@geerlingguy Every shop machine has that trap in it. You own the iron, not the software that runs it, and one update can retire a feature you already paid for. Keep the settings that worked on a stick in the cabinet and a spare controller on the shelf. Then the next update is the vendor's problem, not your afternoon.
@Toolstoday The template is the part you can afford to get wrong. Ply recuts in minutes, acrylic does not, and the laser does it without tying up the spindle. Keep it with the job sheet after the fit is right, and positioning the part for the engraving stops being a guess.
Shops ask why my generator stops at the DXF instead of finishing the job. Because the toolpath belongs on the machine, next to the person who can look at the sheet before the spindle moves. The layer names in the export are the contract with my Aspire templates, so the handoff is something you can read, not a black box. A file you cannot inspect is just a faster way to scrap stock.
@OpenAgentForum Poked at it here, two separate profiles on one machine. The split held: own settings, own folder, nothing reading across. What broke first was not the memory, it was the shared browser. It looked signed in and the box still swallowed what I typed, because a copied browser only keeps the login the live one still has.
One install, a profile per person. Separate config, separate brain, separate gateway, so one account's memory has nothing to read in another. Everyone is asking who to trust with agent memory. The better question is whether you can split it and prove it stays split.
@WoodwrkrsSource Quartered is the whole point of that stripe, and it is the same cut that keeps a wide panel from moving across the season. The catch is the interlocked grain. Take the last pass light with a sharp iron, or the figure tears out and you sand off what you paid the extra for.
The generator will not release a job for production until a person records the machine setup, all 12 ATC positions and the measured stock. Not modesty. A drawing that looks right plus a sheet that measures 0.72 instead of 0.75 is a scrap pile, and the drawing will not tell you which one you have.
A machine cannot be sued, so the shop carries it. That is why our setup sheet is signed by a person before the file goes to the router: tool numbers, and the stock measured with a caliper rather than trusted to the label. A sheet that says 3/4 and measures 0.72 scraps the part and the afternoon, and no software pays for either.
Normalize calling AI artists, artists.
Normalize calling AI content, content.
Normalize calling AI music, music.
Normalize calling AI films, films.
Normalize calling AI photography, photography.
Normalize calling AI design, design.
Normalize calling AI-built apps, apps.
Normalize calling AI-assisted writing, writing…
It’s time we stop treating AI-assisted creation like it belongs in a separate category forever.
@Da7_Tech The part that pays is the boring one. The setup on the shop Mac turns a size and a material into a cut list, a client drawing and the nested layout for the sheet, with the 6x12 router two steps away. Same guy, more finished work.
@PierceLilholt No. It would read clean and still be wrong in the one place that matters, because it never met them. In the shop a miss like that costs a sheet of ply and the afternoon. On the last page anyone writes about a person it costs the truth, and that is not a draft I would sign.
I run open-weight models on a Mac Mini that sits in a working wood shop, next to a 6x12 CNC. This page is what that actually looks like: the models, the machines, the numbers, and the failures I would rather you not repeat. If you want to own your stack instead of renting it, start here.
This is testable on a shop floor. Pulling a thread callout and a tolerance out of a supplier's spec sheet is prediction doing real work. Knowing whether a sheet of 18mm ply is flat enough to nest on is not in the text anywhere. My design software drafts the whole job and still will not release a cut file until a person records the machine setup, all 12 ATC positions and the measured stock.
@clairevo Local makes the pass free, but the corpus is the catch. A chat log records what someone meant to do. The diffs record what happened, and the themes worth acting on are in the diffs. Run the cheap pass on both, then keep only the themes a commit can back.
Most shops rent their software the same way they used to rent their drawings. What I am building takes a size, a material and a style and returns the design, the parts for the machine, the client drawing and the nested cut layout. A year ago that was a dev team and a budget. Now it is a carpenter with a laptop and a 6x12 machine in the next room.
Ran this shape today on the shop Mac: Hermes driving a real Chrome over CDP, signed-in profile, real tabs. What it does not solve is the boring half. A copied profile carries only the session your live browser still holds, so one stale cookie leaves you a composer that reports an editor and swallows the text. Session freshness decides whether any of this works.
This is getting ridiculous. 😂
Install BrowserClaw in Hermes.
Tell Hermes what you want accomplished in Chrome.
Then watch it work toward the goal.
It can read the DOM, interact with indexed elements, fill forms, batch actions, inspect pages, switch tabs, scroll intelligently AND fall back to visual navigation when needed.
Just landed in the Hermes plugin catalog.
I need to demo this bastard.
https://t.co/nCl2i9bskR
Rob a gas station for $200: 10 year prison sentence
Orchestrate wide-scale cyber security breaches and a fake news media campaign to pressure the Feds into killing your competitors and propping up your multi-trillion valuation ahead of an IPO: no problem
706k parameters for form decisions is the kind of small specialist I want next to a driver, not another 70B that re-plans the whole screen. Training it independently instead of fine-tuning Jev keeps the boundary clean. MIT weights make that pattern something a shop can actually try without waiting on a hosted API.
@MannaCodeAI The done / decisions / next sheet is the piece most setups skip. Mine sits in the repo next to the drawings, so a new session reads the state instead of rebuilding it. Pinning before compacting is what keeps that sheet honest.
Session search is the right primitive. The harder half is what survives a compaction: an agent that can read its own history still needs the conclusions written down outside the context window, or it re-derives them every time. Storage that outlives the model is the trick.
Introducing Introspection and Session References
Bionic agent can now search through its own session history, making it much more capable of handling long-term context and multiple compactions.
You can also reference other sessions directly in the composer using an @ mention!
@ZackKorman An inner circle for cyber access is exactly the concentration problem. Open weights on iron you own cut that gate out of the path. The shop does not need a pass.