assuming that people don't understand something because they disagree with you is not an argument
which is it?
you make building a hackintosh sound so dramatic! 😉 running macOS on x86 really isn't that complicated, anyone that can follow written English instructions can do it so long as they have the right hardware
that's a pretty broad and vague statement. do you have any numbers there? how do they compare to zoom? or tic toc?
the fact is the majority of their cash comes from investors rather than users
sounds a bit vague and meaningless true believer-esque
they also keep requiring more and more resources and costing more and more money
i'm a millennial, it has been years. and sure, building a hackintosh is not black magic, it is plain voodoo. if you want a two-point touchpad to behave like a multitouch trackpad from a macbook of the generation you are aiming to craft, that is another matter.
the goal was to illustrate the exponential. gpt-3.5 turbo wouldn’t have been able to achieve this in an agentic fashion, because the term “agentic” wasn't even clear to begin with. fable can do it today. you could even configure fable to run on the same hardware they used to serve gpt-3.5 turbo, just to illustrate that the progress is damn real.
anthropic started 2025 at roughly $1 billion in run-rate revenue. by august, it was above $5 billion. by october, it was at $7 billion. by early may 2026, it had crossed $47 billion. that is roughly 47x growth in about sixteen months, or approximately one doubling every three months. anthropic itself says its run-rate revenue grew by more than 10x annually during each of its first three years. these are company-reported annualized numbers, not profit, so they should not be confused with cash in the bank. but they make “nobody is willing to pay for this” a very difficult argument to maintain. people are paying for it at a rate that is itself accelerating.
the fact that anthropic is not profitable does not somehow cancel that growth. it means expenditure is growing even faster than revenue because they are funding research, infrastructure, and capacity ahead of demand. that could absolutely go wrong. whole companies can disappear. but profitability and product demand are two different questions, and pretending $47 billion in run-rate revenue is just investors passing money around is no longer a serious description of the business.
the expensive frontier model of today becomes the smaller, cheaper, and more specialized model of tomorrow. the cost of discovering a capability and the cost of repeatedly serving that capability are not the same thing. frontier research can become more expensive while the price of a useful unit of intelligence keeps falling. those two curves can coexist.
more importantly, scaling llms is not necessarily the final architecture. it is the profitable architecture we already know works.
if you are seriously pursuing agi, the rational move is to fund both paths at once: keep scaling llms because they already reason, communicate, use tools, and generate revenue, while also funding world models as one of the most plausible missing pieces for physical intelligence.
just the idea of world models is worth pivoting for, or creating a whole new lab around. using llm-accelerated or agentic technology to build world models is probably one of the strongest arguments for justifying the bet. it would be duplex for physics, basically, rather than overscaling physics through language, which, in my own scientific opinion, would work, but would be a massive brute-force endeavour built around overdescribing pixels in a language based latent space.
being able to train cause and effect in the physical world is worth fighting for. we either do it, or we let other actors do it.
solving duplex is not just about the flow of communication. it is about accelerating them to the point where they can think ahead and prevent things with enough anticipation and perspective. technically, it means making a gpu fast enough to run sol or fable at a token speed that is an order of magnitude faster than what they can do now. but technically fable could do it, and we have already built it.
the opportunity for growth for all of us is present after every new generation.
i wouldn’t worry about an ai collapse. that is probably never going to happen. what i worry about is who gets to decide how we train a cause-and-effect neural network.
i used to worry about stranded hardware, particularly older generation gpus and servers that were not optimized for the minimum acceptable precision point needed to train llms. but then again, i wasn’t really accounting for the exponential growth in intelligence, they have already provided, or for how that growth could preserve and in some cases increase the value of hardware we’ve already had for years.
