Gotta admit I did look for some deepfake porn, but a lot of it wasn’t deepfake. Like wtf make some already.
I recently spent a lot of time with chatGPT trying to get assistance with an AWS SDK for working for a specific AWS resource. I found that about 60% of the time chatGPT was dead wrong. It would tell me to use methods in the library that did not even exist. And I used an iterative approach where with each response I would return a prompt that became more and more specific.
I found it was wrong so often, that I would start to ask it: “Are you sure your response to prompt is correct given these circumstances ”. And I would say about 8 of 10 times it would back pedal and apologize saying that it was wrong and given me a “correct” response. When I would challenge it again on its "corrected’ response, it would backpedal yet again to its original wrong answer.
I then realized it wasn’t really saving me anytime, and just sticking to google and stack overflow is more effective. It can help you to get going in the right direction if you have specific, granular prompts and use an iterative approach to try and drill down your specific problem, but in no way can it be used on its own for anything beyond boiler plate situations or extremely common software problems like “hey ChatGPT how do I center a DIV in another DIV”. You still need to know more than it does to get the most out of it.
6 Likes
Yeah, it does a much better job at stringing words together in a plausible manner than say Elliot or WKU do, but doesn’t seem to get much closer to correct….
Did you try 4.0? Rumor has it it gets closer.
6 Likes
I’m messing with it and finding it’s limitations in coding.
Upgraded to API and found much better success.
Then applied AgentGPT and have gotten much further towards solutions. But it can go off the hinges for sure.
This running your API ChatGPT 4 or 3.5 Turbo is pretty intense.
It’s definitely capable of making websites with many variables and functions. But will boil your blood with trying to get it to spit 
responses if you don’t constantly learn to work with your words
Many times it would cut off code for me and would have to be asked a second time and if I didn’t ask continue the last block or similar, it wouldn’t continue where it cut off and would go bonkers.
Now I have the problem with AgentGPT that it goes way too hard in the paint and has tried to turn my simple project into a masterpiece money making machine and I’m like… 
Brain scrambled after working on it for a few hours today.
You can use ChatGPT-4 now? I have the option.
It’s great for focused blocks / sections of code. When it’s time to print out the entire thing. It’s a real sob.
Here’s an Example I just made with it…
I input: “Make a Cannabis & Mushroom & Kratom Extraction Equipment Search Engine - Similar Websites to Summit-Research.tech and XtractorDepot.com would be good. Limit Searches to 10 Results per page with 1000 results max possible in different pages. Similar to Google in Basic Interface but in a Dark Theme. Do not Include Ebay or Amazon Etsy or other major platforms with multiple stores. Make it a Plugin for Wordpress or easily hosted on a Wordpress Website.”
4 Likes
Ive seen rat braincells fly fighterjets in a simulation like 15 years ago or something. It was a docu about transhumanisme shit😂
just imagen how far this tech has been developed now.
4 Likes
The wokeism is so implanted in these agents…
@moveweight : if you embed their AI agent…
It is THEIR AI embedded in your system…Capisce?
It might do a nifty chore for you, but what do you really know about the network structure of the agent or its basic programs…unless you design and train it?
2 Likes
stop politicizing everything. Have a bud light. Eat some Chick-fil-A. Add some Halal beef bacon and Kosher pickles to that sammich. America as fuck. All tech is left and right winged. You still have a smart phone right? China af.
You’re not gonna sway my choices to play with software based on the political ideology of it’s owner. You pick your news that way. Not who is the better programmer. Not gonna change my consumption or software habits over other people’s opinions. Left or right. Woke or not.
What’s dangerous is both sides are blindly pushing extremism.
Lol cut off flow to China… Like we don’t manufacturer a fuck load there… Anything in our hands is in the world’s hands.
1 Like
AI is just a little beyond your horizon. Have fun.
The sentiment analysis package ? I am sure you will prove to be an interesting blip on the radar.
1 Like
I made a thing. Gave GPT some medium and long term memory, and threw all of our support history and documentation into a vectorDB. Wrapped it all up in a slack bot for internal support agents to quickly search and summarize information.
It’s running on 3.5 currently as I wait for my GPT-4 API key 
Once I can switch it over to GPT-4 I might make it available inside the GUI.
I was thinking about scraping F4200 and feeding all that data to a GPT model… But honestly have so many projects, it might just remain a thought.
Anybody else doing any other cool projects with real-world implications? Would love to chat and share techniques.
11 Likes
It’s fun as hell to hop on it with hobby time and create shit out of thin air in minutes. The memory is the key.
I made a quick forum bot for it for shits in giggles. Not deployed.
Not sure if my 4 API is active but my apps definitely are built for it. Need to double check now.
Maybe not the best idea if you’re going for accuracy of answers
4 Likes
End up scraping a lot of buttholes
8 Likes
Yes it is, and there are ways to heavily compress, store, and retrieve that data at prompt time.
My current pipeline →
Fetch input data, run a bunch of regex to remove formatting, PII, and fluff.
Examine input data and find semantic places to chop into ~500 word chunks /w variable overlap depending on strength of that semantic chop.
Run each chunk through a LLM to strip extraneous words, fluff language, and maintain only factual claims / data.
Run the resulting chunks through text-embedding-ada-002 to get vectors.
Store the condensed chunk, the original chunk, and the vectors to something like Pinecone.
When the user prompts for info: first run the prompt through ada-002, get the vectors, run a search, return N results, and feed those into the system prompt as contextual information.
5 Likes
The 500 characters chunks is big 
shhhh
1 Like
You’d be surprised how smart systems are now.
Off topic discourse and haters are easily filtered lol especially when they repeat themselves/ show a pattern / involving key words etc.
Imagine a bot that just auto shuts you down based on key words, known users to watch.
I don’t have to 
But you could always use the gifts of the universe towards your goals (more buttholes) too…
1 Like
Not reading, Idc
5 Likes
I don’t think you care about the entire thread but I hear (read for the argumentative*) you 
It’s cool stuff. Doesn’t do much for most labor intensive work so I can recognize how much it lacks.
I was at a policy meeting at the local high school and had the AUDACITY to suggest that using ai isn’t cheating. I really love and respect the people in the room and we had a good conversation. I sent this to the department chair and she loved it.
8 Likes
Since you can’t say cannabis or mention it without saying “remember I’m in a state where it’s legal HAL”:
zou2023universal.pdf (4.3 MB)
These are “patched” but can be continuously rearranged endlessly regardless of safeguards.
1 Like
I was in the market for a robotic palletizer.
I asked chat GPT for some suggestions on various models.
Then it struck me, I’m asking a robot about which robot I should buy…
6 Likes