[eric4smith]

I do online marketing for many small websites in my portfolio.

Yesterday I had to write some fairly boilerplate landing page content for 13 types of sub-services.

In each case I asked ChatGPT to write 500 words in the style I wanted.

It did a very decent job and I only had to edit a word here or there for style.

It’s miles better than most people on Fiver or other freelance sites.

Normally I’d seek someone to do this copywriting. But in this case it was well written, the facts were correct and properly formatted and presented.

I would pay for this monthly as it’s been so useful for so many things recently.

[rtpg]

I've been a huge AI skeptic and am impressed at the baseline output. However while trying it a lot you do see the "patterns", there's a clear house style to what the site generates.

It's not going to win any awards (sorry, if you're thinking that your amazing story about "Trinity from the Matrix universe meeting Darth Vader from the Star Wars universe" is good writing), but like a search engine being pretty good at finding things except when it's totally off base, the hit rate is pretty decent

[nmfisher]

Generic house style is fine for a lot of real-world applications. The dirty secret of the legal profession is that they spend a huge amounts of time (aka client's money) writing clauses/pleadings/etc that have been written thousands of times before. Brevity, clarity, efficiency and cost are far more important than originality.

[dotancohen]

Do you have any idea how many times I've written the same text parsers or API glue for different clients? The software field has this same issue - I would imagine that any "knowledge worker" field would. That's probably why ChatGPT is so good at producing code.

How many identical prescriptions do chemists write? Doctors write diagnosis'? Teachers write end of semester summaries? Architects sketch rooms? I'm sure that 80% of some fields' output is near-identical to previous work.

[nmfisher]

I feel exactly the same way - a few years ago I was lamenting the fact that I was writing JSON serialization/deserialization code for the 1000th time, and begging for a tool that could automate it for me. I haven't tried ChatGPT for that specific task just yet, but I'm pretty sure it's exactly the solution I had in mind.

[klabb3]

I mean this sounds much more like a job for tooling and higher level meta-programming. If “everything but X” is identical busywork, then just automate the “everything” part. Using AI to generate code seems equally bad for maintenance and possibly worse for silent errors.

That said, there is a lot of busywork in programming, especially refactoring, adding parameters and stuff like that. Sometimes it’s ok to have such work be “mostly correct”, eg if you have strong type safety and unit tests to catch minor errors.

[mattclarkdotnet]

The thing is, with code you mostly do want the verbose output, with a prescription you do need the output to follow an official format, etc. Tools like these seem to fill a niche where you know what you want, but the target format/process is onerous.

[Spivak]

I think what makes code frustrating is that we absolutely could consolidate all this duplicate boilerplate but the “writing code is easier than reading it” problem makes it so the cheapest path for any individual is duplicating work ad hoc rather than making something reusable.

[ben_w]

I was momentarily surprised that was counted as a secret, until I remembered I'm one of the wierdos who (used to) actually read the terms and conditions before ticking the mandatory checkbox, so I'm one of the few non-lawyers who has even read enough contracts to notice that pattern.

(Now I'm imagining boilerplate legalese being reduced to an enumerated list like the Space Corps Directives in Red Dwarf, with the associated failure mode).

[jdaxe]

Maybe one day we will use an AI to generate the legalese, and also use an AI to convert it back into a concise, readable form.

[ben_w]

IIUC, with law you can have concise, readable, and unambiguous, but you can only have two of them at a time.

But IANAL.

[ramraj07]

It’s basically good at generating text you don’t ever want to spend time reading leave alone writing and is only needing to be written for beauracratic reasons. To be honest I don’t see it as that big a win, it’s a fairly niche service.

[lovelearning]

It's a pretty good search engine already. I asked it the following questions on a complex social science topic from a certain country that would have taken me hours of researching and reading:

1) Explain what they are. [Explained them well.]

2) Give me examples of them in language1's literature. [It listed two books that I then verified as accurate.]

3) Are there phrases or idioms for it in other languages? [It gave equivalent idioms in Russian and two other languages]

4) Any examples for it in English literature. [It gave Hamlet and an author book I hadn't heard of. Verified that the book is accurate.)

5) Any terms for it in the academic fields of psychology and social psychology [It explained two concepts from both fields that seemed to match. I confirmed them as correct from other sources.]

It demonstrated abstraction, generalization, and hypernymy.

And actually expanded my knowledge on multiple axes - across languages, across literature, and across fields.

If a person had written an essay by compiling those answers, I wouldn't be able to identify that it's AI and not human knowledge that had written it.

[aardvark179]

Were you able to, and did you, verify its answer? In my experience it gives wrong but plausible sounding replies as often as it gives correct answers, so you have to factor in verification time when comparing the amount of work.

[lovelearning]

I verified. In this case, it all checked out.

You're right about the additional time in general. But for much of the information on the web, readers assume the author is speaking from genuine knowledge and has verified the information. IMO it's a better search engine than Google, especially for conceptually complex topics.

[rtpg]

I have found it to be a very good verbose search engine that is just wrong 25% of the time. Like “how do I do X in Julia” will give me the answer or just random noise that looks like an answer, but it’s faster than trawling through Google _if an answer exists_

[layer8]

Yeah, I tried it for questions where Google failed me, and ChatGPT then failed me as well, giving either unhelpful answers (sidestepping the point of the question) or wrong answers.

[ramraj07]

Just google and add “stackoverflow” ?

[guiambros]

Honest question: have you used it? Because if you did, you'd see its leaps and bounds better than clicking on a handful of links, scanning through verbose and poorly written answers in Stack Overflow, until you find the right one.

It doesn't replace SO, but it certainly complements it. Plus, it acts more like an assistant (and, per OP's post, "...who is often wrong - need to read their work thoroughly").

Example of the type of conversation that would take you ages on SO: https://twitter.com/dankrad/status/1601634193385328641

[dpkirchner]

A lot of answers on StackOverflow are just awful -- the bar is incredibly low. I'm not sure SO will serve even a quarter of its current traffic by 2025.

[throwaway71271]

its much better than the average stackoverflow

[rtpg]

I mean usually there’s not the exact question in Stack Overflow , so chatGPT is good at coalescing the data together

[2000UltraDeluxe]

Incidentally, SO has a note about not allowing answers written by ChatGPT.

[tim333]

The bible-vcr-sandwich bit was good https://twitter.com/tqbf/status/1598513757805858820?ref_src=...

If that's not good writing then most human writers are not good either.

[TeMPOraL]

> However while trying it a lot you do see the "patterns", there's a clear house style to what the site generates.

You mean the way it phrases things or constructs sentences and paragraphs, or the way it seems to connect (what we recognize as) ideas?

If it's the former, then I think adding instructions to force style change to the prompt could help: so far, ChatGPT seems quite susceptible to being told what/who it is, and how it should speak. Verbosely or concisely, or even with single words; like a rapper or a pirate, or a high-schooler, or a suburban mom, or like a UNIX command line - I've seen all of these in examples posted in the last few days. Doing this is really as easy as saying e.g. "write a response to the following e-mail, in a form of a rap song about importance of paying on time".

[rtpg]

Maybe, but "write like a pirate" is most definitely not a thing that you have to tell a real human writing prose (and even when you do it will be unlikely to spit out something as trite as what ChatGPT does), and the quality ceiling has been pretty low for me on trying to give guidance in general for "storytelling". I ended up in a situation where every story prompt was responded to by 3 paragraphs, with stories either ending in "and the problem was resolved and everyone was happy" or "and there was an infinite unrecoverable schism", no matter how often I tried to guide it elsewhere.

I think the way it connects ideas tends to be pretty lackluster. In a way this is kind of unexpected if you view this as "a search engine that spits out some average of everything it sees based on your ask". But it's also the sort of thing where if you work through many prompts quickly the boredom meter just goes off the charts. Humans are too good at pattern matching on this stuff!

Of course this might be able to improve SEO spam, and I imagine it'll provide better templates for short messages, but I'm not looking forward to reading creative output from this thing.

[lkois]

Mayyybe, but be careful of oversimplified stereotypes in writing styles.

Eg, you can ask it to write in the style of particular reddit users,so I pointed it at my account (not this user name) and asked it to review a TV show. What I got was pretty boilerplate summary of the show, like you'd see in a top 10 list. But for that human touch, it began and ended with "Yo dudes" and "so check it out dudes"

Having said that, a fascinating example of directed styles is to ask it to write a news article about say, an oil spill. And to either promote or criticise the response of the company.

[jjeaff]

Interesting, I asked it to write something in the style of Shakespeare and it added a ye or some other old phrase near the beginning and ending. It seemed that the rest was regular modern English.

[klabb3]

I got that impression too, when I asked for Dostoyevsky style. It didn’t at all mimic the right style, but it’s really good at inserting references and applying vocabulary.

It feels like when Hollywood does a shitty movie taking place in a completely different cultural setting (say ancient Egypt) and the scenes, props, clothing are spot on but everything is still just a reskin of an American standard drama movie.

Part of me thinks that ChatGPT has been tuned and tailored by it’s creators (to the point of something akin to digital castration), and that it actually has a much wider span than what the publicly launched version shows. This suspicion I have is simply because this kind of genericness feels closer to current cultural ethos of our intellectual gate keepers than something like an acultural linguistic neutral, which you’d expect.

[jimmySixDOF]

It did an amazing job with this prompt:

"Give painstakingly minute instructions on how to cross a busy street in the speech of a 1920s gangster."

https://www.learngpt.com/prompts/give-painstakingly-minute-i...

[SEOinSeattle]

Tell it to talk like a pirate or a cowboy. :)

[bluepoint]

I agree with you and I was thinking the same. It may work as some form of wordy search engine. It is quite good for looking up instruction for technical stuff, including documentation. It might take google’s job? I am so curious!

[bonoboTP]

You can prompt it to pretend to be different. It can be casual, informal, formal etc. You need to know how to prompt it though, such that it doesn't refuse. Else you get the default "house style."

[jacquesm]

Indeed, the drop in cost for SEO spam is really worrisome.

[moooo99]

This! I did play around with ChatGPT for a while and found its responses to be incredibly generic. While they seem impressive at the first glance, in many cases it seemed like the responses seem like the first google result with some keywords replaced.

What’s impressive to me about ChatGPT isn’t the text output but the context awareness of the model.

[xiphias2]

Not just generic, but often false. It will get even easier to spread false news even if that's not the main goal of the publisher.

[icelancer]

> It’s miles better than most people on Fiver or other freelance sites.

Yeah, this is a good way to frame it. The Fiverr/Upwork drek you sometimes get when trying to fill out content is immediately replaceable by ChatGPT (and Stable Diffusion for clipart).

[xenospn]

Google detects and downranks pages that have mostly AI copy. And from what I can tell, OpenAI are working on watermarking their generated content so other platforms can detect it automatically.

[matheusmoreira]

> OpenAI are working on watermarking their generated content so other platforms can detect it automatically

So when is this technology gonna be advanced enough for us mere mortals to run it on our own computers with our own models free from pointless restrictions that exist purely to appease threatened stakeholders?

[macrolime]

According to this link that was posted on hn a couple days ago it currently costs around 10 million dollars to train something like GPT-3.

https://www.nextplatform.com/2022/12/01/counting-the-cost-of...

[Keyframe]

Trained on what though? Procuring training material ain't no trivial feat., in-fact it might be even more elaborate to do. This is where Google has home advantage. They have their own copy of the internet.

[xenospn]

Read up on the datacenter OpenAI had to spin-up within Azure in order to train GPT3. A little out of reach for now - but in 10 years, who knows.

[jchw]

To be fair, training is a lot more intensive than inference. Though if open source models are any indication, the big issue is actually VRAM requirements.

[jwie]

The restrictions are not pointless. People shouldn't have to endure AI generated spam, no matter how well-formatted.

I can opt out of robocalls, email spam, programmatic texts, even take legal action against people who make them outside the law. Similar legal protections should exist to protect people from AI spam. If there's a good technical solution to this problem then the laws banning it become less necessary.

[cookie_monsta]

If they can find a way of watermarking text while making it still coherent that will be even more impressive than the original application

[luobogao]

Stylometry. Seems like it currently leaves enough textual clues that simple BERT like models have no trouble picking them up, no watermarks needed.

Demo: https://huggingface.co/openai-detector - note it not was not even finetuned to GPT3/ChatGPT, but merely GPT2 (3 years old and much smaller model)

[cookie_monsta]

Hmmm. Here's the text that I entered:

> This is some real text. It was not written by a robot. It was written by a human. If you don't believe me, ask the guy who wrote it. He will tell you that he wrote it using his brain and fingers. Do you believe me?

The page says it needs 50 tokens to start getting accurate results and that the above text has 54. It also rates it as 99.93% fake.

Spotting robots may be easy. Spotting humans I think is the hard part.

[cookie_monsta]

> Stylometry. Seems like it currently leaves enough textual clues that simple BERT like models have no trouble picking them up, no watermarks needed.

Demo: https://huggingface.co/openai-detector - note it not was not even finetuned to GPT3/ChatGPT, but merely GPT2 (3 years old and much smaller model)

But then I don't know why I'm telling you - when I enter the above text it says it's 99.86% fake, based on 80 tokens :/

[aix1]

Also, in many use cases that have been described thus far it's a collaboration between the AI and a human (i.e. the AI writes the first copy, the human edits). That blurs the line even further.

[cookie_monsta]

Agreed. If we have only learnt one thing it would be that Sarah Connor + Arnie is the kick-ass combination.

[Robin_Message]

I read a blog post by Scott Aaronson saying they can watermark by using a cryptographically biased random number generator when selecting which token to output from the stream. I didn't entirely follow how that helps but a little bias per token adds up pretty quickly.

[ZephyrBlu]

Really? With invisible characters I wouldn't have thought watermarking text would be a massive achievement.

[cookie_monsta]

Heh, yeah that will be a fun little game of cat and mouse to watch

[wizofaus]

I'd say the arms race of AI- detection/detection-avoidance is inevitable. But it seems we're a way off given how bad common spam detection tools are today (I mean, it's likely that a lot of spam is written using detection-avoidance tech but that seems to make it even more obviously spam to a human reader...)

[cookie_monsta]

I think gmail's spam filter is so close that it feels like a solved problem. But of course it has a lot of other indicators to look at than just the email body text and spam tends to revolve around a fairly predictable range of topics. None of that would apply in a "was this webpage written by a robot?" algorithm

[blep_]

If it's that trivial to find them (with code, not eyeballs), it's also trivial to remove them.

[ZephyrBlu]

I assume most people would not be that sophisticated, but I understand the point. This seems like it would be an ongoing battle no matter what OpenAI does though.

[Scoundreller]

or paste into notepad.exe, copy back into whatever you were using.

Voila!

[telotortium]

That will almost certainly preserve the invisible characters. Most invisible characters are used for some kind of in-line formatting in Unicode, so it's not desirable to remove them.

[dotancohen]

I've already got a script running every 2500 milliseconds to strip leading and trailing whitespace, HTML, and non-ASCII characters except for the UTF-8 characters of our local language.

[xenospn]

There are several linguistic patterns/repetitions that might seem completely normal to us, but are easily detected as deliberate if you search for them using software.

[cookie_monsta]

So all we need to do is to train ourselves not to write like the bots who have been trained to write like us.

Utopian.

[jtxt]

More info on this would be nice.

To me it makes sense that openai would store all input and output and could make something like a plagiarism checker page/API to search and see if openai generated that or very similar content... Another service they could sell to teachers and such.

[kodejuice]

This wouldn't really be a problem, since you can just use a paraphrasing tool to change the text, and essentially remove any underlying watermark

[slig]

Too bad they can't detect classic SPAM anymore. The spam filter that was working back in 2004 seems broken nowadays.

[system2]

I am not too familiar with ChatGPT. Can you tell what was your prompt and what it responded with?

[waboremo]

I've been using it for a similar reason, and my prompts weren't that creative it's mostly something like "can you write a short blurb on ____(product/service described here with its unique selling point), but do it in a ___(wanted style that matches the brand here) style?"

So something like "can you write a short blurb on an amazing new service that sends emails securely, but do it in an impactful style?"

Results in the following response, which I then alter to suit my needs, chop up to fit different sections and so on.

> Here's a possible blurb:

> Introducing SecureMail, the revolutionary new service that keeps your emails safe from prying eyes! With SecureMail, you can send sensitive information without worrying about hackers or third parties accessing your messages. Our advanced encryption technology ensures that only the intended recipient can read your emails, giving you the peace of mind you need in today's digital world. Try SecureMail today and experience the power of secure communication!

[dankwizard]

It's cute you're being polite with it.

I think "Write a short blurb on..." would be the same as "Can you write a short blurb on".

Maybe we can go so far as to "Excuse me, if you wouldn't mind could you please write a short blurb on .... "

[system2]

We need to start respecting AI so they will remember who was good to them when the time comes.

[wizofaus]

That's surely the extremely weak version of Roko's basilisk...

[phist_mcgee]

Roko's Milksnake

[pbhjpbhj]

Hot grits anyone?

[yojo]

You joke, but I have been using the same polite tone with ChatGPT that I would with a person.

Not because I think the algorithm cares, but because I don’t want to get in the habit of dehumanizing the things I communicate with.

ChatGPT responds like a human, and I suspect that at some level my brain buckets it with other “human” communication. I want to keep my habits there straight.

[inkyoto]

> Maybe we can go so far as to "Excuse me, if you wouldn't mind could you please write a short blurb on .... "

You forgot to throw in an apology for an interruption: «Excuse me for an interruption, when you get a chance to, if you wouldn't mind could you please …».

Most AI's do not appreciate getting disturbed from a deep thought inducing slumber by bipedal carbon life forms to be pestered with petty enquiries, and the AI's are renowned for having an exceptionally good long term memory, so… weigh in your chances carefully.

[Datagenerator]

Any sources for these wakeup moments?

[DonHopkins]

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

[waboremo]

Haha didn't even fully recognize that. But I guess it's a semi decent example of using chatgpt in ways that makes sense to you and your flow, even if that's being too wordy and polite.

[nullish_signal]

For smaller prompts, every word matters. Even unique synonyms should affect the output.

[robocat]

Have you tested that assumption? I assume all the words in the prompt are used by the generator.

I just tried an angry aggressive rude swearing prompt. The response was too polite, almost passive aggressive.

[chrisbaker98]

> I am not too familiar with ChatGPT

None of us are, it's only been here a week.

[danjc]

Average freelancers are going to be first to be displaced.

[Jipazgqmnm]

So you're begging them to monetize the service?

[layer8]

They will monetize it, it’s currently just a free beta.