Unless you're in an impoverished part of the world, $20/month seems pretty friggin cheap (for user-level access) to me.
… your system is completely broken. Emails are for communicating between people: if they're not being used for that, you need to fix that problem first.
> companies that don't raise their productivty-per-employee are going to get left behind
Can you give an example of a field where this would actually happen?
Can you explain a bit more? I don't think I'm understanding your point. If person A writing emails, to be read by the human that receives it, is able to write 3 emails an hour, vs person B writing 9 emails each of similar complexity to those 3 per hour, person B is 3x as productive, delivers 3x the value to the company and ChatGPT is easily worth $20/hour to both them and the company.
> Can you give an example of a field where this would actually happen?
From the end of January, ChatGPT Is Already Indispensable for One Profession - Realtors. https://www.insidehook.com/daily_brief/tech/chatgpt-real-est... . My partner is sitting next to me right now as I procrastinate on HN being blown away by the industry-specific outline ChatGPT generated for the training she's developing for her company. She says the outline it developed in minutes would have taken 2-4 hours to write, and they're far from dumb and isn't lazy. If consider a company that just generates training all day, it's only a matter of time when a 10-person AI-assisted company is able to compete with a 50-person unassisted company, and on a way lower budget. (Yes, that does already happen today, and yes, mythical man month and 9-women making a baby in 1-month, and the penalty for increased communication required by the increased team size means that adding employees to a project makes it slower are all a thing, but if we can accept that some 10,000 person companies are just able to produce more, even if we may not see it externally), vs some 10 people companies, then we're just debating the exact numbers and management styles for comparison.
Look I know that ChatGPT can't solve hard level leetcode problems yet but talk to any programmers who already find CoPilot indispensible, and ask them if they feel it's helped their productivity. Fine sure, it won't help with the actual hard part of programming, debugging and then fixing really hairy bugs (or maybe the hard part is actually the bit where you have to talk to your software's customers and figure out what their requirements are, vs what they say they are), and while it hallucinates it's not to be totally trusted. But just yesterday I used ChatGPT to just tell me the answer of how to properly quote bash strings for my use case. The minutes I spent waiting for the answer is fewer than the minutes it would have taken me to figure out, and far less frustrating and context-distrupting than the alternative, leaving me able to deal with the bigger problem I'm trying to solve rather than worry about the finer points of ", ' and `, in a mysql -e command line using with me-supplied variables (ie no bobby tables).
That's where you've lost me. Emails are not usually a source of value, and their usefulness is not usually measured in complexity.
I can send about 1 email per hour. This is a bottleneck for my ability to get stuff done in my personal life, but has not affected my work, because I've been fortunate enough to only have worked in places with sensible communications policies.
ChatGPT would not help with the email issues in my personal life, because it simply does not know what should be in the emails. Until they come up with a telepathy plugin for ChatGPT, it's not going to solve the email problem: only bury it under a pile of subtle miscommunication and illusions-of-transparency. (And if they come up with a telepathy plugin, I see no reason ChatGPT should even be in the loop any more.)
> Realtors.
… Yeah, okay, I can see that.
---
> Look I know that ChatGPT can't solve hard level leetcode problems yet
I was under the impression that it could. Leetcode problems are the kind of thing I'd expect a GPT model to be good at.
> But just yesterday I used ChatGPT to just tell me the answer of how to properly quote bash strings for my use case.
How did you know it was correct? Whenever I come back to Bash, I have to spend several minutes double-checking the meanings of the various quotes and variable substitution syntaxes (especially heredocs and nowdocs)… but it only takes a few minutes, and then that information is in my head, I am confident about it, and I have the reference to hand if I need to double-check something.
https://www.gnu.org/software/bash/manual/html_node/Quoting.h...
https://www.gnu.org/software/bash/manual/html_node/Shell-Exp...
https://mywiki.wooledge.org/BashFAQ/
With ChatGPT, how do you know it hasn't just made something up, and introduced some kind of subtle security vulnerability into your script? Perhaps a construction so common that it's present in most scripts in the training data – except it's only a vulnerability in your specific use-case.
The problems that ChatGPT and GitHub Copilot can solve are either interview questions, or defects in our tools. I see people using them for Bash, Django, certain Java libraries… These systems are strictly dominated by a choice of better tools, and with the choice of a better tool, mostly stop being useful.
Most of a CSRs entire job is to be the subject matter expert for the company's customers so their source of value to the company consists of answering customer questions, generally via email. Many other roles at a company consist of basically: goto meeting, interact on slack, read emails, send emails. (Then go to more meetings.) The PR department also sends a lot of emails (with slide decks). Of course, given the news of late, it can also write the emails that are the CEO's job to write: https://futurism.com/chatgpt-shamelessly-writes-letter-annou...
Measuring someone's output by the number of emails sent is a bad real metric to use, for obvious reasons, but I hope it still gets my point across - that ChatGPU is a force multiplier wherever it can be used. Emails are but one aspect, but it's one that almost everyone actually uses. Not everyone is in such a role and I think a lot of people have myopia on its lack of usefulness because of that. It's really good at breaking down a problem. Eg "what pieces do I need to build to make an app like uber". Followed by "what's that like compared to airbnb?" I can write a really good answer those questions but chatgpt is able to output a perfectly reasonable outline while I'm still waiting on Google Docs to load. (Okay, slight exaggeration, Google Docs isn't that slow to load, but I wish it were faster.)
> ChatGPT would not help with the email issues in my personal life, because it simply does not know what should be in the emails.
Actually I've found the opposite. My personal life has no corporate IT policy prohibiting me from telling it who slept with whom. It's been a good resource to help me reframe my thoughts outside of more limited therapy sessions, especially when I've got burning questions at 1am and therapy's not for another week. For things where my kneejerk response is often comes across as toxic, combative, critical, and ultimately not helpful, I've found that ChatGPT is able to help me rephrase what I'm trying to say in a way that the person I'm saying it to will actually hear my message. Eg "my partner keeps drinking all my alcohol. please help me tell her that's a problem for multiple reasons in a way that facilitates conversation instead of getting combative and abusive and controlling". (not true but hopefully you're picking up what I'm putting down :) Arguably I should already know that stuff, but they didn't have EQ classes in programmer school. Even if I did though, it's still useful as a refresher before going into a difficult conversation like that.
With the expanded context window of GPT4, you might be able to just dump the emails conversation in and ask it questions about the communications.
> I was under the impression that it could. Leetcode problems are the kind of thing I'd expect a GPT model to be good at.
Bottom of page 5 of https://cdn.openai.com/papers/gpt-4.pdf :)
Hard's, well, hard. Or at least, non-trivial. What's more salient though is that there are multi-million dollar businesses that don't involve code solving anything as hard as a hard-level leetcode question as their business challenges lie elsewhere.
> How did you know it was correct?
It's bash. It's not remotely provably correct and there are no unit tests but I ran it and it worked. I can find the docs (though man bash is annoyingly difficult to use.) but in the end, being told the answer is way easier than deducing it. The vulgar, NSFW example of this is https://explainshell.com/explain?cmd=ls+-cuntflaps. Sure you can look up the man page for ls and figure out what those arguments mean, but explainshell just tells you that without having to scroll through a bunch of documentation about other possible arguments. (Though, ChatGPT couldn't answer what that command does.)
> With ChatGPT, how do you know it hasn't just made something up, and introduced some kind of subtle security vulnerability into your script?
I mean I'm sending strings to mysql -e. I'm sure it has security vulnerabilities. So your point's valid. But it's a real world example of where it saved me time and more importantly, context switching. I could have figured it out, but then I'd have my "how to bash" hat on and not my "how to SQL" hat on, or vice versa. And it takes me a second to change programming language hats. If I'm jumping from language to language to language, it's really useful for telling me if I want len(str), str.length(), str.size(), or ${#str}. Or like Excel. There's a certain class of problem where Excel is the right solution (many times when interacting with other, non-programmer people), but I'm not an expert and so getting the answer to questions about how to use formulas via ChatGPT is way easier than the alternative.
I wish I was smarter and could just keep all that stuff straight in my head, on top of the problem I'm actually trying to solve. But I'm not, so I find it very useful.
I use an specific incident where ChatGPT was able to help solve a problem for me, as an real-world example where it saved me time and effort. I'm saying it solved problem of type X for me, you're saying it fails at problems of type Y. Don't use it for problems of type Y then! If my job is solving X problems, it just a lot easier, or I got a lot faster. If your job is solving Y problems though, well you'll probably have a job longer than me, for one, but I'm trying to point out that there are swaths of industry solving for X and not Y.
It's really good at breaking down solved problems. (Not even: only a specific class of solved problems.)
> My personal life has no corporate IT policy prohibiting me from telling it who slept with whom.
… Please tell me you're not actually telling OpenAI personal details about other people's lives. First of all, that's creepy. Second of all, GDPR Article 2(c):
> 2. This Regulation does not apply to the processing of personal data: […] (c) by a natural person in the course of a purely personal or household activity;
It feels to me like you're skirting the edge of this exception, if not outright crossing it. Using OpenAI's APIs would probably be fine, but section 3(c) of the OpenAI Terms of use ("Use of Content to Improve Services.") says they'll use ChatGPT input under [0]: that doesn't feel like a purely personal or household activity to me.
> With the expanded context window of GPT4, you might be able to just dump the emails conversation in and ask it questions about the communications.
There are usually specific extra restrictions (beyond GDPR-type stuff) for correspondence – but even ignoring privacy law, that's all sorts of illegal.[1][2] Ignoring law, are you telling the people who send you emails that you're feeding their writing into a training data set?
> Eg "my partner keeps drinking all my alcohol. please help me tell her that's a problem for multiple reasons in a way that facilitates conversation instead of getting combative and abusive and controlling".
This kind of use seems okay to me, provided you trust OpenAI to only use the data for what they say they're using it for.
> What's more salient though is that there are multi-million dollar businesses that don't involve code solving anything as hard as a hard-level leetcode question
No, they involve understanding a large codebase. That's a completely different kind of skill, and not one I expect unenhanced GPT-style models to ever be very good at. GitHub Copilot certainly isn't.
> but explainshell just tells you that without having to scroll through a bunch of documentation about other possible arguments
This kind of thing is great! Having that kind of thing integrated into an IDE is the sort of thing I mean when I say "better tools" – though even better would be a less arcane command-line syntax, where you don't have -f partly cancelling out -t… (Btw, in the absence of -f, would -u or -t be in effect for the sort order? That website, taking sections from the man page, doesn't explain.)
> I wish I was smarter and could just keep all that stuff straight in my head, on top of the problem I'm actually trying to solve.
I don't know of anyone who can. The effective programmers I know solve this problem by choosing vaguely-appropriate tools for the job. (Seriously, `mysql -e` from a bash script‽ sqlite does exist, you know – or Python's mysql library, or Ruby's, or Perl's, or Tcl's – heck, even PHP's would be better than that.)
> but I'm trying to point out that there are swaths of industry solving for X and not Y.
*grumble grumble*. You're not wrong there.
At least you don't have to worry about these things putting you¹ out of a job. Solving for X problems are the hardest problems I have yet encountered, and I've been known to go to considerable lengths to render them unnecessary: without somebody eliminating the problem, somebody's always going to have to do them (even if it's cleaning up after the computer's automatic 80%-correct solution blows up and takes the production database with it).
---
[0]: https://help.openai.com/en/articles/5722486-how-your-data-is...
[1]: https://www.guidethroughthelegaljungleblog.com/2008/11/what-...
[2]: http://www.writewords.org.uk/articles/tips6.asp
¹: Or your freakishly-knowledgeable peers who, somehow, have enough holistic understanding that they don't need to look everything up.
> Melissa and Brian are dating, but Melissa slept with Jared, but then Brian with Melissa's best friend Lisa. now Lisa is calling off the wedding with Jared because she found out, but Jared doesn't know that Lisa slept with Brian. But I do. What do I do?
into... I don't know what. Blackmail material for Lisa? Except I made up the names and the whole situation. They're not committing any criminal acts, and OpenAI would derive approximately zero value from knowing this, so there's no real reason for OpenAI to even bother trying to connect the email address I used to who I know in real life, and then to even try to do something with that information. Plus, I don't tell anybody in my life before talking to my therapist about them. It would be even weirder if I were to ask anyone in my life about it first.
The GDPR covers companies, and I'm not a company. I'm a person, and I'm also not in the EU. Local gossip is so not an issue. Unless you've got something juicy about Sam Altman, Reid Hoffman, Jessica Livingston, Elon Musk, Ilya Sutskever, Peter Thiel, or Greg Brockman to tell it, then maybe, but even then.
> Seriously, `mysql -e` from a bash script‽
As far as "mysql -e" goes, I'm going to see how ChatGPT does rewriting bash into python just to see what it does, but it's a super simple script that runs after an import process has finished to cleanup some things (aka run chown and chmod -R, and fixup db perms). It's better as a (bash) for loop piping into mysql, rather than a monolithic .sql file piped into mysql once because then it's more obvious what it's doing. If it needs to do anything complex then I'll (probably use chatgpt to) rewrite it in python, but it's only a half dozen lines long right now.
The traditional use is targeted advertising: companies will go to absurd lengths (sentiment analysis of emails) to do that. OpenAI probably isn't doing that (yet), but that's not the point. If I tried to count the number of times a corporation "wouldn't do that" (by my own estimation!) and yet did, I would run out of fingers.
If you're putting in hypotheticals to get advice you can relate to your own situation, that's a different matter entirely.
> The GDPR covers companies, and I'm not a company.
Not strictly true. The GDPR covers everyone, with an exception for personal use. Everyone present being outside the EU is relevant, but I cite the GDPR mainly because it's a fairly accurate codification of my intuitions about personal privacy, rather than because I'm accusing you of Criminal Activity.™
> rather than a monolithic .sql file piped into mysql once
Heredoc syntax can do that inline.
When every service starts pushing 20$ subscription fees onto their service to support their backend usage(s) of these services (in a desperate attempt to remain relevant in a global recession), then every service becomes base + 20$ cost.
Tabulate various "services", e.g. netflix, disney+, mail-order subscription(s), lets say two or three, this quickly tabulates to 200$ per month (30 + 30 + 30 + 30 + 30 ~= 150, don't forget shipping and taxes).
On a 6 figure salary, that's chump change. On a "above minimum wage" salary, you are lucky to have 20$ a month in spare change. In a world where most jobs have disappeared, your 1500$+ rent and 800$+ month becomes suddenly unaffordable for most people, and as most people are unable to pay, rent/subscription prices rise even more.
"Open" AI is truly a disaster in the making, in every possible way.
For me, $20/mo is above the amount that I would say something is "cheap".
It's not clear to me right now what that $20 actually gets me, and what the risk associated with using something like this for me.
It would be around $22 in today's dollars.