Independent Researcher. Founder of Antinomia — a PKM system built around contradiction, not storage. Mapping how thinking changes over time. | The Deeper Layer
Joined May 2022
- Tweets202
- Following781
- Followers501
- Likes449
I photograph a page of a textbook. About 3.5 minutes later it is an interactive 3D lesson.
The AI does not draw images or make video: it writes the lesson script, and 18 hand-built 3D "theatres" stage it on the device. 🎥🧵
Built solo in Italy by orchestrating Claude Code agents: I made the decisions, the agents wrote the code, every task was reviewed. About 1,300 tests.
Open source, MIT.
1
9
It is early: tested on only a few real pages. Teachers, students, publishers: I would love your feedback.
Try 10 lessons in your browser: demartinogiuseppe.github.io/…
Code: github.com/demartinogiuseppe…
10
They say simple is best, so I kept it simple.
A Chrome/Edge extension that copies browser stereo to the rear, side, center and sub of a 5.1/7.1 setup. Four switches, no drivers.
Got a 5.1/7.1 rig? Try it:
github.com/demartinogiuseppe…
ALT Promotional card for the Surround Fill 5.1/7.1 browser extension. On the left, large text: "Surround Fill 5.1 / 7.1. Four switches. No drivers." On the right, the extension's popup showing "Output detected: 5.1" and four checkboxes: Rear (L/R) checked, Side (L/R) greyed out with the note "7.1 only", Center checked, and Subwoofer, "bass < 120 Hz", checked.
Day one: two releases from replies to this thread.
@wazX_7 asked about secrets already in old commits → 0.8.3 searches the git history.
Then: the warning should say rotate the key, because deleting a pushed commit doesn't make it safe → 0.8.4.
Keep them coming.
ALT idle-mod pane in Claude Code listing two possible secrets found in old commits, a GitHub token and an AWS access key, each with file, line and commit, ending in "rotate it if pushed", with Fix with Claude and Ignore buttons. Below: once a commit has been pushed, taking it out of the history doesn't make the key safe again.
Claude writes most of my code now. The part I still worry about is what it leaves behind.
So I built idle-mod, a small pane in Claude Code. If Claude tries to commit a key, the commit doesn't happen. Claude is told the file and the line, never the value.
Something I learned building it: agents write files with the shell far more than with their edit tools. My first version only watched the edit tools and missed a key written with printf. Now any file changed during the session counts, whatever wrote it.
1
2
19
It runs locally, makes no model calls, and it's free and open source:
github.com/demartinogiuseppe…
If you work with coding agents: what do you check before you commit what they wrote?
18
Giuseppe De Martino retweeted
The Commodore 64 hardware simulator is now online: github.com/antirez/c64-hw-si…
A transformer with 688,128 weights runs on an unexpanded Commodore 64.
1 MHz. 64KB RAM. No memory expansion.
One word every 20 seconds.🧵
1
96
The weights are ternary {-1, 0, +1}, 2 bits each.That’s not for compression. The 6502 has no multiply instruction. With ternary weights a matrix-vector product becomes add, skip, or subtract.The quantization isn’t an optimization. It’s the only reason the math runs.
4 layers, dim 128, 8 heads, MQA, 512-token vocab.
168KB of weights in a banked cartridge.
Everything after training is int16 fixed point. Lookup tables for SiLU, exp, 1/sqrt, RoPE.
Zero floating point anywhere in the inference path.
One result I didn’t expect:dim 64 ternary → 2.25 val loss
dim 64 fp32 → 1.81
dim 128 ternary → 1.70At small scale ternary doesn’t cost accuracy, it costs capacity. You pay it back in width before the memory saving is free.
The kernel went from 55 cycles per MAC to 20.Biggest win came from deleting code, I’d left 2048 real multiplications inside a table-build routine, in a project whose entire point is avoiding multiplication.Found it only by timing the phase that looked innocent.
Source, weights and the .crt attach it in any emulator with EasyFlash support and it runs:github.com/demartinogiuseppe…
1
53
I got tired of rebuilding my Claude Code setup on every machine, so I built plugdrop.
/plugdrop:export snapshots your plugins, skills and CLAUDE.md.
/plugdrop:import lets you pick what comes back.
It never overwrites anything. Free and open source. 🧵
6
4
220
Plugins are reinstalled from their own marketplaces, git-cloned skills are cloned again, and your CLAUDE.md only gets the blocks you choose, appended after a backup.
Snapshots go to a private GitHub repo, any git server, or a synced folder.
1
51
Repo: github.com/demartinogiuseppe…
Install from inside Claude Code:
/plugin marketplace add demartinogiuseppe/plugdrop
Tested on Windows only so far, so macOS/Linux feedback is very welcome. Independent project, provided as is.
29
Claude shipped Dashboards this week. I used it to build something a bit different: 3D production lines for an aircraft plant, a shipyard and a rocket factory. One day, invented data.
Input: 4 CSV files per site. Weekly KPIs, stations, delay causes, an NCR log. Claude turned them into dashboards: first pass yield, takt adherence, open NCRs, traveled work.
Then I moved it into a standalone page: three.js for the 3D, d3 for the charts, no backend, no cookies. Every section on the line is clickable and drives the dashboard.
Each site hides a different problem. Departments waiting on each other (aircraft). One crane booked 96% of the time (ship). A weld process drifting toward its limit, 4 weeks before any alarm (rocket).
25 years on production lines taught me the costly problems aren't always where the KPI points. What's the first signal you'd want lit up on your line?
Live demo (fictional data): demartinogiuseppe.github.io/…
119