« first day (4500 days earlier)      last day (678 days later) » 

cyberghost
cyberghost
00:37
@CaffeineAddiction lol.
 
2 hours later…
CaffeineAddiction
CaffeineAddiction
02:07
user image
so close
code actually works, GPT-4's understanding of what the result of it will be ... not so much
 
8 hours later…
Rory Alsop
Rory Alsop
09:46
@CaffeineAddiction I'm not going to see that as a win yet :-)
The copilot type bots I see as quite useful, because of the corpus used for their training. Chat-GPT definitely not so much - that isn't what it is for
Journeyman Geek
Journeyman Geek
10:14
@RoryAlsop I think the problem is really... people using it with no clue what its limitations are
Rory Alsop
Rory Alsop
@JourneymanGeek well, yes - "it's AI, it can do anything"
no - it's some good algorithms, with a wide range of base data inputs
Journeyman Geek
Journeyman Geek
Also vice.com/en/article/jg5qy8/…
 
3 hours later…
CaffeineAddiction
CaffeineAddiction
13:37
yah, no I am approaching it w/ no expectations. I think its pretty obvious by now this is the way of the future ... kinda like when google first came out. Early adopters who built up their google foo before the term "google foo" was even a thing had a major advantage.

I am also playing with it just to see what it is capable of as it grows.
One of the first things I figured out that I am not sure people are fully aware of yet is that the "Free" version that most people are playing with is GPT-3 ... but if you pay for the subscription you get access to GPT-4 which does the fun stuff like acting like a linux terminal.
for example, I recently played a homebrew game of 20 questions with GPT-4 and it worked quite well. Do the same thing w/ GPT-3 and it gets confused halfway through and accuses you of cheating.
user image
 
2 hours later…
Rory Alsop
Rory Alsop
15:12
@CaffeineAddiction Many companies I know are blocking all access to it - for very good reasons. I think as a tool for improving your writing, it's not too bad, but for most things people try to use it it is just awful
We blocked it in my company - there is no use case for it
CaffeineAddiction
CaffeineAddiction
@RoryAlsop I dont know your company, but generally I would consider strait up blocking it as very short sided. I agree that the use cases for gpt-3 are very limited ... but gpt-4 just fully completed all of the standard interview questions I would give a new-hire DevOps candidate. Not only did it pass the test ... it passed it w/ flying colors.
Just like with anything new, its good to take it w/ a grain of salt ... but consider how bad of an idea it would have been to blacklist google.com when it first came out
early days still ... but the potential is there
Rory Alsop
Rory Alsop
@CaffeineAddiction Yes, GPT-4. And that is another reason to block it.
We initially just restricted access for devs, as they would have one of the biggest security risks from it
GPT being able to answer interview questions is a use case you have? The problem is that it isn't an AI. It isn't good at anything except producing text that meets the rules it has learned. And unfortunately that makes it very very bad for all the use cases in my org, except possibly marketing materials
CaffeineAddiction
CaffeineAddiction
all about risk / reward imo. I know python quite well ... but am fairly new to powershell ... I just asked it to translate some code from python to powershell and it did it flawlessly. Similarly I have found it to be crazy useful for writing unit tests.

I agree, security concerns for some things ... but you dont just outright ban fire due to its potential of burning down a building. If we did just blanket ban fire ... restraunts would prob have a huge prob doing business.
Rory Alsop
Rory Alsop
The big problem is the more you play with the model and understand how it works, the worse it is for these sorts of things. In your analogy, it's like banning a random selection of fireworks, including bombs, from being used to light cigarettes in a tobacco shop
It is not an AI, and is not designed to be. It is a chat bot
There are much better ones for specialisms
like coding, for example
there are good reasons more countries are seriously considering copying Italy with its GPT ban
admittedly, most of them are rather repressive: China, Russia, Iran, North Korea...
New York students and teachers can no longer access it
CaffeineAddiction
CaffeineAddiction
I see what you are saying, but I feel I have to disagree. GPT is more than just a chat bot ... its the first stab at generalized AI. Nothing it says should be taken at face value ... but the same is true of anything else you would find on the internet (which is what it was trained on)
Rory Alsop
Rory Alsop
15:24
@CaffeineAddiction and that is the problem - the training corpus.
CaffeineAddiction
CaffeineAddiction
Yep, and the training corpus of a 5 year old is the other snot nosed kids they play w/ ... but you got to start somewhere
It also works for anything you teach someone else to do. "Oh yeah, I trained a pair of neural nets, Emily and Kevin, to respond to support tickets."
Comic is good, but the mouseover is the best part
 
2 hours later…
nobody
nobody
17:13
@RoryAlsop I have found it genuinely useful for a bunch of things. GPT3 rarely completes my task for me, but it does give me a very good starting point
People can say "LLMs just predict text, they don't understand anything", doesn't change the fact that it can be useful if you understand that it isn't a super-intelligence yet.
Do we even know what it means to understand something? Or what the hell the LLM really does under the hood?
Rory McCune
Rory McCune
17:29
@RoryAlsop whilst many companies are blocking the public chatGPT (for data privacy/security concerns) a load will be hitting up MS for subscriptions to their OpenAI service to work on projects, I'm sure.
I can see a number of interesting use cases for things like explaining legacy code bases to help re-factoring and documentation chat bots to help users without asking them to wade through a load of web pages looking for things
Rory Alsop
Rory Alsop
17:47
@RoryMcCune Agreed. And it is a very different beast, both from a potential data leakage risk (corps using MS suite already have a level of implicit trust in MS) and from a specialism perspective
I don't have anything against that "copilot" type of algorithmic ML

« first day (4500 days earlier)      last day (678 days later) »