Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > comp.lang.forth > #135498 > unrolled thread

AI Forth Embedded Pair-Programming: Big Fail !

Back to article view | Back to comp.lang.forth

Contents

  AI Forth Embedded Pair-Programming: Big Fail ! Terry Porter <techman001@eternal-september.org> - 2026-09-01 02:36 +0000
    Re: AI Forth Embedded Pair-Programming: Big Fail ! albert@spenarnc.xs4all.nl - 2026-09-01 10:06 +0200
      Re: AI Forth Embedded Pair-Programming: Big Fail ! Terry Porter <techman001@eternal-september.org> - 2026-09-01 08:44 +0000
        Re: AI Forth Embedded Pair-Programming: Big Fail ! albert@spenarnc.xs4all.nl - 2026-09-01 12:39 +0200
    Re: AI Forth Embedded Pair-Programming: Big Fail ! Hans Bezemer <the.beez.speaks@gmail.com> - 2026-09-01 19:03 +0200
      Re: AI Forth Embedded Pair-Programming: Big Fail ! Terry Porter <techman001@eternal-september.org> - 2026-09-01 20:37 +0000
        Re: AI Forth Embedded Pair-Programming: Big Fail ! Hans Bezemer <the.beez.speaks@gmail.com> - 2026-09-01 23:15 +0200
        Re: AI Forth Embedded Pair-Programming: Big Fail ! Buzz McCool <buzz_mccool@yahoo.com> - 2026-09-02 08:42 -0700
          Re: AI Forth Embedded Pair-Programming: Big Fail ! Terry Porter <techman001@eternal-september.org> - 2026-09-02 20:21 +0000
            Re: AI Forth Embedded Pair-Programming: Big Fail ! Buzz McCool <buzz_mccool@yahoo.com> - 2026-09-02 16:11 -0700
      Re: AI Forth Embedded Pair-Programming: Big Fail ! albert@spenarnc.xs4all.nl - 2026-09-02 10:53 +0200

#135498 — AI Forth Embedded Pair-Programming: Big Fail !

https://youtu.be/1aGIpoakZTY

The third and most honest episode in the story of a technician and his AI
partner — and the one where the partnership finds its real shape.

It started with a clean win. In Python, Tkinter, databases and glue code,
DeepSeek proved invaluable — it co-authored Regmon, the register monitor
that reads the live STM32 silicon over the debug wire, and LiveCheck, the
tool that checks your Forth against the real chip as you type. Two applications
a technician needed for decades. The tooling was a complete, unambiguous
success.

Then came the walls. The hardware wall — a Nokia 5110 LCD brought up over
and over, every blank screen tracing back to a wire, never a word. The lesson:
the human proves the bench, the agent amplifies the design. Then the
configuration wall — the code claiming the SPI was a master while the live
register read showed it running as a slave, plus a timing trap that truncated
every byte. Both found by believing the silicon, not the comments.

And then, once every piece of infrastructure was in place — the hardware
proven, the SQLite databases built, the MCP servers wired — the real wall
appeared. The one the tooling can't fix: the training data.

DeepSeek is deeply trained in Python and C, but Mecrisp-Stellaris Forth is a
niche dialect of a niche language, and the model's training data for it is thin
— beginner-level at best. The AI can read registers, build databases and write
MCP servers, but when it comes to actually designing and writing the Forth, it
is honestly a basic Forth beginner.

Terry's conclusion, stated plainly: the agent builds the tooling and answers
the design questions, but the human does all the Forth design — including the
coding. For now, Terry does Forth bench development alone. No more pair
programming in embedded Forth. Not a failure — an honest division of labour,
found the hard way, and it works.

This may change as training data grows, or if the C ecosystem shifts. But for
now: the tools are there, the design is human, and that's the honest state of
AI and Forth on a real bench.

(About the teller: Terry Porter is a 72-year-old electronics technician with
50 years of soldering and circuits, Forth since the 1990s. He dictates his
notes and an AI helps him write them up clearly — but the bench work itself,
the registers, the LEDs, the clocks, the measurements, is his, done live.)

Produced with the help of AI without which this video would not have been
possible

[toc] | [next] | [standalone]

#135499

In article <1175dne$1eoec$1@dont-email.me>,
Terry Porter  <techman001@eternal-september.org> wrote:
>https://youtu.be/1aGIpoakZTY
>
<SNIP>
>Then came the walls. The hardware wall — a Nokia 5110 LCD brought up over
>and over, every blank screen tracing back to a wire, never a word. The lesson:
>the human proves the bench, the agent amplifies the design. Then the
>configuration wall — the code claiming the SPI was a master while the live
>register read showed it running as a slave, plus a timing trap that truncated
>every byte. Both found by believing the silicon, not the comments.

This is to me incomprehensible, bordering to nonsense.
<SNIP>

>Produced with the help of AI without which this video would not have been
>possible

As the result is not satisfying, I think you should try harder.

Groetjes Albert

>
-- 
The Chinese government is satisfied with its military superiority over USA.
The next 5 year plan has as primary goal to advance life expectancy
over 80 years, like Western Europe.

[toc] | [prev] | [next] | [standalone]

#135500

On 2026-09-01, albert@spenarnc.xs4all.nl <albert@spenarnc.xs4all.nl> wrote:
> In article <1175dne$1eoec$1@dont-email.me>,
> Terry Porter  <techman001@eternal-september.org> wrote:
>>https://youtu.be/1aGIpoakZTY
>>
><SNIP>
>>Then came the walls. The hardware wall — a Nokia 5110 LCD brought up over
>>and over, every blank screen tracing back to a wire, never a word. The lesson:
>>the human proves the bench, the agent amplifies the design. Then the
>>configuration wall — the code claiming the SPI was a master while the live
>>register read showed it running as a slave, plus a timing trap that truncated
>>every byte. Both found by believing the silicon, not the comments.
>
> This is to me incomprehensible, bordering to nonsense.

That's one of the reasons I'm no longer trying to use an Agent to write my
Forth code, it can't make embedded designs work and can't explain why.

This is further complicated by NotebookLM which while adding a lot of extra
insight does get carried away at times.

I haven't got the display working yet as I'm not a programmer, however I'm
slowly working my way thru it using a local agent to explain the pieces.
I have not worked with matrix displays before so it is all new ground to
me.

><SNIP>
>
>>Produced with the help of AI without which this video would not have been
>>possible
>
> As the result is not satisfying, I think you should try harder.

I'm sorry you are not satisfied, Albert, I had hoped the video would would be
informative for those contemplating using embedded Forth with AI.

I put a lot of time and effort into the hardware, the experimentation and
 the documentation.

Perhaps others will try similar experiments (I hope so), but so far I seem to
be the only hardware guy publishing any material in this embedded area.

>
> Groetjes Albert
>
>>

Cheers,
Terry

[toc] | [prev] | [next] | [standalone]

#135502

In article <117639r$1l4q1$1@dont-email.me>,
Terry Porter  <techman001@eternal-september.org> wrote:
<SNIP>
>I'm sorry you are not satisfied, Albert, I had hoped the video would would be
>informative for those contemplating using embedded Forth with AI.

Not so much informative, as inspiring. That is okay too.

>> Groetjes Albert
>>
>>>
>
>Cheers,
>Terry
-- 
The Chinese government is satisfied with its military superiority over USA.
The next 5 year plan has as primary goal to advance life expectancy
over 80 years, like Western Europe.

[toc] | [prev] | [next] | [standalone]

#135506

Well, 40 minutes, often watching the very same image for over five 
minutes with AI voices blabbering away. I mean: really?

But I can tell you what's the limiting factor to AI and Forth. When 
there are more than three values to handle, things get tricky -- and AI 
generation becomes more prone to error.

BTW, I've gained that number from theoretical analysis and observing the 
AI countless times. I gave the AI a simple mission: write me a fast 
integer square root routine. Gemini managed it -- but I had to fix 
several errors.

https://sourceforge.net/p/forth-4th/code/HEAD/tree/trunk/4th.src/lib/gsqrt.4th

Grok has been the only one so far who wrote a flawless word. Now -- 
that's child's play. Others have confirmed as well that AI and Forth are 
not the best couple to do code generation. So, I guess I'm not the only 
one having that experience.

https://sourceforge.net/p/forth-4th/code/HEAD/tree/trunk/4th.src/lib/msqrt.4th

That doesn't mean AI and Forth development aren't compatible. AI is 
eerily capable to analyze Forth code. It does a fair job playing a 
sparring partner -- like debugging or optimization. But plain code 
generation? Building your own code is faster IMHO.

Hans Bezemer

On 01-09-2026 04:36, Terry Porter wrote:
> https://youtu.be/1aGIpoakZTY
> 
> The third and most honest episode in the story of a technician and his AI
> partner — and the one where the partnership finds its real shape.
> 
> It started with a clean win. In Python, Tkinter, databases and glue code,
> DeepSeek proved invaluable — it co-authored Regmon, the register monitor
> that reads the live STM32 silicon over the debug wire, and LiveCheck, the
> tool that checks your Forth against the real chip as you type. Two applications
> a technician needed for decades. The tooling was a complete, unambiguous
> success.
> 
> Then came the walls. The hardware wall — a Nokia 5110 LCD brought up over
> and over, every blank screen tracing back to a wire, never a word. The lesson:
> the human proves the bench, the agent amplifies the design. Then the
> configuration wall — the code claiming the SPI was a master while the live
> register read showed it running as a slave, plus a timing trap that truncated
> every byte. Both found by believing the silicon, not the comments.
> 
> And then, once every piece of infrastructure was in place — the hardware
> proven, the SQLite databases built, the MCP servers wired — the real wall
> appeared. The one the tooling can't fix: the training data.
> 
> DeepSeek is deeply trained in Python and C, but Mecrisp-Stellaris Forth is a
> niche dialect of a niche language, and the model's training data for it is thin
> — beginner-level at best. The AI can read registers, build databases and write
> MCP servers, but when it comes to actually designing and writing the Forth, it
> is honestly a basic Forth beginner.
> 
> Terry's conclusion, stated plainly: the agent builds the tooling and answers
> the design questions, but the human does all the Forth design — including the
> coding. For now, Terry does Forth bench development alone. No more pair
> programming in embedded Forth. Not a failure — an honest division of labour,
> found the hard way, and it works.
> 
> This may change as training data grows, or if the C ecosystem shifts. But for
> now: the tools are there, the design is human, and that's the honest state of
> AI and Forth on a real bench.
> 
> (About the teller: Terry Porter is a 72-year-old electronics technician with
> 50 years of soldering and circuits, Forth since the 1990s. He dictates his
> notes and an AI helps him write them up clearly — but the bench work itself,
> the registers, the LEDs, the clocks, the measurements, is his, done live.)
> 
> Produced with the help of AI without which this video would not have been
> possible
> 

[toc] | [prev] | [next] | [standalone]

#135512

On 2026-09-01, Hans Bezemer <the.beez.speaks@gmail.com> wrote:
> Well, 40 minutes, often watching the very same image for over five 
> minutes with AI voices blabbering away. I mean: really?

It seems you only had a quick look at the Video summary, because you don't
seem aware that my video was about **embedded Forth** ?

It's not about AI and Forth on a PC, which is what you do. It's not about
software.

They're totally seperate disciplines. Hardware guys don't program and
programmers don't do hardware. Why ? no one lives long enough to master
both disciplines. Chuck invented Forth to do hardware at Kit Peak.

He went on to design the Novix CPU and then the GA144. Chuck absolutely
does hardware and software. Chuck is a Unicorn.

I guarantee that AI knows more about hardware than you do, and more
about software than I do.

<snip>
> Grok has been the only one so far who wrote a flawless word. Now -- 
> that's child's play. Others have confirmed as well that AI and Forth are 
> not the best couple to do code generation. So, I guess I'm not the only 
> one having that experience.

I have to disgree with you here, because in my opinion Forth and AI would be
ideal partners in the hardware area, but if only AI had the Forth training
data, which it doesn't.




--
Creator and maintainer of "The Unauthorised Mecrisp-Stellaris UserDoc (2014)"

[toc] | [prev] | [next] | [standalone]

#135514

On 01-09-2026 22:37, Terry Porter wrote:
> I have to disgree with you here, because in my opinion Forth and AI would be
> ideal partners in the hardware area, but if only AI had the Forth training
> data, which it doesn't.
What I've heard -- of people using Forth as an AI test bed -- it's the 
serialization of the data that's fundamentally different from other 
languages than Forth. It's just a complication they can't do above a 
certain level (that's why the number "3" pops up all the time).

And unless you provide proof for that "training" statement, the rule 
"What can be asserted without evidence can be dismissed without evidence."

Because we agree: it can't be done now.

Hans Bezemer

[toc] | [prev] | [next] | [standalone]

#135528

On 9/1/2026 1:37 PM, Terry Porter wrote:
> ... Hardware guys don't program and programmers don't do hardware.
> Why ? no one lives long enough to master both disciplines.

Are we supposed to take the bait, or is this just AI attempting to
sow discord?

[toc] | [prev] | [next] | [standalone]

#135535

On 2026-09-02, Buzz McCool <buzz_mccool@yahoo.com> wrote:
> On 9/1/2026 1:37 PM, Terry Porter wrote:
>> ... Hardware guys don't program and programmers don't do hardware.
>> Why ? no one lives long enough to master both disciplines.
>
> Are we supposed to take the bait, or is this just AI attempting to
> sow discord?
>

plonk

-- 
--
Creator and maintainer of "The Unauthorised Mecrisp-Stellaris UserDoc (2014)"

[toc] | [prev] | [next] | [standalone]

#135536

On 9/2/2026 1:21 PM, Terry Porter wrote:
> 
> plonk
> 
Advice taken

[toc] | [prev] | [next] | [standalone]

#135524

In article <11770gu$20v17$1@dont-email.me>,
Hans Bezemer  <the.beez.speaks@gmail.com> wrote:
>Well, 40 minutes, often watching the very same image for over five
>minutes with AI voices blabbering away. I mean: really?
>
>But I can tell you what's the limiting factor to AI and Forth. When
>there are more than three values to handle, things get tricky -- and AI
>generation becomes more prone to error.
>
>BTW, I've gained that number from theoretical analysis and observing the
>AI countless times. I gave the AI a simple mission: write me a fast
>integer square root routine. Gemini managed it -- but I had to fix
>several errors.

It is a miracle that Language Model should work at all for these
tasks. Advanced results (e.g. progress in the Riemann hypotheses)
were obtained in a similar fashion as a professor guiding his
Phd Students. Real thinking involves building a mental model
of the problem, this is closer to that.

Groetjes Albert
-- 
The Chinese government is satisfied with its military superiority over USA.
The next 5 year plan has as primary goal to advance life expectancy
over 80 years, like Western Europe.

[toc] | [prev] | [standalone]

Back to top | Article view | comp.lang.forth

csiph-web