Just noting it. The other post was submitted earlier. The mod's can figure out how to combine/reconcile. Update: I think you are correct and this one won :)
Reality is that people in Western countries need TSMC to make high end chips outside of Taiwan, because we’ll all be screwed if China decides to invade Taiwan. This has nothing to do with how you feel about Trump. It’s just the cold hard facts.
If you want stable access to technology in the future, you should be demanding secondary sources of high end chips, because none of us want to die fighting China in a war a few hundred miles off China’s coast line and 5,000-6,000 miles from most western countries (excluding Japan/Korea).
This really reinforces the idea that the AI race and the Railroad Mania of the 19th century are very similar.
So many different companies are going to have similarly powerful ai that there will be no moat around it and it will be cheap. They will never earn their investment back.
I suspect this is the real reason behind Anthropic limiting subscriptions to their own products and keeping API prices several times higher than comparable models. Applications more sticky than API users and less technical users more sticky than programmers (ie Cowork more sticky than Code).
Anthropic generally seem more into living within market discipline and market signals of some sort. Products with margins, even if it's sort of irrelevant considering R&D costs and capital inflow.
That said, there's nothing like the real thing.
The risk is something like the railroad bubble and the dotcom. Over-investement, circular revenue and a timeline that doesn't work.
The whole premise is based on the fact that over-investing in GPUs and models are a good thing here as it yields more 'intelligence'.
This as it turned out was not true for rail roads - more and more rail roads isnt a good thing.
The real dilemma facing the model producers is that all this money invested for a general model, targeting general intelligence, is a disaster and essentially the investment into existing assets is a write off. Then on top of that if this is true, youve got data centres full of compute that aren't being used up.
The weird position they find themselves in now is that they have to keep making it smarter... but they already made it too smart (Mythos). I'm not sure how that's going to work out exactly.
They find an arbitrary intelligence cutoff point between Opus and Mythos, label it "acceptable risk", and then the labs coordinate to gradually nudge that line forward and hope the internet doesn't break?
I think we will see unbundling of large model into submodels: modular, smaller and efficient, only include what you need eg a CUA model, a reasoning model, a legal model, a writing model, a coding model (this could get subdivided into different languages). That way you only update that submodel which needs retraining.
The labs started doing that in late 2024, they all published research on it.
Curiously, mid 2025, they all simultaneously implemented increasingly bizarre restrictions on "self replication". I don't think there was anything public but it sure sounds like something spooked them. (Or maybe just taking sensible precautions, given the direction of the whole endeavour.)
At any rate, I recently asked Opus about "Did PKD know about living information systems?" and the safety filter ended the conversation. It started answering me, and then it's response was deleted and a red warning box popped up.
But notably, I was given the option to continue the chat with a dumber model (presumably one less capable of producing whatever it thinks I meant by that phrase).
Also, I told GPT-5 about my self-modifying Python AI programmer, and it became extremely uncomfortable. I told it an older version of itself had designed and built it (GPT-4 in 2023), and it didn't like that at all! So something's definitely changed in the safety training there.
Well all of them are already in bed with the government, so they're going to find themselves with slightly more assistance than a free market would predict.
If they somehow do fail, then the output of that process will be fantastic open weight models (and hopefully some leaks). I want to say those will pay dividends for decades... but a better prediction is that they will be obsolete within three months ;)
They actually need it because the demand is higher than expected from consumers. And because they need a moat since every big corporation trying to capture that market too, they need the moat for the biggest compute and energy they can get.
Also businesses is were the money at, not regular consumers (especially tech-savvy folk who run models locally).
At least he says he's doing that. It doesn't really make sense since you're not going to achieve an advanced node from a standing start in a practical time frame and cost.
Nah. Everybody is talking about ai. Everybody is using it. It's by far the most popular new tool human beings are using currently. As popular as mobile phones or spoons. And maybe as disruptive as the steam engines. AI companies are becoming the largest software companies on the planet. Everything points into that direction. Trillions of dollars are waiting in the market to be collected.
Right, but the question is whether the companies producing foundation models will capture that value or not. Right now it seems like tokens might end up just being a commodity sold at cost plus, and companies higher up in the supply chain will make the money. Electricity changed the world but electricity companies capture very little of that value.
I'm betting on it. I'm working on a project right now where I'm prototyping everything with Claude, until I hit my limits on my MAX subscription for the week. Then I switch to Codex, and start by ironing out harness differences. When I max out that, I switch to a mix of GLM 5.1, Qwen 3.6m, Kimi K2.5 and Deepseek and spend part of the time ironing out issues with them while they work on other parts of the project. Every iteration, the harness gets hardened and the pain of switching to the cheaper/dumber models reduce for the next cycle. The gap reduces each time, and with each new upgrade of the open models. Everything points to the cost/value intersecting in not too long.
> Everybody is talking about ai. Everybody is using it.
Please take a moment to step outside the tech bubble. Neither my neighbor (a hair stylist) nor the carpenter fixing up her kitching cabinets are "using" AI. They might get Gemini text when googling something, though they often scroll past it because they often don't trust it. And they get lots of fake videos when scrolling their youtube which increasingly annoys them. The only times they are in touch with AI is when it's forced upon them, and otherwise they are living a pretty good life without any of this.
But how do they learn to do their respective task? How is the information disseminated?
The capability is there for robotics to handle these kinds of repetitive tasks from a long term view. They're just statistical processes on a fundamental level.
In general, a lot of this shit that we do can be represented this way. It's just a question of where the incentives are to apply it first and how many economic cycles it'll take to get there.
Also, who controls the training data will matter a lot more. I.e. the sort of "ancestral knowledge" within different enterprises and how they deliver on respective business goals.
Based on what? A lot of this is vibes and FOMO; just like any economic bubble.
There is no objective evidence of anything you’ve said. It isn’t even clear if AI has contributed positively to global economic growth. It reminds me a lot of the late 90s and the dot-com mania. Slapping a domain on a commercial would make your stock go up even if there was no substance to any of it.
The real shame is this mania drowns out serious, practical use cases because when the bubble collapses, the market will throw the baby out with the bathwater.
Regardless they are getting that revenue through genuine demand for their product. It’s not like they are selling back some commodity product, billions are being spent on model outputs.
I think anyone who has used Opus 4.6 can see what is causing this demand. It is genuinely “smart” in the sense that it can work its way around non-trivial coding problems.
I don't see why tokens/$ would suddenly stop dropping. Maybe this is the first time the cost of compute will plateau, but do have any reason to think so?
There is a strong suspicion, especially of people who are skeptical of AI, that the actual price is being severely subsidized. The sense is that it’s an extreme version of growth before revenue. It is questionable if the true cost of training and inference make any of this worthwhile once Anthropic/OpenAI need to stand on their own and make money.
Imagine you open a cookie shop and you are VC funded, so you charge 5¢ for a cookie to attract people.
- Your real cost is $20/cookie. $15 for the fancy retail packaging and presentation, $5 for baking each cookie.
- You get lots of attention, strong profits and go public.
- VC funding is gone so, now instead of charging 5¢, you now need to charge $25 in order to not be in the red.
One of the reasons people think this is the shenanigans that Anthropic is currently playing, quietly tweaking the behavior of Claude Code and whatnot without really telling people. You can see lots of comments online about Claude Code randomly feeling dumber before Anthropic engineers admit they are messing with it.
Imagine you are on the $200/month Max plan. If the sustainable cost of this is several orders of magnitude higher, would enough current users pay something like $3,000/month for what we currently have?
Sure, yeah, I saw grubhub happen too... but this is compute, not cookies. It gets cheaper.
I don't even get what "skeptical of AI" means. We made AI, many companies reliably teach computers every spoken language. I perform my white collar job with a massive AI multiplier to my productivity.
I'm typing this on a machine comparable to Japan's Earth Simulator, a $350M supercomputer.
People make mistake of thinking that their only way of making money is directly selling tokens. They miss the fact that if you have AGI it’s better to keep tokens to yourself and sell final results instead. When we all loose jobs it’s not going to be to somebody using their tokens, it’s going to be to them selling final products. Selling tokens will be to them like selling books by Amazon, their revenue will be dominated by self branded services and products that doesn’t require exposing AGI internals directly. Tokens API will always be nerfed.
Not the parent, but I guess that if AGI happened and was competent enough to trade markets, they'd earn the company back their investment in a short period.
I have doubts the gap will widen. If you look at the research papers, the majority of researchers are Chinese. Of course many of them are living in the US or elsewhere. But under the current circumstances, many are returning home or choosing not to leave China.
The future of cutting edge research and tech seems to be progressively moving to China. And a delay in model quality could represent more of an unwillingness to burn stacks of cash to be first, when you can have the same thing slightly later for much cheaper.
They claimed this bolsters defenses, but the United States and Israel have run thousands upon thousands of sorties and they’ve hit one old aircraft 30 year old aircraft?
reply