^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

it's insane to me that we're talking about post-AGI already



[Edited on January 19, 2025 at 3:59 PM. Reason : https://ia.samaltman.com/]

1/19/2025 3:58:46 PM

My definition of AGI is I can ask it to give me a schematics and code for a device to stabilize a fusion plasma

If it can do this I’ll be impressed

Generally speaking though OpenAI doesn’t do anything welll beyond what the academic researchers are looking at, and the research community hasn’t seen anything like what that video describes

But for years now it’s be been theorized that models should be able to extrapolate new points in the latent space that represent new Knowledge. I’ve actually been surprised we haven’t seen this already… if OpenAI has a new attention mechanism that solves this thats pretty amazing on its own.

1/19/2025 11:42:37 PM

^ I'd settle for,.doesn't have to enter the number from the sticker on the apple when it's weighed at the grocery store.

1/20/2025 7:23:02 AM

I did a Google search last week about how to calculate the distance across the points of a hexagon for a known distance across the flats, and the generative AI response was completely wrong. Almost the opposite of what I asked, though it stated the result confidently. It was giving the calculation for distance across the flats for a known distance across the points. It was suggesting that the distance across the points was LESS than the distance across the flats. SMH!

Understood this is a known weakness of generative AI and there are disclaimers and whatnot, but seems like a lot of work still to do.

1/20/2025 8:49:48 AM

Sama saying the AGI talk as all hype

https://www.reddit.com/r/singularity/comments/1i5lxtn/out_of_control_hype_says_sama/?utm_source=share&utm_medium=mweb3x&utm_name=mweb3xcss&utm_term=1&utm_content=share_button

1/20/2025 11:16:45 AM

he's only saying "not next month"

2 weeks ago on his own blog

Quote :

"We are now confident we know how to build AGI as we have traditionally understood it. We believe that, in 2025, we may see the first AI agents “join the workforce” and materially change the output of companies. We continue to believe that iteratively putting great tools in the hands of people leads to great, broadly-distributed outcomes. We are beginning to turn our aim beyond that, to superintelligence in the true sense of the word. We love our current products, but we are here for the glorious future. With superintelligence, we can do anything else. Superintelligent tools could massively accelerate scientific discovery and innovation well beyond what we are capable of doing on our own, and in turn massively increase abundance and prosperity. This sounds like science fiction right now, and somewhat crazy to even talk about it. That’s alright—we’ve been there before and we’re OK with being there again. We’re pretty confident that in the next few years, everyone will see what we see, and that the need to act with great care, while still maximizing broad benefit and empowerment, is so important. Given the possibilities of our work, OpenAI cannot be a normal company."

1/20/2025 1:06:48 PM

But when will it be able to do basic geometry?

1/20/2025 1:49:31 PM

^^
Yeah that makes sense. But he’s using a circular definition of AGI there. They’re saying AGI is when an agent can work a job on its own, but that’s now how colloquially most practitioners use AGI— not that there’s any formal definition of the word

That’s a big milestone but not what most people are thinking of. There’s probably people this year using an ai agent to replacing substantial parts of peoples jobs. Most frontline Tech support can be done by ai right now.

Overall I like Altman but he’s in major CEO mode, looking out for sales and funding. He’s not a science guy most of the time

Edit: there’s not a clean line between AGI and ASI anyway… I’d argue current LLMs are ASI in the sense you can literally ask them about any topic, they can translate, they can tell kids stories, cutting edge multimodal models can intermix this with audio and image generation, they can imitate any voice (although most providers block this capability for obvious reasons). What AI can’t do is learn on the fly and integrate this with its body of knowledge, and do complex analysis and reasoning. I bet we solve the science of both of these things things year though, and we’ll see some spectacular engineering projects later this year and next year based on this. Sci-fi like humanoid robots the following year. I also would bet these big advancements don’t come from OpenAI either.

The wildcard is going to be Groq/Cerebras. These ai accelerators could be game changers and be an inflection point with the right software.

[Edited on January 20, 2025 at 2:13 PM. Reason : ]

1/20/2025 2:07:51 PM

Quote :

"not that there’s any formal definition of the word"

yeah this is the main problem with where we're at. AGI was such a pie in the sky concept for so long that we weren't anywhere close to being ready when the possibility that it could even be approaching starting to materialize.

even in 2019 when I was working for Yang and trying to sound the alarm to the country on what could and would start to happen once AI got to a certain level, I wasn't expecting the possibility of AGI-level stuff for at least another decade or longer.

[Edited on January 20, 2025 at 4:27 PM. Reason : will be interesting to see where arcprize goes in 2025 though ]

1/20/2025 4:26:25 PM

Looks like OpenAI wants a new network of data centers for their models rather than to continue to rely on cloud services providers

Could be a great opportunity for people wanting entry level server admin jobs. Biden admin had a great plan to pair ai infra build out with green energy. I’m guessing this is not a main concern anymore

https://openai.com/index/announcing-the-stargate-project/

1/21/2025 6:22:46 PM

^^
My personal date for “in the lab” AGI was 2030, just from scaling up whole brain simulation research with moores law — thats an upper bound so I’m delighted to see algorithmic approaches possibly succeeding before then

1/21/2025 6:25:11 PM

what's $500B between patriots?

1/21/2025 10:28:36 PM

^i was actually looking for a thread to ask your opinion on the 500b for AI. Is that good? Seems way too much money for me. Can't we put like 500b into fusion? Seems like that would be more awesome

1/22/2025 2:45:10 AM

half of it will go into electricity generation, however they can get it. We don't have nearly enough capacity. Fusion, solar, fossil fuels, doesn't matter. I'm not sure we'll make enough progress on fusion in the timeframe necessary for this investment to accomplish its goals.

Last year there was chatter about Sam Altman wanting to raise $7T, so this is a good start, but there's a long way to go.

And while I support this move, it's also worth pointing out that this investment may not and you might even say probably won't generate a profit in the traditional sense. It really is a we have to get there first for national security kind of thing, so they'll spend whatever it takes.

1/22/2025 5:51:32 AM

^ a former prominent soapboxer that no longer posts told me it's like during eniac asking for a huge sum to build a computer as big as New Mexico with vacuum tubes. They need some advancements in power and scaling to make it justified.

I'd be more happy if some of it was directed at miniaturization and mass ASICs and GPUs.

1/22/2025 10:07:47 AM

https://chat.deepseek.com/

Finally read about reasoning. It's basically chain of thought prompting generated as part of the training data.

Operator looks rad, but don't really want to pay $200 a month for it.

1/26/2025 5:21:21 PM

The investment was inevitable

You can basically assume buildout to 1 gpu at least per person in the country (have to consider trainium/TPUs/apple/AMD in this too)

Once the tech gets a bit better, the demand is basically going to be inelastic. Ai capability is going to have to grow beyond the number of knowledge-based problems to solve before it slows down

For reference nvidia is shipping like 5 million gpus per year now

1/26/2025 6:05:28 PM

https://www.datacenterfrontier.com/machine-learning/article/55244139/the-colossus-ai-supercomputer-elon-musks-drive-toward-data-center-ai-technology-domination?utm_source=perplexity

2.21 Gigawatts! /Doc Brown.

Apparently the power surge when GPUs idle can break shit so they either have to off load into batteries (XAI) or burn cycles calculating nothing (Meta). Most of the companies involved have dropped their climate pledges over the excessive fossil fuel burn.

2/4/2025 10:05:05 PM

https://www.reuters.com/markets/deals/elon-musk-led-group-makes-974-billion-bid-control-openai-wsj-reports-2025-02-10/

Quote :

"Feb 10 (Reuters) - A consortium of investors led by billionaire Elon Musk is offering $97.4 billion to buy the nonprofit that controls artificial intelligence startup OpenAI, the Wall Street Journal reported on Monday. The offer intensifies a longstanding battle between OpenAI CEO Sam Altman and Musk over the future of the startup at the heart of a boom in generative AI technology. "

2/10/2025 4:04:09 PM

Quote :

"Apparently the power surge when GPUs idle"

how does that work

2/11/2025 4:33:05 PM

^ when they are calculating the GPUs are using lots of power. These giant models span multiple GPUs. When they calculate the updates in one part of the model they need to be sent to another to update. While it's waiting for updates it's idle and not using power. The grid needs somewhere to send the excess power.

There are some other times it may need to idle (like loading the model from disk to CPU to GPU).

2/13/2025 2:06:37 PM

Spent a lot of time with Grok 3. Here’s what stood out:

Image generation is terrible—Midjourney is vastly superior. Grok can't generate images in specific styles like "Japanese woodblock print" or "8-bit NES." Some basic photorealistic style is ok, but beyond that weak

Hallucinates a lot—More than any other LLM I regularly use. When evaluating a DOGE story, it falsely claimed DOGE was Elon Musk’s cryptocurrency and that the Treasury was investigating it.

Fabricates arguments—I had it analyze a forum debate, and it completely made up the argument and declared a winner based on false reasoning.

Blazing-fast inference—It can summarize dozens of pages instantly.

Impressive research skills—I gave it a Home Depot shopping list, and while it couldn’t retrieve exact aisle locations, it searched blogs and Reddit to estimate aisle ranges. If Home Depot’s site didn’t block it, it likely would have nailed it.

Great at uncovering forum details—As noted message_topic.aspx?topic=654531&page=1#16689515, I asked about "the wolf web" and political discussions, and it quickly identified key posters and the political forum. The research ability is amazing—faster than Perplexity, so I’d start here now.


Haven’t tested math, logic, or coding yet, but neutral reviews suggest it’s among the best.

2/20/2025 4:28:27 PM