LLMs as parlor trick to raise money for compute infrastructure
| Richard Ames | 05/05/26 | | The Redwall Cinematic Universe | 05/05/26 | | Dr. Samuel-Auguste Tissot | 05/05/26 | | banks? COCKSUCKERS. | 05/05/26 | | OYT and the Indie Reprieve | 05/05/26 | | cowgod | 05/05/26 | | OYT and the Indie Reprieve | 05/05/26 | | ...,.,.,.,....,..,.,.,,.,.,,..,. | 05/05/26 | | Cave Johnson | 05/05/26 | | Dr. Samuel-Auguste Tissot | 05/05/26 | | Cave Johnson | 05/05/26 | | Dr. Samuel-Auguste Tissot | 05/05/26 | | Richard Ames | 05/05/26 | | cowgod | 05/05/26 | | Dr. Samuel-Auguste Tissot | 05/05/26 | | Richard Ames | 05/05/26 | | Hello, World! | 05/05/26 | | Cave Johnson | 05/05/26 | | icanseemyneurons | 05/05/26 | | The Redwall Cinematic Universe | 05/05/26 | | Richard Ames | 05/05/26 | | Cave Johnson | 05/05/26 | | Richard Ames | 05/05/26 | | Cave Johnson | 05/05/26 | | Richard Ames | 05/05/26 | | Cave Johnson | 05/05/26 | | The Redwall Cinematic Universe | 05/05/26 | | icanseemyneurons | 05/05/26 | | The Redwall Cinematic Universe | 05/05/26 | | icanseemyneurons | 05/05/26 | | Cave Johnson | 05/05/26 | | ,.,....,..,.,.,,,,..,..,.,..,.,.,.,... | 05/05/26 | | Cave Johnson | 05/05/26 |
Poast new message in this thread
Date: May 5th, 2026 6:35 PM Author: Richard Ames
Huge amount of "AI will never work" doomers focus on the unprofitability of LLMs. "ChatGPT is losing money! It can't pay its bills!" Etc. etc. I have started to view LLMs as the equivalent of J.P. Morgan turning on the lights in his Manhattan estate to show off the power of electricity. Or World's Fairs of old with their dazzling lighting displays.
Whether or not LLMs can ever truly be profitable is besides the point. They have served as proof for the compute layer which will power the much more mundane uses of AI that will go much, much farther from an economic standpoint.
Shit like AI for trucking compliance (to save an unbelievable amount of paperwork), AI for internal firm coordination, AI for financial reporting. You can look at basically every industry and find layers of work that amount to moving information from one bucket to another with a modification or two -- specialized AI which runs on the compute layer will do this whether or not ChatGPT / Claude / Grok can maek it.
None of this will "replace all jobs" or automate everything via AGI. But it will dramatically change the landscape of white collar work such that firms will have fewer and fewer employees. They will still HAVE employees, but huge numbers of people who have simply existed to do basic information processing / formatting / transferring will be toast.
I am probably stating the obvious to most here. But I was lisening to Ed Zitron go on and on about how LLMs can never be profitable and I just thought to myself that he, and many like him, do not see the full picture at all.
(http://www.autoadmit.com/thread.php?thread_id=5864029&forum_id=2:#49868029) |
Date: May 5th, 2026 6:38 PM Author: The Redwall Cinematic Universe
it's just another financialized ponzi scheme. the data center/infrastructure buildout is just another dimension of the ponzi
lots of companies "aren't profitable" but their valuations keep going up and up and up. AI isn't all that different from other VC ponzi scams. it's just the scale of it that's unprecedented
(http://www.autoadmit.com/thread.php?thread_id=5864029&forum_id=2:#49868033) |
Date: May 5th, 2026 6:53 PM
Author: ...,.,.,.,....,..,.,.,,.,.,,..,.
the 'data centers' are just empty buildings like the Somali daycares. check the tapes.
(http://www.autoadmit.com/thread.php?thread_id=5864029&forum_id=2:#49868069) |
Date: May 5th, 2026 6:55 PM Author: Richard Ames
I am also open to the idea that it's just another ponzi to keep the debt bubble going. I am a big "the debt bubble drives everything" believer. But I can also see both things being true.
I can imagine doomers back in the day saying "World Fairs will never be profitable!!" and totally missing The Point.
(http://www.autoadmit.com/thread.php?thread_id=5864029&forum_id=2:#49868078) |
Date: May 5th, 2026 9:15 PM Author: Cave Johnson
Here's what's really happening. The demand is there. People want to use the AI but the computer isn't there, so the SOTA cloue providers - the companies that ONLY do AI like OpenAI and Anthropic, are doing workarounds and keeping everything a trade secret. Here's what that means:
The "full" Claude might be trained on 1.5 trillion parameters. It might require a terabyte of RAM to run the full uncompressed model. There's no way in HELL Anthropic is giving you that for $20/mo. People are saying they need trillion dollar data centers because that's what it would take to meet current demand for the full uncompressed dense models. The SOTA providers are using quants and MoE to get around this. That's why by some measures those models are getting shittier now. Users are reporting errors that didn't used to be there.
At the same time, open weight models with just 300 billion or maybe just 30 billion parameters are getting hosted by new maalrket entrants like Ollama, Opencode and Nous Research, and people are finding that they're "good enough,": especially as agentic harnesses
The open weight models are DOMINATED by CHINA with Alibaba, Deepseek,Z, Kimi just DOMINATING. The only American companies with their toe in the door are Meta and Google, although IBM just showed up too. Six months ago this was unthinkable, now it's the future.
(http://www.autoadmit.com/thread.php?thread_id=5864029&forum_id=2:#49868331) |
 |
Date: May 5th, 2026 9:41 PM Author: Richard Ames
Interesting. If this proves to be true, this seemingly means a lot of the data center buildout is a total waste of time and money.
If reporting is to be believed, it does seem like the buildout is slowing (or may be a bit of a house of cards to begin with.)
This would have meaningful consequences. Though I'm not sure what exactly at the moment.
(http://www.autoadmit.com/thread.php?thread_id=5864029&forum_id=2:#49868375) |
Date: May 5th, 2026 9:27 PM
Author: ,.,....,..,.,.,,,,..,..,.,..,.,.,.,...
cr. Guys like Ed Zitron aren’t worth paying attention to at all. He was calling a top of the bubble in 2024 and saying models wouldn’t get much better than they were then. Agentic coding is here now, models are much better, revenue is exploding because the models actually work, and there are non-trivial model improvements still occurring frequently. There are a number of unexplored ideas still out there to make models better and training compute is exploding, which will likely facilitate model progress.
(http://www.autoadmit.com/thread.php?thread_id=5864029&forum_id=2:#49868353) |
|
|