2024-10-23

Last modified: 2024-10-24 19:36:31

• Sovereign Scales
• CAM Workflow
• LLM Steganography
• Stuff to 3d print

Sovereign Scales

The closer one has been in the tumbler overnight, the further one has had no tumbling.

There's not a massive difference, but the tumbled one is nicer.

I tried balancing them both on the edge of a plastic pot with the quarter sovereign in, and they both balance the wrong side of the pivot hole! So they can't be calibrated by removing material from the counterweight. This is a blunder.

I think the explanation is that the polished one has the counterweight slightly short, because of a coordinate system error (?), and the other one has the "beam" slightly too thick.

Probably I should work out how many of these can be made to work, and then make matching stands, tumble everything, and see what Paul thinks? And then see about making a sheet-metal version, with either a bolted or folded counterweight.

Well I definitely know that I'm going to want to make a sheet-metal version, so I can buy a sheet of 3mm brass now and work out what I'm going to cut out of it later. I paid £31.20 for a 300mm * 200mm sheet, that's the same size as the machine table.

Also a 3mm aluminium sheet the same size, £8.10.

I'm thinking that I'll manually drill mounting holes in both sheets, and bolt them both down, using the aluminium sheet as a sacrificial layer to cut into.

If I do the "folded counterweight" idea then I'll first use the V-bit to put score marks in the back side of the material, then flip it over and machine from the top. And I'm not super sensitive to the alignment of the score marks, as long as the axes are aligned so the bends are straight. And if I come up with a suitable bolt to use as a counterweight, then I can just use that.

I wondered about making/acquiring an abrasive sanding pad for the CNC machine and using it to grain the top surface?

For making stands like this:

I think we clamp the bar in the vice on very-tall parallels (<1mm clamped), and surface off the top, and cut out 6 tessellated stand shapes almost all the way through, and drill the first pivot hole. Then separate them into pairs using the bandsaw. Then surface off the other side of 2 at a time to separate the pieces. Then machine the slot one at a time. Then drill the second hole.

So the plan is:

• clamp bar in vice with <1mm in jaws (tall parallel + shim?)
• load 6mm tool
• X0 at left edge; Y0 at top edge; Z0 at top
• surface top: surface-top-6.nc
• cut out 6 stands almost all the way through: outline-6.nc
• load 1mm drill
• drill 6 pivot holes: pivot-hole-6.nc
• separate into pairs with bandsaw
• load 6mm tool
• load a pair of parts in the vice
• surface off the top down to 14mm total thickness: surface-top-2.nc
• repeat for all 3 pairs
• load a single part in the vice
• X0 approx. at left hand edge; Y0 in centre; Z0 at top
• machine the slot: slot.nc
• repeat for all 6 parts
• load 1mm drill
• load a single part in the vice
• X0 at left; Y0 at top edge; Z0 at top
• drill the hole: drill-1.nc
• repeat for all 6 parts
• deburr by hand
• vibratory tumble

Gcode needed:

• surface top for bar of 6 parts: surface-top-6.nc
• cut out a pair of stands, gcodetile 3x: outline-6.nc
• drill a pair of pivot holes, gcodetile 3x: drill-6.nc
• surface top for pair of parts: surface-top-2.nc
• cut out a single slot: slot.nc
• drill a single hole: drill-1.nc

I made quite a lot of blunders in this, only endedu p with 3 passable stands, none "perfect".

Here is a fully-assembled one: https://www.youtube.com/watch?v=qRj19s1ggK8

The centre of mass didn't end up where I wanted it, it was the wrong side of the pivot, and by a long way! Which meant that to balance it I had to drill a load of holes in the side that the coins go in instead of one small hole in the bottom of the counterweight...

Also I didn't get the pivot hole very perpendicular so the pivot is slightly wonky.

Also the slots are too narrow.

This whole thing is just a long series of blunders.

Ruari had an idea: stand the coins vertically! That way you can pack the much closer together. We also get the thickness test and diameter test for free in one go. This one would be a big hassle to machine, but would be perfectly easy to 3d print.

CAM Workflow

I should write about my CAM workflow, and make it more rigorous.

Tooling:

• FreeCAD
• Inkscape gcodetools
• Meshmill
• gcodetile
• gcodecat
• custom programs

And then each project has a directory for each part, and the directory contains a "master" CAD model of that part, plus a separate CAD model for each CAM job, named after the CAM job (so my-part.FCStd for the main part, and top-surface.FCStd for the top surface, which has a Path Job to create top-surface.nc), and then a shell script to create the "final" gcode collected in the "out/" directory.

Techniques:

• the individual Job files have the part moved into the same coordinate system as the Gcode
• for special toolpaths, you can model a subset of the geometry required and make the Body 50% transparent so you can see the "real" part through it; or otherwise modify the model in the individual Job files
• you can model the vice jaws, clamps, etc. to make sure you won't hit them
• import the "final" gcode for each operation into a FreeCAD project to check that it all looks like it will work
• you can create multiple Path Jobs in one FreeCAD project, but switching between them is inconvenient, recommend avoiding; if you must do it, you need to toggle off the Job, and its Operations, and its Model, to make a Job completely hidden
• use a checklist to make sure you select the right tool, feeds/speeds, coordinate system, postprocessor on every job
• don't bother building up a tool library in FreeCAD, just change the "TC: Default Tool" in each job; you only need to set the tool diameter, horizontal feed, vertical feed, spindle speed
• when you create a Job, toggle off the underlying model and toggle on the Model inside the Job, to make sure you are looking at the right one (but if you put the part in the right coordinate system it oughtn't make any difference)
• write yourself a bullet point list of instructions for how you're going to make the part before you even start on the CAM; the list should say when to load tools, which tools, when to load workpieces, and how, and what the reference edges and coordinate systems are; then use this list t work out what gcode you'll need; then for each gcode file you need work out whether you're going to write it manually, or generate it with some CAM tool, or a combination of the two; then make each file, going back to make your list of instructions more precise as you go; then make your parts
• put your project in git, even the big binary FreeCAD files; it's fine to commit copies of gcodetile and gcodecat
• at the end of each job make it move the tool away from the work so that you can load/unload workpieces without having to manually jog the tool out of the way
• get Cursor to help you write the custom programs
• you can select more than one Body to include in a Path Job

LLM Steganography

I want to generate some example texts to complete "Once upon a time, "

• always take rank 0
• always take rank 1
• choose randomly between rank 0 and rank 1
• always take the worst token with probability >1%
• always take the worst token with probability >5%

Could we make something like https://tokenvisualizer.netlify.app/ but it tells you the probabilities and ranks etc. instead of just the token boundaries? Also, that tool suggests the tokenisation is not unambiguous, if both "L" and "LM" are real tokens?

All of these examples are completing from "Once upon a time, " with Llama3.2-1b on a RTX 3060 12GB, and excluding special tokens and non-ASCII tokens.

Always take rank 0

userankn.py with n = 0.

Once upon a time, 20 years ago, I was a young, single, working woman. I was living in a small town in the Midwest, and I was working at a small, local newspaper. I was a reporter, and I was covering the local government. I was covering the city council meetings, and I was covering the city manager meetings. I was covering the city manager meetings, and I was covering the city manager meetings, and I was covering the city manager meetings, and I was covering the city manager meetings, and I was covering the city manager meetings, and I was covering the city manager meetings, and I was covering the city manager meetings, and I was covering the city manager meetings, [...]

This gets stuck in a loop. The writing has a blandness.

Always take rank 1

userankn.py with n = 1.

Once upon a time, 10, 000 miles from home and in the midsts a blinding storm. The sky is black, and there are flashes and thunder, but no wind, no hail. I was on my bike and the rain came in a steady, heavy stream, but it didn&apos;nt stop my ride. The wind came in gusting at times and I had a few moments where the wind blew my hat away, and the wind was cold, and the wind made the trees shake, but the wind was not strong enough to blow me off my bike. The rain was heavy and the wind was strong, but the rain did not stop my ride. I rode on and on, and on, and on, and on, and on, and on, and on, and on, and on, and on, and on, and on, [...]

This also gets stuck in a loop. The writing is more exciting than the rank-0 writing.

Choose randomly between rank 0 and rank 1

01rank.py.

Once upon a time, 10, 20 or 50 miles from a city centre, you could still find a pub. But not any pub, mind you. The pubs that survived were those with character, with history and with a real sense of place. Now, with a few exceptions, those places are long-gones, their owners having sold up and moved to a new life in the suburbs. The pubs that remain, the ones with the real ale and a real atmosphere, are often in the most unlikely places, [...]

This doesn't get stuck in a loop because of the random element. This feels more like "real" writing than either of the first 2 example texts.

Always take the worst token with probability >1%

abovethreshold.py with threshold=0.01.

Once upon a time, 1993 and in 1961 I moved my mother (her daughter's child and then child in 4 generations in two and in six steps away, then three generation ago. Her grandparents on mother side married cousins with parents 40 steps, in between in her paternal greatgrand parents). When in that step removed me the parents to each step was born one to seven siblings of all. (And two steps further was another ten years or in all twenty-seven and my step-grandfater of four [...]

Word salad.

Always take the worst token with probability >5%

abovethreshold.py with threshold=0.05.

Once upon a time, 10,000 years ago in a place known as Mesopotamia (which was later named Sumer, Akkade and finally Iraq) people started building the world we live in. They had the idea to create the first city in order to live there, so that they would never get hungry. But, as soon as the people had a house they realized it wasn't a very safe place. The house had walls made from wood. So the first city had walls. Then people had the [...]

Almost believable text.

We have:

• rank 0: 14 tokens
• rank 1: 23 tokens
• rank 2: 24 tokens
• rank 3: 26 tokens
• rank 4: 9 tokens
• rank 5: 4 tokens

Stuff to 3d print

• better spinning top for Lucy
• swiveling vice stop, bolt down to table
• house model
• gingerbread man cutter
• VFD cable cover
• tablesaw lifting gear
• spanner for ballscrew locknuts