30 Aug 2025
A lot of us haters know that bubble's gonna' pop, but what does that mean for everyone?
Over on Bluesky, red made a thread about how combating LLMs and their usage by calling them useless isn’t a salient point. Their mother asked ‘em if they tried ChatGPT yet right before making meatballs and meatloaf using a recipe that ChatGPT gave her. The verdict on mom’s meatballs: delicious. The everyday prompter isn’t going to be swayed when the computer is so confidently and mostly-accurately giving answers.
This line of thinking got me, because it’s true. A lot of people are at home using ChatGPT by asking it a question and getting a response back. However, ChatGPT didn’t always have this status as the quick-question oracle. That belonged to Alexa a decade ago.
“Hey Alexa” gets you whatever information you requested or task you wanted performed. Hey Alexa, how long is a football field? Wow, it’s that long, cool! Hey Alexa, add a calendar appointment tomorrow at 5 PM. Thank you! This is a system that was set up to ensure that those using it had a smooth experience, and that whatever they asked could be answered. Special voice commands called “skills” can be custom-programmed to do different actions! Part of this helpful ecosystem is Alexa’s creator, Amazon, here to help you order via Amazon Fresh the ground beef you need for the recipe you asked Alexa for. There’s a synergy to keep Amazon’s customers buying via a helpful assistant.
ChatGPT does not have this same synergy. While Alexa is subsidized by Amazon and is used as an advertisement for Amazon products and services, ChatGPT is in service of itself. For free you are limited to how much you can request of it, and asking it more will require either a subscription or the purchase of Tokens. Tokens being the currency used when a prompt is computed. These are sold in packs of millions or are provided with monthly subscriptions. If you run out of tokens, you’ll either need a higher tier subscription or to buy more tokens.
General LLMs are able to do a large number of things, but they are unable to do them as well as an existing product for the same task. Unlike the Alexa of the past, you now have a usage limit. If you ask it something and you know that what’s returned is incorrect, it will cost extra to take another chance at a correct answer. If Alexa used a token format, it never would have received the adoption it did. People adopted it because it was so easy and seamless to ask it and get an answer back. ChatGPT is subsidized by VC money injections and does not turn a profit as it is now. If OpenAI were to start charging the public for asking random questions and they had to work out how many millions of tokens they needed, people would bounce back to an Amazon Echo so goddamn fast.
Aside from Amazon putting up a “BACK IN BUSINESS” sign? Well, this won’t kill machine learning as a concept. Tensor core PCI-E cards existed over five years ago, and now most video cards come with them or something similar integrated. Nothing’s stopping someone from doing local machine learning and creating a bespoke model. But the instant that OpenAI and CodeWeave and Anthropic and Perplexity vanish into the ether, the free-to-use general LLM business concept is dead in the ground.
These models require retraining to correct any issues or to introduce new information into the model. In large models this is an astronomical amount of data to process. This is not feasible at home, but this processing can be handled at one of many large 100MW+ datacenters being created across the United States. Without the flow of VC money pouring into these companies to run and maintain expensive datacenters, these large models can’t be retrained. Either the unsubsidized cost of the datacenters and their energy will get passed to consumers, or general LLMs will disappear over time.
I say “over time” because while you need all that energy to train, models can be ran offline and locally. The average person will have no issue using a general model that’s a year or so out of date, but over time the age of its training will show. There’s a great short story called MMAcevedo that’s got similar tones in it about the memory backup of a person, I highly recommend reading it.
Once these are out of date, they won’t be of much use. Maybe they were trained to link to external services that no longer operate. Maybe who they think the current president is is wildly wrong by 5 administrations. It’s going to be easier to convince someone to go back to an Alexa that Just WorksTM rather than convince them to use this new startup LLM that’s starting at $100/mo for 15 tokens and hey where are you going.
It’s the only way this goes. Machine learning existed before this LLM craze but was for focused tasks. You’re never killing that. The everyday person will find something else to take the role of knowledge machine when the general LLM becomes completely paygated or disappears entirely. If they want, Alexa still operates in the space and has long before these came around.
But the business model of the Computer Oracle is unsustainable. It’s a hungry beast that must suck up money at unprecedented rates for mediocre results. It’s what’s being held accountable when a company goes “sorry, but it said we gotta fire you”. Without that constant firehose of cash and data being pumped into it, it makes no profits and tanks an entire economy that decided to make it a lynchpin of investments. It’s going to be so fucking ugly when it happens, people’s lives were ruined in pursuit of this technology. People are dead because of it. Hopefully we don’t try and rebuild the machine that was kinda’ right sometimes.
Also this isn’t intended to be an ad for Alexa, the second that stops driving revenue for Amazon I wholly guarantee you they drop it like a hot rock.
