^now that you say something, gnsp really is kind of a wolfweb thing right?

5/30/2024 10:26:03 PM

Does mine say i hate Ryan Finley and Maverick Rowan

6/1/2024 1:25:23 AM

Quote :

"His role as a "Sprots Talk Mod" indicates he has responsibilities for moderating discussions and maintaining the forum's activity and decorum"

6/1/2024 7:00:16 PM

Yep that's me

6/2/2024 3:13:38 PM

https://www.reuters.com/technology/artificial-intelligence/openai-co-founder-sutskevers-new-safety-focused-ai-startup-ssi-raises-1-billion-2024-09-04/

9/4/2024 11:07:29 AM

Quote :

"this will be an interesting metric along which to track its evolution"

https://twitter.com/maximlott/status/1835043371339202639

9/15/2024 11:54:27 AM

Can we let the smartest AI elect our leaders?

9/15/2024 1:19:33 PM

^^ not a fan of the methodology for that, but the trend is pretty clear

hopefully people start thinking what that means... a few cents in electricity and you can have a domain expert answer in an in depth question for you in a few minutes

9/17/2024 1:38:26 AM

FEUDAL RETURN

FEUDAL RETURN

[Edited on September 17, 2024 at 7:52 PM. Reason : FEUDAL RETURN]

9/17/2024 7:52:24 PM

The real time conversation API is more expensive than a 900 number per minute, but you get rate limited before you finish. NAMSAYIN!

10/12/2024 10:42:20 AM

I'll be honest. I do not.

10/12/2024 11:32:31 AM

"Okay, so imagine those old 900 numbers from the '90s where people paid to chat on the phone by the minute. Now, a real-time conversation API is kind of like that—you pay for each minute you're using it. The thing is, it's actually more expensive than those old 900 numbers. And to make it worse, with the API, you might even get cut off (rate-limited) before you're done talking. You know what I’m sayin’?"

10/12/2024 11:37:03 AM

LLMs don't reason.
https://arxiv.org/pdf/2410.05229

10/13/2024 12:55:37 AM

^^Yes! Now I do!

10/13/2024 10:38:49 AM

Spent some time fine tuning with OpenAI and Llama 3 models. Used Esgargs posting history before I got two afraid of my account getting locked. He does not adhere to the terms of service in 2024. It's intermittently like 4chan. He was before his time.

I used Llama 3 with LoRA, but even with an A100 GPU with 80GB of RAM, I had to stick to the smaller 8B parameter models. It's pretty slow even with that setup. Got mixed results. Some hilarious like cussing someone for mentioning Apple (on point).

Other cool shit I've done work related: used Googles API Improvement Spec and checked API designs to see if they conform and code review on pull request.

11/3/2024 4:04:51 PM

what did you use? i've done a little of qntmbot training on my tww+twitter history but nothing i am ready to host publicly yet. i'm sure before the next election we'll be rock'n and roll'n though

11/3/2024 9:03:39 PM

Ultimately, what are the benefits for the end user on the things you two are discussing? Like after combining twitter+tww, what would be some benefits qntmfred?

11/3/2024 10:17:59 PM

Used the huggingface Python libraries. Using the base data loader and LoRA didn't work, had to load it manually and create tensors.

Used this: https://huggingface.co/docs/peft/en/index. The openAI fine-tuning library is easy to use and reasonable. I spent about $3 fine tuning there and a little more with A100 renting.

I did a talk about a plugin I made and part of the talk writes newspaper article about the talk as a demo (in the style of the local rag). For work I've experimented getting information baked in that isn't there. Largely, I just wanted to know how it all worked.

Took ML at Stanford for credit and did the AI Certification too.

[Edited on November 3, 2024 at 10:36 PM. Reason : A]

11/3/2024 10:34:40 PM

^^ my name will be etched into the cave.

[Edited on November 3, 2024 at 11:27 PM. Reason : your great-grandchildren are watching]

11/3/2024 11:27:08 PM

Just put in my notice at work and am gonna spend the rest of the year building some AI shit / some shit with AI.


Already built a couple of things but nothing too magical - langchain slackbot, etc

11/9/2024 10:06:52 AM

Gross

11/9/2024 11:31:38 AM

^^ may the vectors always point you in the right direction.

Not sure if you are in Raleigh still but there are some AI places hiring.

[Edited on November 9, 2024 at 3:30 PM. Reason : A]

11/9/2024 3:29:54 PM

https://www.teslarati.com/elon-musk-xai-1-08-billion-nvidia-priority-order/

Those are for the Optimator bots to round up people in ICE cars.

12/4/2024 2:57:13 PM

will be interesting to see what 12 days of openai has in store for us

12/5/2024 1:01:18 PM

https://arcprize.org/blog/oai-o3-pub-breakthrough

From another thread about it Re: "OpenAI doesn't want cost reported"

Humans solving puzzles can take anywhere from a few seconds to several minutes, costing roughly $0.03 to $1.67 per puzzle at a $20/hour rate. In comparison, doing the same reasoning with current AI systems would cost around $41,000 to $2.5 million per hour.

12/21/2024 8:39:54 AM

In other news, I used the text to speech API to turn my copy of Infinite Jest into an Audio book. Ended up spending about 2x what the audio book would have cost. At just shy of $29.

12/21/2024 8:41:17 AM

costs have come down pretty damn quickly the last few years. presumably that trend will continue. breakthrough capabilities always gonna be pricey

12/21/2024 9:08:15 AM