The author seems unaware or at least uninterested in the fact that writing style can have a dramatic impact on output. It's unclear if they wish to not be emailed "by ChatGPT" or if they wish to not have to sift through unnecessary and flowery pose.
I'll spare you the details, but a follow up "That's great. However, please rewrite it in the style of Ernest Hemingway." delivers concise, yet obviously Hemingway-esque emails. Example here: https://pastebin.com/rrkCMd8c It works much better in a two-step process. If "Write this email in the style of Ernest Hemingway" is affixed to the original prompt, the model will generate prose at length, defeating the purpose of being concise.
"That's great. However, please rewrite it in the concise style of Paul Graham," of course, works even better.
For me, it's a matter of decency. If you're sending me an email and you're expecting that I'll read it, you're asking me to invest my time (presumably for your benefit). But, you're unwilling to make the same investment with your time by using a tool to simulate a human connection. Having excessively long, unedited prose adds insult to injury since you've doubled down on the decision to spend someone else's time as long as it saves yours.
I love these multi-step prompt 'hacks'. They very much take advantage of the fact that this is still 'just' a model predicting the next token.
Asking a model to write an email as if it were written by Hemingway requires the model to generate a probability distribution based on the context of an email it needs to write + the style it needs to write it.
In the second approach, you've changed the model weights/inputs by including the email in the context window, so the task of predicting the next token is fundamentally different (and possibly easier) for the model.
It's also why models are sometimes bad at answering a factual question, but good at judging whether their own answer is correct.
It took more work to write the prompt that it would to write the script email.
“Please write a brief email from an employee to a boss giving an update on the api. The email should include 1. The /customers and /address endpoints are complete, but we're still waiting on the architecture team to finish the /orders spec. I'll also be taking Friday off.”
And then you had to adjust it. Just include the information you out into the prompt. If chatgpt can understand it, so can I.
I can’t wait to live in a world where someone enters bullet points into chatgpt to generate an email. Then I have to run the email through chatgpt to transform it back into bullet points.
The point isn't to save time or energy - the point psychological displacement.
Whether we want to admit it or not, not everyone has done the work to handle critique. Some people rely on ChatGPT as a digital scapegoat[1]. Rather than subject their own abilities and decisions to critique, they can launder them through ChatGPT and the psychological distance it affords, allows them to avoid feelings of anxiety - they can blame any negative response on ChatGPT.
Never under estimate the amount of time an energy people will spend on coping mechanisms.
The prompt is a much better email than the stuff it produced:
> The /customers and /address endpoints are complete, but we're still waiting on the architecture team to finish the /orders spec. I'll also be taking Friday off.
vs. garbage about gazelles, and the fact that telephones exist.
> The human race has one really effective weapon, and that is laughter. The artificial race has one really ineffective weapon, and that is imitation. - Mark Twain (paraphrased)
> Please if you want to send me a message and feel compelled to use GPT, please just send me whatever you wrote in your prompt instead. I promise I will still read it!
Best bit of the article.
Question for someone who knows more about this stuff. How likely is it to get the same response to the same prompt with gpt ? Does it have some kind of random seed applied behind the scenes?
There is a parameter called "temperature" that determines randomness. I think there is never a guarantee of the same response, but generally with a low temperature (something like 0.0, the lowest but still not deterministic) I get nearly identical responses, if not actually identical. When temperature is increased though to e.g. default of 1.0 or max of 2.0, the same prompt will give very different responses. The higher the temperature the more creativity that is injected because of the randomness, but also the less determinism.
Honestly, as someone who does this, Bard seems to be much better at writing natural emails. It's just easier to dump some thoughts in to bard and ask it to redraft, then fluff it a little to sound like me.
It's not a random seed, but it will always be slightly different. It's similar to "prompting" a human... you'll likely get the same general result if you ask the same person the same question a dozen times, but it will have enough variability that it won't be identical.
Pretty sure this is entirely dependent upon the "temperature" parameter used when inferring. The higher the temperature the more uniform the response will be.
It's certainly not in any way whatsoever similar to "prompting" a human. And yes there are definitely random processes involved. For an extremely objective prompt you may get very similar results but for general stuff like "write an e-mail about this or that" it will be significantly different each time.
"Please if you want to send me a message and feel compelled to use GPT, please just send me whatever you wrote in your prompt instead. I promise I will still read it!"
Maybe this is a personal experience. As a non-native English speaker who works in academia. I learned formal English since I was a child. ChatGPT/bard will produce a much quicker formal professional email than I will do. But if I took my time to check the email, I would produce a similar email, which is actually how I normally write if I'm sending an email to a professor or someone else in the field. AI like bard or ChatGPT will be a huge time saver in such cases. However, living in the US for half a decade changed a lot.
From what I've seen so far, the best writers (and, I suppose, the best _editors_) tend to get the best results from ChatGPT. I have seen some examples of people whose writing I already liked drive it in ways I never imagined; but like the author suggest "normies" tend to get piles of drivel.
- "I love that line a little bit, hmm actually I read it a bit more and it started to stink"
- "This line feels a little too Vogue"
- "Can we make this a little more Arnold in Kindergarten Cop"
And it will distill that into something distinctly actionable. You have an interface to communicate your edits in the lowest friction way ever exposed by mankind and people are still trying to enter a prompt and Ctrl+C Ctrl+V.
My experience, both writing and coding non-boilerplate so far, has been that this back and forth almost always takes longer than doing the work myself.
There are situations that's not true, namely information retrieval, but I've yet to find it highly practical beyond being a fun experience.
I've gotten several emails from a recruiter that address me as my last name, all lower caps (Hi swanson, this job..."). It was for an AI company of course.
This is just the beginning. Ever since late last year, I estimate that around 25-30% of everyone I ever talked to via some online session where we had to share the screen for something or another, had a ChatGPT tab open. For students this jumps to around 50%.
It's not possible. The current AI detectors are extremely unreliable, whatever people selling them might tell you. You'd probably get fewer false positives by filtering the people who send mail that you think is AI generated.
There is no need to use ChatGPT for that. I suppose running the emails through LanguageTool (or, if you don't care about privacy, Grammarly) and counting the suggestions to fix "too wordy sentences" would be a good approximation.
Readit News
I'll spare you the details, but a follow up "That's great. However, please rewrite it in the style of Ernest Hemingway." delivers concise, yet obviously Hemingway-esque emails. Example here: https://pastebin.com/rrkCMd8c It works much better in a two-step process. If "Write this email in the style of Ernest Hemingway" is affixed to the original prompt, the model will generate prose at length, defeating the purpose of being concise.
"That's great. However, please rewrite it in the concise style of Paul Graham," of course, works even better.
Asking a model to write an email as if it were written by Hemingway requires the model to generate a probability distribution based on the context of an email it needs to write + the style it needs to write it.
In the second approach, you've changed the model weights/inputs by including the email in the context window, so the task of predicting the next token is fundamentally different (and possibly easier) for the model.
It's also why models are sometimes bad at answering a factual question, but good at judging whether their own answer is correct.
“Please write a brief email from an employee to a boss giving an update on the api. The email should include 1. The /customers and /address endpoints are complete, but we're still waiting on the architecture team to finish the /orders spec. I'll also be taking Friday off.”
And then you had to adjust it. Just include the information you out into the prompt. If chatgpt can understand it, so can I.
I can’t wait to live in a world where someone enters bullet points into chatgpt to generate an email. Then I have to run the email through chatgpt to transform it back into bullet points.
Whether we want to admit it or not, not everyone has done the work to handle critique. Some people rely on ChatGPT as a digital scapegoat[1]. Rather than subject their own abilities and decisions to critique, they can launder them through ChatGPT and the psychological distance it affords, allows them to avoid feelings of anxiety - they can blame any negative response on ChatGPT.
Never under estimate the amount of time an energy people will spend on coping mechanisms.
1. https://www.psychologytoday.com/us/blog/hide-and-seek/201312...
> The /customers and /address endpoints are complete, but we're still waiting on the architecture team to finish the /orders spec. I'll also be taking Friday off.
vs. garbage about gazelles, and the fact that telephones exist.
Turns into:
> Come Friday, I won't be in. Just one day. Things will continue, though. Any emergencies, I'm a call away.
Off is off, don't call me. Must have been trained on a lot of linkedin posts to volunteer your personal time like this.
Although, I don't know how he would feel about being mimic'ed million-fold.
> The /customers and /address endpoints: born, alive, breathing. The /orders endpoint: still trapped in the womb of the architecture team's thoughts.
> The human race has one really effective weapon, and that is laughter. The artificial race has one really ineffective weapon, and that is imitation. - Mark Twain (paraphrased)
Best bit of the article.
Question for someone who knows more about this stuff. How likely is it to get the same response to the same prompt with gpt ? Does it have some kind of random seed applied behind the scenes?
-edit- Thank you for the responses. TIL.
You might read it, but you won't answer the same way as you do when "choose a nice wording." was added to the prompt.
Seems like we should all just be communicating via bullet points grug or Kevin Malone style.
"Tell the person that he's being an idiot asking me to turn on my cameras during the meetings but be subtle about it, I don't want to offend them."
That'll work.
We can dream.
Dead Comment
Deleted Comment
- "I really really hate this line"
- "I love that line a little bit, hmm actually I read it a bit more and it started to stink"
- "This line feels a little too Vogue"
- "Can we make this a little more Arnold in Kindergarten Cop"
And it will distill that into something distinctly actionable. You have an interface to communicate your edits in the lowest friction way ever exposed by mankind and people are still trying to enter a prompt and Ctrl+C Ctrl+V.
There are situations that's not true, namely information retrieval, but I've yet to find it highly practical beyond being a fun experience.