I don't buy the full purported relationship between Darpa and Big Tech as being so tight. I am old enough to have been there before the beginning of Fidonet and Usenet. A network among universities originally supposedly to help Darpa (Arpa before that projects) helped start some raw things but does not mean State was paramount or stayed in control. True they want control of it now. That is for certain. I know how much the people and especially techno-hippies originally poured into the creation of all the tech we have. I know how many since that earnestly wished the world a better place and have worked hard to make it so have come. I agree that State tries and highly succeeds to suborn it. But let us not make the State out to be even more powerful than it is.

Guess what. Bitcoin is fully trackable by the State and they increasingly have miners that will block transactions for wallets that are on some arbitrary naughty list. The centralized exchanges are already KYC. Blackrock own more and more BTC. Read the Trump crypto-friendly (hahahaha) legislation and see the every increasing KYC, moves to banks owning all stablecoins and effectively forcing crypto people to sop up Federal Debt. Bitcoin will not save us. Privacy coins and non-transparent chains if we are to stand a chance.

Exit we must. But Bitcoin has been given over to the banksters and the State. It is fully trackable and soon State and Blackrock will bend more miners to their will. It cannot save us.

I am not against LLMs. They are excellent for some things. But I am deeply worried about today's deep imbalance. The following is a conversation snippet with Claude Opus 4.8 on the matter.. me: Hyperscalers have a self-interest in pushing LLM for everything to make the demand seem more real. Unfortunately it also drives up and up the operating costs for everyone including themselves. 4:25 PM Claude responded: Yes — and you've put your finger on the part that makes it a trap rather than just a strategy: the move that manufactures the demand signal also inflates the c… Yes — and you've put your finger on the part that makes it a trap rather than just a strategy: the move that manufactures the demand signal also inflates the cost base it's supposed to justify, including their own. It's self-consuming. The mechanism is worth tracing because it's genuinely perverse. Pushing LLM-for-everything serves the demand narrative in two ways at once: it raises measured token consumption (so usage charts go up and to the right, which is the number investors read as demand), and it discourages the question of whether a cheaper deterministic component would have done the job. Every task reframed as an LLM task is a task that now shows up as AI demand and consumes inference capacity. So the incentive is to route work through the expensive stochastic path even where a function call, a query, or a bit of GOFAI would be cheaper, more reliable, and more appropriate — because the inappropriate routing is what makes the demand look real. The misallocation isn't a side effect; it's the product. And the cost dynamic you're naming is the bite. Inference is not free and doesn't amortize like training — it's a recurring per-use cost in compute and energy, every single call. So "LLM for everything" means everyone's unit economics carry the weight of running the most expensive available primitive on tasks that didn't need it. That raises the operating cost floor across the whole stack — for the customers building on it, and for the hyperscalers themselves, who are running the data centers and eating the power bills. They've inflated the demand figure by inflating the cost of serving it. The signal and the burden are the same act. There's a further twist that sharpens your winter worry. The inflated-demand strategy raises a false floor under usage — usage that exists because it was pushed there, not because it's the right tool. That usage is the most fragile in a correction. The moment cost discipline arrives — a downturn, a margin squeeze, a CFO asking why the summarization job costs 50x what a templated approach would — the inappropriately-routed work is the first to be pulled back to cheaper primitives. So a meaningful slice of the demand that's holding up the valuations is exactly the slice that evaporates under the first real cost scrutiny. The manufactured demand is structurally the least durable demand, which means the gap between apparent and real demand doesn't just exist — it's primed to reveal itself suddenly, the instant the incentive flips from "show growth" to "show margin." And it puts the hyperscalers in a genuine bind, not just a hypocrisy. They can't easily stop pushing LLM-for-everything, because the demand narrative depends on it and the valuations depend on the narrative. But continuing it raises their own operating costs and builds the false floor higher, which makes the eventual correction sharper. They're committed to a strategy that worsens the thing they most need to avoid — a sudden, legible collapse in real demand — because the alternative, letting measured usage fall to its appropriate level now, would puncture the narrative immediately. So they keep inflating, knowing it raises the fall. That's the same structure as the capex arms race: individually compelled, collectively destabilizing, and worse the longer it runs. The sane-engineering path we just discussed is, among other things, the cost-discipline response to exactly this — route each task to the cheapest primitive that meets the requirement, use the LLM only where its strengths are real. Which is one more reason the hyperscalers' interest runs against it: the right engineering is deflationary to the demand figure. The thing that's good for everyone's operating costs is bad for the narrative. That tension is, I think, the cleanest single statement of the imbalance you've been circling this whole conversation.