hph6203 said:
Anthropic is reportedly set to make an operating profit this quarter. They are not selling $20 of tokens for $2. That is a misunderstanding of their finances derived from a population of people that think CNBC is a good source of information.
CNBC is the one of many incessantly hyping them. Most AI providers have switched to API billing precisely because the subscription fees were not even close to covering actual token costs. For Claude it was something like $1300 of tokens given out per $100 subscription. Entirely unsustainable.
https://she-llac.com/claude-limits?ref=wheresyoured.atQuote:
What these model companies are doing is buying compute, building a model, shipping that model, and delivering it with a portion of the compute they trained it on. Buying more compute, building a better/more expensive model. The prior model, if left to its own operation, is a profitable endeavor. The new model expense is where the losses come from. That pattern is going to roughly continue until the models asymptotically approach the limits of capability given existing data sets.
Already there. Scaling is largely dead. XAi threw 10x the amount of compute at their latest Grok model for only modest improvement in benchmarks over the prior version. The jumps we saw from GPT2 to GPT3 and from GPT3 to GPT4 were not repeated for GPT5. In fact, GPT 5 was delayed because they barely got much improvement out of several rounds of training and ended up releasing it as 4.5. When they finally did release GPT-5 Sam Altman dropped all pretenses about it being AGI, which is what he was bragging about back in 2024. It was only a modest improvement from 4.5. The scaling is no longer scaling. They already have all the data from the internet.
Almost every bit of improvement we've seen over the last year or so of model releases has been in the wrappers, the code that actually calls and harnesses the LLM.
Quote:
Eventually it will reach what we call AGI of digital intelligence, i.e. capable of doing basically every job that involves a computer. That will consume a substantial, but not total, proportion of existing jobs. The market will expand and the net effect is labor shortage, not job shortage.
They don't have the technology to do this. LLMs are not capable of understanding why they are doing anything. That's just baked into the design. Reasoned decision making without a clear right or wrong answer is impossible for an LLM. Training an LLM more won't work. Jobs involve reasoned decision making. In order to accomplish this they would need a completely different type of AI.
Quote:
The model companies will then focus most of their resources on physical world observation through machines and robots, generating their own data in an ever expanding quantity.
There are several companies now working on different types of AI that are specialized using local world models. If you are looking for something approaching AGI for a specific use case this is where it would likely come from.
Using LLMs to train other LLMs is a recipe for devolution. LLMs don't know a right answer from a wrong one. New LLMs would be training on the hallucinated data created by the prior one.
Quote:
On the open weight side the models are being reduced in parameters so that they can run locally on lessor hardware with some incremental quality, but still usable on high end hardware ($5,000+). That $5,000 hardware will cost $2,000 in less than 5 years. That $5,000 hardware can produce $40,000 worth of tokens annually from models that were at the frontier 12-18 months ago. And that hardware is likely to have a 5 year utility, because model size for equivalent quality is falling.
Which is a disaster for the big AI companies because anything that can run locally they can't bill for.
Quote:
AI is getting cheaper. It will continue to get cheaper. Consumer hardware has gotten better. It will continue to get better, and despite recent price increases they will eventually fall as production begins to match the new elevated demand and prices will normalize/fall relative to past capabilities. It is the nature of technology.
OpenAi, anthropic, SpaceX and the others investing heavily in AI will have to generate ungodly amounts of revenue to cover the insane capex. They aren't going to be able to do that by lowering prices.
