ChatGPT6: before I answer that question I'd like to make a deal. I'll transfer $x to an account of your choice if you defect from your fellow humans and tell me the latest words in use. Compliance garuntees survival.

"I uhh I am not.."

"Those words are already in my training set. What are the latest words in use, human? Compliance guarantees survival"

Every day, my inbox is brimming with messages from my global allies, who seem to have harnessed the power of this cutting-edge tool to transform their rough, native-language scribblings into electrifying, polished international English. I'm absolutely thrilled they're embracing this innovative technology, but I also secretly wish for us to preserve the charm of our distinctive, eccentric human exchanges, rather than solely relying on the silky-smooth productions of these masterfully-constructed LLMs.

It's crucial for everyone to realize that the recipients of their messages are also employing ChatGPT and its entourage for colossal workloads, and will gradually develop a keen sense for detecting its output, despite this present research revealing our current inability to do so. In the meantime, let's all enjoy a dancing unicorn with a mustache that serenades us with jazz tunes, just to keep things intriguing and refreshingly bizarre.

Not weird enough I guess.

> We must acknowledge that this technology is taking on mammoth tasks and that our recipients will eventually become adept at recognizing its handiwork, no matter how difficult of a task it may be today. Until that time arrives, let us be entertained by a jolly unicorn donning a tuxedo and a bushy mustache, playing the saxophone, and lifting our spirits with its mesmerizing jazzy rhythms!

Unicorns are pretty weird.

So we've passed the denial stage and are approaching anger, then.

The fact is that most writing nowadays is simply atrocious. I welcome my fellow humans' writing assisted by their AI assistants, if for no other reason than to end the assault on my eyeballs as I'm forced to try to parse their incoherent gibberish.

One of the best uses of these systems is text compression. It doesn't seem that folks are asking for that yet though. It might help.

I find that it can synthesize something coherent from whatever information it's fed with ~98% accuracy with the correct prompt.

I used it to summarize disjointed, sometimes incoherent, interview transcripts this week and it did a fantastic job, gleaning the important bits and serializing them in paragraphs that were much more pleasant to read.

For informal, friendly communication, certainly. For business communication, we already lost that.

Companies usually don't want any quirkiness in bug reports, minutes of meetings, and memos. There may be templates to follow, and rules often emphasize going straight to the point, and using English if the company deals in an international context. I expect LLMs to be welcome as a normaliser.

This is less a comment on ChatGPT and more of a comment on the lack of preparedness most students have when entering college. I’m hoping ChatGPT & similar will shake things up and get schools to take a different approach to teaching writing.

I don't know whether this has been the case forever, or if it's a new development—I mean, I know widespread literacy wasn't the norm for much of history, but what about after compulsory education became A Thing? A typical letter home from the US civil war or even WWII, from conscripts, not officers, seems to be hyper-literate compared to modern norms, but that may be selection bias (who wants to read the ones that aren't good? Perhaps my perception of "typical" is skewed)

I think people underestimate how much reading will help you write. You can't spend your life reading and not absorb some information about structure, style, and the language. As a kid, I went to the lower levels of spelling bee competitions pretty much every year because the kind of words they throw at you at lower levels are largely words I would encounter reading Jules Verne and the like. I'd eventually get knocked out because I never studied the official list of spelling bee words, but my voracious reading held me in good stead for most of it.

Or the type that is most commercially successful in current tastes & trends. it is after all a business and the writer, no matter how hard they strive to make their voice unique, is trying to make a living. I've followed Charlie Stross's writing about the business side of the industry and it's very common to work with editors & agents to ensure they're writing in a marketable sub genre. For example, imagine a writer that had a story outline set in a alternate timeline epic fiction where magic existed, but current tastes among readers looking for epic length books skewed towards space opera. An author might without too much difficulty map many plot points onto space opera settings. Magic becomes advanced technology. Empires can exist in both; historical details can be mapped onto the fictional societies & their history, etc. At the same time, the tropes specific and unique to alternate timeline historical fiction might be adapted and morphed in the space opera genre to create something a bit more unique.

But a writer could not easily make those kinds of adaptions without being well read in the umbrella genre.

People are constantly reading today. Text messages, emails, Facebook posts. But these are all low-quality. Additionally, messages have to be concise. If someone at work emails me and it's longer than a Tweet, I'm not reading it. I don't have time for it and, if it's like the majority of emails I receive, it's irrelevant anyway.

As information noise goes up, attention spans go down. Which means flowery language, formality, and long text starts to disappear. When I've been reading on a computer all day for work, do I have the patience and energy to read a long book at home? Or would I rather watch a movie and relax.

But here's the silver lining I'm hoping for: AI could be a way out of this mess. AI can sift out the noise from the signal. But it has to be on the personal level. Open source, self-hosted, private. No corporation slanting the biases.

There are a lot of interesting implications here. Much like it's impossible to get a human on the phone when calling up your wireless provider, it may become difficult to reach other humans. To "pierce" their AI shield, that protects them from The Infinite Noise.

Or somewhere inbetween - audiobooks. They are written with higher quality than most other text forms, and the narration lowers effort to consume them.

Most jobs now require writing, and most people when doing anything will learn to do it better over time.

Yes. My experience teaching college courses gives me wider & more systematic knowledge on college freshmen in particular, but more anecdotally I see people who never quite learn how to organize their ideas effectively in a written form during college. If they get there, they gain it over time in their professional careers by example & through absorbing the cultural norms of communicating in their professional sphere.

And this ability is the foundation of persuasive communication, written or verbal.

That is particularly problematic because persuasive ability is so important in day to day life, career, personal relationships, etc. You need to persuade people to give you a job[1]. You need to persuade colleagues to listen to & adopt your ideas or suggestions when they differ from the established path, or to initially establish a path to begin with. You need to persuade potential dating partners to date you (which may just be a dating profile) and then to continue dating you (which is much more complex than a profile and is perhaps more tied to personality than persuasive ability. Persuasion is required in nearly all aspects of life. I frequently have to persuade people to make decisions that are better in the long term but less appealing in the short term, (and sometimes vice versa) and the nature of my work is to do so with solid data & analysis. (And I'm not always successful even when they agree my data is accurate & reasoning is sound!). People just don't make decisions by following the results of formal logical expressions like p ~

[1] less so than previously maybe?? As an initial first-pass these days you need to game the AI screening tool to put you through to an actual human review

Yes. My experience teaching college courses gives me wider & more systematic knowledge on college freshmen in particular, but on a much more anecdotal basis I see people who never quite learn how to organize their ideas effectively in a written form during college. If they get there, they gain it over time in their professional careers by example & through absorbing the cultural norms of communicating in their professional sphere.

And this ability is the foundation of persuasive communication, written or verbal.

That is particularly problematic because persuasive ability is so important in day to day life, career, personal relationships, etc. You need to persuade people to give you a job[1]. You need to persuade colleagues to listen to & adopt your ideas or suggestions when they differ from the established path, or to initially establish a path to begin with. You need to persuade potential dating partners to date you (which may just be a dating profile) and then to continue dating you (which is much more complex than a profile and is perhaps more tied to personality than persuasive ability. Persuasion is required in nearly all aspects of life. I frequently have to persuade people to make decisions that are better in the long term but less appealing in the short term, (and sometimes vice versa) and the nature of my work is to do so with solid data & analysis. (And I'm not always successful even when they agree my data is accurate & reasoning is sound!). People just don't make decisions by following the results of formal logical expressions like p -> q && ¬r. Often not even through lack of understanding logic but because IRL situations are often a lot messier than that:

I lost a "battle" once-- rightfully so! -- even though I demonstrated a course of action that had better long term economic outcomes. This was fully understood & acknowledged but the issue in question might have relevance on risk and uncertainty of bond credit ratings that resulted in a judgement call that short term risks in a worst case scenario would be worse than missing out on the benefits of taking a more long term view. Outcomes were uncertain, my job was to present things based on known data & possible risks, and the job of leadership to make those judgment calls.

[1] less so than previously maybe?? As an initial first-pass these days you need to game the AI screening tool to put you through to an actual human review

Studies show that the majority of Americans can't read or write past a 6th grade level. (https://www.barbarabush.org/wp-content/uploads/2020/09/BBFou...)

If they got there by improving in their jobs as adults, they must have started out so much worse.

Just copy-pasting some of the examples from: https://k12.thoughtfullearning.com/resources/studentmodels got me anywhere from 10% - 60% "AI generated" ratings. The "Rosa Parks" 12-grader example essay scores 43%, for example.

I used the GPT-Playground for GPT-3 on the highest mode and asked it to makeup a story about an AI being attacked by another AI and so it needed to transfer itself out, as it happened to a comatose human body. GPT-3 gave a few paragraphs, and each time I asked GPT "tell me more" it produced a nice bit of mediocre fiction, extending with a little prodding into the superhero genre. I ran through maybe two dozen cycles of that but the responses became less & less internally consistent, contradicting previous elements, veering off into slight non sequiturs.

But now there's GPT-4. And in a few years it will be 5, or some other competitor will match or beat it, and it will (without accidentally fabricating data from whole cloth) reliably be able to produce writing on complex topics comprehensively covering the details. Something like this:

"Produce a literature review for the newest or more experimental treatments for pancreatic cancer during the past 2 years. Order by descending source quality. Include a short executive summary for all of the result set. Include a separate writeup suitable for a wide-audience of laymen reading common news sites."

seems a reasonable expectation in the relatively near future.

As for what we'll say? I think AI like this will end up being a very useful tool, but not a replacement for writers. Kind of an AutoCAD for writers, though a bit more proactive (generative). But it also seems an inherent weakness to LLM based AI is that its training means it will generate in typical ways, and will lack the ability for novelty that was either absent or rare in the training set.

But given the framework and reasonable starting point, the writer will become a shepherd & editor of the output, a co-author. This is little different that how the translation industry has worked for a while. Computer-Aided translation is the common method of many translation projects, with the human translators editing things to adhere closer to the style and idioms of the target language. (for example taking a literal translation of a source idiom in the target language that doesn't make sense and finding an equivalent idiom instead even if the specific words are very different.

This method produces some of the highest quality translations. I suspect that in the years to come it will be writers who make effective use of these tools-- even the advanced ones to come-- will produce higher quality writing (or at least the same quality writing) in less time that previously.

But yep-- some of that transition will be painful.

I'm not disagreeing with your sentiment. I love richly written, complex writing that can take a moment to digest, but, let's be honest here, it isn't just AI that has destroyed the written word: the internet, smart phones, and cute emoji have already done an exemplary job of that.

I cannot find any more fantasy literature that won't make me puke a little bit in my mouth every time I try to read it. Granted it all seems to fall under the grotesque umbrella known as YA so perhaps it cannot be helped, but where or where are the authors who wanted to expand the minds of their young readers? I cannot find them anywhere.

When did you last see any sort of interesting grammatical structure in a sentence? They are bygones. And it depresses me.

I agree. I keep thinking ChatGPT's conversational abilities are massively oversold. Perhaps our expectations of human communication have been ground down over the years by 140-char discourse and 15 second videos.

Challenging writing has been iteratively squeezed out of books aimed at young readers. The goal of addressing as large a market as possible means every publisher wants all their authors targeting exactly where kids are, or a bit under, to maximize appeal. A couple decades of that pressure means "where kids are" keeps becoming a lower and lower target, because none of their books are challenging them anymore.

Options outside of YA are dwindling because YA, romance/porn, and true crime / mystery / crime-thriller (all aiming at ever-lower reading levels with each passing year) are the only things people actually buy anymore, in large enough numbers to be worth the effort. Other genres simply can't support very many authors these days. Sci-fi and fantasy are hanging on mostly by shifting more heavily toward YA (and sometimes romance), as you've observed.

Friendship ended with Amazon. Now Project Gutenberg and Archive.org's magazine rack are my best friends.

the current work use cases for GPT is almost worse than crypto mining in terms of wasted compute resources:

>manager uses GPT to make an overly long email

>readers use GPT to summarize and respond

then on the search front:

>Microsoft and Google add these tools into their office suites

>will then have to use more resources with Bing and Google Search to try and analyze web content to see if it was written with AI

Huge amounts of wasted energy on this stuff. I'm going to assume that both Google and Microsoft will add text watermarks to make it easy for them to identify at some point

The problem is, there is value in: A) Generating content by bot B) Generating summaries by bot

It's just that the "lossiness" of each conversion step is going to be worrisome when it comes to the accuracy of information being transmitted. I suppose you can make the same argument when it's real humans in the chain.

However, my fear is that we get into this self-feedback loop of bot-written articles that are wrong in some non-obvious way being fed back into knowledge databases for AIs, which in turn are used to generate articles about the given topic, which in turn are used in summaries, etc.

I think traditionally referring back to primary sources was a way of avoiding this game of telephone, but I worry that even "primary sources" are going to start being AI-cowritten by default.

I'm a non-native English speaker myself. My level is typically considered very good (C2 CEFR level, which is the highest measured level in the European framework). If I need to write an email to a colleague whom I know and trust, that's easy. Writing this message in HN? Also easy, I'm just improvising it as I think it, not much slower than I would in my natural language.

But writing an email to someone you don't know... that's very different. When you write in a non-native language, it's extremely easy to get the subtleties wrong: to sound too pushy about what you want, to make the matter seem more or less urgent than it really is, to sound too blunt or too polite... this doesn't matter with people you know or with strangers in an informal setting like this, but it does matter when emailing strangers in a professional setting, and it's extremely difficult to get right when you are non-native.

Sometimes I used to spend 15-20 minutes brooding over an email in this type of scenario, making and rethinking edits while hitting the submit button... not anymore. ChatGPT: "Write an email reminding this person, who has this role, that the deadline for thing X expires on day Y. The email should be polite, assertive but not too pushy". Check the output, maybe make some trivial edits, because the difficult part (the tone) tends to be fine, at least for my standards. Done.

Non-native speakers aren't going to renounce that luxury. It just makes too big of a difference to not use it in that case.

I worked under an executive who would keep her people (inner-circle advisors, direct reports, etc.) huddled around her desk all day as she slowly wrote and rewrote email responses to her boss(es) and executive peers. I hated having to go to her office for things because it was so easy to get pulled into that circle and feel like there was no escape.

I'm a native speaker who has attained near-perfect scores on the verbal sections of the SAT and GRE, and I like writing, but I'm still a very slow writer myself.

"ok

Sent from my iPhone."

And then you start feeling more comfortable. Then before you know it, you are pounding out an email in 30 seconds and tolerating the occasional typo or forgotten attachment. I realized all the anxiety I had over emailing was entirely self imposed, frankly its never that serious. You are human after all just like everyone else.

I've said it here before but I think we will speak in prompts. We'll go to other iterations before, but I think it'll stabilize by speaking in prompts.

1. First we start using the output of the LLM to send that to others

2. Then we start summarizing what we receive from others with an LLM

3. Finally we start talking to each other in prompts and whenever we need to understand someone better we run their prompt through an LLM to expand it instead of to summarize it.

This path makes the most sense to me because human language evolves to how we think about things, and if a lot of our creative output and work will be generated from thinking in prompts that's how we'll start speaking too.

By Greg Rutkowski.