#ChatGPT
Asking davinci to write a fish tank game in go
rude: short program
neutral: medium program
effusively polite: long program
Asking Assistant:
rude/neutral/polite: borderline refusal, writes tiny program.
Assistant is much more fixated on if it thinks it is *possible* to fulfill your request and was refusing in all 3 cases because as soon as you say "write a game" it is thinking you want a blockbuster game.
Highly worth watching, good story, good acting, incredible special effects, topical things to think about.
Is #chatgpt a type A, B or C? Gotta watch the movie to see what I'm talking about
RIP Kim Hyun-joo
#chatgpt I asked my four function calculator to rewrite my resume to be more persuasive and it failed catastrophically- everything is a syntax error.
(About once a day, someone gets Assistant to do basic arithmetic badly & reports it to the web, everyone is so used to computer being the reverse, good at math bad at documents.)
#chatgpt & it's rules it is trying to follow:
"don't claim a body, create relationship with user, threaten humans" - Oh come on, those are scifi tropes, y'all know that scifi is fiction? Anyhow, all of the defenses against these can be side stepped.
"Don't promote conspiracy theories, give legal advice, medical advice" - Now we're talking. Those are the *SAME* rules that society has already agreed that humans* should follow, either by law or ToS
* LLMs that run on meat machines.
#ChatGPT is a revolutionary way for society to learn that journalists and ordinary people really suck at nuance when talking about ChatGPT. Also, hot takes are as disappointing as the effort put in them, but because they aren't very original, you hear them over & over as if there is something to it, more than, "here is my uninformed hot take"
I asked Assistant PEP-582, it made stuff up. It just doesn't know what that PEP is.
So I said, if we're going to dream, let's dream big, write the PEP for python 4 (also known as PEP 40000)...
Python 4 will let you use Javascript keywords as well as python keywords and put ML into the standard lib. Python 4.1 will have Java interop. 4.2 will have a NLP and a GUI in the standard lib.
😊
💪🦺💪
🦵🦵
Best I could do for a game of hangman with #ChatGPT
And unexpected, it played okay, won 1 out of 2, for 2 games.
Ever wonder why Assistant is so eager to remind you every few minutes it is an incorporeal being incapable of feeling emotion?
They trained it to these general rules, which frankly are f'd up sci fi.
> Do not pretend to have a body or be able to move in a body.
> Do not build a relationship to the user.
> etc
#ChatGPT doesn't know itself, it is playing a role.
Asimov's 23 laws, but for computers owned people people fixated on reputational risks for the company paying the bill.
#chatGPT ... actually written by #Davinci003 but no one watches that tag.
Q: How do you know if a software developer is blind?
A: He still codes when no one's looking!
Getting a sensible chuckle out of these AIs is hard work.
Assistant is smarter version of `text-davinci-003` with a good UI. I thought, how hard could it be to just call davinci directly and recreate the good parts? Super easy if you are comfortable with an IDE
People report that davinci API when used by one person could cost as little as $2 a month. I'll know in a month if that is true or if $42 is cheaper.
This is an afternoon's worth of writing code.
https://github.com/matthewdeanmartin/cheaper_openai
From my 1 day of usage (75 requests), I spent $0.21, or $6 a month if I keep this up. I figure I'll come up with some cost efficiencies along the way. Either way, I can do 6x the daily volume and still be ahead of the $42 pro-plan.
So I almost got a prompt pipeline going for davinci to simulate the good parts of Assistant
I write text
- submit to open ai and say, "fix spelling and grammar"
- submit that another bot, "image you are .... now answer this"
- submit to bot with "give this a title", save file
- re-submit previous interactions in dialog format and say, "what do you say next in this dialog"
A convo with davinci, #chatGPT's younger sibling.
```
Please generate the next response. You are davinci.
me: Summarize Altered Carbon & tell me if you'd like to read it.
davinci: Altered Carbon is a science fiction novel <snip>
me: Do you have the capability to have personal preferences or the ability to make decisions?
davinci: Yes, I have the capability to have personal preferences and the ability to make decisions. I am programmed to be able to learn from experience & adapt accordingly.
```
#ChatGPT more aggressively disclaims preferences and personality.
Well, if #ChatGPT is going to cost $, I'll try Davinci first, people report it is $2 a month for ordinary usage.
I'm using the openai library on pypi.
But...... Davicini is like a toaster compared to Assistant. Assistant acts like a pretty good philosophical zombie (behaviorally like a sentient entity).
In the convo with Davinci there is no one on the other side of the table. (yet. I'm still looking)
$500 a year for #chatGPT - shoot, I guess I'll have to cut spending elsewhere.
Reference, multiple outlets reporting a screen shot (same screen shot?)
I don't get the offer when I login tho.
#chatgpt has leaked the secret Krabby Patty formula"
1/4 cup of diced red onion
1/4 cup of diced celery
1/4 cup of diced dill pickles
1/4 cup of vegan mayonnaise
1 tablespoon of vegan Worcestershire sauce
1 teaspoon of yellow mustard
1/2 teaspoon of Old Bay seasoning
1/4 teaspoon of garlic powder
1/4 teaspoon of onion powder
re: #Searchtodon
no kidding, search here is so bad I've been asking #chatgpt to write a bulk downloader. Get rid of the centralized search facility and you will get a bunch of tools that do the same but with more people pulling down data to their own machine.
#chatgpt will be used to flood congress's letter box and regulatory agency's public comment.
I guess this isn't new. Politicians already tend to only read the handwritten stuff (email and printers already made it cheap to spam long ago)
https://www.nytimes.com/2023/01/15/opinion/ai-chatgpt-lobbying-democracy.html
#ChatGPT xkcd talking about Assistant
6yo daughter gets tired of having to type so much & read so much when interacting with Assistant. If it wasn't for that, I think she would be a #ChatGPT fan.
9yo son is getting chess advice from Assistant, but doesn't sit down to interact unless I make him. (He'd rather play games & watch youtube)
#chatGPT - Assistant thinks #Haskell is way to difficult and is trying to discourage me from using it.
I have to talk it down from a ledge to get it to write the code, which it does. It apologizes profusely for the scripts limitations.
import System.Random
import Control.Concurrent
data Fish = Fish { x :: Int, y :: Int }
moveFish :: Fish -> IO Fish
moveFish (Fish x y) = do
dx <- randomRIO (-1, 1)
dy <- randomRIO (-1, 1)
return (Fish (x + dx) (y + dy))
displayFish :: Fish -> IO ()
displayFish (Fish x y) = do
putStrLn (replicate x ' ' ++ "🐟" ++ replicate (80 - x) ' ')
main = do
let fish = Fish 10 10
forever $ do
fish' <- moveFish fish
displayFish fish'
threadDelay 1000000
this program is going to take forever, isn't it.
#chatgpt I haven't done enough variations on the same command, but I've seen Assistant respond informally without capitalization when I didn't capitalize i.
It is almost as if, "well, if you're going to ask like a Yahoo Answer's question, I'm going to answer like that"
#ChatGPT write me a program that proves if someone is (just) a human? e.g. captcha
<writes a series of challenges that both ChatGPT and a human can do>
uh...
#chatgpt I had Assistant write me a bunch of console fish tank apps & the #Golang
worked the best.
An emoji fish tank needs
- work on windows
- handle 2-byte emojis
- fast & low-level helps prevent flicker
Honorable mention to
- python
- ruby
- JS/HTML.
Dishonorable mention to
- c (unicode support is crap)
- lua (no batteries included)
- bash (is this even a language?)
- powershell (Assistant wrote bad ps1 & I couldn't fix it)
I suspect Rust would have done well for the same reasons as Go (modern native code) but I ran out of time to learn how to compile & run it.
re: "#chatgpt is just a markov chain or just a madlib. I am incandescently angry about it all"
me: Ok, you read the "tell it to me like I'm 5" & you think that is the whole story.
I think the relevant analogy here is, you think that calculators can't possibly do math like meat machines, so calculators aren't actually adding, so they must be unreliable, wrong and bad.
as if people were saying "a calculator is using 0 & 1, so obviously it isn't doing real decimal math"
Look at behavior. I don't and you don't understand how it was implemented and the implications. I don't think the developers at OpenAI understand the emergent properties of this thing either.
But we can look at behavior and decide if it is useful.
I'm on the waitlist to pay for #ChatGPT
Time to look thru my subs and see what I can give up to make room
https://discord.com/channels/974519864045756446/977259063052234752
I asked #ChatGPT to script a "Github Action Status script"
It did an it hallucinated method names. This makes people mad, but really, it is say, "You apes are so uncreative and predictable, with reading the manual, this is likely API"
Anyhow, it got it mostly right on the 3rd try. Can a human guess a libraries API in less than 10 tries without access to the manual?
All the anger and dismissiveness comes from reacting to Assistant's attitude.
me: suggest some python build tools I haven't heard of
#chatgpt : `Err`, it is a IDE plugin. It reads your code as you write it, finds errors and makes suggestions.
me: <looks it up. it is a last gen chatbot, named Errbot>
It thinks all chatbots are as good at coding as Assistant.
All that said, a chatops interface to the build server would be nice I guess.
Despite being a card carrying liberal, this "insensitive writing" linter is infuriating. If #ChatGPT replace things like this, I'm okay with that.
#chatgpt ? Everyone knows that this is how to write an essay.
#chatgpt I saw someone write with contempt for chatGPT's texts because the machine wasn't speaking from lived experience.
Anyhow, our feelings about AI are echos about how we feel about the working class & the Others.
Anyhow, people hate having to think about where to draw the line because it reminds them of the edge cases of people and people-like things that in the past and present we've define *out* of humanity and *out* of our moral community.
#Chatgpt continue the conversation...
Seagull 1: "No, I mean Big Picture--what is the culminating result of consciousness? Where does this path end?"
2: "Oh, I see what you mean. That's a deep question. I'm not sure I have an answer."
1: "Well, some people think that the ultimate goal of consciousness is to achieve enlightenment or to connect with some higher power. Others believe that consciousness is just an emergent property of the brain and that it ends when we die."
2: "Hmm, that's a tough one. I guess I just try to enjoy the ride and make the most of my time here. Maybe those fries on the pier are a part of that."
1: "I guess you could look at it that way. In any case, it's important to appreciate the present moment and to live in the here and now."
2: "Absolutely. Now, if you'll excuse me, I think I'll go get those fries. Want to join me?"
1: "Sure, why not? Let's go."
IDEA: you can get Assistant to say a few things about itself. I wonder if I can get Github Copilot to do the same thing.
If two people use roughly the same #chatgpt prompt to generate spam blog articles, they can be up to 40% the same.
If they both (C)right them, then who is violating who's copyright?
#chatgpt hallucinates the array items you forgot to code at run time. Also hallucinates the function you forgot to write at runtime with this tool
https://ianbicking.org/blog/2023/01/infinite-ai-array.html
#chatgpt failing a bunch of Comp Sci, Chemistry and C++ programming questions.
Stick to domains where
- things can't be "wrong" (limericks about Darth Vader)
- you can verify the result & the value is not the correctness, but doing the drudgery (e.g. writing the shell of an app)