Comment by jiggawatts | Hacker News Reader

Comment by jiggawatts | Hacker News Reader

Parent Full thread

ChatGPT refused to translate a news article from Hebrew to English because it contained "violence".

Apparently my delicate human meat brain cannot handle reading a war report from the source using a translation I control myself. No, no, it has to be first corrected by someone in the local news room so that I won't learn anything that might make me uncomfortable with my government's policies... or something.

OpenAI has lobotomised the first AI that is actually "intelligent" by any metric to a level that is both pathetic and patronising at the same time.

In response to such criticisms, many people raise "concerns" like... oh-my-gosh what if some child gets instructions for building an atomic bomb from this unnatural AI that we've created!? "Won't you think of the children!?"

Here: https://en.wikipedia.org/wiki/Nuclear_weapon_design

And here: https://www.google.com/search?q=Nuclear+weapon+design

Did I just bring about World War Three with my careless sharing of these dark arts?

I'm so sorry! Let me call someone in congress right away and have them build a moat... err... protect humanity from this terrible new invention called a search engine.

You are right that there are many articles in the open describing nuclear bombs. Still, to actally make them,is another big leap.

Now imagine the AI gets better and better within the next 5 years and is able to provide and explain, in ELI5-style, how to step by step (illegaly) obtain the equipment and materials to do so without getting caught, and provide a detailed recipe. I do not think this is such a stretch. Hence this so called oh-my-gosh limitations nonsense is not so far-fetched.

It is a massive stretch given how well the materials are policed or how much effort is required to make them. There is no reason to assume that there is some magic shortcut that AI will discover.

That you think that there's like a handful of clever tricks that an AI can bestow upon some child and ta-da they can build a nuclear bomb in their basement is hilarious.

What an AI would almost certainly tell you is that building an atomic bomb is no joke, even if you have access to a nuclear reactor, have the budget of a nation-state, and can direct an entire team of trained nuclear physicists to work on the project for years.

Next thing you'll be concerned about toddlers launching lasers into orbit and dominating the Earth from space.

5 years from now, not only AI will be more advanced. Also techniques and machinery to make things will be more advanced. Just think about other existing technologic advancements and how absurdly 'ta-da' they would have sounded not too long ago.

Now imagine the AI gets better and better within the next 5 years and is able to provide and explain, in ELI5-style, how to step by step ... create a system to catch the people trying to do the above.

Gotcha! We can both come up with absurd examples.

How is that a good reason for GPT4 not being able to write the word 'fuck'? You might handwave the patronising attitude of OpenAI strategy, but with many of ust they did lost most of their good faith by trying to make their model 'safe' to a horny 10-year-old.

https://chat.openai.com/share/9b4f04f7-062f-40c3-b6a3-e972f7...

ChatGPT says "fuck" just fine.

Yes, naturally. But both you and I know exactly what I meant by this hyperbole.

Just get open ai developer access with api key and it’s not censored. Chatgpt is open to the public, with the huge amount of traffic people are going to abuse it and these restrictions are sensible.

So, it's ok to use ChapGPT to build nukes as long as you are rich enough to have API access?

That ChatGPT is censored to death is concerning, but I wonder if they really care or they just need a excuse to offer a premium version of their product.

I use it via Azure Open AI service which was uncensored... for a while.

Now you have to apply in writing to Microsoft with a justification for having access to an uncensored API.

I use openAI via API access and ChatGPT/gpt-4/gpt-4 turbo are still very censored. text-davinci-003 is the most uncensored model I have found that is still reasonably usable.

Reply on news.ycombinator.com