2024-10-07

Last modified: 2024-10-08 09:39:26

• Sovereign Scales
• Cursor devs on Lex Fridman

Sovereign Scales

Just thinking some more about how we'd make these.

I'm thinking we surface the "top" of part 1, and machine in the features and profile of the top side, then flip over and do the same for part 2, making sure the surfacing brings the overall piece to 15mm thick.

But the tricky part is aligning part 1 and part 2 in the Y axis, because once you make the profile for part 1, the Y reference surface is lost. And if you don't take great care to align them, then it won't sit nicely flat later on when we come to separate the two pieces.

So to make these parts, I think:

• load 6mm end mill
• chop the metal into chunks large enough to make 2 parts (150mm long chunks)
• set a vice stop that roughly centres the chunks in the vice
• run "top-side-1" on every chunk
• set the Z offset so that the top surface will end up 15mm above the parallels
• set the Y offset so that the "part 2 top side" will be aligned to "part 1 top side", pick it up by centering on the "other" counterweight
• run "top-side-2" on every chunk
• load drag engraver
• run "engraving" on both sides of every chunk
• load 1mm end mill
• run "slots" on both sides of every chunk
• set up a fixture for clamping the pieces down on their side in a repeatable place
• pick up coordinates by pivot hole locations?
• load 1mm drill
• run "drill-both-pivot-holes" on every chunk
• load 6mm end mill
• run "separate-parts" on every chunk

And the stand needs I think 30mm long chunks to make 2 parts, so in total we use 180mm per 2 parts of both types, so in 1 metre we get 5 chunks, or 10 parts.

Gcode needed:

• top-side-1: form the profile of the entire part, making sure to go deeper around the counterweight, and make pockets for coins, also square off the front and back edge almost all the way down to the vice jaws, but stay clear of the 2nd part's geometry
• top-side-2: same as "top-side-1" but don't bother squaring off front/back edges any deeper than necessary
• engraving: engrave text with drag engraver
• slots: make slots with 1mm end mill, mind that part of it needs to cut at the very top surface, even though most is lower down
• drill-both-pivot-holes: use M1 optional stop so that I can check locations?
• separate-parts: machine a big profile/pocket in the centre of the material, and then gingerly escape out the sides to separate the parts, where the workholding is weak

So this evening if there is time I'll try to start on the gcode, and tomorrow I plan to start on making these.

OK, I've worked out what to do about alignment.

In the first part "top-side-1" we square off the left and right end of the material so that the reference surface is consistent.

And we have the material sitting on tall parallels, and we machine down the front and back edge almost down to the vice jaws, so that they're consistent too.

And then yeah pick up the new Y zero by finding the centre of the other half's counterweight.

Cursor devs on Lex Fridman

Following on from yesterday.

They talk about storing the KV cache on the GPU, and they talk about it in terms of "requests", and Cursor is fast even on machines without CUDA. Does that mean that while you're using Cursor there is a GPU in the cloud that is basically dedicated to you? Or at least some fraction of a GPU, if you assume one GPU has enough memory to store N users' KV caches. Is this really costing less than $20/mo per user for them to run? Or is it just burning VC money?

As good as they are at writing code, they're surprisingly extremely poor at finding bugs in code, if you just ask them to find bugs.

They're talking about different programming contexts as "calibration". Like the difference between how acceptable bugs are in a quick experiment versus in actual production code in Postgres or Linux. That's something I've been thinking about a lot lately as well. They're pointing out that (good) human programmers tend to "know" the context and are calibrated accordingly, but the LLMs don't have that holistic knowledge of what they're doing.

They use AWS for hosting, think it's really good.