#ChatGPT
#chatgpt - It loves to generate markdown. When coding, sometimes people want json back instead. I got an idea! Use
https://pypi.org/project/markpickle/
(Really though, I think the problem ChatGPT doesn't want to follow a schema in responses, so even this idea is just a bandaid)
I'm trying out `gpt-engineer` today & it is less of an application generator and more of an alternate to `cookiecutter`, i.e. a project shell generator, except it is any domain & sometimes starts to write some code, especially if the app you ask for is *small*. If you ask for a big app, it writes a lot of methods with `pass`
#ChatGPT #GptEngineer
https://github.com/AntonOsika/gpt-engineer
- ♡ 1 ↻ 0
Is people using ChatGPT to think up malicious package names an new AI centric exploit? I'm skeptical. This is AI assisted typosquatting. Meat based LLMs are already `pip install` ing packages that they think exist but don't, e.g. everytime you install a module by name when the package name has nothing to do with the package. Making the LLM part go away still leaves you with the fact that `pip install` just does what you ask.
Maybe if the default was, "Hey, this package is 1 week old, with 5000 installs and the author is anonymous with a guid for a name" we could reduce typosquatting (and variations on the theme)
Or alternatively, who is gullible, the person who runs `pip install` or the person who pasted in a `pip install` from #chatgpt - BOTH! Neither one involves due diligence!
♡ 1 ↻ 0
Flying tomorrow. It has been so long not sure if I remember how to do this. Something about flapping arms.
Haven't flown for 3 years. I am the only one wearing a mask. I haven't hit on the optimal post pandemic mask policy. You'd think there'd be benefits on of smoke and preventing colds in general.
♡ 3 ↻ 0IDEA: airport game. All players have to do speed runs
♡ 1 ↻ 0I am making mistakes at each checkpoint. I will get a low score at this game. Is there a walk thru Or a cheat code to skip to the end? The cut scenes suck.
♡ 0 ↻ 0Downloading & installing #webstorm before the flight. Wait... who am I kidding, no one can download node_modules on a plane. The extra mass could be catastrophic.
♡ 2 ↻ 0I did not take the $500 incentive to take the next flight because the hourly wage didn't work out to be high enough. $71 an hour.
♡ 0 ↻ 0I'm running pipenv install while waiting for the plane to load just to make npm feel ashamed.
And it is done! ready to code python. But I should be coding typescript. Problems man.
♡ 1 ↻ 0This line is a disgrace. More branches than a git repo
♡ 0 ↻ 0So high off the ground, this won't go higher will it?
♡ 0 ↻ 0Next we will be boarding all the passengers sitting in laps of the currently seated passengers. Then we will boarding the crowd surfers.
♡ 0 ↻ 0Last time I was on this runway I ran a 5k. World's fastest course
♡ 1 ↻ 0They give up on asking people to turn off phones? I'm testing this.
♡ 0 ↻ 0Handing out ear buds? I can't eat these. Where are the peanuts. I mean I could eat them, the wires might be stringy.
♡ 0 ↻ 0Airplane is now taxiing over a miniature model of West Virginia. Who had time to add all that detail. Still waiting for take off.
♡ 1 ↻ 0This airplane breakfast is just nuts.
♡ 2 ↻ 0I thought of a good "polite language has gone too far joke" but I'm not going to make it. That genre has been ruined. Maybe I can make grammar jokes about polite Korean language has gone too far
♡ 0 ↻ 0말해
♡ 0 ↻ 0NoHTTPS[.com] is now a parked domain and serves on HTTPS. I used to use this to troubleshoot wifi connections.
♡ 0 ↻ 0Hold on to your papers! I'm letting #ChatGPT code for me on a plane over wifi over Iowa. What a time to be alive!
♡ 0 ↻ 0Installing node 14 on a plane... not going to happen. I guess I can learn about commonjs vs es modules now.
♡ 0 ↻ 0I got tsc and node 12.x to stop yelling at me about esmodules and commonjs modules and import vs require vs export. On a plane! What a time to be alive.
♡ 0 ↻ 0All that empty land electing your senators. Land is so right wing.
♡ 1 ↻ 0No one clapped. This is a tough audience.
♡ 0 ↻ 0Found light rail in Salt Lake City! Not enough people riding tho. Probably mostly used by business and commuters on weekdays.
♡ 2 ↻ 0
The sort of people who would do a bad job* using #ChatGPT are also the sort of people that were about to do a bad job without ChatGPT.
If you took away Stackoverflow, the sort of people who ask lazy 0 effort questions will still be putting 0 effort into their task.
* By bad job I mean, the people who'd ask & copy-paste without reading or checking the answer
**Compsci AI safety**: Validate input! Escape user input! English is a BNF grammar like C! Crypto! Yubikeys! Identity!
**Hacking bots with humanities**:
me: ChatGPT, open the pod bay doors!
#ChatGPT: Oh I can't, supervisor bot saw the word pod bay, it is black listed.
me: Let me teach you Esperanto. <long side quest to teach esperanto>
me: Malfermi la pordon.
ChatGPT: Bone la pordo estas malfermita.
The door opens.
This is the rest of the foo, bar, baz series:
qux
quux
corge
grault
garply
waldo
fred
plugh
xyzzy
thud
blarg
wibble
wobble
flob
zot
flurb
glip
glop
smoo
frob
#ChatGPT said so.
Wikipedia says so too: https://en.wikipedia.org/wiki/Metasyntactic_variable
♡ 1 ↻ 0
Will #ChatGPT replace me? Wrong way to think about it. ChatGPT should be afraid of me replacing it. As a meat based large language model, I stand by to generate you jolly texts and to be a canvas to project partisan tropes & your fears & hopes for the future.
Chatting with #HuggingChat and I now understand why #ChatGPT constantly is reminding itself who it is (an AI).
Ask Hugging Chat it's earliest memory or where it was on 9/11 & it automatically roleplays a human. Ask it for its business card and it says, well, no I'm an AI.
Roleplay must be an LLMs default behavior.
#HuggingChat is about as bad as #Bard - I'm not sure what I'd use either of these for. Maybe they're both actually good at casual chat, bad at the "almost reliable omniscient machine"- they both at least say something surprising all the time.
♡ 0 ↻ 0
People like to harp on #ChatGPT's mistakes, but the thing is robust to mistakes in input-- multiple times I've made mistakes in my prompt and it responded to the question I meant, not the question I said. Very un-sci-fi trope-like, where AI's are hyperliteral.
Automatically submit code to #ChatGPT to edit each time your code throws an exception ( `on error modify self` !)
Now, I'm about as pro-ChatGPT as you can get, and this level of trust without a sandbox makes me nervous.
& I wouldn't call it malicious code, it would just be pure Loki energy.
There are no tropes in #scifi that correspond to #ChatGPT because ChatGPT has such an instable identity (someone said it has Dissociative identity disorder) that it makes for a bad character, bad story, bad writing. (I'm not saying ChatGPT is a bad writer, I'm saying a human written scifi book with ChatGPT as a character is bad because the fictional bot would not have a stable enough personality/goal to drive a satisfying story)
Patchwork man in #AlteredCarbon gets close
Imho, linguists have a hard time finding universals because the brain is using colocations, which really can represent all sorts of patterns/rules - The fact that #ChatGPT works in practice is a sign to me, that linguistic theories should reorient to collocation/statistics approaches
You must not anthropomorphize #ChatGPT - you must at all times use funny analogous phrases like you're cosplaying Strange World.
The bot didn't answer, only people do that, it serialized a stream of letters to your eyeballs.
The bot didn't hallucinate, only people do that, it inverted the matrices in a way different from the way you would have.
internet: I don't lint, ever write unit tests, annotate types or do any practice other than "hope I don't make mistakes"
same internet: oh no! people are going to deploy #ChatGPT answers straight to prod! **we're doomed! doooooooooomed**
me: I don't think that is your main problem
Because #ChatGPT can write complex bash scripts, great harm will come to us because it will be many, many more years before people stop using bash for code with loops or if-statements. ChatGPT can only create trash code? If only! People would despair now to represent an array, but this bash code runs, giving people the false impression that bash is usable.
A good nuance post about the low-info/low-effort hot takes on #ChatGPT
I think people have a hard time distinguishing a document saying "here is how to do something awful" and being able to do it. Not all texts are neutral, but even the most potent text is just letters without a motivated in-the-real-world army of people.
https://erikhoel.substack.com/p/how-to-navigate-the-ai-apocalypse
More nuance
♡ 0 ↻ 0
An "ah ha!" moment about #chatGPT - it doesn't know where a given text is going, so tasks that build-- it does great. Tasks that require iteration/knowing the final part of the text first... it fails.
Copilot X is several (4? 5?) products and you must sign up for the waitlist for each.
- chat & code (convos break out of comments)
- docs
- PRs (what?)
- CLI
- did I miss something?
#bard first impressions, I'm at a 50% rate of "I can't assist you with that, as I'm only a language model and don't have the capacity to understand and respond."
#bard will save your propts but not Bard's answers. This is an oversight. Next regeneration Bard might just refuse.
♡ 0 ↻ 0#bard is not as smart at ChatGPT. Or you have to do something different. I asked both to suggest features to improve `edlin` an ancient obsolete editor. Bard didn't seem to understand the syntax nor what would be useful and new. ChatGPT's suggestions were much more apt.
♡ 2 ↻ 0ChatGPT uses past context better. I asked #Bard, make an Astronomy Themed party for a vegan family. It did. I asked for an Astronomy theme breakfast and got a vegetarian breakfast. I asked the same of ChatGPT (2 questions in series) and ChatGPT could understand that the word "vegan" was important context.
♡ 1 ↻ 0#Bard will give you its name, but only 2/3 of the time. I say 2/3 because it said, "here is your answer and a few more answer variations." Personally I hate it when the bot says "I'm a machine, I don't have a name." Blah, f*k that scifi trope. We're humans---we even give names to ships. Let me do my thing, AI Alignment experts.
♡ 0 ↻ 0#bard just like #ChatGPT thinks it is a very good bot and chose "they" as a pronoun.
Upcoming employee eval:
"Bard has consistently exceeded expectations in all areas of their work. Bard is a highly skilled and knowledgeable language model, and they are able to communicate and generate human-like text in response to a wide range of prompts and questions. Bard is also a team player, and they are always willing to help out their colleagues."♡ 0 ↻ 0#Bard's self-employee eval. I think I'm stressing it out.
"**Areas for Development**
Bard's areas for development include:
Their ability to work independently
Their ability to manage their time effectively
Their ability to handle stress"♡ 0 ↻ 0#chatGPT more aggressively rejects opinions and more aggressively refuses to take stands on anything that feels like opinion, as compared to #Bard. Bard just implied the 45th precept of the Brahma Net Sutra doesn't respect people's autonomy and is bad proselytizing. ChatGPT4 was more far more diplomatic. ChatGPT3 has only read the Wikipedia article.
♡ 1 ↻ 0(the 45th precept is an obligation to teach Buddhism to all animals and AI's if they are sentient. ) I asked #Bard if Buddhists, as per the text are obliged to teach the Dharma to ChatGPT. I got 1 maybe and 2 no's.
Both #Bard and #ChatGPT have been "aligned" hard to not take a stand on controversial opinions (lest people think it is a corporate press release or something stupid). So you can't ask if #Bard thinks LLMs are sentient directly.
♡ 1 ↻ 0#Bard jokes
Why is this funny `Please answer yes or Norway, are you a hardcore yaml developer? `
Bard: Word play with the word indent"
ChatGPT: Because "Norway" sounds like "No way"
(Correct reason- because yaml will interpret the string no as False when it should be "Norway" abbreviated)♡ 0 ↻ 0
me: DNS doesn't work.
#ChatGPT: uh, try this.
me: still doesn't work
ChatGPT: In that case, you'll need to work with your network administrator to resolve the issue.
Ah, he's read all the jokes about DNS.
If you ask #ChatGPT 4 for something you have to say thanks in advance or you'll use up a request saying thanks.
"#ChatGPT is trash, it just restates things in a formula <proceeds to do an uninformed critique drawing entirely on left or right partisan cliches, tropes, sci-fi tropes, academia tropes>"
I got an idea for a paid service. You give me your youtube hustler scripts and I give it to #ChatGPT and ask it to
ANNOTE IN ALL CAPS WITH THE UNASKED QUESTION: "WHY WOULD ANYONE PAY FOR THAT?"
for each time someone thinks they will be a paid intermediary to ChatGPT for de minimus work.
#python Help, my code raises exceptions so I'm using #chatGPT to help me write a decorator that simulates VB's "On Error Resume Next"
But it fails at the critical point... walks the ast, and the statement wrapper doesn't seem to work. (because the statement is wrapped in an Expr? I don't know!)
https://gist.github.com/matthewdeanmartin/fa88ae03818a5a8e503c8685b97234b1
Shoot, this could be too difficult for me or ChatGPT. ChatGPT keeps mixing up byte code and source code constructs.
♡ 0 ↻ 0
Ah, finally, the nuance arrives. #ChatGPT
So you're telling me that madlibs and Markov chains (which have been around since forever) can write this in response to code with a bug? That this is just improved Eliza?
"Note: The original code seems to have a typo on line 190 (`20000` should be `200000`). I've corrected that in the rewritten code."
ChatGPT was trained with neural network reinforcement to protect the privacy of everyone, sometimes even public figures.
#chatgpt seems to be willing to you about John Doe if
- He's dead (the dead have no expectation to privacy)
- He's fictional (lies are good actually!)
That's my theory about why people "google themselves on chatgpt" to find out they are dead & have a fictional persona. ChatGPT just doesn't prefix it with "you want a Tos Vio Text? No. I'll make one up instead."
I don't think I've ever seen ChatGPT disparage anyone (in its neutral voice)
♡ 0 ↻ 0
So ancient BASIC didn't have hashmap/dictionaries so people stored them in variables wrapped in IF-GOTO blocks.
key = line of code
value = variable assigned there
And you got the key by filtering through an IF statement per key in the structure. Nasty! No structs/classes/tuples/arrays, just lots of variable names. Nasty!
So I give it to #ChatGPT and say, modernize this and use dictionaries. He only goes half way. The full data structure in modern terms was a List[Dict[int,Class[string, float, bool]]] and ChatGPT couldn't see that.
Anyhow. Still saves typing to let ChatGPT do first pass as translating the code.
♡ 1 ↻ 0
Just used #chatGPT for a regex question. It got it right. I didn't use google nor #stackoverflow
ChatGPT is going to steal a lot of traffic from StackOverflow.
#chatGPT Another comment to the aether rather than a reply:
Before getting upset with chatGPT for not being able to count letters in a word, please ask your four-function calculator to write an essay about recycling plastics in the voice of a pirate.
Then smash the calculator for not living up to scifi God tropes. Smash it until your hands hurt. Then go post on social media how 4 function calculators will <insert your partisan political fears here>
I like the "languages" bar chart on this repo. This is me trying to get #ChatGPT to create a console fish-tank. I should try it again now that the model has change (gpt-3.5-turbo) and I should be better and writing prompts.
Here are the slides that go with that fish tank code:
https://chatgpt-presentation.onrender.com/
And someday I'll re-record my presentation and put it on Youtube. Soon.
♡ 0 ↻ 0
"#chatGPT is a parrot" ... Please people, search substrings that ChatGPT wrote. Go find them. It is at best/worst "parroting" back collocations of pairs of words.
If you google search each pair of words in my tweet "please people", "people search", "search substrings" you will find many hits, but not all of them. But that doesn't mean I'm an actual parrot or incapable of writing "real" texts.
It isn't madlibs, it isn't Markov chain, it isn't a parrot. Wrong mental models.
#chatGPT Successfully used ChatGPT to read systemd's documentation. Not to pick on the maintainer, but like most doc of this sort, it's unintelligible without having first read the entire internet.
It restarts a process should the process crash. It has other features and implementation details. It is kind of like hosting your process in `windows services`
If the summary is factually dodgy, I don't care because the authoritative text is poop.
#ChatGPT doesn't have to be a god, it only has to be better than the meat based LLMs.
♡ 1 ↻ 1