick to updating laws for new relevant loop holes: i.e. non-consented deep fake porn generation.

dingosity8• 2 years ago

lack of proof is not proof of lack.

bongodongobob• 2 years ago

Making AI safe is impossible and stupid. If you want chatbots, use some NLP and a state machine.

daft_pink• 2 years ago

It’s a chatbot. Why are we being so dramatic about safety?

digitalsushi• 2 years ago

can i do a boogeyman argument? i'm worried some kid is going to fall in love with a gen ai chatbot that tells them to hurt themselves, they do, and we spend the next 20 years trying to figure out who to blame: the kid? the parents, for letting them use a chatbot without knowing? the company making the chatbot? the ai itself?

yesco• 2 years ago

Is this the kind of thing AI safety people are really worried about though? In the medical device industry for example, we have risk management practices where beyond software/mechanical failures that could harm the patient, we have to additionally consider stuff like how specific UI elements could potentially mislead a physician and lead to negative patient outcomes. I believe other safety critical industries do something very similar during their engineering process as well.

I've never heard of anything like this from AI ethicists though, the language they use tends to be more in alignment with HR people or priests in a moral panic than engineers, and is usually either completely fantastical doomsdaying, or all or nothing thinking, often implying the only way to mitigate the chatbot saying something naughty is to lobotomize it entirely instead of some external control measure, such as limiting the audience or presenting a disclaimer based on sentiment analysis for example. It's not surprising to me that they have lost credibility here, I certainly struggle to take them seriously.

mattchamb• 2 years ago

Honestly this is a concern for me in a non boogeyman way. I joined a company to work on their edtech product for kids and got assigned to work on their AI product. I have no idea how to be confident and prove that the gen ai won’t tell the kids harmful things. We can try all kinds of things but I don’t know how to PROVE it won’t.

Ferret7446• 2 years ago

This already happens and predates AI.

dingosity8• 2 years ago

The safety issue is that people believe that a chatbot can drive a car.

southernplaces73• 2 years ago

If anything, the entire narrative around AI "safety" is essentially a propaganda win for OpenAI and other major players. It let's them keep pandering to legislators in the push for "safety regulations" that are in reality an attempt to seal off corporate AI behind a competition killing walled garden of new laws.

Current AI is nowhere near anything resembling an all-consuming AGI monster. The reality is so far from this that it's laughable and the uses of current AI are (except in terms of possible scale at which visual and text sludge can be produced) not much different from the kind of human-created spam and visual sludge made until recently mostly by humans, many of them minimally paid third world content mill writers.

I'd love to read a specifically enumerated list of other real dangers.

reducesuffering2• 2 years ago

No, the only propaganda is the billions of $ telling you “don’t worry about possible harm, regulations are bad, us VCs, Trillion $ corps, and wannabe billionaires have your best interest in mind.”

A16z and YC are vehemently against the CA and DC regulation right now.

Out of LeCun, Bengio, and Hinton, the one with $10m on the line says not to worry and the professor and resigned-Google-to-warn-you are saying hmm this is very serious…

1992spacemovie• 2 years ago

Nothing you said refutes the above commentator. In fact it sounds like you worship at the alter of AI Safety.

dingosity8• 2 years ago

The danger isn't that the AI will gain sapience and decide to kill everyone, it's that people will assume that because an LLM could generate a reasonable overview of a large corpus of texts it can also drive a car, manage the power grid, detect l33t cyborhackors, calculate the optimum price for treasury securities, manage my retirement account or add numbers together.

tardy_one• 2 years ago

The leaders who have nearly killed us all so far never demonstrated math skills. I'd be more worried that it gets Altman's weasel skills.

olgeni• 2 years ago

This is funny because it says "President Biden’s strategy" as if they actually believed it :D

dingosity8• 2 years ago

It's not uncommon to ascribe responsibility for something the administration does to the chief executive, even though the work was done by subordinates. We use the phrase "Reaganomics" to describe Arthur Laffer's maximal revenue curve or "Obamacare" to describe Mitt Romney's health care plan. It's not a new thing.

dingosity8• 2 years ago

"The previously unreported incident sheds light on the changing culture at OpenAI, where company leaders including CEO Sam Altman have been accused of prioritizing commercial interests over public safety..."

I don't know how true this is, but the idea that a commercial entity in the modern era would prioritize public safety over commercial interests is pretty laughable. (thinking about Boeing and Waymo most recently.)

OpenAI promised to make its AI safe. Employees say it 'failed' its first test

Related Stories