Apple Restricts Employee Use of ChatGPT, Joining Other Companies Wary of Leaks

Apple has restricted the use of ChatGPT and other external artificial intelligence tools for some employees as it develops its own similar technology, WSJ repors, citing an internet document and people familiar with the matter. From the report: Apple is concerned workers who use these types of programs could release confidential data, according to the document. Apple also told its employees not to use Microsoft-owned GitHub's Copilot, which automates the writing of software code, the document said. ChatGPT, created by Microsoft-backed OpenAI, is a chatbot derived from a so-called large language model that is able to answer questions, write essays and perform other tasks in humanlike ways. When people use these models, data is sent back to the developer to enable continued improvements, presenting the potential for an organization to unintentionally share proprietary or confidential information. OpenAI disclosed in March that it took ChatGPT temporarily offline because a bug allowed some users to see the titles from a user's chat history.

Slows people down

[S_Stout]

Now they have to use ChatGPT on their other computer and will not be able to copy/paste large chunks of code.

Re:

[AmazingRuss]

... which is probably the correct way to use it for coding.

Re:

[anonymouscoward52236]

You could always choose to not use ChatGPT on your personal computer, and be 10 steps behind the rest of your programming team, LOL!

About time

[Ol Olsoc]

That some outfit figured out that ChatGPT is an immense security issue. And not just for business information.

Re:

[sean-it-all]

Pretty sure most companies that have a security department have started to figure that out. Especially when all logs are public. I guess the rest thought they could save money by removing Security.

Design our first new product since Jobs

[ebonum]

I'm Afraid I Can't Do That, Dave.

It doesn't seem to learn from users.... yet.

[Seven Spirals]

Let me share a specific incident with ChatGPT, which leads me to believe it doesn't learn anything from user's, yet. I asked it about volume management for Tru64, which is a topic I'm an expert on. It gave a lot of wrong information in it's response. It claimed LVM was available for Tru64 and it ignored the fact that it came with LSM (an early version of what later became VxVM). I corrected it and it took the correction, acknowledged it had given incorrect/wrong info about LVM and then rephrased it's origin

Re: It doesn't seem to learn from users.... yet.

[Logger]

It will even tell it doesnâ(TM)t, if you ask. It uses the trained model plus the history of your current conversation. It canâ(TM)t even refer to your own other conversations. But thatâ(TM)s not the risk here. Everything you chat with it is permanently stored on their servers. So you have now entrusted OpenAI to protect your data from leaks to other users, and nosy OpenAI employees who decide to go mining for inside information.

Re:

[Seven Spirals]

So you have now entrusted OpenAI to protect your data from leaks to other users

Good point. They certainly *could* go back and mine that text, too. They just don't seem to do it.... yet.

Re: It doesn't seem to learn from users.... yet.

[unami]

Of course they do, but one user won't change the whole model. Afaik, if you pay for it, it doesn't mine your text.

Trust

[RitchCraft]

Companies don't trust these language models but are more than willing to inflict them on users. Proof that they're nothing more than tracking/anti-privacy/money grabbing gimmicks at the moment. These companies are just proving that no one should trust these things.

Re:

[tlhIngan]

Companies don't trust these language models but are more than willing to inflict them on users. Proof that they're nothing more than tracking/anti-privacy/money grabbing gimmicks at the moment. These companies are just proving that no one should trust these things.

Huh? Apple isn't using ChatGPT for anything. The only thing ChatGPT related on Apple devices is that OpenAI released an app for the platform, something Apple can't really restrict.

In fact, I think the only company using ChatGPT for anything public

Re:

[RitchCraft]

I said language models, which encompasses all of these half wit programs.

Use it sensible and its great

[sapgau]

Well don't copy hundreds of lines of code into the ChatGPT text box and ask it to add new features or to figure out what's wrong. The (human) developer should be able to plan for new features or troubleshoot the code.

What ChatGPT is useful for is to retrieve basic algorithms, code syntax and EXAMPLES of how to do something.
One thing I was having trouble with and couldn't google the exact answer was how to connect to Active Directory and verify credentials using NodeJS.
ChatGPT gave an example of how to

Maybe they should restrict Web use too

[Tony Isaac]

If Apple is worried about employees leaking confidential data, they should probably restrict use of Google, and Wikipedia, and, you know, web browsers. These are all systems where data could be leaked.

Re:

[anonymouscoward52236]

Might as well shut down all communication in a 3 mile radius around the Apple complex. The cellphone / RF jammers will be installed tomorrow, and all the fiber lines are getting back-hoe cut immediately! Shut it down, shut it all down! LOL.

In house

[stikves]

Given the recent explosion in LLM based engineering, it is not too difficult to have a "light" GPT "at home".

There are already generic models you can download that matches GPT-2, and those you can train that could match GPT-3. GPT-4 level still requires roughly a billion dollar investment, so out of reach at the moment. But if all you are looking for is coding advice, a simple model would be more than enough.

And... there is already an open and pretty capable model you can download for coding: https://huggin [huggingface.co]

Re:

[anonymouscoward52236]

Or, if you have Apple-level of money, you could easily train your own models. It wouldn't take billions. A few millions, at most.

Re:

[stikves]

The estimated cost, including research is literally $1 billion: https://twitter.com/sytelus/st... [twitter.com]

GPT-4 level models are pretty large. Unimaginably large. Even if you are "renting" the hardware on the cloud, it still costs hundreds of millions on training alone, much more in iterations and fine tuning. A billion might even be considered cheap.

Why do you think many companies went into alarm mode? (hint: did they make the same calculation?)

On the other hand, as I discussed above, simpler coding models will onl