Dragging people screaming and kicking into the future

[Bill E]

I have added a Claude skill that knows about the pdp-1 and my am1 assembler. Yes, I know some purists will have heart failure, but I find this a fascinating blend of modern technology with historic computing, and it should definitely make it easier for people to get involved with the pdp-1, much easier to write a program. It's checked in under the Claude directory.

Note, though, that the training is a work in progress. As with anything done coding-wise with an LLM 'ai', you can' t take what it says blindly. I have used Claude to optimize t30dpy and t30dpy3, it did an excellent job but it was an interactive process that required my knowledge.

Anyway, all hail our new machine masters.

Bill

https://github.com/wjenh/pidp1-mods.git

PS - I remember seeing somewhere, maybe here, a LLM context document for the pdp-1. I haven't searched yet, but if anyone remembers such a thing, let me know.

[Whit Turner]

Sigh. So glad I retired over a decade ago. :-)

Whit

--
You received this message because you are subscribed to the Google Groups "[PiDP-1]" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pidp-1+un...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/pidp-1/79d23fdf-2781-4df1-90b0-d1c43875b490n%40googlegroups.com.

[Bill E]

I used to think the same thing, but it has turned out to be a very powerful tool. It is NOT a replacement for knowing what you're doing, though.

Example, I pointed it at the mods repository, fed it various pdfs and such about the pdp-1. It went through the whole mess, figured out am1 from its yacc, lex, and user documentation, then reported than an opcode am1 was generating didn't match what pdp1.c in the emulator expected. Wow. And it was correct. An obscure problem I wouldn't have noticed until someone used that specific instruction. It also confirmed that all the other code generation does match what a pdp-1 would expect. It also learned my coding and commenting style and when it suggests code, it follows my conventions with amazing accuracy.  Its comments are spot-on.

What did it not figure out correctly? Only one thing so far. It didn't completely understand that all the skips can have their sense flipped by the indirect bit. It got that correct for most of the skips, but not all. I told it, it now understands. This is why its a tool, not an authority. So far.

Bill

[R Clark]

"So glad I retired over a decade ago" ... I am working up to it in a few more years.  I don't see a use for AI here.  I 'like' programming and figuring out things for myself. The age of Idiocracy is coming though :) .  Let the AI do it....  And when we become dependent on it ... watch out.  Give me a 'C' compiler, Assembler, and Python for some tasks and I am good!   

[Milo Velimirović]

Retired for a few years now.

s/ython/erl/g and I’m good.

It’s definitely time to revisit

my pidp-11 and BSD.

--
You received this message because you are subscribed to the Google Groups "[PiDP-1]" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pidp-1+un...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/pidp-1/4e309fd1-6ec3-4c72-809c-5b10629aba4fn%40googlegroups.com.

[sunnyboy010101]

I'm still working full time (at over 70) at a university teaching computer science - programming (C++, Java) and robotics (all done behind the smoke and mirrors in C, of course!)

In the past year I went from maybe 2% of assignments done by AI to over 80% done by AI. The total enshitification of education by AI is almost complete. I have fellow academics writing courses by AI now, and papers by AI, so why not students using AI to do their work.

At least I won't have to formally retire - I'll have an AI marker mark all the AI generated assignments and can just sit back and drink coffee while I play with old hardware (Z80 stuff) and all Oscar's fine kits (I have all the PiDP ones).

[Bill E]

I fully agree that there is a developing overdependence on ai, and of course that's a misnomer, it's not artificial intelligence. It's a terrible trend and I don't see it ending well. However, ai can be a useful tool and that's how I view it. I've used C for over 50 years, all of the 25K+ lines of stuff I've written here has been all done without ai assistance using straight c, yacc, lex, and pdp-1 assembler. What Claude is doing now that's really useful is finding things like the aforementioned opcode bug. While I have plenty of test cases for everything, I can't think of everything, and not enough people are using the mods versions (or apparently the original version) to find the obscure cases. Think of Claude as a fancy test framework and an independent code-reviewer.

Secondarily, Oscar has a vision of getting younger generations, even kids, interesting in real programming down at the level where they can actually see what a computer is doing, not hidden by layer after layer of obfuscation. But, he and I both think that's not going to happen if they have to use development tools from the '60s, it's just too painful. So, again, this can be a pathway to at least get people trying something. BTW, that was one of the reasons I did am1. In addition to actual error checking, the added features take nothing away from the pdp-1 but make developing software for it much more producive. Even back when people were using real PDP-1's, there were much better assemblers than macro developed and used, not sure why they have disappeared. Ok, I'll get off my soapbox now.

Bill

[Glenn Babecki]

I like the hands-on tinkering as well as the next person.  I find that it's one of the better ways to learn something.  Thus the difference comes down to "doing and learning" versus using AI "to just get something done."

That said, I agree with Bill that it's another tool for the toolbox, and seems to be getting better as it evolves.  Of course with that comes having to learn yet another tool, but like any tool it should be the right tool for the job.  AI assistance is a complex tool so it requires more effort to learn (if so desired); kind of like the difference between learning how to use a file and a milling machine.

I've dabbled a bit with using AI-based searches/chats and agents, however I don't have projects to justify the entry cost of using such tools.  Basic Claude access is something like ~$20/month (or about $17/month with an annual subscription).  While that won't break the bank on its own, it's easier to "tinkering" money for other expenses - bucket list expenses can add up. 😉

To view this discussion visit https://groups.google.com/d/msgid/pidp-1/1c032762-390c-40ac-94c7-eae1a1de2678n%40googlegroups.com.

[Bill E]

There is a free Claude version which is what I was using initially. Its problem is that it is a bit limited for any major work. But, the free one is what I used to figure out some optimizations in t30dpy. It was good at explaining things I never would have thought of, like the best arrangement of data so that the cpu would optimally use it without causing excessive cache misses. 

Bill

[Glenn Babecki]

Bill,

Thanks for your experience.  Is it possible to further clarify the threshold of what constitutes "major work?"  I know a lot involves how many tokens are required for queries, so I'm sure it's hard to quantify.  I guess I have to just explore on some kind of software chore and see where I hit a limitation of the free account.

P.S.

Came up to Boston this week on a whim to visit friends/classmate and tour.  We decided to skip the fun and frivolity of reunion week so I believe I missed your 50th reunion event (?) and a potential opportunity to meet up. My 50th was last year.

Anyway, I couldn't resist to stop by the Institute.  I actually have been on a mission to uncover some archieved PDP-9 related information related to my undergrad thesis.  This quest has been going on about 2 years now ever since I recovered my DECtape files.

I had some initial progress searching the library archives online.  Today I made a face-to-face connection with someone in RLE that's going to try and help dig into the library archives more deeply. 🤞 ChatGPT couldn't help "penetrate" into the MIT inner sanctum! 😉

Glenn

On Fri, Jun 12, 2026, 3:33 PM Bill E <wjegr...@gmail.com> wrote:

There is a free Claude version which is what I was using initially. Its problem is that it is a bit limited for any major work. But, the free one is what I used to figure out some optimizations in t30dpy. It was good at explaining things I never would have thought of, like the best arrangement of data so that the cpu would optimally use it without causing excessive cache misses. 

Bill

--

You received this message because you are subscribed to the Google Groups "[PiDP-1]" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pidp-1+un...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/pidp-1/d9158237-4958-4303-a860-713d39d22e98n%40googlegroups.com.

[Bill E]

I don't have any real understanding of how many tokens something takes, or even how a token maps to anything other than paying money. I did sign up for the lowest level $17/month because it has more flexibility in maintaining context over sessions, but all you can find out is how much of your daily allowance you've used. The whole process if it going thru all the  am1 code, documentation, and all the pdp-1 architecture and instruction set documentation took 3% of my allowance, so it appears you can do quite a lot. I think there is some correlation between a token and the amount of power you are consuming. It says itself that a modern data center is a vast energy sink, hundreds of megawatts, enough to run cities.

As for MIT, sorry i missed you. I did run into a number of people that worked with the pdp-1, including the guy that wrote a lot of the BBN timesharing stuff, as well as Mike C from the Computer History Museum. I got an earful about p7sim, one reason I've done t30dpy and modeled the CRT and phosphor differently.

If you didn't know, the engineering library has been closed, all the thesis stuff is gone. Not sure what happened to it, I think it's been digitized. I hope. It was a bit of a shock going back, haven't been back since I graduated except for one brief pass. Everything is changed. But I still have my original COOP card from 1972. :)

Bill

[Glenn Babecki]

First, apologies to others for the alumni chit-chat on this thread, but this response should wrap it up from my perspective.

Yeah, on the face of it cost = time(compute-cost/time), but it's hard to equate how much compute cost it takes to execute a task right now.  The tools are still very new to assess.  However, all great data points you've noted.  Thanks for your experiences thus far.

I've had occasion to ask Mike C. about PDP-1 engineering drawings, just cause, and he was kind enough to give me a dump of what was used in the restoration process.  Not sure what I can make from it, but very cool.

I haven't had occasion to get back to the Bay Area since grad school at Stanford, but the CHM is on my bucket list.  I have been to the RICM to get insight from Michael Thompson and crew about one of the only currently working PDP-9s in existence.  His knowledge was invaluable to enable me to reconstruct the software development environment I used for my thesis.

I don't know about all physical media in the MIT archives, but a lot if not all materials have been digitized.  I found relevant (albeit somewhat insufficient) documents from the CIPG lab.  Even though I have a couple of paper copies of my thesis, I also managed to find the digitized version in the archives.  What I'm searching for now is documentation (dare I say source listing) for the lab image processing library used by my software.  My initial attempts to disassemble the library binaries has still left me with unanswered questions; hence, my attempt to establish the RLE contacts.  As all of us in this space have experienced, "Technical Archeology" can be challenging.  Sometimes it's hard to find that Rosetta Stone! 😉

--

You received this message because you are subscribed to the Google Groups "[PiDP-1]" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pidp-1+un...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/pidp-1/ad0e6ff8-3bfb-47a7-bb2e-3ed960e8e020n%40googlegroups.com.

[Norbert Landsteiner]

I don't really see the point in using "AI" tools for PDP-1 programming.

The entire point of the PDP-1 is that it features a rather nice instruction set that has been exclusively designed for human programming (i.e., before "hardware support for higher languages" was introduced), and it's probably one of the last systems to do so. (It may be even unique in its kind of excellence in this.). It even features an assembler in hardware for self-modifying programs (dap, dip)! Moreover, it pairs all this with visual feedback (i.e., a scope). And it's not a fantasy console, but a real system with realistic constraints of the time.

The point should be really in enjoying bare metal programming on a system, which is fairly easy to understand (and not without beauty), a system that you totally own, without any foreign context whatsoever, and seeing what kind of structures you can build and what effects may be achieved. So, if you're not doing this yourself…? (If you enjoy prompting for a hobby, LLMs are probably better with Python or React, where there is plenty of context.)

– Norbert

[Bill E]

I'm not using it for pdp-1 assembler work, getting it to understand am1 was my first experiment. I'm basically using it do do code reviews and spot potential problems. It's been highly effective for that. However, I have to disagree with your statement. It's unfortunately the case that recent generations don't appear to have the attention span to deal with all of the overhead of programming the pdp-1 in the original way, load multiple tapes, etc. There is no reason modern tools can't be used. If pdp-1's were still common, we wouldn't still be using '60s tools with them; software evolves (frequently not in a good way, though). I suppose the strongest defense for my position is that I'm producing large amounts of code for the pdp-1, although a lot of it is on the C side adding peripherals and such. But I have provided a fair number of programs written in assembler without any assistance other than using a modern cross-assembler, which I also wrote. What's everyone else doing? (Norbert, you're exempt from the 'what are you doing' group, you've done quite a lot!)

As for current pdp-1 development, there seems to be only one or two other people doing anything, which kind of shoots down the 'fun with bare metal' idea, although that is probably the thing I like most about the -1, no endless layers of operating system and libraries to hide everything. But again, I want to enjoy that, not spend time loading tapes and dealing with an error-prone and very limited assembler.

Oh well, sorry, it's first thing in the morning, I'm not awake yet. But another day of pdp-1 work coming up, I'm starting to tackle the panel driver, it is in sore need of attention.

Bill

[Rob Foster]

I am 62, have been coding since the 70s as a kid when my Dad bought an Apple II when they came out, and have been working as a professional software developer since graduating college with a CS degree.

Having said that, I am all in on AI. I have a mix of vintage computers and modern AI servers running at my house.

I have done it all programming wise over the last 40 years, and have no problem leaning on AI to generate code for me that I want to run on my vintage gear.

To me, it is getting results, and not the 'process of coding' these days. Let the AI bang out the boilerplate stuff and then I fine tune it as needed.

--

You received this message because you are subscribed to the Google Groups "[PiDP-1]" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pidp-1+un...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/pidp-1/8cb670ca-f3d5-4817-8350-67f7e57bc033n%40googlegroups.com.

[Rob Foster]

I am 62, have been coding since the 70s as a kid when my Dad bought an Apple II when they came out, and have been working as a professional software developer since graduating college with a CS degree.

Having said that, I am all in on AI. I have a mix of vintage computers and modern AI servers running at my house.

I have done it all programming wise over the last 40 years, and have no problem leaning on AI to generate code for me that I want to run on my vintage gear.

To me, it is getting results, and not the 'process of coding' these days. Let the AI bang out the boilerplate stuff and then I fine tune it as needed.

My AI runs local on my AI servers, so I don't have 'token' costs with Claude or OpenAI as I use it, just the initial hardware costs.

On Sat, Jun 13, 2026 at 7:01 AM Bill E <wjegr...@gmail.com> wrote:

--

You received this message because you are subscribed to the Google Groups "[PiDP-1]" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pidp-1+un...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/pidp-1/8cb670ca-f3d5-4817-8350-67f7e57bc033n%40googlegroups.com.

[Bradford Miller]

On Jun 13, 2026, at 1:00 AM, Norbert Landsteiner <n.land...@gmail.com> wrote:

(i.e., before "hardware support for higher languages" was introduced)

I’m not sure that’s the defining criteria you think it is. The PDP-6 and PDP-10 also have very nice instruction sets, but were enhanced with instructions to make it easier/more efficient to run Lisp, for example, based on suggestions from MIT to DEC. And later processors (e.g. the MIT Scheme-79 chip, though just an experiment) were microprogrammed in a Lisp language, to implement a higher level Lisp, again thought to make it easier on the developers at the time to leverage higher level language experience even when working in the guts of the frob. MIT’s Lisp Machine was more of a blend, with clear roots in the PDP-10, but much more optimized at the microcode and hardware level to implement Lisp, but this made it more accessible to folks in the AI lab, not less so.

[MICHAEL GARDI]

I have mixed feelings about using AI in programming. On one hand I'm definitely not into vibe coding. Like others here I enjoy the process of coding. I especially like writing assembler code, though I didn't get to do much of it in my career. I guess that goes back to my first personal "computer," a SYM-1 SBC. 

 

Back then, I was hand assembling 6502 code and using the built in monitor to input the hex. (Pretty soon I added a serial port and got a terminal plus a cassette for storage). 

When I got my PiDP-1 and decided I wanted to jump onto the democoding bandwagon for it I wished to "understand" the 1960s programming experience. So I worked through the simple 'draw a circle' example in Obsolescence Guaranteed's guide, "PROGRAMMING THE DEC PDP-1 - A QUICK WAY TO GET STARTED."  I used ET to "edit" the program code, punched the source to "tape", used MACRO's two-pass assembler process (loading the source tape twice), punched the binary and start block to another tape, then loaded that tape and pressed READ IN. All of this is virtual, of course except the READ IN switch. A lot of steps for sure, and they must have burned through a ton of paper tape. However it was truly iterative for them. My first real programming experience involved punch cards and a two day turnaround for execution. So who's to say which is better?

I did not use ET or MACRO in my workflow. I ended up using the built-in simple text editor "mousepad" to edit my code, MACRO1_1 to assemble the source from the command line, and stuck with loading the "tape" in the virtual tape reader, and pressing READ IN. I found I rather enjoyed watching the tape "load", and the anticipation of the program starting. Enough of the 60s "experience" for me without going crazy ;-)

PDP-1 assembler is so elegantly simple that I was always able to figure out my "bugs" without the aid of a debugger. Figuring out "issues" really helped with the learning process. I wish, however, that I had been more aware of Flags earlier in my learning curve. The DEC PDP-1 computer has six independent "program flags," which are user-addressable flip-flops that function as software-controlled switches or synchronizers. In addition these flags are mapped to 6 lights on the Console. They would have been an easy way to display runtime debugging "codes".

OK, no vibe coding. Initially, I relied heavily on coding examples from the documentation for Norbert Landsteiner's entry to Retrochallenge 2016/10: Ironic Computer Space Simulator (ICSS)., where he implements a version of Computer Space on the PDP-1. After I felt more comfortable with PDP-1 opcodes though I found myself asking Google things like "How do I implement a jump table in PDP-1 assembly language?", or "Remind me how the PDP-1 dpy opcode works." and using the resulting AI Overview (with caution). 

So, I'm not opposed to using AI, but I want to do most of the coding myself.

Mike

If you are interested in more details of my PDP-1 programming journey check out: Lunar Lander for the PDP-1.

On Sat, Jun 13, 2026 at 7:01 AM Bill E <wjegr...@gmail.com> wrote:

--

You received this message because you are subscribed to the Google Groups "[PiDP-1]" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pidp-1+un...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/pidp-1/8cb670ca-f3d5-4817-8350-67f7e57bc033n%40googlegroups.com.

[MICHAEL GARDI]

I wanted to mention that the most head scratching "bugs" I had to track down had nothing to do with my code, but were a result of not having RTFM!  I had an imperfect understanding of the screen coordinate layout and how to pass the X,Y coordinates to the dpy opcode. Pretty rookie mistake for an old hand like me.

M

[Bill E]

Another real-world example of what AI is good for. I used claude to figure out the really complex and obscure cause of the hardware panel flicker when not setting real time thread capability on it. It's something I doubt most people would ever have figured out, a complex interaction between the linux scheduler, the timing loop in  pdp1 and both the old and new panel internal timing loops, Pi power management, and threading. It was an interactive process. I posed the problem, claude asked for various system command outputs, added some instrumentation in the programs. After some iterations, problem identified and solved by modifying timing parameters in the new panel. I didn't change the old panel. And, it was actually fun figuring it out this way.

Bill

[Jon Jackson]

I still have my SYM-1 SBC with a keyboard, upgraded ROMs including a Digitalker ROM, and an S-100 bus adapter with memory card attached.  Still have the cassette tape player used for program storage and playback (program loading)... and all os the SYM-physis newsletters 

At the time my young kids loved listening to "this is Digitalker" and all of its words, but ESPECIALLY all of the address locations with boingy-sounding gibberish from a simple program I had written to play everything in the address range !!

To view this discussion visit https://groups.google.com/d/msgid/pidp-1/CAH8fXAELXKdqg%3Dc-Zp1gCJBUEnt_E8wWpGpGpP1CiNWS79Zm6Q%40mail.gmail.com.