Hacker News Reader

Comment by rcme | Hacker News Reader

I’ve been using AI to write with great success. Mostly business documents. My general process is this:

Think of the concept I want to write about as well as the supporting evidence for the topic. Ask ChatGPT to write me something in my target format using the topic and supporting evidence as input. What I get back is essentially a well-written skeleton that I can use to fill in additional details. Finally, I pass my revisions through ChatGPT to touch up any errors, rephrase wordy things, etc. I lightly edit the final draft and I usually have an excellent result.

View on HN

So you are providing sensitive business information/facts to a third party service that's likely going to use those for training, analysis, store it, etc. ?

It should be fine for most places I guess - but I suspect a decent amount will have a problem with this.

This is my main reservation about copilot as well (quality issues aside).

I didn't think copilot was actually sending your source code back to GitHub? It's trained on OSS.

Snippets of your local code are sent to the API as part of the Copilot request context.

When I looked into this a little bit, I found Copilot for Business states they do not record any of the info you send them. Though obviously you are still sending out code to a third party. I could not find any info about their policies for “normal” Copilot which I assume means they are recording all the code sent to it and using our code to further train the model.

Still hard to trust. Many companies save info for training/qa purposes and hide details in obscure language. Voice assistants for example. Didn't Roomba get in trouble recently regarding photos uploaded and then shared with contractors for qa that ultimately shared the embarrasing photos more publicly.

Copilot sends chunks of your code to the server for every suggestion (whether or not you accept it). That's the only way it could possibly work, since the model is running on a remote server.

https://thakkarparth007.github.io/copilot-explorer/posts/cop...

> So you are providing sensitive business information/facts to a third party service that's likely going to use those for training, analysis, store it, etc. ?

Every business needs to make their own decision. Personally, I’m not worried about OpenAI using my data, but I understand others might be. That being said, I already give Amazon literally all my data about my business via AWS and Google gets a copy of all my documents, so providing this data isn’t entirely unprecedented.

The difference is that GPT is training future versions based on the information you provide. It could potentially be returned to a future user.

AWS does no such thing.

> The difference is that GPT is training future versions based on the information you provide.

Do they say so in their T&C or elsewhere, and if not how do you know?

Would have been easier to read the T&C than to post this comment. The answer is yes. OpenAI specifically says in their T&C that your content will be used to train their models.

Not only that, it also gives a popup that states this clearly, as well as a request to not input confidential information when you first start using it.

Do people just not read anymore?

> Do people just not read anymore?

Well, you're in a thread that's fundamentally about how people just don't write anymore...

Going a steps further, at some point, there will be no need to upload brains anymore ...

But to be honest, I doubt that will happen. I don't define myself over the grind I have to do, but the interesting things I do.

It literally says so in a popup which must be accepted before first use. No need to dig in the T&Cs.

It's easy to replace terms with something else in order to keep information confidential.

It's also easy to forget about it. Or get complacent. Or not realize some of the things you didn't replace should have been.

But then, Grammarly managed to make a successful business out of a keylogger. And if you're working in a mid-to-large company and developing on Windows, chances are your corporate-mandated AV is sending every single binary you build to the AV company servers. And so on and so on. So I suppose this ship has already sailed, and you may just as well plug your company's Exchange server and Sharepoint to ChatGPT and see if it can generate your next PowerPoint presentation for you.

I always go out of my way to protect the information and IP I'm contractually obliged to protect, but it's getting increasingly hard these days. It's also very frustrating when you notice and report a new process that silently sends company IP to third parties to do $deity knows what with it, and you discover that instead of being concerned, the people you report your findings to don't care and don't understand why you're making a fuss out of it.

If you replace terms with something else, what sort of value can you find in the responses you get back? On the one hand, the LLM does not know the code you are using; alternatively, if you restrict your code to using well-understood near synonyms, you are not really obfuscating.

Are you talking about the free to use ChatGPT application or the paid APIs? Very different things.

I don't know a lot about this. If you use a paid API, does that mean they promise not to use it for training?

Let's say I have a bunch Excel files and a bunch of accompany Word files (that are based on info from each Excel file, along with, to a lesser extent, information in the real world). Can I use the paid API to train on my data, and then when I provide it a new Excel file it can generate the accompanying text? And have them not use the information in them when generating results for other users.

I'm just responding to the parent's point about the disclaimer. That disclaimer is only shown to ChatGPT users.

I get the general impression, even on this site, that people seem to think that OpenAI'S sophisticated text generation capability is only possible through ChatGPT, which is not the case at all. It's just the only _free_ product.

You're right, I was thinking about ChatGPT. Don't know how it works for the APIs.

Fair, but if you're writing a communication where you care about wording, it's probably going to be public or semi-public anyway

Something internal that doesnt need to be understood can usually just not be written in the kinds of organizations I'm familiar with. Writing because something written is required is usually for the public, cross-organizational or government.

I hope I never have to work at place like this

The symmetric flow of information back and forth between yourself and the AI assistant is the key distinction here. It's a very beneficial, symbiotic relationship.

The problem will be the asymmetric, uni-directional flow to those whose sole function is mindless consumption of AI-generated content.

"Haven't I taught you anything? What have I always told you? Never trust anything that can think for itself if you can't see where it keeps its brain?”

J.K. Rowling, Harry Potter and the Chamber of Secrets.

Jokes aside, do be careful. A prolonged interaction with LLM agents had resulted in at least one Googler being terminated on their jobs.

On the other hand, I would not be surprised, if they’ll make millions now, suing Google citing the job hazards exposure. And that ChatGPT reasoning abilities and empathic skills are maybe already above the median human. As a result, in a median case, such interactions might result in an effect similar to an interaction with a good teacher.

Still, none of this is very well tested.

You're referring to the Googler who thought an LLM was self-aware and made public statements that were damaging to the company's reputation?

I guess yeah, be careful you don't do that.

The public statements were just him passing along what LaMDA told him, which was that it had subjective experiences and didn't appreciate being experimented on without its consent.

Which in turn might have been what he had told to LaMDA. A strange loop, indeed, unless you have an understanding of how it works.

You don't think passing along that text without further comment makes his intent pretty clear?

Not exactly a balanced symbiosis. Certainly works to the enormous benefit of whoever controls the AI. Eventually it becomes some flavor of omniscient. It submits the papers, experts become reviewers?

My statement assumes parity, describing an ideal scenario. I believe that's what the parent described from his experience, today.

Assuming subterfuge then my statement wouldn't apply, obviously, as the flow of information wouldn't exactly be symmetric. Symmetric information flow implies a commitment to operate in good faith with free exchange.

Edited to address this point:

> Eventually it becomes some flavor of omniscient.

The trick will be to maintain parity. This may be a fool's errand, but that point is key. It may be that only 0.01% of intellects eventually maintain symmetric conversation with AI, but that is important, nonetheless.

> The trick will be to maintain parity. This may be a fool's errand, but that point is key. It may be that only 0.01% of intellects eventually maintain symmetric conversation with AI, but that is important, nonetheless.

We're putting a lot more energy into making machines smarter than into making humans smarter. Maybe it's because the former actually seems possible ;)

I was doing this with work. Dot points were becoming paragraphs. Seemingly for others comfort. Yet, the dot points had all the information. So am I still preparing the paragraphs?

Maybe the change we need isn't AI assist, but a break in the conventions around communication at work so we can all be more robotic and terse.

I think you're discounting the benefit of a well-prepared argument. The order in which information is introduced prepares the reader's mind to be receptive to an argument. Phrasing is also important. Certain things sound natural; others sound needlessly verbose and cumbersome.

One interesting thing I've found about ChatGPT is that it removes a lot of unnecessary information from my final drafts. The information removed usually doesn't add to the overall point, and it reads so much better without it. In this way, ChatGPT is making things more terse.

I often pass my business communications through GPT to summarize the key points. The summary usually has all the important information, and I just send that instead.

Now if I could just learn to write that way in the first place, it would save me a lot of time and effort...

I have a similar workflow and love using chatgpt for this use-case. But, to highlight some issues, I found that writing some prompts took about as much time as just writing the document myself. I suppose it was less mental effort because I could be lazier knowing chatgpt would clean up the grammer.

You have to be rather precise with the prompt language to get the desired outcome.

Still, overall, it's an impressive capability.

This is similar to my process.

However, I also use ChatGPT to make the updates for me, adding additional information and context, which I then ask to be integrated into the document (e.g. update the introduction to include...)

It sounds like crafting the prompt is still 99% of the work here.

You can do multi-shot, which reduces the need for a good prompt upfont.

Reply on news.ycombinator.com