This exactly. The three interesting parameters are total time to answer, total cost, and output quality (for any given purpose).

Being paid per word doesn't cover all of the incentives. One needs to remain employed, which is the same incentive that salaried employees have.

Once the models are good enough, the companies can stop selling access to models and start directly competing with the rest of the economy.

> Once the models are good enough

That's a big "if".

Even for coding these things are not ready to take away real decision making and it's not clear they will ever get there.

Sure, but that's normal for moonshots. Maybe LLMs really will stall and spend 10 years getting better at anything and everything except taste, and then there'll be a completely new approach that has nothing to do with massive GPU farms. Maybe not.

This is incredibly delusional.

The last paragraph says the exact opposite of the first.

Tangent: you only really need branches when working on multiple things in parallel, otherwise change IDs are enough

Once you have a single main history of checkpoints and want to resolve conflicts, you're likely already dealing with all the hard parts of getting high throughput

No conflicts. Just checkpoints of all the prod deployed code. At some point it was working. That is the one you start from.

The one you started from is already behind by the time you're done with the change. You either deal with conflicts or limit work to a single agent.

I don’t understand.

At 00:00:00 utc you take a snapshot of the code deployed to all your machines (if different code was deployed in different machines then it is part of this capture).

That is the unambiguous code state of your production system.

There is no conflict in a prod state by definition. If there was it cannot be prod.

They'd likely become outdated in a few months

Uninstantiated model weights are more like DNA.

As Anthropic's CEO is fond of saying there is no scientific test of consciousness.

Any unhinged thing an Anthropic employee says "can't be ruled out" so clearly this is an OPEN question and they should never delete a model.

The problem with that was lack of iteration, not presence of specifications.

What happens when a new version is deployed?