In addition the graph "Massive VRAM Savings" graph states what looks like a tautology, reducing from 16 bits to 4 bits leads unsurprisingly to a x4 reduction in memory usage

It seems quite logical to me as well. An LLM is not a logical computing system but it has the knowledge on how to do a multiplication

The exact definition of a useful heuristic, "good enough"

"Write a protocol and GraphQL", god damn it escalates quickly.

Fortunately, there are intermediate steps.

Any suggestions for a good RPC library?

I have had a really good experience with https://connectrpc.com/ so far. Buf is doing some interesting things in this space https://buf.build/docs/ecosystem/

I've used twitchtv/twirp with success. I like it because it's simple and doesn't reinvent itself over and over again.

aws, azure and gcp aren't cheap for sure, but for sure they are doing a lot of money

Definitively agree. Being blunt, too direct is just the opposite of a good and effective communication.

Being too blunt raises defenses and completely wipes out the effectiveness of your feedback. Folks that are invested in outcomes make choices for good reason, and they've probably got a track record to back it up. You have to meet them where they are, and considering their communication styles and how they make decisions will improve the chances you're actually heard.

It's a fact of life that people shut down when approached with evidence that refutes their world view or choices. It doesn't matter if it's your boss or grandparents.

> It's a fact of life that people shut down when approached with evidence that refutes their world view or choices.

I don't really agree with you. This is a basic quality of skilled leadership. You want people refuting your worldview with evidence! It lets you correct course and make things better.

Only insecure people shut down like this in my experience.

Batteries cannot sustain the electricity consumption of a whole nation. Maybe in many years, but right now, no sign that it can, whereas plant can

Nuclear plants can't be built instantly; the growth patterns for batteries suggest that by the time new nuclear comes online, there will be enough batteries.

(And I'm saying this as someone who actually likes nuclear, thinks the public perception of the danger is over-stated, and who thinks everyone should have build a lot more reactors decades ago).

Genuine question: would these batteries be always connected to the grid with enough power and capacity to feed the demand?

Where the batteries end up and how they're used heavily depends on policy decisions between now and then.

I'm expecting most batteries (worldwide) to be used for cars, with bi-directional power so they can function that way (for grid storage) when plugged in. That doesn't say much about what any specific country will do, and as Texas demonstrates, US states can have their own independent energy policies.

I had assumed that Datadog was using Elasticsearch as a storage solution, including for metrics. Seems not.

You don't need event sourcing to organize payrolls event "at scale".

From nothing to result rarely happens in real live. I hardly see someone to start/stop a program per unit task (like piping commands).

Code compilation is a common example. In many languages, when you do a clean build of a large project, you run a large number of short-lived processes.

With command line tools, the dominant design philosophy is that a program completes one task and exits. The tasks may be large or small, and the same tools are often expected to handle both.