they don't have to make a profit on the investment. The investment is defensive. The goal is not to be outside the moat.

instead of explaining to my boss why he caught me spinning on my chair and stacking office supplies into Jenga towers as "I'm waiting for the compiler to finish" I can now tell him I'm waiting for claude to get back to me.

I was looking forward to tokenising the Swiss citizen. There would be fewer of them than bitcoins and not growing in populations, so by the same logic bitcoinETF use they would appreciate. And they are even backed by gold. Dang.. I'd have made a fortune.

Last I checked, networking was a four-letter word, and those letters were CCNA.

Especially when this particular "anything new" is the subject of pending copyright litigation. See Doe v. GitHub.

Probably to the effect "We failed to poach the people who developed this framework in the first place. Can we import some immigrants, pretty please?"

How long can you trust that the operator of the server offering a particular LLM wll let you continue to choose a seed?

There is no reason bitcoin can't slide back to being worth a dollar a coin. There no guarantee of value behind it. You can argue about whether fiat money is any better once the US went off the gold standard, but there is still a bit of irreplaceable value there; "the full faith and credit" means you can use it to transact business with the government, both by contracting to do public works for which the government pays you, and paying taxes and fees which are used to perform government functions and give you tokens such as licenses which show that you've contributed your share. Money's very liquidity for these purposes is a source of value, even if you can't redeem your picture of a President for precious metal. By contrast, Bitcoin literally isn't worth anything unless you can find someone (for some reason the phrase "bigger fool" comes to mind) to trade you something for it that does have value.

A compiler is a deterministic process that runs locally on a modest home computer in reasonable time without needing an expensive NVIDIA GPU, doesn't regurgitate memorized copyrighted code from its training set, and doesn't boil the oceans in a datacenter for training plus use.

I'm guessing gift cards were handier for unbanked and underbanked people, such as teens who earn a cash allowance or earn cash from yard work for neighbors but aren't old enough to have a bank account in their own name.

Why did you not list their lack of profits? Nobody cares about revenue growth, it looks impressive but is made up nonsense full of circle jerk deals among AI companies. Investors care about ROI and don't like capex unless there's a big payday coming very soon.

Here's a better number: Anthropic announced their first ever operating profit just ten days ago for Q2. Very suspicious timing given the IPO. And interestingly the profit claim happens to coincide with a compute discount for said Q2 from Elon. No discount for Q3+Q4. Elon has to make his profit claims.

Like I said: Circle jerk all around.

I stand by what I said. The *hype* IS dying. Investors are getting impatient.

double rot13 is the way to go!

Google and Anthropic are perfectly able to build their own compute-centers, but it may not be economically a good idea at this point. Remember, the AI hype is dying and the investors are starting to smell a rat. Leasing GPUs from Elon means Google and Anthropic are not stuck with tons of hardware they can't monetize in the future. If things go bad, they just stop paying for access and Elon is left holding the bag.

It's important to distinguish the model from the generated sequence dynamics. The "model" (in the traditional scientific sense) is a fixed function of the input data. It cannot improve itself, except when the scientist replaces it. But that's not improvement, that's revision.

Most regular people don't understand the word model, and think it refers to their interaction with the tool. That can indeed evolve, in the same way that if you repeatedly run an input output loop with a given transformation, and add some random perturbations (aka user control), you will get a trajectory that evolves. It will probably converge to some fixed point, but that could take a long time to detect by users, who think it's evolving before their eyes.