{"slug": "show-hn-making-an-mit-hacker-replicant-for-a-pdp-1", "title": "Show HN: Making an MIT hacker replicant for a PDP-1", "summary": "A developer released an open-source AI companion for the PiDP-1, a replica of the Digital Equipment Corporation PDP-1, that installs a tutor, guide and copilot able to see the display and operate the front panel. The project ships in three parts — markdown skill files that turn any LLM into a PDP-1 expert, a companion update to the pidp1 GitHub repo giving the AI the same typewriter, front-panel and Type 30 graphics access as the user, and an optional turn-key Hermes Agent install — and requires only an external LLM, with the author suggesting DeepSeek as of November 2026, where $10 buys months of use with no subscription. The author states the goal is tutoring rather than code generation, noting the agent ported Hopalong from JavaScript in 20 minutes without human intervention but arguing retrocomputing is best left AI-free for the sense of achievement.", "body_md": "**An AI companion for the PiDP-1** — installs a tutor, guide, and capable copilot\non your replica PDP-1. It sits next to you, and can see the display/operate the front panel just like you. Learn how to use the PDP-1 faster, and more enjoyably.\n\n*The PiDP-1 is a replica of the original Digital Equipment Corporation\nPDP-1 computer.\nSee the [PiDP-1 web site](https://obsolescence.dev/pdp1.html)*\n\nIt is great to have a proper PDP-1 Hacker sit next to you at the front panel. Perhaps having replica Hackers around is just as important as having replicas of their old computers?\n\nI wanted to make an AI companion to the PiDP-1. It is able to vibe-code, but the real purpose is to help people learn how to use the PDP-1 — more quickly, and more enjoyably. Any user, from beginner to advanced.\n\nIt is not too hard in 2026. This is a first attempt; I hope it will be picked up and improved. So, feel free to fork, recycle, replace, whatever.\n\nIt can install a complete open-source Hermes Agent setup on your PiDP-1\n(or any Linux machine). All it needs is an external LLM to connect to.\nAs of November 2026, I'd suggest the cheap, good-enough **DeepSeek**, but it is up to you.\nThe PDP-1 \"problem set\" is small enough, and $10 at Deepseek gives you\nmonths of playtime. And there is no subscription: when the money runs out,\nyou are off the hook, unless you want more playtime. I like that. The install script will set it up in a few minutes.\n\nThis project comes in three independent parts:\n\n- **The skills** — markdown files that make any AI you like into a PDP-1\nexpert.\n- **Code to share the PiDP-1** — a companion update in the[pidp1 repo](https://github.com/obsolescence/pidp1) itself, so\nyour AI has the same access to the PDP-1 as you: it can see and type on the\ntypewriter, see and control the front panel, and watch the Type 30 graphics\ndisplay. The install script will do that update for you.\n- And if you want it: this also installs a **turn-key Hermes Agent** on the\nPiDP-1 to use the above.[Hermes](https://github.com/nousresearch/hermes-agent) is an\nopen-source AI agent that runs locally, and it is charming because it learns\nas you use it: your preferences, and its own past mistakes. It really gets\ngood after an hour or two! You chat with your expert through its own window, whilst it can handle the PDP-1 just as you can.\n\n|  Rotating cube |  Tic-tac-toe |        [Hopalong radiant attractor](https://camo.githubusercontent.com/14ba1ade3a8e48497a16a7cff1d2b7469e300d62be35110e37cc20523e6fc030/68747470733a2f2f6f62736f6c657363656e63652e6465762f70696470312d73772f6167656e742d706470312f686f70732e77656270)(click to zoom!) | Hopalong was ported without human intervention from Javascript in 20 minutes. The agent wrote [this page](https://obsolescence.dev/pidp1-sw/agent-pdp1/hops.html) . It depressed me: a sign that the end of democoding is near? | \n\nAdmittedly, as a side-effect, the AI becomes a competent PDP-1 coding agent. It was trained doing code reviews of major PDP-1 programs, and it wrote some nice demo code, like the tic-tac-toe game above.\n\nBut, personal view: retrocomputing is best left AI-free in that sense. There is no sense of achievement or historical value in letting an AI write an amazing new program on a PDP-1.\n\nInstead, its goal is to be a **tutor**, click to zoom the screen shot. It can teach you to operate\nthe machine, teach assembly programming, help find bugs. And also, **give you\nan interactive tour** through the process of writing code in ET, assembling it,\nusing DDT. Or show you how to work in Lisp. Or take you through a tour of the\nbest graphics demos.\n\nAlmost nobody still has an experienced PDP-1 hacker nearby. So it makes sense to have an AI take that role; it will make learning about the PDP-1 much faster and much more fun.\n\nTo see some nice demonstrations:\n\n- ask for a **code review of something like the DDT source code** — the\ncode-review skill is an amazing teacher, and you will see exactly how\npowerful this tutor can be. Here is an[example](https://obsolescence.dev/pidp1-sw/agent-pdp1/ddt_phase7_changelog.md) of what studying the DDT source code added during the training process.\n- new to the PDP-1? Ask for the **assembly tour** ;\n- or just ask about the **history of demo coding** on the PDP-1.\n\n```\ncd /opt\nsudo git clone https://github.com/obsolescence/agent-pdp1.git\ncd agent-pdp1\n./install.sh\n```\n\nNote\n\nThe install script will download and install Hermes for you through their headless install option. The user-friendly way. But Hermes is a bit of a moving target, it depends on - well, lots of Linux dependency hell it seems. So at some point in the future: if the Hermes install through my install script breaks: not a problem. Then, just install Hermes yourself, manually. Easily googled. And rerun the install script afterwards to let it complete the Hermes setup. But this install script is tested to work during 2026 and saved some hassle from a manual install.\n\nFrom now on, to start working with this: do the usual `pdp1control start` command, switch the power on (top right of the front panel). Then type `hermes --tui` any time you want your companion alongside you. **For now, use the PiDP-1's GUI setup, the web server alternative is not tested thoroughly yet**.\n\nWorks on a PiDP-1 (Pi 4/5, 4 GB RAM; untested: 2 GB RAM). If you don't have\none, install this after the [pidp1](https://github.com/obsolescence/pidp1)\npackage, on just any Linux laptop — you get a virtual front panel, and\neverything else stays the same.\n\nThe installer presents a menu. Every step explains itself, and every step can be declined:\n\n1. **Basic setup** — update the pidp1 package, install the agent tools.\n2. **Hermes Agent** — setup Hermes, connect DeepSeek (or defer\nthat), add the skills into the agent.\n3. **Update/restore the core skills** — an update function. Pull the latest from GitHub. Edits you made to\nthe core skills are replaced, but your PDP-1-Learnings are left alone.\n\n`DRY_RUN=1 ./install.sh` simulates the install without changing anything\n— handy as a preview, or for testing.\n\nThe Hermes step asks for a DeepSeek API key — type/paste it in, or point the installer at a file that holds one (like ~/my-api-key.txt). There is no particular reason to use DeepSeek: it is just, at the time of writing, the cheapest model that is good enough. Hermes runs any model — add others later, things will evolve.\n\n**`skills/`** — core of the project. Readable markdown files with PDP-1 knowledge.\n\n- **pdp1-assembly** — instruction set, MACRO-1, programming patterns\n- **pdp1-debugging** — the new debugger protocol, recipes, a pdp1dbg helper client\n- **pdp1-plumbing** — knowledge of the simulator itself\n- **pdp1-type30-vision** — reading the screen (through a the pdp1_dpy tool)\n- **pdp1-tutor** — guided tours of PDP-1 applications (still WIP, but OK for now)\n- **pdp1-code-review** — structured code review workflow\n\n`hermes-specific/`\n\n- **SOUL.md** - the rules that Hermes should keep in mind\n- **pdp1-learnings** template: the agent's own learning file. Read it and change it to your liking. Perhaps share it with us so we can improve the knowledge base.\n\n**`install.sh`** — the installer script. Tested on Raspberry Pi OS and Ubuntu.\n\nNone of this needs to be Hermes-specific — and nothing depends on DeepSeek, either. The skills are plain markdown, and most any AI will absorb them; Claude Code, for instance, will happily adopt them. The API that lets an agent work the machine the same way you do is described in the skills files themselves.\n\nAs part of this project, the base [pidp1 simulator](https://github.com/obsolescence/pidp1) was updated so an AI can\nshare the PiDP-1 with you. Useful for future projects as well. And the\nPDP-1 simulator gained a massively improved debugger: useful for the AI, so\nit is not forced to debug over the front panel all the time; useful for\nhuman PDP-1 hackers, too. Again, described in the skills files.\n\nWe noticed that a self-learning agent like Hermes will, over time, not just learn new things — it will also get caught in its own previous mistakes. Anecdote: when programming Tic-Tac-Toe, it looked online for more example code, found some helpful PDP-8 code, and then got itself confused for a week about the \"bug in the subroutine instructions\" of the PDP-1…\n\nSo, we decided that the self-learning is best kept curated. Instead of\nediting the canonical skills, agents are instructed to write what they learn\nto their own `pdp1-learnings` skill file\n(`~/.hermes/skills/pdp1-learnings/`). Read this file in an editor some time —\nit is interesting.\n\nWe hope people will submit their `pdp1-learnings` files, so we can use them\nto improve the core skills. You can also do that locally:\n\n1. `/opt/agent-pdp1/skills/curate.sh` — disable the write protection.\n2. Tell your Hermes agent to curate your `pdp1-learnings` into the core\nskills, and ask it to go through each learning with you, so you can judge\nwhat is useful and what is not. Or, of course, tell Hermes to do it for\nitself.\n3. `/opt/agent-pdp1/skills/protect.sh` — restore the write protection when you're done.\n\nThere is also `/opt/agent-pdp1/skills/update.sh` (or menu option 3 of the\ninstaller): it restores your local core skills to the originals from the\nGitHub repo — so there is no risk of breaking anything.\n\nPlease use the **PiDP-1 Google Group**: [https://groups.google.com/g/pidp-1](https://groups.google.com/g/pidp-1)\n\nAnd send your `pdp1-learnings` files if you want to contribute — through\nGitHub, as attachments in the Google Group, or by email. You will have my\nemail if you are a PiDP-1 user ;-)\n\n- Tell Hermes to be interactive with you, and not to go off figuring things out alone. You want to say things like that regularly, so it remembers.\n- Set its `/reasoning` to medium or low for regular chats, maybe to`max` or`high` when you want it to write a program. Tell it to be concise.\n- Now and then, tell it to curate its new learnings; I recommend you ask it to that together with you.\n- Give it an hour or two of use to get settled in (proving I have some work to do still!).\n\nThere is no claim to greatness here. This is experimental, zero copyright or ownership claims — I am just a hobbyist. AI agents will evolve rapidly, so by the time you read this, there may be much better ways to use these skills files. But it is a start: fun, cheap, open source, subscription-free.\n\nWe do believe these skills files have some *eternal value* — they are a\nthorough knowledge base for anyone, and any AI, to build on. A lot of time\nwas spent getting them effective; improving this curated PDP-1 knowledge is\nwhat we think the real value here is.\n\nTo close, again, our intention. Preserving PDP-1 legacy is not about vibe coding — this agent can do that, but computer history should be free from AI, IMO.\n\nBut...\n\nAs people with knowledge of the historical machine start to become scarce, an AI tutor and guide is a very legitimate way of keeping PDP-1 history alive.\n\nUgh. Lawyers.\nMIT — see [LICENSE](https://github.com/obsolescence/agent-pdp1/blob/main/LICENSE). Fork it, recycle it, replace it.", "url": "https://wpnews.pro/news/show-hn-making-an-mit-hacker-replicant-for-a-pdp-1", "canonical_source": "https://github.com/obsolescence/agent-pdp1/", "published_at": "2026-10-06 11:26:42+00:00", "updated_at": "2026-10-06 11:49:37.866387+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "artificial-intelligence"], "entities": ["PiDP-1", "PDP-1", "Digital Equipment Corporation", "Hermes Agent", "DeepSeek", "Nous Research", "obsolescence.dev", "DDT"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/show-hn-making-an-mit-hacker-replicant-for-a-pdp-1", "markdown": "https://wpnews.pro/news/show-hn-making-an-mit-hacker-replicant-for-a-pdp-1.md", "text": "https://wpnews.pro/news/show-hn-making-an-mit-hacker-replicant-for-a-pdp-1.txt", "jsonld": "https://wpnews.pro/news/show-hn-making-an-mit-hacker-replicant-for-a-pdp-1.jsonld"}}