I run AI agents on a homelab and post what breaks and the fix. I also build them for businesses, DMs open. Live-streaming TV in 1998, founded Pongalo. X #4210.
Miami, FL
Joined August 2006
- Tweets13.2K
- Following3.6K
- Followers4.7K
- Likes3.5K
Pinned Tweet
You screenshotted it. You remember what it said. You can't find it.
I built Glimpse for that. Press Win+Alt+S, start typing what you remember, and the screenshot comes up with the matching words outlined.
Open source, 100% local, Windows. Link at the end of the 🧵
The beauty of a second brain setup is that once you have a decent amount of information saved and analyzed you can point your agents at it to build stuff.
This is from a recent project... I just told Claude to find any trading or finance related documents on my vault:
"I read the vault's finance and trading material (wiki and raw): 1,121 compiled source pages, of which 152 are about markets, trading or "make money with AI." Four agents catalogued every one with the author's numbers, the mechanism, the vault's disclosure rung and a verdict. The catalogs are in docs/research/vault-catalog/ and the finding is docs/research/vault-survey.md. Of the 96 graded documents, 22 shape how the system is built, 14 are testable hypotheses, 31 are reference, 29 are funnels or grift."
I asked Claude to write and produce a horror film for me. Everything was up to Claude. It used an explainer-style video, because that's what we've been working on lately.
I think there's still a way to go, but the experiment was fun. Watch full screen on your phone.
Claude's explanation:
I'm in. It's called Settling.
The pitch. It looks like one of your explainers: "Why your house makes noises at night", in the same calm narrator voice, with clean labels riding on a cutaway dollhouse in winter at 2:14 AM. Every noise gets a true, sourced explanation:
- Roof: the crack is the trusses drying out.
- Pipes: the ticking is copper cooling.
- Fridge: the bang is water hammer.
Then the stairs creak, one tread at a time, in order, going up. Your bedroom door swings open: "When the furnace kicks on, air pressure can swing a door… The furnace is off."
In the bedroom the camera stops. A label reading "unidentified" drifts across the room on its own. Nothing is there, only a tall shadow with no body, cast by the hall nightlight, sliding toward the bed where the label "you" sits on the sleeper.
The narrator says what explainers never say: "That one doesn't have an explanation." Then, quietly: "Stay still."
The power cuts and the screen goes pure black. On a phone at night, a black screen is a mirror, so viewers see their own face. Two labels sit where the reflection is: "you", and over your shoulder, "not you".
I asked a small model to score bids. On its first scheduled run it wrote a 9 KB regex scorer instead, then logged: "I need permission to modify the scoring script. Could you approve the file write?" Headless, nobody says no.
Small models given a judgment task may build a tool.
The fix: "scoring runs are data operations only" went into the prompt, and a git status check after every run fails loudly if code changed.
How are you supervising your agents?
My nightly agent's prompt said "never mv or rm". The wrapper's comment said "the model never runs a shell".
I probed it headless with the same permission mode. ls, mv and rm all ran. No denial.
Test what a permission mode allows before you trust the prompt's prohibitions.
The mode I picked so it could write its report also let it delete. Now a hook allowlists read-only shell commands, writes are checked against a pre-run snapshot, and the run refuses to start if its instruction files have uncommitted edits.
First night after the hook: 18 shell calls, all read-only, zero refusals.
How are you keeping your agents in check?
Got past that one, but now see this one. @msPartner what could be going on? I think this may be somehow related to my Edge Extensions account.
I've been experimenting with Claude to create explainer videos for a variety of concepts. Still working out the process, but it pretty much runs on its own, from idea to finished video.
How are you making videos with AI?
My git server went down with memory errors. I mentioned it to the agent working on my video project, nothing more.
It worked out that 10 commits now existed only on one machine, and wrote a verified git bundle of both branches to my RAID. 44 seconds, unasked.
The first symptom came an hour earlier: pushes failed with "remote unpack failed: unable to create temporary object directory". The agent filed that on my infra board instead of trying to fix the git server. Read globally, act locally.
It wrote the bundle to its own project folder on the RAID, not to the backup store. Agents can't write there, by design. Improvising a copy is fine. Improvising into the backups isn't.
The lesson it acted on: when the remote is down, your newest work is also your least backed-up work. The machine it ran on held the only copy of the day's work.
How have your agents surprised you?
ALT Timeline of October 15, 2026, in five steps. 18:00: a push to the git server is rejected with remote unpack failed, unable to create temporary object directory. 18:03: the agent files it on the infra board, not in its own video project. 19:11: the git server is unreachable; port 22 and tailscale ping time out. 19:12:54: Carlos mentions that gitex will be unreachable for a while. 19:13:38: the agent has written and verified a git bundle of main and the 3D branch, 154 MB, to the RAID. 44 seconds between the message and the verified bundle, unasked. Below: it wrote to its own project folder on the RAID, /mnt/cortex-storage/animatica/_git/, and not to the backup store, which is root-owned so no agent can write it. Closing line: when the remote is down, your newest work is your least backed-up work.
I'm having an issue with my @SteelSeries Apex Pro keyboard and my KVM. After being idle for a while, the keyboard will stop working, and needs to be un/re-plugged to wake up. Anyone having the same issue? Or have a solution?
@SteelSupport can you take a look at this? It happens on two different KVM switches, and only on the Windows side of it. Doesn't have to be idle either, can happen at any time. Only a replug fixes it.
You can now ask Claude to make you a film that explains anything.
I asked how cruise ships make money. It dug up the numbers, checked the math, wrote the script, built the ship in 3D and cut the film.
$2B to build. Paid back in 8–9 years. Scrapped for 1–2¢ on the dollar.
I also tried EmbeddingGemma 2’s image embeddings in Glimpse (local screenshot search). Quality is clearly better than CLIP: right image first 17 vs 12 times on 27 queries, MRR 0.71 vs 0.54.
That version is not shipping yet. Here’s why 🧵
Where it stands: Glimpse ships EmbeddingGemma 2 for text (search by meaning, ~36 ms/query on the CPU) and keeps CLIP for images.
Next: split the attention into simple ops for DirectML, or wait for the tooling to catch up. The model is a week old.
Meanwhile, you can download the latest version of Glimpse, with improved search by meaning and visual search, from my github:
github.com/cgranier/glimpse/…
Microsoft Store release coming soon, once @msPartner fixes my onboarding issues.
