[pushfoo]

TL;DR: I love this idea and have two questions:

1. For early iterations / similar projects, could all-metal screw terminals[1] accept power internally?

2. Would supporting UXN/Varvara[2] be an option?

More on these questions below after initial comments.

## Initial Comments

> if you aren't worried about going batteryless, or about mass production, you could literally buy an alphasmart neo and wire its keyboard and display up to an esp32 or something

I could, but it's a limited in availability. The keyboard is pretty important in my opinion.

> memory-in-pixel lcd displays

I vaguely remember hearing about this, but don't have the EE knowledge to judge the benefits of these. Data sheet for anyone interested: https://www.sharpsde.com/fileadmin/products/Displays/Specs/L...

## 1. The screw terminal:

Disclaimer: I'm not an EE specialist.

My understanding is that if the power loss isn't too great, an all-metal internal screw terminal[1] might improve device durability:

* The power source could be replaceable. AA, AAA, LiPo, solar, etc

* If you don't solder to the metal part, even sandpaper or a conveniently shaped rock could remove oxidation

* For a case, internal screw terminals could turn a charging port into an easily replaceable component

## 2. UXN/Varvara:

From your blog, I see an "artemis apollo 3 board"[3] is being used. From a Sparkfun page[4], it seems to have enough ram to host graphical Varvara.

I was initially doubtful of UXN despite loving the idea, but:

1. The UXN community seems to have built a self-hosting ecosystem[5] of:

   * development and GUI ROM launching tools

   * text, font, and graphics editors

   * music sequencing tools

2. The UXN VM is light: 64k system ram, 4-bit 2-plane color, and misc state and debug

3. The core UXN VM is simple: a minimal implementation fits in under 150 lines[6] of C89

[1]: https://en.wikipedia.org/wiki/Screw_terminal

[2]: https://wiki.xxiivv.com/site/varvara.html

[3]: https://blog.za3k.com/zorchpad-update-cardboard-mockup-mk1/

[4]: https://www.sparkfun.com/artemis

[5]: https://github.com/hundredrabbits/awesome-uxn

[6]: https://wiki.xxiivv.com/etc/uxnmin.c.txt

[kragen]

i'm glad to hear you find it appealing! i'll explain some of my thinking on these issues, some of which is specific to my own eccentric objectives (nobody designs computers to last decades) and some of which is more generally applicable

screw terminals or post terminals are a good idea if you're going to have charging and want longevity. but charging implies batteries. and batteries themselves induce lots of failures; they typically have a design lifetime of two years, more than an order of magnitude too low for the zorzpad. by itself that's not a fatal flaw, because you could replace them, but it becomes fatal in combination with either of the two supplementary flaws:

1. battery shelf life is limited to about 10 years, except for lithium thionyl chloride and so-called 'thermal batteries' that use a molten-salt electrolyte (kept frozen during storage). consequently, if your computer needs batteries, after 10 years you're dependent on access to not just a market but a full-fledged industrial economy to keep your computer running. a friend of mine in rural romania has been trying to build a usable lead-acid battery from scratch to provide energy storage for his solar panels, and it's motherfucking difficult without that supply chain

2. batteries have a lot of failure modes that destroy the devices they power. overheating, swelling, leaking corrosive electrolytes, and too much voltage are four of them. if these failure modes were very rare (like once in 100 battery-years) that might be okay, but they aren't

for me these are fatal flaws, but for others they may not be. if you can afford batteries, you can run for one hell of a long time on a single 18650 lipo on a milliwatt. a coin cell can deliver a milliwatt for a month

as for uxn, uxn is the first attempt at frugal "write once, run anywhere" that's good enough to criticize. everyone should read devine's talk about its aspirations (https://100r.co/site/weathering_software_winter.html)

however, uxn is not frugal with cpu, or fast, nor is it good for programmer productivity, and it’s not all that simple. It’s designed for interpretation, which inherently means you’re throwing away 90% or 95% of even a single-threaded cpu, and it’s not designed as a compilation target, which means that you’re stuck programming at the assembly-language level. This is pretty fucking fun, but that’s because it makes easy things hard, not hard things easy; the offgrid notes lionize oulipo (https://100r.co/site/working_offgrid_efficiently.html)

as a result of the efficiency problems, the platforms with complete uxn implementations (see https://github.com/hundredrabbits/awesome-uxn#emulators) are the luxurious ones: microsoft windows, linux, amigaos, the essence operating system, anything that can run sdl or lua and love2d, gameboy advance, nintendo 64, nintendo ds, playdate, and the nook ereader: all 32-bit or 64-bit platforms with ram of 4 mebibytes or more, except that the gba only has 384 kibibytes, and there were amigas with as little as 512k (though i don’t know if they can run uxn). more limited machines like the gameboy, ms-dos, and the zx spectrum do not have complete uxn implementations

(384k is the same amount of ram as the apollo3, so it's probably feasible, as you say)

the uxn machine code is basically forth. forth is pretty aesthetically appealing to me, and i've gotten to the point where forth code is almost as easy for me to write as c code (though i have to look things up in the manual more often and have more bugs). however, it's much, much harder for me to read forth than it is to read c. it's harder for me to read forth code than to read assembly for any of 8086, arm, risc-v, i386, or amd64. possibly i just haven't written enough forth to have the enlightenment experience yet, but currently my hypothesis is that forth is just hard to read and that's it, and that the really great thing about forth is not actually the language but the interactive development experience

the varvara display model is not a good fit for the memory-in-pixel displays, which only have one bit per pixel, while varvara requires two. also, evaluating things 60 times a second is not a good fit for low-power systems of any kind. and i'm not sure how well the existing varvara applications will work at 400×240 or tolerate a requirement to use fonts many pixels tall in order to be readable

(a lot of the above notes are from xpasm.md in http://canonical.org/~kragen/sw/pavnotes2.git/ if you're interested in the context)

as for tools for software development, program launching, editing text, editing fonts, editing other graphics, and sequencing music, i'm confident i can write those myself if i have to. i've written a compiler in a subset of scheme that compiles itself, a compiler in a forth-like language that compiles itself, various toy text editors, various virtual machines in under 150 lines of code, and various pieces of music synthesis software, and i've designed a few fonts of my own (though not using software i wrote)

maybe i should set up a patreon for this or something, maybe people would be interested in supporting the project

[pushfoo]

TL;DR: Thank you for confirming some mismatch in display hardware and project goals

> also, evaluating things 60 times a second is not a good fit for low-power systems of any kind.

Agreed. As to a terminal-mode UXN, you pointed out:

> as for uxn, uxn is the first attempt at frugal "write once, run anywhere" that's good enough to criticize. everyone should read devine's talk about its aspirations

I think it's the primary use case like Java Applets and Flash preceded JS as iterations of sorta-portable and low efficiency tools which make up for it with sheer volume of high-familiarity material.

> nor is it good for programmer productivity

I'm curious about this. Could you please elaborate?

When trying earlier uxntal versions, the worst part seemed to be the string syntax. If it wasn't because it seemed easier at the time, my guess is the pain point might be intentional nudges away from making things the designer doesn't like.

> and it’s not all that simple.

This is where I get confused:

* Can you explain more about your views on this?

* Do you favor a specific structure of hardware, OS code, and user scripts, or whatever ends up offering the best use of power and long-term device durability?

> i'm not sure how well the existing varvara applications will work at 400×240 or tolerate a requirement to use fonts many pixels tall in order to be readable

Counting mW and decades implies text > GUI. Despite 100r & fans favoring GUI, there's an UXN terminal mode and a few math and terminal libraries written for it.

As to GUI, many of the ecosystem's GUI programs are FOSS with room to shrink the graphics. There's a decent 4x6 px ASCII font which is CC0 and could help, but as fair warning, it's awful at accents and umlauts.

> possibly i just haven't written enough forth to have the enlightenment experience yet, but currently my hypothesis is that forth is just hard to read and that's it, and that the really great thing about forth is not actually the language but the interactive development experience

This seems to be a universal opinion since REPL isn't rare anymore. Being portable and easy to bootstrap are probably the main draw now. In addition to 100r's work, I think remember hearing ${NAME}boot or another project used it to simplify porting.

> the varvara display model is not a good fit for the memory-in-pixel displays, which only have one bit per pixel, while varvara requires two.

Ah, this is what I suspected. I wasn't sure I understood the display you mentioned, so thank you for confirming the hardware mismatch.

Although your blog showed you'd considered two identical side-by-side displays[1], I have a radical idea which might not be a good one:

1. A "Full" display of W x H characters (the Sharp display or whatever you replace it with)

2. A "Fallback" 1 or 2 line display (2nd line reserved for editor info / UI?)

A single 1-bit editable line is so hard to work with that ed[2] isn't even installed by default on many Linux distros. However, people did once got real work done with it and similar tools.

As a funny sidenote, this is where non-colorForth[3] forths may align with UXN's dev. He's a Plan9 fan and uses its Acme editor[4]. So Varvara is theoretically capable of syntax highlighting, but he refuses to implement it as far as I know.

> as for tools for software development, program launching, editing text, editing fonts, editing other graphics, and sequencing music, i'm confident i can write those myself if i have to. i've written a compiler in a subset of scheme that compiles itself, a compiler in a forth-like language that compiles itself, various toy text editors, various virtual machines in under 150 lines of code, and various pieces of music synthesis software

You may be pleasantly surprised by how many people will show up if you start building and releasing tools which:

* let them solve problems or have fun

* allow extending the tool

* can be used to build their own tools

Many will will even contribute back. To paraphrase the Gren[5] maintainer's recent livestream:

> we've had more engagement in our time on Discord than all our years on Zulip.[6]

> maybe i should set up a patreon for this or something, maybe people would be interested in supporting the project

Based on your blog, I think you can safely skip to the footnotes:

* GitHub sponsors might be worthwhile if you're willing to use non-fully FOSS platforms

* Don't make people feel like they're paying for their own work

* Do let the system do real-ish creative work and self-host, even if only as emulators

If you'd like, I can provide more specific suggestions.

[1]: https://blog.za3k.com/zorchpad-update-cardboard-mockup-mk1/

[2]: https://en.wikipedia.org/wiki/Ed_(software)

[3]: https://concatenative.org/wiki/view/colorForth

[4]: https://wiki.xxiivv.com/site/acme.html

[5]: https://gren-lang.org/ (Disclaimer: I sorta contribute to this now)

[6]: https://www.youtube.com/watch?v=PO8_pV7r168 (Sorry, I don't have an exact timestamp, but I'm pretty sure it was in here)

[kragen]

i'll reply in more detail later but for the moment i just want to clarify that i'm not za3k, although i've been collaborating with him; his priorities for the zorchpad are a bit different from my priorities for the zorzpad

you said:

> I think it's the primary use case like Java Applets and Flash preceded JS as iterations of sorta-portable and low efficiency tools which make up for it with sheer volume of high-familiarity material.

i wasn't able to parse this sentence. could you unpack it a bit?

you said:

> If you'd like, I can provide more specific suggestions.

yes, please! even if our priorities aren't exactly aligned i can surely learn a lot from you

[kragen]

correction!

> although i've been collaborating with him; his priorities for the zorchpad

should be

> although i've been collaborating with them; their priorities for the zorchpad

[pushfoo]

> 'll reply in more detail later

np! I'll comment now in case it gets weird about a double-comment on the same parent by the same user.

> > I think it's the primary use case like Java Applets and Flash preceded JS as iterations of sorta-portable and low efficiency tools which make up for it with sheer volume of high-familiarity material.

> i wasn't able to parse this sentence. could you unpack it a bit?

Software that doesn't care about being well-made or efficient. It doesn't matter because it's either fun or useful.

> > If you'd like, I can provide more specific suggestions.

> yes, please! even if our priorities aren't exactly aligned i can surely learn a lot from you

This is a long topic, but some of it comes down to Decker[1] vs Octo[2]'s differences:

* Decker can be used to make and share things that process data

* The data itself can be exported and shared, including as gifs

* Octo can't really, but it does predate the AI commnity's "character cards" by shoving game data into a "cartridge" gif

* Octo has LISP Curse[3] issues

I'm serious about the LISP curse thing. In addition to awful function pointers due to ISA limitations, everyone who likes Octo tends to implement their own emulator, tooling, etc and then eventually start down the path of compilers.

I haven't gone that far yet, but I did implement a prototyping-oriented terminal library[4] for it. Since Gulrak's chiplet preprocessor[5] is so good, I didn't bother with writing my own.

> i just want to clarify that i'm not za3k

Ty for the reminder. On that note, the larger font sizes you brought up are seeming more important in this moment. I don't think I can deal with 4 x 6 fonts on tiny screens like I once could. HN's defaults are already small enough.

[1]: https://github.com/JohnEarnest/Decker

[2]: https://github.com/JohnEarnest/Octo

[3]: https://winestockwebdesign.com/Essays/Lisp_Curse.html

[4]: https://github.com/pushfoo/octo-termlib

[5]: https://github.com/gulrak/chiplet

[kragen]

with respect to winestock's account of the 'lisp curse', i think he's wrong about lisp's difficulties. he's almost right, but he's wrong. lisp's problems are somewhat social but mostly technical

basically in software there's an inherent tradeoff between flexibility (which winestock calls 'expressiveness'), comprehensibility, and performance. a glib and slightly wrong way of relating the first two is that flexibility is when you can make software do things its author didn't know it could do, and comprehensibility is when you can't. bugs are deficiencies of comprehensibility: when the thing you didn't know your code could do was the wrong thing. flexibility also usually costs performance because when your program doesn't know how, say, addition is going to be evaluated, or what function is being called, it has to check, and that takes time. (it also frustrates optimizers.) today c++ has gotten pretty far in the direction of flexibility without performance costs, but only at such vast costs to comprehensibility that large codebases routinely minimize their use of those facilities

comprehensibility is critical to collaboration, and i think that's the real technical reason lisp programs are so often islands

dynamic typing is one example of these tradeoffs. some wag commented that dynamic typing is what you should use when the type-correctness of your program is so difficult to prove that you can't convince a compiler, but also so trivial that you don't need a compiler's help. when you're writing the code, you need to reason about why it doesn't contain type errors. in c, you have to write down your reasoning as part of the program, and in lisp, you don't, but you still have to do it, or your program won't work. when someone comes along to modify your program, they need that reasoning in order to modify the program correctly, and often, in lisp, they have to reconstruct it from scratch

this is also a difficulty in python, but python is much less flexible in other ways than lisp, and this makes it much more comprehensible. despite its rebellion against curly braces, its pop infix syntax enables programmers inculcated into the ways of javascript, java, c#, c, or c++ to grasp large parts of it intuitively. in lisp, the meaning of (f g) depends on context: it can be five characters in a string, a call to the function f with the value g, an assignment of g to a new lexical variable f, a conditional that evaluates to g when f is true, or a list of the two symbols f and g. in python, all of these but the first are written with different syntaxes, so less mental processing is required to distinguish them

in traditional lisp, people tend to use lists a lot more than they should, because you can read and print them. so you end up with things like a list of a cons of two integers, a symbol, and a string, which is why we have functions like cdar and caddr. this kind of thing is not as common nowadays, because in common lisp we have defstruct and clos, and r7rs scheme finally adopted srfi-9 records (and r6rs had its own rather appalling record system), although redefining a record type in the repl is pretty dodgy. but it's still common, and it has the same problem as dynamic typing, only worse, because applying cadr to the wrong kind of list usually isn't even a runtime error, much like accessing a memory location as the wrong type in forth isn't

this kind of thing makes it significantly easier to get productive in an unfamiliar codebase in python or especially c than in lisp

40 years ago lisp was vastly more capable than the alternatives, along many axes. smalltalk was an exception, but smalltalk wasn't really available to most people. both as a programming language and as a development environment, lisp was like technology from the future, but you could use it in 01984. but the alternatives started getting less bad, borrowing lisp's best features one by one, and often adding improvements incompatible with other features of lisp

as 'lightweight languages' like python have proliferated and matured, and as alternative systems languages like c++ and golang have become more expressive, the user base of lisp has been progressively eroded to the hardest of hardcore flexibility enthusiasts, perhaps with the exception of those who gravitate to forth instead. and hardcore flexibility enthusiasts sometimes aren't the easiest people to collaborate with on a codebase, because sometimes that requires them to do things your way rather than having the flexibility to do it their own way. so that's how social factors get into it, from my point of view. i don't think the problem is that people are scratching their own itches, or that they have less incentive to collaborate because they don't need other people's help to get those itches thoroughly scratched; i think the social problem is who the remaining lisp programmers are

there are more technical problems (i find that when i rewrite python code in scheme or common lisp it's not just less readable but also significantly longer) but i don't think they're relevant to octo

[pushfoo]

TL;DR: You may be right about LISP issues, but my intent was show how IO and social factors seem key to platform success

I'll skip discussing Python for the moment because I think there are a lot of things wrong with it. If I try, we'll be here forever, but they mostly come down to not enough consistency and it being pointless to discuss for reasons which will become clear below.

Whatever the nature of the cause/effect of the LISP ecosystem, I see Octo's issues as similar in result:

1. There is no way to load and persist useful data in any popular CHIP-8 variant

2. Once people get a taste, the lack of IO means they do one of two things:

   * leave

   * become emulator or assembly developers

3. They few who stay in Octo or CHIP-8 are interested in writing their own tools, libraries, and ISA variants

Keep in mind, I still see this as success. This is based on my understanding of the Octo author's post saying farewell to running OctoJams[1] and my time as a participant. It got me interested in assembly and has helped many other people learn and begin to use low-level tools.

Compare this to Uxn:

* Although flawed, it has useful IO capabilities

* People have built applications which they use to perform real work every day

* An already-familiar ecosystem of tools exists to cater to the interests of similarly-minded people

> 40 years ago lisp was vastly more capable than the alternatives, along many axes. smalltalk was an exception, but smalltalk wasn't really available to most people. both as a programming language and as a development environment, lisp was like technology from the future, but you could use it in 01984.

Yes, I think that's getting at what I mean. One of the motivating factors for 100r was that Xcode is bad fit for people who live on a boat with solar-powered computers. So are Visual Studio or PyCharm.

Although the Uxn ecoysystem is far from optimal, it's still better than those heavy IDEs. Better yet, it was already mostly here a few years ago. Even to this day, it feels like it's from some point(s) in a better timeline where the world wasn't full of leftpad/polyfill/AI or whatever the next ridiculous crisis will be.

In other words, Uxn is a pidgin[2] of ideas "solarpunk" or even just small, understandable computer types like. At the same time, it does this without too much purism holding it back:

* Simple enough and familiar enough mish-mash of forth and low-color, plane-based display

* Low enough power that its cheaper than Xcode to run without more effort

* Enough IO capabilities to get things done

Since it already had all of this as of a few years ago, it's had time for social effects to take off. For example, you can share a picture, a music track, or text made using Uxn-based software. Those files will still work even if the Uxn spec evolves in backward-incompatible ways again. Its design choices are also inherently good for (or limited to) producing art aligned with the aesthetics of the creators. That seems to be part of a self-reinforcing popularity loop at this point.

Although it's theoretically possible to make similar things using CHIP-8 based tools, nobody bothers. Even if you record the output of music sequencers which have been written for it, you can't really save your work or load in data from elsewhere. In theory, there's the 16 bytes of the persistent state registers in XO-CHIP which could be repurposed from a calculator-like thing into IO, but the fact I'm mentioning it should underline the main issue in that community: purity concerns. Those limit it more than having "real" 16 buttons do. Yes, you could do some interesting multiplexing or even just bit-shift every other button press to make a pretend terminal-like keyboard, but the purity focus inadvertently kills interest more than silly function pointer syntax.

Decker is the complete opposite of Octo in this. It also goes even farther than Uxn in giving up purity concerns. Sure, it's complicated and inefficient, but it is inherently built around sharing widgets with others in a mishmash of HyperCard and Web 1.0 revival spirit:

1. You can make gifs with it, and this is probably its most popular use in the form of Wigglypaint[3]

2. When asked "um, can I borrow source from it?", the answer is "Absolutely!"[4]

This is why as much as certain Python projects frustrate me, I'm not trying to fix their inherent Python-ness right now. I just accept they're what they are. As you pointed out, it's better than the alternatives in many cases. Before you ask, I won't go as far as calling JavaScript is good, but I'll admit it's shareable. :P

[1]: https://en.wikipedia.org/wiki/Pidgin

[2]: https://beyondloom.com/blog/octojam.html

[3]: https://internet-janitor.itch.io/wigglypaint

[4]: https://itch.io/post/10183760

[kragen]

with respect to software that doesn't need to be efficient, sure, there are lots of things that are usable without being efficient. but in rek and devine's notes on working offgrid, which are part of the background motivation for uxn/varvara, they say:

> Computers are generally power-sucking vampires. Choosing different software, operating systems, or working from machines with a lower draw (ARM) or even throttling the CPU, are some of the many things we do to lower our power requirements. The way that software is built has a substantial impact on the power consumption of a system, it is shocking how cpu-intensive modern programs can be.

so uxn/varvara is not intended for software that doesn't need to be efficient, very much the contrary. so, from my point of view, the fact that logic-intensive programs running in uxn consume 20× more energy than they need to is basically a mistake; rek and devine made choices in its design that keep it from meeting their own goals. which isn't to say it's valueless, just that it's possible to do 20× better

what i mean by 'logic-intensive' is programs that spend most of their time running uxn code doing some kind of user-defined computation, like compilation, npc pathfinding, or numerical integration. if your program spends most of its cpu blitting sprites onto the screen, well, that's taken care of by the varvara code, and there's nothing in the varvara definition that requires that to be inefficient. but uxn itself is very hard to implement efficiently, which creates pressure to push more complexity into varvara, and varvara has ended up fairly complex and therefore more difficult than necessary to implement at all. and i think that's another failure of uxn/varvara to fulfill its own ideals. or anyway it does much worse according to its own ideals than a better design would

how much complexity does varvara impose on you? in https://news.ycombinator.com/item?id=32219386 i said the uxn/varvara implementation for the nintendo ds is 5200 lines of c, so roughly speaking it's about 20× more complexity than uxn itself

and that's what i mean by 'and it's not that simple'. in the comment i linked above, i pointed out that chifir (the first archival virtual machine good enough to criticize, which is a somewhat different purpose) took me 75 lines of c to implement, and adding graphical output to it with yeso required another 30 lines of code. you might reasonably wonder how much complexity i'm sweeping under the carpet of yeso, since surely some of those 5200 lines of code in the uxn/varvara implementation for the nintendo ds are imposed by the ds platform and not by varvara. the parts of yeso used by my chifir implementation compiled for x-windows are yeso.h, yeso-xlib.c, and yeso-pic.c, which total 518 lines of code according to either sloccount or cloc.

still, uxn, even with varvara, is much better than anything else out there; like i said, it's the first effort in this direction that's good enough to criticize

i don't understand what you mean about octo vs. decker but possibly that's because i haven't tried to use either one

you definitely can't deal with a 4×6 font on the ls027b7dh01 display i'm using. it's 35mm tall, including the pixel-less borders, and 240 pixels tall. so 6 pixels is 0.88 mm, or a 2.5 point font. even young people and nearsighted people typically need a magnifier to read a 2.5 point font.

[kragen]

i just found out that my critique of chifir was apparently one of the influences on the design of uxn: http://forum.malleable.systems/t/uxn-and-varvara/107/7

[kragen]

> You may be pleasantly surprised by how many people will show up ...

that would be great! on the other hand, i don't want to count on it; i want it to be a pleasant surprise :)

[kragen]

you write:

> I have a radical idea which might not be a good one:

> 1. A "Full" display of W x H characters (the Sharp display or whatever you replace it with)

> 2. A "Fallback" 1 or 2 line display (2nd line reserved for editor info / UI?)

i've thought about this, especially with respect to e-ink screens. some versions of the barnes & noble nook work this way: there's a small color lcd touchscreen for interactive stuff, and a large black-and-white e-paper display for bulk information display. i thought that a small reflective lcd like cheap scientific calculators use would be enough for a line or two of text, permitting rapid interaction without putting excessive demands on the e-paper's slow and power-hungry refresh

but i think that sharp's memory lcds are so much better than e-paper that there's no real benefit to that approach now

> A single 1-bit editable line is so hard to work with that ed[2] isn't even installed by default on many Linux distros. However, people did once got real work done with it and similar tools.

it's not that command lines are hard to work with; it's that ed's user interface is designed for 110-baud teletypes, where erasing is impossible, and typing out a 10-character error message means forcing the user to wait for an entire second, so it's preferable to just say '?\r\n'. user interfaces designed for more capable hardware, such as an adm-3a connected over a 1200-baud modem, can be immensely easier to use, without abandoning the command-line paradigm. you'll note that ex(1) is installed by default on every linux distribution except alpine (which does, oddly enough, have an ed(1)), and i know a programmer who still uses it as his preferred editor, preferring it even to vi

i've certainly done plenty of real work with tcsh and bash, where the primary user interface paradigm is editing a single line of text to your satisfaction, submitting it for execution, meditating on the results, and then repeating. not to mention the python repl, gdb's command line, octave, units(1), r, pari/gp, sqlite, psql, swi-prolog, guile, ngspice, etc. i like direct-manipulation interfaces a lot, but they're harder to design, and especially hard to design for complex problem-solving activities. it's just that correcting a typo a few words ago isn't really a complex problem-solving activity, so, by my lights, even a relatively dumb direct-manipulation text editor can beat a command-line editor like ex most of the time. especially if their bandwidth to your eyes is measured in megabits per second rather than kilobits per second

[pushfoo]

> you'll note that ex(1) is installed by default on every linux distribution except alpine

On my system, it seems to be set up as an alias of vim ("Entering Ex mode"). Huh. A bit more reading after that has been an interesting historical detour.

> i know a programmer who still uses it as his preferred editor, preferring it even to vi

I've never tried ed before this comment. Others might see it as strange and clearly teletype-oriented, but I think it's refreshingly explicit about every action taken. Even more than ex, it's a vim ancestor where commands and state are persistently visible instead of being hidden.

Thank you for mentioning this, but now I'm a little worried I might start wanting to use ed too.

> i've certainly done plenty of real work with tcsh and bash, where the primary user interface paradigm is editing a single line of text to your satisfaction, submitting it for execution, meditating on the results, and then repeating.

I think you've hit on something important here. Web-oriented workflows with live reload (and Flutter[1]) inherit this. I think it means the form of the REPL itself isn't what we should be focusing on, but easy iteration and maybe cached results to save power and time.

On the topic of accessibility, do you have any opinions on Hedy[2] and whether a long-lived, low-power system might have room for something like it? The site is annoyingly JS-filled, but the idea is interesting: a language where the names in the symbol table are more overtly a user-localizable value.

[1]: https://flutter.dev/

[2]: https://www.hedycode.com/

[kragen]

(trying not to respond at too great length to avoid being overwhelming)

vi has been a mode of ex since the inception of vi; early versions of vim (up to 3.0.0 i think) didn't implement ex mode, but later it was added. the : commands of vi are exactly the ex commands (: is the ex prompt; one of the improvements over ed enabled by higher-bandwidth terminals was explicit feedback to distinguish command mode from insert mode)

i don't think of ed as very strange. teco, now, teco is strange. ed is just a more taciturn vi. but i would never describe ed as making state persistently visible!

if you're finding yourself tempted by ed you should probably try sam instead

i don't understand the nature of hedy but it sounds like just another scripting language; what would make it especially difficult or costly to implement? i don't understand what you mean by 'user-localizable' or 'more overtly'

[pushfoo]

> ed is just a more taciturn vi.

I might quote you on this if I have to explain it to others.

> if you're finding yourself tempted by ed you should probably try sam instead

Ty, I may try it then.

> i don't understand the nature of hedy but it sounds like just another scripting language; what would make it especially difficult or costly to implement? i don't understand what you mean by 'user-localizable' or 'more overtly'

The language drop-down on the editor / try-it page translates not just page text, but the scripting language keywords itself. I haven't had time to dig into it yet, but it's inherently an internationally oriented scripting language by design since the goal is to allow adding your own language to the supported set of keyword display languages.

[kragen]

you ask what i mean about programmer productivity. consider this python code from https://norvig.com/spell-correct.html:

    def edits1(word):
        "All edits that are one edit away from `word`."
        letters    = 'abcdefghijklmnopqrstuvwxyz'
        splits     = [(word[:i], word[i:])    for i in range(len(word) + 1)]
        deletes    = [L + R[1:]               for L, R in splits if R]
        transposes = [L + R[1] + R[0] + R[2:] for L, R in splits if len(R)>1]
        replaces   = [L + c + R[1:]           for L, R in splits if R for c in letters]
        inserts    = [L + c + R               for L, R in splits for c in letters]
        return set(deletes + transposes + replaces + inserts)

in seven lines of code, it computes all the potentially incorrect words that can be made from a given correct word with a single edit. so, for example, for 'antidisestableshmentarianism', it returns a set of 1482 words such as 'antidisestauleshmentarianism', 'antidisestableshmentarianlism', 'antidiseitableshmentarianism', 'antidisestablesjhmentarianism', and 'antidiseptableshmentarianism', totaling 42194 bytes. how would you do this in uxntal?

here's another part of norvig's program. this part tabulates the case-smashed frequency of every word in its 6-megabyte training set (which presumably consists only of correctly spelled words):

    import re
    from collections import Counter

    def words(text): return re.findall(r'\w+', text.lower())

    WORDS = Counter(words(open('big.txt').read()))

this takes about 340 milliseconds one core of on my laptop here, which runs at about 6000 MIPS, so it would take about 34 seconds on a machine running at 60 MIPS, maybe a little longer on the apollo3. there are 32198 distinct words in the training set, totaling 244015 characters; the most common word ('the') occurs 79809 times, and the longest word ('disproportionately') is 18 characters. so plausibly you could represent this hash table without any compression in about 500k, though cpython requires about 70 megabytes. ram compression could plausibly get those 500k down to the 384k the apollo3 has without trying to swap to offchip flash

finding the best correction for a word requiring two corrections like 'slowlyyy' takes 70ms, so plausibly it would take 10 seconds or so on the apollo3. (you could maybe do this in the background in a text editor.) (if it were compiled to efficient code, it would probably be closer to 300 milliseconds on the apollo3, because cpython's interpretive overhead is a factor of about 40.) 'disproportionatelyyy' takes 370ms. here's the rest of the correction code:

    def P(word, N=sum(WORDS.values())): 
        "Probability of `word`."
        return WORDS[word] / N

    def correction(word): 
        "Most probable spelling correction for word."
        return max(candidates(word), key=P)

    def candidates(word): 
        "Generate possible spelling corrections for word."
        return (known([word]) or known(edits1(word)) or known(edits2(word)) or [word])

    def known(words): 
        "The subset of `words` that appear in the dictionary of WORDS."
        return set(w for w in words if w in WORDS)

    def edits2(word): 
        "All edits that are two edits away from `word`."
        return (e2 for e1 in edits1(word) for e2 in edits1(e1))

note that this requires you to have two such `edits1` sets in memory at once, though you could plausibly avoid that problem by tolerating more duplicates (double letters provoke duplicates in deletes, transposes, and replaces)

norvig doesn't tell us exactly how long it took him to write the code, but he did it in a single airplane flight, except for some minor bugs which took years to find. more importantly, though, it's very easy code to read, so you can easily understand how it works in order to modify it. and that's the most important factor for programming productivity

here are some things in this code that are more difficult to write and much more difficult to read in uxntal:

- managing more than 64k of data (uxn's memory addresses are 16 bits)

- dynamically allocating lists of things such as the (left, right) tuples in splits

- dynamic memory allocation in general

- string concatenation

- eliminating duplicates from a set of strings

- iterating over the words in a text file

- generating a sequence of outputs from a sequence of inputs with a filtering predicate and a transformation function [f(x, y) for x, y in xys if p(x, y)]

- generating a lazy flat sequence of outputs from a nested loop (return (z for y in f(x) for z in f(y)))

- hash tables

- incrementally eliminating duplicates from a sequence of candidates that turn out to be valid words (set(w for w in words if w in WORDS))

- counting the number of occurrences of each string in a lazy sequence of strings

- floating-point arithmetic (which would be fairly easy to eliminate in this case, but not in many other cases; this deficiency in uxn is especially galling since the apollo3 has fast hardware floating point)

- finding the highest-rated item of a lazy sequence of candidates according to some scoring function

and all of that is on top of the general readability tax imposed by postfix syntax, where even figuring out which arguments are being passed to which subroutine is a mental challenge and a frequent source of bugs

note that these are mostly not deficiencies you can really patch with a library. i didn't mention that the program uses regular expressions, for example, because you can certainly implement regular expressions in uxntal. they're things you probably need to address at the level of language semantics, or virtual machine semantics in the case of the address space problem. and they're not tightly tied to cpython being implemented grossly inefficiently; pypy implements all the necessary semantics, and common lisp and c++ have similar facilities in most cases, though their handling of lazy sequences is a weakness that is particularly important on hardware with limited memory like the apollo3

so that's what i mean when i say that uxn is designed to make easy things hard, rather than making hard things easy

you say:

> the pain point might be intentional nudges away from making things the designer doesn't like

the thing is, i don't really care whether rek and devine think that autocorrecting misspellings is a bad thing to do; i want the computer to be a means of expression for my ideas, indeed for everyone's ideas, not for the ideas of a singular designer. that's the apple walled-garden mindset, and it's anathema to me. and, though i could be wrong about this, i think rek and devine would probably agree

[kazinator]

TXR Lisp:

  (defun edits1 (word)
    (hash-list (build
                 (each ((i 0..(len word)) (j 1))
                   (let ((le [word 0..i]) (rj [word j..:]) (ri [word i..:]))
                     (add (join le rj))               ;; deletes
                     (each ((c "a".."z"))
                       (add (join le c rj))           ;; replacements
                       (add (join le c [word i] rj))) ;; prefixes + inserts
                     (or (empty le) (empty ri)        ;; transposes
                         (add (join [le 0..-1] [ri 0] [le -1] [ri 1..:])))))
                 (each ((c "a".."z"))
                   (add (join word c))))))            ;; suffixes

[kragen]

this doesn't look bad at all! considerably better than common lisp, in particular. but i think the flatter structure of the python improves readability, and the independence of the different clauses facilitates interactive incremental testing:

    >>> word = 'the'
    >>> letters    = 'abcdefghijklmnopqrstuvwxyz'
    >>> splits     = [(word[:i], word[i:])    for i in range(len(word) + 1)]
    >>> splits
    [('', 'the'), ('t', 'he'), ('th', 'e'), ('the', '')]
    >>> deletes    = [L + R[1:]               for L, R in splits if R]
    >>> deletes
    ['he', 'te', 'th']

but lisps are generally pretty good at that kind of thing, so i imagine you could formulate it slightly differently in txr lisp to support that kind of thing (i just don't know txr lisp)

as a semantic question, is this materializing the whole list (as the python does) or are the `add` calls inserting into the hash table as the loops run, thus eliminating duplicates?

[kazinator]

I had a bug somewhere, so I selectively off some of the add expressions. They can be commented out with #; or by flipping add to list or identity to throw away the value.

The add is something which pairs with build. Lisp doesn't have "bag-like" lists. For those times when they are helpful, we can have procedural list building syntax. The build macro creates an environment in which a number of operators like add that build up or otherwise operate on an implicit list. When the build form terminates, it returns the list. (Its sister buildn returns the last form like progn).

In this function, I could just have used (push expr stack) because we don't care about the order; there would be no nreverse. That would be a better idea, actually.

We could also add the strings to a table directly, like (set [h (join ...)] t).

The hash table is built by the hash-list call. It associates the elements of the list with themselves, so if "a" occurs in the list, the key "a" is associated with value "a".

[kragen]

> Counting mW and decades implies text > GUI.

so, i don't want to be unkind, but i think this is a misconception, or a pair of misconceptions, and i think i know where one of them comes from. i'll try to steelman it before explaining why it's wrong

the first assertion is that counting milliwatts (or actually microwatts in this case) implies that text beats guis. the second assertion is that counting decades implies that text beats guis. i have no idea where the second assertion comes from, so i'll focus on the first

a first plausible justification for the first assertion is that power usage (over leakage) is proportional to the number of toggling bits (or perhaps cubic or quadratic in the number of toggling bits if voltage scaling is an option) and a character generator toggles many fewer bits than all the computation needed to draw a gui, so applications written for a character-cell text terminal can use much less power than gui applications. a second plausible justification is that we have many examples of applications written for character-cell text terminals which use much less computation (and therefore power) than any existing gui alternative

the first justification has two problems: first, it's comparing solutions to the wrong problem, and second, character-cell text terminals aren't even the best solution to that problem; they're the best solution to a significantly different problem which we don't have

the problem that the first justification is trying to solve is the problem of generating a video signal that is useful according to the goals of the application. in that context, the character generator is running all the time as the beam refreshes the phosphors on the crt over and over again. it is true that the character generator can compute the pixels to send to the crt much faster and with less transistors than a cpu can. but in the zorzpad, the display itself is bistable, like the plato gas plasma terminals from the 01960s. there's a flip-flop printed on the glass in amorphous silicon next to each pixel. so we don't have to consume any computational power refreshing a static display (though we do have to reverse the polarity to it a few times a second to prevent gradual damage to the lcd). we only have to send it updated lines of 400 bits. we don't have to generate a continuous video signal

moreover, the fact that a character generator would toggle less bits than the cpu to generate a line of pixels is somewhat irrelevant, because ambiq doesn't sell a character generator chip, and a character generator implemented in conventional cmos from another vendor (or on an fpga) would use orders of magnitude more power than the subthreshold-logic ambiq microcontroller, so we'd have to do the character generation process on the cpu anyway. at that point, and especially considering that the memory-in-pixel lcd display hardware doesn't impose any timing constraints on us the way an ntsc or vga signal does, there's no extra cost to generating those lines in a more flexible way. even without a framebuffer, generating lines of pixels on demand, we could support glyphs positioned with pixel precision rather than character-cell precision, proportional fonts, multiple sizes, polygons, rectangles, stipple patterns, overlaid sprites, and so on

also, though, generating a line of pixels or a vga video signal from a framebuffer requires less bit toggling than generating it from a character generator. with the framebuffer you literally just read sequential words from ram and send them to the display with no processing. what character generators save isn't power; it's memory. with a character generator you can drive a 1280×1024 display with an 8×16 font (256 characters, say, so 4096 bytes of font memory) displaying 64 lines of 160 columns, requiring 10240 bytes of glyph indices, for a total of 14336 bytes of memory. for the traditional 80×25 you only need 2k of modifiable memory and a font rom. and you can do it with strictly sequential memory access (many early character-cell text terminals used delay-line memory or shift-register memory) and meeting the strict hard-real-time guarantees required by the crt hardware being driven

in this case, 400×240 1-bit pixels would require 12000 bytes of framebuffer per screen, and 24k is about 6% of the apollo3's total 384k of sram. in theory the apollo3 needs about 2.1 cycles per 32-bit word to copy data into the framebuffer without rotating it, a number i haven't tested, so 12600 cycles to redraw both screens from scratch; at 48 megahertz that's about 260 microseconds, so it doesn't use much of the chip's available horsepower even at submilliwatt power levels. the display datasheet only guarantees being able to draw two megapixels per second, so redrawing the screen should take 48000 microseconds, although others have reported that it works reliably at three times that speed. hopefully i can use the cpu's built-in spi peripheral to spew out buffered spi data to the screen while the cpu goes back into deep-sleep mode. if i use a character generator implemented in software to avoid having a framebuffer, i have to wake up the cpu periodically to generate new lines of data, using more power, not less. (which might be worth it for the memory savings, we'll see.)

even without enough memory for a framebuffer, though, character generators weren't the only way to generate hard-real-time video signals. the nintendo (nes) only had 2k+256 bytes of vram, for example, and you could do pretty elaborate graphics with that. it used tiles (which are sort of similar to font glyphs and can in fact be used for font glyphs) and sprites (which aren't), plus scrolling registers. some other hardware supported display lists of straight lines instead of or in addition to glyph indices

the second justification is basically a cohort confounder. yes, vs code uses a lot more cpu than vim. but vim was originally written for the amiga, and vs code is written in electron. applications that run in character-cell text terminals do empirically tend to be a lot faster than gui applications, but that's largely because they're older, and so they had to be written with more attention to efficiency. remember that geos provided a standard wimp interface on the commodore 64, with mouse pointers, sliders, pulldown menus, proportional fonts, overlapping windows, and so on, and all that in color without a color framebuffer, adding extra work. it was slow, especially when it had to load data from the floppy drive, but it was usable. the commodore 64 is about 0.02 dhrystone mips (https://netlib.org/performance/html/dhrystone.data.col0.html) and the apollo3 is about 60, so running something like geos on it at 10× the commodore speed ought to use about .3% of its full-speed cpu. or 3% if it's running in a simple bytecode interpreter

so that's why i think submilliwatt computing doesn't require abjuring the gui

does that make sense? have i just totally misunderstood you?

[kragen]

you ask:

> Do you favor a specific structure of hardware, OS code, and user scripts, or whatever ends up offering the best use of power and long-term device durability?

well, i think it's critical for the zorzpad to be able to rebuild its own operating system code, because otherwise it's dependent on external computers to modify its own operating system, and the external computers might have become incompatible or untrustworthy at some point. this poses tricky problems of recovery from bad operating-system updates; probably having multiple redundant microcontrollers in the device is the best option, so that on bricking one of them, i can restore its old configuration using the other

ideally, though, the core code capable of bricking it would be a small tcb, not everything that runs on the system the way it is in ms-dos or templeos. that kind of isolation is usually implemented with an mmu, but the apollo3 doesn't have an mmu, just an mpu. so currently my plan is to implement it with a low-level virtual machine instead; initially an interpreter (my prototype seems to have about 10× interpretive overhead, but probably providing a small number of extra primitives will be adequate to eliminate the worst hotspots from things like updating the display and digital signal processing) but maybe later a simple jit compiler. like most microcontrollers, the apollo3's memory/throughput quotient comes to about a millisecond, rather than the second or so that has been typical for human-interactive computation for the last 60 years, so i suspect i'll be leaning pretty hard to the side of sacrificing speed to reduce memory usage

in addition to providing the protection needed to write experimental code without risk of bricking the machine, the virtual machine can also provide services like virtual memory (including compressed ram), transactional concurrency, and transparent persistence

the nand flash chips i'm using (s34ms01g2) are supposed to consume 27 milliwatts when running, but their power budget at the nominal 1-milliwatt full power is only 300 microwatts. at full speed they're capable of 133 megabytes per second, so at 300 microwatts they might be capable of 1.5 megabytes per second. i can't write to them nearly that fast without wearing them out early, but i can load data from them that fast. in particular that means that the 24 kilobytes of a full screen can be loaded from flash every 16 milliseconds, so you can do 60fps video from flash. reloading all 384k of ram, for example to switch to a completely different workspace, would take a quarter of a second. i feel like this is enough of a quantitative difference from the speed of floppy disks on a commodore 64 or a macintosh 128 that it becomes a qualitative difference, and for applications like the spell corrector example i gave in the other comment, it might be possible to use the flash almost as if it were ram, using the ram as a cache of what's in the flash

but maybe when i gain more experience with the hardware, all that stuff will turn out to have been the wrong approach, and i'll have to go with something else