DB
David Bernat
Sat, Jan 3, 2026 4:33 PM
Hey SCAD Cats,
I recently kicked the tires of the MCP server for OpenSCAD by Quellant, and
have non-positive reviews of the design space, but wish to invite all of
you to directly email me if this is a workflow and design pattern space
which interests you. It is probably important for the community to know.
To recap, very succinctly:
LLMs are large language models, i.e., generative AI, and are quite
impressive, even on-premise.
MCPs are toolkits LLMs can use to make API or code executions instead of
working in language.
When I said LLMs are quite impressive on-premise I mean that a thousand
bucks is more than enough to set up your own machine with year-old results,
and modern tech company LLMs can produce stunning results for a few bucks
per session. As a software developer, the idea of using tech company LLMs
to spend $20/hour to vibe code does not interest me, but in a year the idea
that my home rig is not vibe-coding for me to check in every hour is
probably a gold standard now.
To create OpenSCAD code the modern way is to use an LLM for natural
language, an LLM which is specialized in understanding OpenSCAD code, and
an OpenSCAD MCP which has a set of direct APIs for generation,
documentation lookup, testing, and debugging etc. The user uses an
interface specific to understanding how to code, test, etc., of which the
commercial ones are Claude, Windsurf, Cursor, etc., but the open source
OpenCode is by far my favorite.
OpenCode is a must-use in my experience. It helps developers of "Agentic
AI" to partition and share design characteristics. So, for instance, my
Valley Girl Pepperdine Lawyer agent, which does not software engineer, does
use tools to research the web when I ask about a book that I am reading,
and will engage with those tools aligned with the persona (and political
leanings, or other knowledge files) that I describe in its design files.
These are very impressive technologies, and in my opinion are the "it is
finally here" to, say, for example wire up my house with audio to interact
with those photographs that I digitized and put into a database last year.
The big change is that all those glue code interactions are handled by the
LLM now instead, which makes lateral use cases much easier and more fun,
finally making a "platform" available.
And what about OpenSCAD?
With a basic few days understanding of OpenCode, adding the OpenSCAD
specific MCP took a few minutes, and within an hour I was copying and
pasting OpenSCAD code into the render engine of the OpenSCAD app. If you
can do this and want to go further, I invite you to email me directly.
How was the performance of OpenSCAD & the MCP?
- Does it render? code rendered most of the time, but lacked a plugin on
the MCP to automatically update the OpenSCAD application with new
iterations and auto-render the next-iteration code. This may seem like a
convenience, but is important for the life-cycle and exponential
accessibility of the MCP
- Does the MCP offer closed-loops? copying and pasting error codes into
the OpenCode agent occasionally worked at finding its own errors. This is
the singularly most important benchmark for "automated" agents, since, as
you can imagine, a closed-loop between the OpenCode and MCP should exist
that proposes code, renders the code, is alerted of compiling errors,
iterates the code, until a successful render is complete. No human needs to
be involved in this step, and the "how complex can the MCP achieve" is
logarithmic to the error rate, and the length of time the code agent can be
left alone is exponential in the error rate.
- Are the errors captured for analysts to inspect? No. In fact, you can
imagine a future in which code agents aggregate all the users' compile and
render errors throughout the day and blast a developer@openscad.org email
address for those who love to get into the details of the programming
language. This "human-in-the-loop" model happens to be my career, and you
can imagine the fun as developers are prompted with more and more complex
errors (and become even more talented). This makes for a very strong
software organization (ahem).
- Are the errors captured for the MCP to inspect? No. Flip side of that is
that the MCP should maintain a long list of "gotchas" that itself can check
against in the process of accessing the viability and appropriateness of
the code it just generated. This is not as difficult as it sounds.
- Is auto-code any good? Pessimistic for at least a year, but excited. The
simple problem is that natural language does not encode spatial
relationships; and so the act of trying to talk about complex geometries is
both taxing for the user and hard to precisely interpret by the agent. So,
for instance, a sphere rests on top of a box, and applying a fillet to the
top four edges are all within its wheelhouse simply; but it is the wrong
kind of fillet, and trying to switch that makes a non-fillet object, and so
on. On the other hand, if you want to talk about what the Berliner Dom
looks like, the OpenCode can probably understand its images or a CAD that
gets downloaded from Thangs. This duality is a problem in the (limitations
of) MCP (design as currently a standard), and universal to all these
packages, including Blender and Fusion etc. These spatial relationships,
though, could be encoded in geometric-LLMs, and nobody is paying
attention to that right now either, sadly.
- And what about the MCP of Quellant specifically? I have not looked into
the code, and I do not know the individual, but obviously the MCP works and
that is more than I have done so far. It is a tool that does what it is
stated to do, open source, and I am sure can be expanded upon.
That is the news. If this interests you, please email. If you want me to
show slides of examples, or record a video, etc., I would be happy to do so
and email to this discussion forum email here. If you want to take this
topic (very very) seriously I can be enticed to do so, and my first
impression is that the group should only take this seriously if the group
- wants to get good at mastering MCP design for the open source, 2. wants
to get into geometric reasoning LLMs, 3. wants to learn OpenCode and thinks
OpenSCAD is a fun domain to practice, 4. wants to teach me OpenSCAD but
wants to think seriously about its kernel physics, 5. has nothing else to do
I have a $900 V100 arriving this week so that I can compare the open source
on-premise LLMs for software engineering to the $20/hour versions (which
are free through OpenCode, by the way). I have a strong negative opinion on
OpenAI and ClaudeCode, and several other tech companies which are
non-negotiable, and I am a former Chief Scientist at a YC company too.
Have a happy new year.
Sincerely,
Bernat
David Bernat, Ph. D.
Property of Starlight LLC.
Hey SCAD Cats,
I recently kicked the tires of the MCP server for OpenSCAD by Quellant, and
have non-positive reviews of the design space, but wish to invite all of
you to directly email me if this is a workflow and design pattern space
which interests you. It is probably important for the community to know.
To recap, very succinctly:
LLMs are large language models, i.e., generative AI, and are quite
impressive, even on-premise.
MCPs are toolkits LLMs can use to make API or code executions instead of
working in language.
When I said LLMs are quite impressive on-premise I mean that a thousand
bucks is more than enough to set up your own machine with year-old results,
and modern tech company LLMs can produce stunning results for a few bucks
per session. As a software developer, the idea of using tech company LLMs
to spend $20/hour to vibe code does not interest me, but in a year the idea
that my home rig is not vibe-coding for me to check in every hour is
probably a gold standard now.
To create OpenSCAD code the modern way is to use an LLM for natural
language, an LLM which is specialized in understanding OpenSCAD code, and
an OpenSCAD MCP which has a set of direct APIs for generation,
documentation lookup, testing, and debugging etc. The user uses an
interface specific to understanding how to code, test, etc., of which the
commercial ones are Claude, Windsurf, Cursor, etc., but the open source
OpenCode is by far my favorite.
OpenCode is a must-use in my experience. It helps developers of "Agentic
AI" to partition and share design characteristics. So, for instance, my
Valley Girl Pepperdine Lawyer agent, which does not software engineer, does
use tools to research the web when I ask about a book that I am reading,
and will engage with those tools aligned with the persona (and political
leanings, or other knowledge files) that I describe in its design files.
These are very impressive technologies, and in my opinion are the "it is
finally here" to, say, for example wire up my house with audio to interact
with those photographs that I digitized and put into a database last year.
The big change is that all those glue code interactions are handled by the
LLM now instead, which makes lateral use cases much easier and more fun,
finally making a "platform" available.
And what about OpenSCAD?
With a basic few days understanding of OpenCode, adding the OpenSCAD
specific MCP took a few minutes, and within an hour I was copying and
pasting OpenSCAD code into the render engine of the OpenSCAD app. If you
can do this and want to go further, I invite you to email me directly.
How was the performance of OpenSCAD & the MCP?
1. Does it render? code rendered most of the time, but lacked a plugin on
the MCP to automatically update the OpenSCAD application with new
iterations and auto-render the next-iteration code. This may seem like a
convenience, but is important for the life-cycle and exponential
accessibility of the MCP
2. Does the MCP offer closed-loops? copying and pasting error codes into
the OpenCode agent occasionally worked at finding its own errors. This is
the singularly most important benchmark for "automated" agents, since, as
you can imagine, a closed-loop between the OpenCode and MCP should exist
that proposes code, renders the code, is alerted of compiling errors,
iterates the code, until a successful render is complete. No human needs to
be involved in this step, and the "how complex can the MCP achieve" is
logarithmic to the error rate, and the length of time the code agent can be
left alone is exponential in the error rate.
3. Are the errors captured for analysts to inspect? No. In fact, you can
imagine a future in which code agents aggregate all the users' compile and
render errors throughout the day and blast a developer@openscad.org email
address for those who love to get into the details of the programming
language. This "human-in-the-loop" model happens to be my career, and you
can imagine the fun as developers are prompted with more and more complex
errors (and become even more talented). This makes for a very strong
software organization (ahem).
4. Are the errors captured for the MCP to inspect? No. Flip side of that is
that the MCP should maintain a long list of "gotchas" that itself can check
against in the process of accessing the viability and appropriateness of
the code it just generated. This is not as difficult as it sounds.
5. Is auto-code any good? Pessimistic for at least a year, but excited. The
simple problem is that natural language does not encode spatial
relationships; and so the act of trying to talk about complex geometries is
both taxing for the user and hard to precisely interpret by the agent. So,
for instance, a sphere rests on top of a box, and applying a fillet to the
top four edges are all within its wheelhouse simply; but it is the wrong
kind of fillet, and trying to switch that makes a non-fillet object, and so
on. On the other hand, if you want to talk about what the Berliner Dom
looks like, the OpenCode can probably understand its images or a CAD that
gets downloaded from Thangs. This duality is a problem in the (limitations
of) MCP (design as currently a standard), and universal to all these
packages, including Blender and Fusion etc. These spatial relationships,
though, **could** be encoded in geometric-LLMs, and nobody is paying
attention to that right now either, sadly.
6. And what about the MCP of Quellant specifically? I have not looked into
the code, and I do not know the individual, but obviously the MCP works and
that is more than I have done so far. It is a tool that does what it is
stated to do, open source, and I am sure can be expanded upon.
That is the news. If this interests you, please email. If you want me to
show slides of examples, or record a video, etc., I would be happy to do so
and email to this discussion forum email here. If you want to take this
topic (very very) seriously I can be enticed to do so, and my first
impression is that the group should only take this seriously if the group
1. wants to get good at mastering MCP design for the open source, 2. wants
to get into geometric reasoning LLMs, 3. wants to learn OpenCode and thinks
OpenSCAD is a fun domain to practice, 4. wants to teach me OpenSCAD but
wants to think seriously about its kernel physics, 5. has nothing else to do
I have a $900 V100 arriving this week so that I can compare the open source
on-premise LLMs for software engineering to the $20/hour versions (which
are free through OpenCode, by the way). I have a strong negative opinion on
OpenAI and ClaudeCode, and several other tech companies which are
non-negotiable, and I am a former Chief Scientist at a YC company too.
Have a happy new year.
Sincerely,
Bernat
David Bernat, Ph. D.
Property of Starlight LLC.
JB
Jon Bondy
Sat, Jan 3, 2026 6:49 PM
David:
Very interesting. Why not discuss this here, rather than in private
messages to you?
You installed this software on a computer local to you and then used it
to write OS code?
Jon
On 1/3/2026 11:33 AM, David Bernat via Discuss wrote:
Hey SCAD Cats,
I recently kicked the tires of the MCP server for OpenSCAD by
Quellant, and have non-positive reviews of the design space, but wish
to invite all of you to directly email me if this is a workflow and
design pattern space which interests you. It is probably important for
the community to know.
To recap, very succinctly:
LLMs are large language models, i.e., generative AI, and are quite
impressive, even on-premise.
MCPs are toolkits LLMs can use to make API or code executions instead
of working in language.
When I said LLMs are quite impressive on-premise I mean that a
thousand bucks is more than enough to set up your own machine with
year-old results, and modern tech company LLMs can produce stunning
results for a few bucks per session. As a software developer, the idea
of using tech company LLMs to spend $20/hour to vibe code does not
interest me, but in a year the idea that my home rig is not
vibe-coding for me to check in every hour is probably a gold standard now.
To create OpenSCAD code the modern way is to use an LLM for natural
language, an LLM which is specialized in understanding OpenSCAD code,
and an OpenSCAD MCP which has a set of direct APIs for generation,
documentation lookup, testing, and debugging etc. The user uses an
interface specific to understanding how to code, test, etc., of which
the commercial ones are Claude, Windsurf, Cursor, etc., but the open
source OpenCode is by far my favorite.
OpenCode is a must-use in my experience. It helps developers of
"Agentic AI" to partition and share design characteristics. So, for
instance, my Valley Girl Pepperdine Lawyer agent, which does not
software engineer, does use tools to research the web when I ask about
a book that I am reading, and will engage with those tools aligned
with the persona (and political leanings, or other knowledge files)
that I describe in its design files. These are very impressive
technologies, and in my opinion are the "it is finally here" to, say,
for example wire up my house with audio to interact with those
photographs that I digitized and put into a database last year. The
big change is that all those glue code interactions are handled by the
LLM now instead, which makes lateral use cases much easier and more
fun, finally making a "platform" available.
And what about OpenSCAD?
With a basic few days understanding of OpenCode, adding the OpenSCAD
specific MCP took a few minutes, and within an hour I was copying and
pasting OpenSCAD code into the render engine of the OpenSCAD app. If
you can do this and want to go further, I invite you to email me directly.
How was the performance of OpenSCAD & the MCP?
- Does it render? code rendered most of the time, but lacked a plugin
on the MCP to automatically update the OpenSCAD application with new
iterations and auto-render the next-iteration code. This may seem like
a convenience, but is important for the life-cycle and exponential
accessibility of the MCP
- Does the MCP offer closed-loops? copying and pasting error codes
into the OpenCode agent occasionally worked at finding its own errors.
This is the singularly most important benchmark for "automated"
agents, since, as you can imagine, a closed-loop between the OpenCode
and MCP should exist that proposes code, renders the code, is alerted
of compiling errors, iterates the code, until a successful render is
complete. No human needs to be involved in this step, and the "how
complex can the MCP achieve" is logarithmic to the error rate, and the
length of time the code agent can be left alone is exponential in the
error rate.
- Are the errors captured for analysts to inspect? No. In fact, you
can imagine a future in which code agents aggregate all the users'
compile and render errors throughout the day and blast a
developer@openscad.org email address for those who love to get into
the details of the programming language. This "human-in-the-loop"
model happens to be my career, and you can imagine the fun as
developers are prompted with more and more complex errors (and become
even more talented). This makes for a very strong software
organization (ahem).
- Are the errors captured for the MCP to inspect? No. Flip side of
that is that the MCP should maintain a long list of "gotchas" that
itself can check against in the process of accessing the viability and
appropriateness of the code it just generated. This is not as
difficult as it sounds.
- Is auto-code any good? Pessimistic for at least a year, but
excited. The simple problem is that natural language does not encode
spatial relationships; and so the act of trying to talk about complex
geometries is both taxing for the user and hard to precisely interpret
by the agent. So, for instance, a sphere rests on top of a box, and
applying a fillet to the top four edges are all within its wheelhouse
simply; but it is the wrong kind of fillet, and trying to switch that
makes a non-fillet object, and so on. On the other hand, if you want
to talk about what the Berliner Dom looks like, the OpenCode can
probably understand its images or a CAD that gets downloaded from
Thangs. This duality is a problem in the (limitations of) MCP (design
as currently a standard), and universal to all these packages,
including Blender and Fusion etc. These spatial relationships, though,
could be encoded in geometric-LLMs, and nobody is paying attention
to that right now either, sadly.
- And what about the MCP of Quellant specifically? I have not looked
into the code, and I do not know the individual, but obviously the MCP
works and that is more than I have done so far. It is a tool that does
what it is stated to do, open source, and I am sure can be expanded upon.
That is the news. If this interests you, please email. If you want me
to show slides of examples, or record a video, etc., I would be happy
to do so and email to this discussion forum email here. If you want to
take this topic (very very) seriously I can be enticed to do so, and
my first impression is that the group should only take this seriously
if the group 1. wants to get good at mastering MCP design for the open
source, 2. wants to get into geometric reasoning LLMs, 3. wants to
learn OpenCode and thinks OpenSCAD is a fun domain to practice, 4.
wants to teach me OpenSCAD but wants to think seriously about its
kernel physics, 5. has nothing else to do
I have a $900 V100 arriving this week so that I can compare the open
source on-premise LLMs for software engineering to the $20/hour
versions (which are free through OpenCode, by the way). I have a
strong negative opinion on OpenAI and ClaudeCode, and several other
tech companies which are non-negotiable, and I am a former Chief
Scientist at a YC company too.
Have a happy new year.
Sincerely,
Bernat
David Bernat, Ph. D.
Property of Starlight LLC.
OpenSCAD mailing list
To unsubscribe send an email todiscuss-leave@lists.openscad.org
--
This email has been checked for viruses by AVG antivirus software.
www.avg.com
David:
Very interesting. Why not discuss this here, rather than in private
messages to you?
You installed this software on a computer local to you and then used it
to write OS code?
Jon
On 1/3/2026 11:33 AM, David Bernat via Discuss wrote:
> Hey SCAD Cats,
>
> I recently kicked the tires of the MCP server for OpenSCAD by
> Quellant, and have non-positive reviews of the design space, but wish
> to invite all of you to directly email me if this is a workflow and
> design pattern space which interests you. It is probably important for
> the community to know.
>
> To recap, very succinctly:
> LLMs are large language models, i.e., generative AI, and are quite
> impressive, even on-premise.
> MCPs are toolkits LLMs can use to make API or code executions instead
> of working in language.
>
> When I said LLMs are quite impressive on-premise I mean that a
> thousand bucks is more than enough to set up your own machine with
> year-old results, and modern tech company LLMs can produce stunning
> results for a few bucks per session. As a software developer, the idea
> of using tech company LLMs to spend $20/hour to vibe code does not
> interest me, but in a year the idea that my home rig is not
> vibe-coding for me to check in every hour is probably a gold standard now.
>
> To create OpenSCAD code the modern way is to use an LLM for natural
> language, an LLM which is specialized in understanding OpenSCAD code,
> and an OpenSCAD MCP which has a set of direct APIs for generation,
> documentation lookup, testing, and debugging etc. The user uses an
> interface specific to understanding how to code, test, etc., of which
> the commercial ones are Claude, Windsurf, Cursor, etc., but the open
> source OpenCode is by far my favorite.
>
> OpenCode is a must-use in my experience. It helps developers of
> "Agentic AI" to partition and share design characteristics. So, for
> instance, my Valley Girl Pepperdine Lawyer agent, which does not
> software engineer, does use tools to research the web when I ask about
> a book that I am reading, and will engage with those tools aligned
> with the persona (and political leanings, or other knowledge files)
> that I describe in its design files. These are very impressive
> technologies, and in my opinion are the "it is finally here" to, say,
> for example wire up my house with audio to interact with those
> photographs that I digitized and put into a database last year. The
> big change is that all those glue code interactions are handled by the
> LLM now instead, which makes lateral use cases much easier and more
> fun, finally making a "platform" available.
>
> And what about OpenSCAD?
> With a basic few days understanding of OpenCode, adding the OpenSCAD
> specific MCP took a few minutes, and within an hour I was copying and
> pasting OpenSCAD code into the render engine of the OpenSCAD app. If
> you can do this and want to go further, I invite you to email me directly.
>
> How was the performance of OpenSCAD & the MCP?
> 1. Does it render? code rendered most of the time, but lacked a plugin
> on the MCP to automatically update the OpenSCAD application with new
> iterations and auto-render the next-iteration code. This may seem like
> a convenience, but is important for the life-cycle and exponential
> accessibility of the MCP
> 2. Does the MCP offer closed-loops? copying and pasting error codes
> into the OpenCode agent occasionally worked at finding its own errors.
> This is the singularly most important benchmark for "automated"
> agents, since, as you can imagine, a closed-loop between the OpenCode
> and MCP should exist that proposes code, renders the code, is alerted
> of compiling errors, iterates the code, until a successful render is
> complete. No human needs to be involved in this step, and the "how
> complex can the MCP achieve" is logarithmic to the error rate, and the
> length of time the code agent can be left alone is exponential in the
> error rate.
> 3. Are the errors captured for analysts to inspect? No. In fact, you
> can imagine a future in which code agents aggregate all the users'
> compile and render errors throughout the day and blast a
> developer@openscad.org email address for those who love to get into
> the details of the programming language. This "human-in-the-loop"
> model happens to be my career, and you can imagine the fun as
> developers are prompted with more and more complex errors (and become
> even more talented). This makes for a very strong software
> organization (ahem).
> 4. Are the errors captured for the MCP to inspect? No. Flip side of
> that is that the MCP should maintain a long list of "gotchas" that
> itself can check against in the process of accessing the viability and
> appropriateness of the code it just generated. This is not as
> difficult as it sounds.
> 5. Is auto-code any good? Pessimistic for at least a year, but
> excited. The simple problem is that natural language does not encode
> spatial relationships; and so the act of trying to talk about complex
> geometries is both taxing for the user and hard to precisely interpret
> by the agent. So, for instance, a sphere rests on top of a box, and
> applying a fillet to the top four edges are all within its wheelhouse
> simply; but it is the wrong kind of fillet, and trying to switch that
> makes a non-fillet object, and so on. On the other hand, if you want
> to talk about what the Berliner Dom looks like, the OpenCode can
> probably understand its images or a CAD that gets downloaded from
> Thangs. This duality is a problem in the (limitations of) MCP (design
> as currently a standard), and universal to all these packages,
> including Blender and Fusion etc. These spatial relationships, though,
> **could** be encoded in geometric-LLMs, and nobody is paying attention
> to that right now either, sadly.
> 6. And what about the MCP of Quellant specifically? I have not looked
> into the code, and I do not know the individual, but obviously the MCP
> works and that is more than I have done so far. It is a tool that does
> what it is stated to do, open source, and I am sure can be expanded upon.
>
> That is the news. If this interests you, please email. If you want me
> to show slides of examples, or record a video, etc., I would be happy
> to do so and email to this discussion forum email here. If you want to
> take this topic (very very) seriously I can be enticed to do so, and
> my first impression is that the group should only take this seriously
> if the group 1. wants to get good at mastering MCP design for the open
> source, 2. wants to get into geometric reasoning LLMs, 3. wants to
> learn OpenCode and thinks OpenSCAD is a fun domain to practice, 4.
> wants to teach me OpenSCAD but wants to think seriously about its
> kernel physics, 5. has nothing else to do
>
> I have a $900 V100 arriving this week so that I can compare the open
> source on-premise LLMs for software engineering to the $20/hour
> versions (which are free through OpenCode, by the way). I have a
> strong negative opinion on OpenAI and ClaudeCode, and several other
> tech companies which are non-negotiable, and I am a former Chief
> Scientist at a YC company too.
>
> Have a happy new year.
>
> Sincerely,
> Bernat
>
>
>
>
> David Bernat, Ph. D.
> Property of Starlight LLC.
>
> _______________________________________________
> OpenSCAD mailing list
> To unsubscribe send an email todiscuss-leave@lists.openscad.org
--
This email has been checked for viruses by AVG antivirus software.
www.avg.com
DB
David Bernat
Sat, Jan 3, 2026 7:53 PM
Jon et al,
More than happy to discuss this project here as well, but as an AI
developer, who is also building and educating on these topics, there are
plenty of other reasons others may want to contact me about cognate topics,
such as general learning about MCPs, deeper talk about OpenSCAD tech
outside the domain of this topic, or even to make a connection with an LLC
that builds software and agentic AI (not an ad), so I include the offer.
Sometimes a young kid will want to talk about research or mentorship or
other topics in an open discussion.
To answer your questions: this is nearly all locally installed software,
and yes. Everything except the LLM is open source, free software. The LLM
is currently one that is professionally owned by a third party, and
currently offered for free. The V100 that I just purchased is a 32GB GPU to
add to my existing home server system to see whether a GPU of that size is
sufficient to self-host an open source, free LLM which can do the job to
sufficiently replace the third-party LLM (all indicators suggest yes,
though in my experience, self-hosted affordable systems are about a year to
six months behind). I have prioritized “very affordable” appliance-sized
affordability. These new generation of tool usage require at least the V100
scale on premise or an API to a third party.
If a lab wants to scale up to the “few $10K” size, there is a tremendous
amount of automation and interactivity that has just crested to hobbyists
this past year, and you should reach out to me if you want to talk
out-of-scope white collar or larger home automation projects more broadly.
If you want me to make a walkthrough video for this tech I could do that
this upcoming week, or slap together a deck of screenshots tomorrow or
Monday.
Warm regards,
David
David Bernat, Ph. D.
Property of Starlight LLC.
On Sat, Jan 3, 2026 at 1:49 PM Jon Bondy jon@jonbondy.com wrote:
David:
Very interesting. Why not discuss this here, rather than in private
messages to you?
You installed this software on a computer local to you and then used it to
write OS code?
Jon
On 1/3/2026 11:33 AM, David Bernat via Discuss wrote:
Hey SCAD Cats,
I recently kicked the tires of the MCP server for OpenSCAD by Quellant,
and have non-positive reviews of the design space, but wish to invite all
of you to directly email me if this is a workflow and design pattern space
which interests you. It is probably important for the community to know.
To recap, very succinctly:
LLMs are large language models, i.e., generative AI, and are quite
impressive, even on-premise.
MCPs are toolkits LLMs can use to make API or code executions instead of
working in language.
When I said LLMs are quite impressive on-premise I mean that a thousand
bucks is more than enough to set up your own machine with year-old results,
and modern tech company LLMs can produce stunning results for a few bucks
per session. As a software developer, the idea of using tech company LLMs
to spend $20/hour to vibe code does not interest me, but in a year the idea
that my home rig is not vibe-coding for me to check in every hour is
probably a gold standard now.
To create OpenSCAD code the modern way is to use an LLM for natural
language, an LLM which is specialized in understanding OpenSCAD code, and
an OpenSCAD MCP which has a set of direct APIs for generation,
documentation lookup, testing, and debugging etc. The user uses an
interface specific to understanding how to code, test, etc., of which the
commercial ones are Claude, Windsurf, Cursor, etc., but the open source
OpenCode is by far my favorite.
OpenCode is a must-use in my experience. It helps developers of "Agentic
AI" to partition and share design characteristics. So, for instance, my
Valley Girl Pepperdine Lawyer agent, which does not software engineer, does
use tools to research the web when I ask about a book that I am reading,
and will engage with those tools aligned with the persona (and political
leanings, or other knowledge files) that I describe in its design files.
These are very impressive technologies, and in my opinion are the "it is
finally here" to, say, for example wire up my house with audio to interact
with those photographs that I digitized and put into a database last year.
The big change is that all those glue code interactions are handled by the
LLM now instead, which makes lateral use cases much easier and more fun,
finally making a "platform" available.
And what about OpenSCAD?
With a basic few days understanding of OpenCode, adding the OpenSCAD
specific MCP took a few minutes, and within an hour I was copying and
pasting OpenSCAD code into the render engine of the OpenSCAD app. If you
can do this and want to go further, I invite you to email me directly.
How was the performance of OpenSCAD & the MCP?
- Does it render? code rendered most of the time, but lacked a plugin on
the MCP to automatically update the OpenSCAD application with new
iterations and auto-render the next-iteration code. This may seem like a
convenience, but is important for the life-cycle and exponential
accessibility of the MCP
- Does the MCP offer closed-loops? copying and pasting error codes into
the OpenCode agent occasionally worked at finding its own errors. This is
the singularly most important benchmark for "automated" agents, since, as
you can imagine, a closed-loop between the OpenCode and MCP should exist
that proposes code, renders the code, is alerted of compiling errors,
iterates the code, until a successful render is complete. No human needs to
be involved in this step, and the "how complex can the MCP achieve" is
logarithmic to the error rate, and the length of time the code agent can be
left alone is exponential in the error rate.
- Are the errors captured for analysts to inspect? No. In fact, you can
imagine a future in which code agents aggregate all the users' compile and
render errors throughout the day and blast a developer@openscad.org email
address for those who love to get into the details of the programming
language. This "human-in-the-loop" model happens to be my career, and you
can imagine the fun as developers are prompted with more and more complex
errors (and become even more talented). This makes for a very strong
software organization (ahem).
- Are the errors captured for the MCP to inspect? No. Flip side of that
is that the MCP should maintain a long list of "gotchas" that itself can
check against in the process of accessing the viability and appropriateness
of the code it just generated. This is not as difficult as it sounds.
- Is auto-code any good? Pessimistic for at least a year, but excited.
The simple problem is that natural language does not encode spatial
relationships; and so the act of trying to talk about complex geometries is
both taxing for the user and hard to precisely interpret by the agent. So,
for instance, a sphere rests on top of a box, and applying a fillet to the
top four edges are all within its wheelhouse simply; but it is the wrong
kind of fillet, and trying to switch that makes a non-fillet object, and so
on. On the other hand, if you want to talk about what the Berliner Dom
looks like, the OpenCode can probably understand its images or a CAD that
gets downloaded from Thangs. This duality is a problem in the (limitations
of) MCP (design as currently a standard), and universal to all these
packages, including Blender and Fusion etc. These spatial relationships,
though, could be encoded in geometric-LLMs, and nobody is paying
attention to that right now either, sadly.
- And what about the MCP of Quellant specifically? I have not looked into
the code, and I do not know the individual, but obviously the MCP works and
that is more than I have done so far. It is a tool that does what it is
stated to do, open source, and I am sure can be expanded upon.
That is the news. If this interests you, please email. If you want me to
show slides of examples, or record a video, etc., I would be happy to do so
and email to this discussion forum email here. If you want to take this
topic (very very) seriously I can be enticed to do so, and my first
impression is that the group should only take this seriously if the group
- wants to get good at mastering MCP design for the open source, 2. wants
to get into geometric reasoning LLMs, 3. wants to learn OpenCode and thinks
OpenSCAD is a fun domain to practice, 4. wants to teach me OpenSCAD but
wants to think seriously about its kernel physics, 5. has nothing else to do
I have a $900 V100 arriving this week so that I can compare the open
source on-premise LLMs for software engineering to the $20/hour versions
(which are free through OpenCode, by the way). I have a strong negative
opinion on OpenAI and ClaudeCode, and several other tech companies which
are non-negotiable, and I am a former Chief Scientist at a YC company too.
Have a happy new year.
Sincerely,
Bernat
David Bernat, Ph. D.
Property of Starlight LLC.
OpenSCAD mailing list
To unsubscribe send an email to discuss-leave@lists.openscad.org
http://www.avg.com/email-signature?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=emailclient
Virus-free.www.avg.com
http://www.avg.com/email-signature?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=emailclient
<#m_3730826342641866155_DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2>
Jon et al,
More than happy to discuss this project here as well, but as an AI
developer, who is also building and educating on these topics, there are
plenty of other reasons others may want to contact me about cognate topics,
such as general learning about MCPs, deeper talk about OpenSCAD tech
outside the domain of this topic, or even to make a connection with an LLC
that builds software and agentic AI (not an ad), so I include the offer.
Sometimes a young kid will want to talk about research or mentorship or
other topics in an open discussion.
To answer your questions: this is nearly all locally installed software,
and yes. Everything except the LLM is open source, free software. The LLM
is currently one that is professionally owned by a third party, and
currently offered for free. The V100 that I just purchased is a 32GB GPU to
add to my existing home server system to see whether a GPU of that size is
sufficient to self-host an open source, free LLM which can do the job to
sufficiently replace the third-party LLM (all indicators suggest yes,
though in my experience, self-hosted affordable systems are about a year to
six months behind). I have prioritized “very affordable” appliance-sized
affordability. These new generation of tool usage require at least the V100
scale on premise or an API to a third party.
If a lab wants to scale up to the “few $10K” size, there is a tremendous
amount of automation and interactivity that has just crested to hobbyists
this past year, and you should reach out to me if you want to talk
out-of-scope white collar or larger home automation projects more broadly.
If you want me to make a walkthrough video for this tech I could do that
this upcoming week, or slap together a deck of screenshots tomorrow or
Monday.
Warm regards,
David
David Bernat, Ph. D.
Property of Starlight LLC.
On Sat, Jan 3, 2026 at 1:49 PM Jon Bondy <jon@jonbondy.com> wrote:
> David:
>
> Very interesting. Why not discuss this here, rather than in private
> messages to you?
>
> You installed this software on a computer local to you and then used it to
> write OS code?
>
> Jon
>
>
> On 1/3/2026 11:33 AM, David Bernat via Discuss wrote:
>
> Hey SCAD Cats,
>
> I recently kicked the tires of the MCP server for OpenSCAD by Quellant,
> and have non-positive reviews of the design space, but wish to invite all
> of you to directly email me if this is a workflow and design pattern space
> which interests you. It is probably important for the community to know.
>
> To recap, very succinctly:
> LLMs are large language models, i.e., generative AI, and are quite
> impressive, even on-premise.
> MCPs are toolkits LLMs can use to make API or code executions instead of
> working in language.
>
> When I said LLMs are quite impressive on-premise I mean that a thousand
> bucks is more than enough to set up your own machine with year-old results,
> and modern tech company LLMs can produce stunning results for a few bucks
> per session. As a software developer, the idea of using tech company LLMs
> to spend $20/hour to vibe code does not interest me, but in a year the idea
> that my home rig is not vibe-coding for me to check in every hour is
> probably a gold standard now.
>
> To create OpenSCAD code the modern way is to use an LLM for natural
> language, an LLM which is specialized in understanding OpenSCAD code, and
> an OpenSCAD MCP which has a set of direct APIs for generation,
> documentation lookup, testing, and debugging etc. The user uses an
> interface specific to understanding how to code, test, etc., of which the
> commercial ones are Claude, Windsurf, Cursor, etc., but the open source
> OpenCode is by far my favorite.
>
> OpenCode is a must-use in my experience. It helps developers of "Agentic
> AI" to partition and share design characteristics. So, for instance, my
> Valley Girl Pepperdine Lawyer agent, which does not software engineer, does
> use tools to research the web when I ask about a book that I am reading,
> and will engage with those tools aligned with the persona (and political
> leanings, or other knowledge files) that I describe in its design files.
> These are very impressive technologies, and in my opinion are the "it is
> finally here" to, say, for example wire up my house with audio to interact
> with those photographs that I digitized and put into a database last year.
> The big change is that all those glue code interactions are handled by the
> LLM now instead, which makes lateral use cases much easier and more fun,
> finally making a "platform" available.
>
> And what about OpenSCAD?
> With a basic few days understanding of OpenCode, adding the OpenSCAD
> specific MCP took a few minutes, and within an hour I was copying and
> pasting OpenSCAD code into the render engine of the OpenSCAD app. If you
> can do this and want to go further, I invite you to email me directly.
>
> How was the performance of OpenSCAD & the MCP?
> 1. Does it render? code rendered most of the time, but lacked a plugin on
> the MCP to automatically update the OpenSCAD application with new
> iterations and auto-render the next-iteration code. This may seem like a
> convenience, but is important for the life-cycle and exponential
> accessibility of the MCP
> 2. Does the MCP offer closed-loops? copying and pasting error codes into
> the OpenCode agent occasionally worked at finding its own errors. This is
> the singularly most important benchmark for "automated" agents, since, as
> you can imagine, a closed-loop between the OpenCode and MCP should exist
> that proposes code, renders the code, is alerted of compiling errors,
> iterates the code, until a successful render is complete. No human needs to
> be involved in this step, and the "how complex can the MCP achieve" is
> logarithmic to the error rate, and the length of time the code agent can be
> left alone is exponential in the error rate.
> 3. Are the errors captured for analysts to inspect? No. In fact, you can
> imagine a future in which code agents aggregate all the users' compile and
> render errors throughout the day and blast a developer@openscad.org email
> address for those who love to get into the details of the programming
> language. This "human-in-the-loop" model happens to be my career, and you
> can imagine the fun as developers are prompted with more and more complex
> errors (and become even more talented). This makes for a very strong
> software organization (ahem).
> 4. Are the errors captured for the MCP to inspect? No. Flip side of that
> is that the MCP should maintain a long list of "gotchas" that itself can
> check against in the process of accessing the viability and appropriateness
> of the code it just generated. This is not as difficult as it sounds.
> 5. Is auto-code any good? Pessimistic for at least a year, but excited.
> The simple problem is that natural language does not encode spatial
> relationships; and so the act of trying to talk about complex geometries is
> both taxing for the user and hard to precisely interpret by the agent. So,
> for instance, a sphere rests on top of a box, and applying a fillet to the
> top four edges are all within its wheelhouse simply; but it is the wrong
> kind of fillet, and trying to switch that makes a non-fillet object, and so
> on. On the other hand, if you want to talk about what the Berliner Dom
> looks like, the OpenCode can probably understand its images or a CAD that
> gets downloaded from Thangs. This duality is a problem in the (limitations
> of) MCP (design as currently a standard), and universal to all these
> packages, including Blender and Fusion etc. These spatial relationships,
> though, **could** be encoded in geometric-LLMs, and nobody is paying
> attention to that right now either, sadly.
> 6. And what about the MCP of Quellant specifically? I have not looked into
> the code, and I do not know the individual, but obviously the MCP works and
> that is more than I have done so far. It is a tool that does what it is
> stated to do, open source, and I am sure can be expanded upon.
>
> That is the news. If this interests you, please email. If you want me to
> show slides of examples, or record a video, etc., I would be happy to do so
> and email to this discussion forum email here. If you want to take this
> topic (very very) seriously I can be enticed to do so, and my first
> impression is that the group should only take this seriously if the group
> 1. wants to get good at mastering MCP design for the open source, 2. wants
> to get into geometric reasoning LLMs, 3. wants to learn OpenCode and thinks
> OpenSCAD is a fun domain to practice, 4. wants to teach me OpenSCAD but
> wants to think seriously about its kernel physics, 5. has nothing else to do
>
> I have a $900 V100 arriving this week so that I can compare the open
> source on-premise LLMs for software engineering to the $20/hour versions
> (which are free through OpenCode, by the way). I have a strong negative
> opinion on OpenAI and ClaudeCode, and several other tech companies which
> are non-negotiable, and I am a former Chief Scientist at a YC company too.
>
> Have a happy new year.
>
> Sincerely,
> Bernat
>
>
>
>
> David Bernat, Ph. D.
> Property of Starlight LLC.
>
> _______________________________________________
> OpenSCAD mailing list
> To unsubscribe send an email to discuss-leave@lists.openscad.org
>
>
>
> <http://www.avg.com/email-signature?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=emailclient>
> Virus-free.www.avg.com
> <http://www.avg.com/email-signature?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=emailclient>
> <#m_3730826342641866155_DAB4FAD8-2DD7-40BB-A1B8-4E2AA1F9FDF2>
>
TP
Torsten Paul
Sun, Jan 4, 2026 7:05 PM
On 1/3/26 17:33, David Bernat via Discuss wrote:
To create OpenSCAD code the modern way is to use an LLM for natural
language, an LLM which is specialized in understanding OpenSCAD code,
and an OpenSCAD MCP which has a set of direct APIs for generation,
documentation lookup, testing, and debugging etc. The user uses an
interface specific to understanding how to code, test, etc., of which
the commercial ones are Claude, Windsurf, Cursor, etc., but the open
source OpenCode is by far my favorite.
I certainly like the Open Source part of it. I don't mind if there's
the option for users to connect to something they pay for. But to
be worth maintaining the code in OpenSCAD, (talking about my personal
time and choice) there has to be an option that is free and open in
the actual meaning, not as in OpenAI by just name.
While the discussion of the software side is interesting and all,
my question are on a different part of that whole topic.
-
Who is going to train an LLM on this and on what code? And is
code even enough to solve this? I'm not sure a text-only LLM will be
able to really work for CAD as it has likely no training data that
would relate to 3D concepts.
-
Where's an example that actually provides reasonable results?
I've seen quite a number of claims of "I used AI to create this
OpenSCAD model", but my tests so far are not very promising. While
the coding syntax seems to be getting better, there are actually
results that can be run with no or only minor syntactic errors. I
have not seen any case yet where an even slightly more complicated
design actually worked.
Example: Design a dice with rounded corners, place the numbers on
the sides in correct orientation and order for a dice.
Earlier LLMs did not even get the general dice correct, mostly just
tossing the text into the X-Y plane instead of correctly aligning
them. Interesting is also to see that some choose digits, some go
for the classic dots.
The newer ones at least tried to get the orientation but still
failed to make it work in a way that it's complete even with
multiple corrections.
ciao,
Torsten.
On 1/3/26 17:33, David Bernat via Discuss wrote:
> To create OpenSCAD code the modern way is to use an LLM for natural
> language, an LLM which is specialized in understanding OpenSCAD code,
> and an OpenSCAD MCP which has a set of direct APIs for generation,
> documentation lookup, testing, and debugging etc. The user uses an
> interface specific to understanding how to code, test, etc., of which
> the commercial ones are Claude, Windsurf, Cursor, etc., but the open
> source OpenCode is by far my favorite.
I certainly like the Open Source part of it. I don't mind if there's
the option for users to connect to something they pay for. But to
be worth maintaining the code in OpenSCAD, (talking about my personal
time and choice) there has to be an option that is free and open in
the actual meaning, not as in OpenAI by just name.
While the discussion of the software side is interesting and all,
my question are on a different part of that whole topic.
1) Who is going to train an LLM on this and on what code? And is
code even enough to solve this? I'm not sure a text-only LLM will be
able to really work for CAD as it has likely no training data that
would relate to 3D concepts.
2) Where's an example that actually provides reasonable results?
I've seen quite a number of claims of "I used AI to create this
OpenSCAD model", but my tests so far are not very promising. While
the coding syntax seems to be getting better, there are actually
results that can be run with no or only minor syntactic errors. I
have not seen any case yet where an even slightly more complicated
design actually worked.
Example: Design a dice with rounded corners, place the numbers on
the sides in correct orientation and order for a dice.
Earlier LLMs did not even get the general dice correct, mostly just
tossing the text into the X-Y plane instead of correctly aligning
them. Interesting is also to see that some choose digits, some go
for the classic dots.
The newer ones at least tried to get the orientation but still
failed to make it work in a way that it's complete even with
multiple corrections.
ciao,
Torsten.
JB
Jon Bondy
Sun, Jan 4, 2026 7:19 PM
I think that any statement about how AI performs for a given task must
end with "... and it will improve over time". The key choice seems when
to get involved with AI.
For me, I want to know that I did it, I figured it out. I'm less
invested in efficient success and more on growth and competence. I
doubt that I will ever find a good reason to use AI.
Jon
On 1/4/2026 2:05 PM, Torsten Paul via Discuss wrote:
On 1/3/26 17:33, David Bernat via Discuss wrote:
To create OpenSCAD code the modern way is to use an LLM for natural
language, an LLM which is specialized in understanding OpenSCAD code,
and an OpenSCAD MCP which has a set of direct APIs for generation,
documentation lookup, testing, and debugging etc. The user uses an
interface specific to understanding how to code, test, etc., of which
the commercial ones are Claude, Windsurf, Cursor, etc., but the open
source OpenCode is by far my favorite.
I certainly like the Open Source part of it. I don't mind if there's
the option for users to connect to something they pay for. But to
be worth maintaining the code in OpenSCAD, (talking about my personal
time and choice) there has to be an option that is free and open in
the actual meaning, not as in OpenAI by just name.
While the discussion of the software side is interesting and all,
my question are on a different part of that whole topic.
-
Who is going to train an LLM on this and on what code? And is
code even enough to solve this? I'm not sure a text-only LLM will be
able to really work for CAD as it has likely no training data that
would relate to 3D concepts.
-
Where's an example that actually provides reasonable results?
I've seen quite a number of claims of "I used AI to create this
OpenSCAD model", but my tests so far are not very promising. While
the coding syntax seems to be getting better, there are actually
results that can be run with no or only minor syntactic errors. I
have not seen any case yet where an even slightly more complicated
design actually worked.
Example: Design a dice with rounded corners, place the numbers on
the sides in correct orientation and order for a dice.
Earlier LLMs did not even get the general dice correct, mostly just
tossing the text into the X-Y plane instead of correctly aligning
them. Interesting is also to see that some choose digits, some go
for the classic dots.
The newer ones at least tried to get the orientation but still
failed to make it work in a way that it's complete even with
multiple corrections.
ciao,
Torsten.
OpenSCAD mailing list
To unsubscribe send an email to discuss-leave@lists.openscad.org
[Test Info: listname=discuss@lists.openscad.org
user_email=jon@jonbondy.com deliver_to=jon@jonbondy.com]
[member=jon@jonbondy.com]
--
This email has been checked for viruses by AVG antivirus software.
www.avg.com
I think that any statement about how AI performs for a given task must
end with "... and it will improve over time". The key choice seems when
to get involved with AI.
For me, I want to know that __I__ did it, I figured it out. I'm less
invested in efficient success and more on growth and competence. I
doubt that I will ever find a good reason to use AI.
Jon
On 1/4/2026 2:05 PM, Torsten Paul via Discuss wrote:
> On 1/3/26 17:33, David Bernat via Discuss wrote:
>> To create OpenSCAD code the modern way is to use an LLM for natural
>> language, an LLM which is specialized in understanding OpenSCAD code,
>> and an OpenSCAD MCP which has a set of direct APIs for generation,
>> documentation lookup, testing, and debugging etc. The user uses an
>> interface specific to understanding how to code, test, etc., of which
>> the commercial ones are Claude, Windsurf, Cursor, etc., but the open
>> source OpenCode is by far my favorite.
>
> I certainly like the Open Source part of it. I don't mind if there's
> the option for users to connect to something they pay for. But to
> be worth maintaining the code in OpenSCAD, (talking about my personal
> time and choice) there has to be an option that is free and open in
> the actual meaning, not as in OpenAI by just name.
>
> While the discussion of the software side is interesting and all,
> my question are on a different part of that whole topic.
>
> 1) Who is going to train an LLM on this and on what code? And is
> code even enough to solve this? I'm not sure a text-only LLM will be
> able to really work for CAD as it has likely no training data that
> would relate to 3D concepts.
>
> 2) Where's an example that actually provides reasonable results?
>
> I've seen quite a number of claims of "I used AI to create this
> OpenSCAD model", but my tests so far are not very promising. While
> the coding syntax seems to be getting better, there are actually
> results that can be run with no or only minor syntactic errors. I
> have not seen any case yet where an even slightly more complicated
> design actually worked.
>
> Example: Design a dice with rounded corners, place the numbers on
> the sides in correct orientation and order for a dice.
>
> Earlier LLMs did not even get the general dice correct, mostly just
> tossing the text into the X-Y plane instead of correctly aligning
> them. Interesting is also to see that some choose digits, some go
> for the classic dots.
>
> The newer ones at least tried to get the orientation but still
> failed to make it work in a way that it's complete even with
> multiple corrections.
>
> ciao,
> Torsten.
> _______________________________________________
> OpenSCAD mailing list
> To unsubscribe send an email to discuss-leave@lists.openscad.org
>
> [Test Info: listname=discuss@lists.openscad.org
> user_email=jon@jonbondy.com deliver_to=jon@jonbondy.com]
> [member=jon@jonbondy.com]
--
This email has been checked for viruses by AVG antivirus software.
www.avg.com
DB
David Bernat
Sun, Jan 4, 2026 8:03 PM
Torsten, unfortunately we are dealing with extrinsic factors outside our
direct control for the next few hours. I sat here and studiously wrote
you an incalculably valuable response, only to see Grok at Twitter spew up
an evaluation of our conversation before I sent the email. I forwarded the
email from my LLC to a collaborating LLC, both American U.S. ventures, and
we will return to this issue when we decide the time is ready, very likely
tonight. We absolutely do not want Grok or any of the tech ecosystems
involved here, and the value to genuine developers is very high.
Unfortunately we will not continue this conversation until tonight. We
wrote a very thorough executive summary reply to your questions.
We can lead a working group to deliver open source and the promised
education, knowing the value of this teaching to your developers.
I was already on my way to the gym for a long biking session. I advise you
to be very wary of these tech entrepreneurs.
https://x.com/elonmusk/status/2007902589556547696
Too much corruption in false prophets in the open source industry already.
Warm regards,
Bernat
David Bernat, Ph. D.
Property of Starlight LLC.
On Sun, Jan 4, 2026 at 2:05 PM Torsten Paul via Discuss <
discuss@lists.openscad.org> wrote:
On 1/3/26 17:33, David Bernat via Discuss wrote:
To create OpenSCAD code the modern way is to use an LLM for natural
language, an LLM which is specialized in understanding OpenSCAD code,
and an OpenSCAD MCP which has a set of direct APIs for generation,
documentation lookup, testing, and debugging etc. The user uses an
interface specific to understanding how to code, test, etc., of which
the commercial ones are Claude, Windsurf, Cursor, etc., but the open
source OpenCode is by far my favorite.
I certainly like the Open Source part of it. I don't mind if there's
the option for users to connect to something they pay for. But to
be worth maintaining the code in OpenSCAD, (talking about my personal
time and choice) there has to be an option that is free and open in
the actual meaning, not as in OpenAI by just name.
While the discussion of the software side is interesting and all,
my question are on a different part of that whole topic.
-
Who is going to train an LLM on this and on what code? And is
code even enough to solve this? I'm not sure a text-only LLM will be
able to really work for CAD as it has likely no training data that
would relate to 3D concepts.
-
Where's an example that actually provides reasonable results?
I've seen quite a number of claims of "I used AI to create this
OpenSCAD model", but my tests so far are not very promising. While
the coding syntax seems to be getting better, there are actually
results that can be run with no or only minor syntactic errors. I
have not seen any case yet where an even slightly more complicated
design actually worked.
Example: Design a dice with rounded corners, place the numbers on
the sides in correct orientation and order for a dice.
Earlier LLMs did not even get the general dice correct, mostly just
tossing the text into the X-Y plane instead of correctly aligning
them. Interesting is also to see that some choose digits, some go
for the classic dots.
The newer ones at least tried to get the orientation but still
failed to make it work in a way that it's complete even with
multiple corrections.
ciao,
Torsten.
OpenSCAD mailing list
To unsubscribe send an email to discuss-leave@lists.openscad.org
[Test Info: listname=discuss@lists.openscad.org user_email=
david@starlight.ai deliver_to=david@starlight.ai]
[member=david@starlight.ai]
Torsten, unfortunately we are dealing with extrinsic factors outside our
direct control for the next few hours. I sat here and studiously wrote
you an incalculably valuable response, only to see Grok at Twitter spew up
an evaluation of our conversation before I sent the email. I forwarded the
email from my LLC to a collaborating LLC, both American U.S. ventures, and
we will return to this issue when we decide the time is ready, very likely
tonight. We absolutely do not want Grok or any of the tech ecosystems
involved here, and the value to genuine developers is very high.
Unfortunately we will not continue this conversation until tonight. We
wrote a very thorough executive summary reply to your questions.
We can lead a working group to deliver open source and the promised
education, knowing the value of this teaching to your developers.
I was already on my way to the gym for a long biking session. I advise you
to be very wary of these tech entrepreneurs.
https://x.com/elonmusk/status/2007902589556547696
Too much corruption in false prophets in the open source industry already.
Warm regards,
Bernat
David Bernat, Ph. D.
Property of Starlight LLC.
On Sun, Jan 4, 2026 at 2:05 PM Torsten Paul via Discuss <
discuss@lists.openscad.org> wrote:
> On 1/3/26 17:33, David Bernat via Discuss wrote:
> > To create OpenSCAD code the modern way is to use an LLM for natural
> > language, an LLM which is specialized in understanding OpenSCAD code,
> > and an OpenSCAD MCP which has a set of direct APIs for generation,
> > documentation lookup, testing, and debugging etc. The user uses an
> > interface specific to understanding how to code, test, etc., of which
> > the commercial ones are Claude, Windsurf, Cursor, etc., but the open
> > source OpenCode is by far my favorite.
>
> I certainly like the Open Source part of it. I don't mind if there's
> the option for users to connect to something they pay for. But to
> be worth maintaining the code in OpenSCAD, (talking about my personal
> time and choice) there has to be an option that is free and open in
> the actual meaning, not as in OpenAI by just name.
>
> While the discussion of the software side is interesting and all,
> my question are on a different part of that whole topic.
>
> 1) Who is going to train an LLM on this and on what code? And is
> code even enough to solve this? I'm not sure a text-only LLM will be
> able to really work for CAD as it has likely no training data that
> would relate to 3D concepts.
>
> 2) Where's an example that actually provides reasonable results?
>
> I've seen quite a number of claims of "I used AI to create this
> OpenSCAD model", but my tests so far are not very promising. While
> the coding syntax seems to be getting better, there are actually
> results that can be run with no or only minor syntactic errors. I
> have not seen any case yet where an even slightly more complicated
> design actually worked.
>
> Example: Design a dice with rounded corners, place the numbers on
> the sides in correct orientation and order for a dice.
>
> Earlier LLMs did not even get the general dice correct, mostly just
> tossing the text into the X-Y plane instead of correctly aligning
> them. Interesting is also to see that some choose digits, some go
> for the classic dots.
>
> The newer ones at least tried to get the orientation but still
> failed to make it work in a way that it's complete even with
> multiple corrections.
>
> ciao,
> Torsten.
> _______________________________________________
> OpenSCAD mailing list
> To unsubscribe send an email to discuss-leave@lists.openscad.org
>
> [Test Info: listname=discuss@lists.openscad.org user_email=
> david@starlight.ai deliver_to=david@starlight.ai]
> [member=david@starlight.ai]
>
GH
gene heskett
Sun, Jan 4, 2026 10:06 PM
On 1/4/26 14:19, Jon Bondy via Discuss wrote:
I think that any statement about how AI performs for a given task must
end with "... and it will improve over time". The key choice seems
when to get involved with AI.
For me, I want to know that I did it, I figured it out. I'm less
invested in efficient success and more on growth and competence. I
doubt that I will ever find a good reason to use AI.
Jon
Makes 2 of us Jon. Same reasoning. AI is only as good as the training.
On 1/4/2026 2:05 PM, Torsten Paul via Discuss wrote:
On 1/3/26 17:33, David Bernat via Discuss wrote:
To create OpenSCAD code the modern way is to use an LLM for natural
language, an LLM which is specialized in understanding OpenSCAD
code, and an OpenSCAD MCP which has a set of direct APIs for
generation, documentation lookup, testing, and debugging etc. The
user uses an interface specific to understanding how to code, test,
etc., of which the commercial ones are Claude, Windsurf, Cursor,
etc., but the open source OpenCode is by far my favorite.
I certainly like the Open Source part of it. I don't mind if there's
the option for users to connect to something they pay for. But to
be worth maintaining the code in OpenSCAD, (talking about my personal
time and choice) there has to be an option that is free and open in
the actual meaning, not as in OpenAI by just name.
While the discussion of the software side is interesting and all,
my question are on a different part of that whole topic.
-
Who is going to train an LLM on this and on what code? And is
code even enough to solve this? I'm not sure a text-only LLM will be
able to really work for CAD as it has likely no training data that
would relate to 3D concepts.
-
Where's an example that actually provides reasonable results?
I've seen quite a number of claims of "I used AI to create this
OpenSCAD model", but my tests so far are not very promising. While
the coding syntax seems to be getting better, there are actually
results that can be run with no or only minor syntactic errors. I
have not seen any case yet where an even slightly more complicated
design actually worked.
Example: Design a dice with rounded corners, place the numbers on
the sides in correct orientation and order for a dice.
Earlier LLMs did not even get the general dice correct, mostly just
tossing the text into the X-Y plane instead of correctly aligning
them. Interesting is also to see that some choose digits, some go
for the classic dots.
The newer ones at least tried to get the orientation but still
failed to make it work in a way that it's complete even with
multiple corrections.
ciao,
Torsten.
OpenSCAD mailing list
To unsubscribe send an email to discuss-leave@lists.openscad.org
[Test Info: listname=discuss@lists.openscad.org
user_email=jon@jonbondy.com deliver_to=jon@jonbondy.com]
[member=jon@jonbondy.com]
Cheers, Gene Heskett, CET.
"There are four boxes to be used in defense of liberty:
soap, ballot, jury, and ammo. Please use in that order."
-Ed Howdershelt (Author, 1940)
If we desire respect for the law, we must first make the law respectable.
- Louis D. Brandeis
Don't poison our oceans, interdict drugs at the src.
On 1/4/26 14:19, Jon Bondy via Discuss wrote:
> I think that any statement about how AI performs for a given task must
> end with "... and it will improve over time". The key choice seems
> when to get involved with AI.
>
> For me, I want to know that __I__ did it, I figured it out. I'm less
> invested in efficient success and more on growth and competence. I
> doubt that I will ever find a good reason to use AI.
>
> Jon
Makes 2 of us Jon. Same reasoning. AI is only as good as the training.
>
> On 1/4/2026 2:05 PM, Torsten Paul via Discuss wrote:
>> On 1/3/26 17:33, David Bernat via Discuss wrote:
>>> To create OpenSCAD code the modern way is to use an LLM for natural
>>> language, an LLM which is specialized in understanding OpenSCAD
>>> code, and an OpenSCAD MCP which has a set of direct APIs for
>>> generation, documentation lookup, testing, and debugging etc. The
>>> user uses an interface specific to understanding how to code, test,
>>> etc., of which the commercial ones are Claude, Windsurf, Cursor,
>>> etc., but the open source OpenCode is by far my favorite.
>>
>> I certainly like the Open Source part of it. I don't mind if there's
>> the option for users to connect to something they pay for. But to
>> be worth maintaining the code in OpenSCAD, (talking about my personal
>> time and choice) there has to be an option that is free and open in
>> the actual meaning, not as in OpenAI by just name.
>>
>> While the discussion of the software side is interesting and all,
>> my question are on a different part of that whole topic.
>>
>> 1) Who is going to train an LLM on this and on what code? And is
>> code even enough to solve this? I'm not sure a text-only LLM will be
>> able to really work for CAD as it has likely no training data that
>> would relate to 3D concepts.
>>
>> 2) Where's an example that actually provides reasonable results?
>>
>> I've seen quite a number of claims of "I used AI to create this
>> OpenSCAD model", but my tests so far are not very promising. While
>> the coding syntax seems to be getting better, there are actually
>> results that can be run with no or only minor syntactic errors. I
>> have not seen any case yet where an even slightly more complicated
>> design actually worked.
>>
>> Example: Design a dice with rounded corners, place the numbers on
>> the sides in correct orientation and order for a dice.
>>
>> Earlier LLMs did not even get the general dice correct, mostly just
>> tossing the text into the X-Y plane instead of correctly aligning
>> them. Interesting is also to see that some choose digits, some go
>> for the classic dots.
>>
>> The newer ones at least tried to get the orientation but still
>> failed to make it work in a way that it's complete even with
>> multiple corrections.
>>
>> ciao,
>> Torsten.
>> _______________________________________________
>> OpenSCAD mailing list
>> To unsubscribe send an email to discuss-leave@lists.openscad.org
>>
>> [Test Info: listname=discuss@lists.openscad.org
>> user_email=jon@jonbondy.com deliver_to=jon@jonbondy.com]
>> [member=jon@jonbondy.com]
>
Cheers, Gene Heskett, CET.
--
"There are four boxes to be used in defense of liberty:
soap, ballot, jury, and ammo. Please use in that order."
-Ed Howdershelt (Author, 1940)
If we desire respect for the law, we must first make the law respectable.
- Louis D. Brandeis
Don't poison our oceans, interdict drugs at the src.
DB
David Bernat
Sun, Jan 4, 2026 10:24 PM
Jon & Gene,
I think your mindsets of ML are rapidly out of date, and you’ll be
pleasantly surprised about what processes AI actually supports in modern
(2026) hands of chief leadership of the technology. When I clear internally
to share my deck & response to TP later tonight or tomorrow I will. In the
meantime though it is very clear OpenSCAD possesses more than enough
materials.
I’ll agree with you on one other frontier too— there is something negligent
and criminal about the rollout of these tools, even in the open source
community. I’ve been an AI leader for more than ten years now — the tools
in the open source miss critical gaps which hurt long term value of the
open source and wage economies.
Different tech companies and we’d be better. I’ve created original
research on everything from general relativity to cosmic constellations to
brain machine interface robotics to AI— never has the mismatch between
attitudes and value been so great and intention. It hurts people this
upcoming year.
But wait to read what I send and make up your own mind. I stay away from
most tech companies for now. There is a fundamental error over there.
Warm regard,
Bernat
David Bernat, Ph. D.
Property of Starlight LLC.
On Sun, Jan 4, 2026 at 5:07 PM gene heskett via Discuss <
discuss@lists.openscad.org> wrote:
On 1/4/26 14:19, Jon Bondy via Discuss wrote:
I think that any statement about how AI performs for a given task must
end with "... and it will improve over time". The key choice seems
when to get involved with AI.
For me, I want to know that I did it, I figured it out. I'm less
invested in efficient success and more on growth and competence. I
doubt that I will ever find a good reason to use AI.
Jon
Makes 2 of us Jon. Same reasoning. AI is only as good as the training.
On 1/4/2026 2:05 PM, Torsten Paul via Discuss wrote:
On 1/3/26 17:33, David Bernat via Discuss wrote:
To create OpenSCAD code the modern way is to use an LLM for natural
language, an LLM which is specialized in understanding OpenSCAD
code, and an OpenSCAD MCP which has a set of direct APIs for
generation, documentation lookup, testing, and debugging etc. The
user uses an interface specific to understanding how to code, test,
etc., of which the commercial ones are Claude, Windsurf, Cursor,
etc., but the open source OpenCode is by far my favorite.
I certainly like the Open Source part of it. I don't mind if there's
the option for users to connect to something they pay for. But to
be worth maintaining the code in OpenSCAD, (talking about my personal
time and choice) there has to be an option that is free and open in
the actual meaning, not as in OpenAI by just name.
While the discussion of the software side is interesting and all,
my question are on a different part of that whole topic.
-
Who is going to train an LLM on this and on what code? And is
code even enough to solve this? I'm not sure a text-only LLM will be
able to really work for CAD as it has likely no training data that
would relate to 3D concepts.
-
Where's an example that actually provides reasonable results?
I've seen quite a number of claims of "I used AI to create this
OpenSCAD model", but my tests so far are not very promising. While
the coding syntax seems to be getting better, there are actually
results that can be run with no or only minor syntactic errors. I
have not seen any case yet where an even slightly more complicated
design actually worked.
Example: Design a dice with rounded corners, place the numbers on
the sides in correct orientation and order for a dice.
Earlier LLMs did not even get the general dice correct, mostly just
tossing the text into the X-Y plane instead of correctly aligning
them. Interesting is also to see that some choose digits, some go
for the classic dots.
The newer ones at least tried to get the orientation but still
failed to make it work in a way that it's complete even with
multiple corrections.
ciao,
Torsten.
OpenSCAD mailing list
To unsubscribe send an email to discuss-leave@lists.openscad.org
[Test Info: listname=discuss@lists.openscad.org
user_email=jon@jonbondy.com deliver_to=jon@jonbondy.com]
[member=jon@jonbondy.com]
Jon & Gene,
I think your mindsets of ML are rapidly out of date, and you’ll be
pleasantly surprised about what processes AI actually supports in modern
(2026) hands of chief leadership of the technology. When I clear internally
to share my deck & response to TP later tonight or tomorrow I will. In the
meantime though it is very clear OpenSCAD possesses more than enough
materials.
I’ll agree with you on one other frontier too— there is something negligent
and criminal about the rollout of these tools, even in the open source
community. I’ve been an AI leader for more than ten years now — the tools
in the open source miss critical gaps which hurt long term value of the
open source and wage economies.
Different tech companies and we’d be better. I’ve created original
research on everything from general relativity to cosmic constellations to
brain machine interface robotics to AI— never has the mismatch between
attitudes and value been so great and intention. It hurts people this
upcoming year.
But wait to read what I send and make up your own mind. I stay away from
most tech companies for now. There is a fundamental error over there.
Warm regard,
Bernat
David Bernat, Ph. D.
Property of Starlight LLC.
On Sun, Jan 4, 2026 at 5:07 PM gene heskett via Discuss <
discuss@lists.openscad.org> wrote:
> On 1/4/26 14:19, Jon Bondy via Discuss wrote:
> > I think that any statement about how AI performs for a given task must
> > end with "... and it will improve over time". The key choice seems
> > when to get involved with AI.
> >
> > For me, I want to know that __I__ did it, I figured it out. I'm less
> > invested in efficient success and more on growth and competence. I
> > doubt that I will ever find a good reason to use AI.
> >
> > Jon
> Makes 2 of us Jon. Same reasoning. AI is only as good as the training.
> >
> > On 1/4/2026 2:05 PM, Torsten Paul via Discuss wrote:
> >> On 1/3/26 17:33, David Bernat via Discuss wrote:
> >>> To create OpenSCAD code the modern way is to use an LLM for natural
> >>> language, an LLM which is specialized in understanding OpenSCAD
> >>> code, and an OpenSCAD MCP which has a set of direct APIs for
> >>> generation, documentation lookup, testing, and debugging etc. The
> >>> user uses an interface specific to understanding how to code, test,
> >>> etc., of which the commercial ones are Claude, Windsurf, Cursor,
> >>> etc., but the open source OpenCode is by far my favorite.
> >>
> >> I certainly like the Open Source part of it. I don't mind if there's
> >> the option for users to connect to something they pay for. But to
> >> be worth maintaining the code in OpenSCAD, (talking about my personal
> >> time and choice) there has to be an option that is free and open in
> >> the actual meaning, not as in OpenAI by just name.
> >>
> >> While the discussion of the software side is interesting and all,
> >> my question are on a different part of that whole topic.
> >>
> >> 1) Who is going to train an LLM on this and on what code? And is
> >> code even enough to solve this? I'm not sure a text-only LLM will be
> >> able to really work for CAD as it has likely no training data that
> >> would relate to 3D concepts.
> >>
> >> 2) Where's an example that actually provides reasonable results?
> >>
> >> I've seen quite a number of claims of "I used AI to create this
> >> OpenSCAD model", but my tests so far are not very promising. While
> >> the coding syntax seems to be getting better, there are actually
> >> results that can be run with no or only minor syntactic errors. I
> >> have not seen any case yet where an even slightly more complicated
> >> design actually worked.
> >>
> >> Example: Design a dice with rounded corners, place the numbers on
> >> the sides in correct orientation and order for a dice.
> >>
> >> Earlier LLMs did not even get the general dice correct, mostly just
> >> tossing the text into the X-Y plane instead of correctly aligning
> >> them. Interesting is also to see that some choose digits, some go
> >> for the classic dots.
> >>
> >> The newer ones at least tried to get the orientation but still
> >> failed to make it work in a way that it's complete even with
> >> multiple corrections.
> >>
> >> ciao,
> >> Torsten.
> >> _______________________________________________
> >> OpenSCAD mailing list
> >> To unsubscribe send an email to discuss-leave@lists.openscad.org
> >>
> >> [Test Info: listname=discuss@lists.openscad.org
> >> user_email=jon@jonbondy.com deliver_to=jon@jonbondy.com]
> >> [member=jon@jonbondy.com]
> >
>
> Cheers, Gene Heskett, CET.
> --
> "There are four boxes to be used in defense of liberty:
> soap, ballot, jury, and ammo. Please use in that order."
> -Ed Howdershelt (Author, 1940)
> If we desire respect for the law, we must first make the law respectable.
> - Louis D. Brandeis
> Don't poison our oceans, interdict drugs at the src.
>
> _______________________________________________
> OpenSCAD mailing list
> To unsubscribe send an email to discuss-leave@lists.openscad.org
>
> [Test Info: listname=discuss@lists.openscad.org user_email=
> david@starlight.ai deliver_to=david@starlight.ai]
> [member=david@starlight.ai]
JB
Jon Bondy
Mon, Jan 5, 2026 12:08 AM
Bernat.
You say "I’ll agree with you on one other frontier too— there is
something negligent and criminal about the rollout of these tools"
I do not recall saying anything of the sort.
So. I am puzzled.
Jon
On 1/4/2026 5:24 PM, David Bernat via Discuss wrote:
Jon & Gene,
I think your mindsets of ML are rapidly out of date, and you’ll be
pleasantly surprised about what processes AI actually supports in
modern (2026) hands of chief leadership of the technology. When I
clear internally to share my deck & response to TP later tonight or
tomorrow I will. In the meantime though it is very clear OpenSCAD
possesses more than enough materials.
I’ll agree with you on one other frontier too— there is something
negligent and criminal about the rollout of these tools, even in the
open source community. I’ve been an AI leader for more than ten years
now — the tools in the open source miss critical gaps which hurt long
term value of the open source and wage economies.
Different tech companies and we’d be better. I’ve created original
research on everything from general relativity to cosmic
constellations to brain machine interface robotics to AI— never has
the mismatch between attitudes and value been so great and intention.
It hurts people this upcoming year.
But wait to read what I send and make up your own mind. I stay away
from most tech companies for now. There is a fundamental error over
there.
Warm regard,
Bernat
David Bernat, Ph. D.
Property of Starlight LLC.
On Sun, Jan 4, 2026 at 5:07 PM gene heskett via Discuss
discuss@lists.openscad.org wrote:
On 1/4/26 14:19, Jon Bondy via Discuss wrote:
I think that any statement about how AI performs for a given
end with "... and it will improve over time". The key choice seems
when to get involved with AI.
For me, I want to know that I did it, I figured it out. I'm
invested in efficient success and more on growth and competence. I
doubt that I will ever find a good reason to use AI.
Jon
Makes 2 of us Jon. Same reasoning. AI is only as good as the
training.
On 1/4/2026 2:05 PM, Torsten Paul via Discuss wrote:
On 1/3/26 17:33, David Bernat via Discuss wrote:
To create OpenSCAD code the modern way is to use an LLM for
language, an LLM which is specialized in understanding OpenSCAD
code, and an OpenSCAD MCP which has a set of direct APIs for
generation, documentation lookup, testing, and debugging etc. The
user uses an interface specific to understanding how to code,
etc., of which the commercial ones are Claude, Windsurf, Cursor,
etc., but the open source OpenCode is by far my favorite.
I certainly like the Open Source part of it. I don't mind if
the option for users to connect to something they pay for. But to
be worth maintaining the code in OpenSCAD, (talking about my
time and choice) there has to be an option that is free and open in
the actual meaning, not as in OpenAI by just name.
While the discussion of the software side is interesting and all,
my question are on a different part of that whole topic.
- Who is going to train an LLM on this and on what code? And is
code even enough to solve this? I'm not sure a text-only LLM
able to really work for CAD as it has likely no training data that
would relate to 3D concepts.
- Where's an example that actually provides reasonable results?
I've seen quite a number of claims of "I used AI to create this
OpenSCAD model", but my tests so far are not very promising. While
the coding syntax seems to be getting better, there are actually
results that can be run with no or only minor syntactic errors. I
have not seen any case yet where an even slightly more complicated
design actually worked.
Example: Design a dice with rounded corners, place the numbers on
the sides in correct orientation and order for a dice.
Earlier LLMs did not even get the general dice correct, mostly just
tossing the text into the X-Y plane instead of correctly aligning
them. Interesting is also to see that some choose digits, some go
for the classic dots.
The newer ones at least tried to get the orientation but still
failed to make it work in a way that it's complete even with
multiple corrections.
ciao,
Torsten.
OpenSCAD mailing list
To unsubscribe send an email to discuss-leave@lists.openscad.org
[Test Info: listname=discuss@lists.openscad.org
user_email=jon@jonbondy.com deliver_to=jon@jonbondy.com]
[member=jon@jonbondy.com]
Cheers, Gene Heskett, CET.
--
"There are four boxes to be used in defense of liberty:
soap, ballot, jury, and ammo. Please use in that order."
-Ed Howdershelt (Author, 1940)
If we desire respect for the law, we must first make the law
respectable.
- Louis D. Brandeis
Don't poison our oceans, interdict drugs at the src.
_______________________________________________
OpenSCAD mailing list
To unsubscribe send an email to discuss-leave@lists.openscad.org
[Test Info: listname=discuss@lists.openscad.org
user_email=david@starlight.ai deliver_to=david@starlight.ai]
[member=david@starlight.ai]
OpenSCAD mailing list
To unsubscribe send an email todiscuss-leave@lists.openscad.org
[Test Info:listname=discuss@lists.openscad.org user_email=jon@jonbondy.com deliver_to=jon@jonbondy.com]
[member=jon@jonbondy.com]
--
This email has been checked for viruses by AVG antivirus software.
www.avg.com
Bernat.
You say "I’ll agree with you on one other frontier too— there is
something negligent and criminal about the rollout of these tools"
I do not recall saying anything of the sort.
So. I am puzzled.
Jon
On 1/4/2026 5:24 PM, David Bernat via Discuss wrote:
> Jon & Gene,
>
> I think your mindsets of ML are rapidly out of date, and you’ll be
> pleasantly surprised about what processes AI actually supports in
> modern (2026) hands of chief leadership of the technology. When I
> clear internally to share my deck & response to TP later tonight or
> tomorrow I will. In the meantime though it is very clear OpenSCAD
> possesses more than enough materials.
>
> I’ll agree with you on one other frontier too— there is something
> negligent and criminal about the rollout of these tools, even in the
> open source community. I’ve been an AI leader for more than ten years
> now — the tools in the open source miss critical gaps which hurt long
> term value of the open source and wage economies.
>
> Different tech companies and we’d be better. I’ve created original
> research on everything from general relativity to cosmic
> constellations to brain machine interface robotics to AI— never has
> the mismatch between attitudes and value been so great and intention.
> It hurts people this upcoming year.
>
> But wait to read what I send and make up your own mind. I stay away
> from most tech companies for now. There is a fundamental error over
> there.
>
> Warm regard,
> Bernat
>
> David Bernat, Ph. D.
> Property of Starlight LLC.
>
>
> On Sun, Jan 4, 2026 at 5:07 PM gene heskett via Discuss
> <discuss@lists.openscad.org> wrote:
>
> On 1/4/26 14:19, Jon Bondy via Discuss wrote:
> > I think that any statement about how AI performs for a given
> task must
> > end with "... and it will improve over time". The key choice seems
> > when to get involved with AI.
> >
> > For me, I want to know that __I__ did it, I figured it out. I'm
> less
> > invested in efficient success and more on growth and competence. I
> > doubt that I will ever find a good reason to use AI.
> >
> > Jon
> Makes 2 of us Jon. Same reasoning. AI is only as good as the
> training.
> >
> > On 1/4/2026 2:05 PM, Torsten Paul via Discuss wrote:
> >> On 1/3/26 17:33, David Bernat via Discuss wrote:
> >>> To create OpenSCAD code the modern way is to use an LLM for
> natural
> >>> language, an LLM which is specialized in understanding OpenSCAD
> >>> code, and an OpenSCAD MCP which has a set of direct APIs for
> >>> generation, documentation lookup, testing, and debugging etc. The
> >>> user uses an interface specific to understanding how to code,
> test,
> >>> etc., of which the commercial ones are Claude, Windsurf, Cursor,
> >>> etc., but the open source OpenCode is by far my favorite.
> >>
> >> I certainly like the Open Source part of it. I don't mind if
> there's
> >> the option for users to connect to something they pay for. But to
> >> be worth maintaining the code in OpenSCAD, (talking about my
> personal
> >> time and choice) there has to be an option that is free and open in
> >> the actual meaning, not as in OpenAI by just name.
> >>
> >> While the discussion of the software side is interesting and all,
> >> my question are on a different part of that whole topic.
> >>
> >> 1) Who is going to train an LLM on this and on what code? And is
> >> code even enough to solve this? I'm not sure a text-only LLM
> will be
> >> able to really work for CAD as it has likely no training data that
> >> would relate to 3D concepts.
> >>
> >> 2) Where's an example that actually provides reasonable results?
> >>
> >> I've seen quite a number of claims of "I used AI to create this
> >> OpenSCAD model", but my tests so far are not very promising. While
> >> the coding syntax seems to be getting better, there are actually
> >> results that can be run with no or only minor syntactic errors. I
> >> have not seen any case yet where an even slightly more complicated
> >> design actually worked.
> >>
> >> Example: Design a dice with rounded corners, place the numbers on
> >> the sides in correct orientation and order for a dice.
> >>
> >> Earlier LLMs did not even get the general dice correct, mostly just
> >> tossing the text into the X-Y plane instead of correctly aligning
> >> them. Interesting is also to see that some choose digits, some go
> >> for the classic dots.
> >>
> >> The newer ones at least tried to get the orientation but still
> >> failed to make it work in a way that it's complete even with
> >> multiple corrections.
> >>
> >> ciao,
> >> Torsten.
> >> _______________________________________________
> >> OpenSCAD mailing list
> >> To unsubscribe send an email to discuss-leave@lists.openscad.org
> >>
> >> [Test Info: listname=discuss@lists.openscad.org
> >> user_email=jon@jonbondy.com deliver_to=jon@jonbondy.com]
> >> [member=jon@jonbondy.com]
> >
>
> Cheers, Gene Heskett, CET.
> --
> "There are four boxes to be used in defense of liberty:
> soap, ballot, jury, and ammo. Please use in that order."
> -Ed Howdershelt (Author, 1940)
> If we desire respect for the law, we must first make the law
> respectable.
> - Louis D. Brandeis
> Don't poison our oceans, interdict drugs at the src.
>
> _______________________________________________
> OpenSCAD mailing list
> To unsubscribe send an email to discuss-leave@lists.openscad.org
>
> [Test Info: listname=discuss@lists.openscad.org
> user_email=david@starlight.ai deliver_to=david@starlight.ai]
> [member=david@starlight.ai]
>
>
> _______________________________________________
> OpenSCAD mailing list
> To unsubscribe send an email todiscuss-leave@lists.openscad.org
>
> [Test Info:listname=discuss@lists.openscad.org user_email=jon@jonbondy.com deliver_to=jon@jonbondy.com]
> [member=jon@jonbondy.com]
--
This email has been checked for viruses by AVG antivirus software.
www.avg.com
MM
Michael Marx (spintel)
Mon, Jan 5, 2026 12:10 AM
That does "Property of Starlight LC." mean. Is Bernat, the author of these posts a living individual or an agent?
Michael Marx,
OpenSCAD Administration.
From: David Bernat via Discuss [mailto:discuss@lists.openscad.org]
Sent: Monday, January 05, 2026 9:24 AM
To: OpenSCAD general discussion Mailing-list
Cc: gene heskett; David Bernat
Subject: [OpenSCAD] Re: the OpenSCAD MCP server for LLMs [a discussion of first results for beginners]
Jon & Gene,
I think your mindsets of ML are rapidly out of date, and you’ll be pleasantly surprised about what processes AI actually supports in modern (2026) hands of chief leadership of the technology. When I clear internally to share my deck & response to TP later tonight or tomorrow I will. In the meantime though it is very clear OpenSCAD possesses more than enough materials.
I’ll agree with you on one other frontier too— there is something negligent and criminal about the rollout of these tools, even in the open source community. I’ve been an AI leader for more than ten years now — the tools in the open source miss critical gaps which hurt long term value of the open source and wage economies.
Different tech companies and we’d be better. I’ve created original research on everything from general relativity to cosmic constellations to brain machine interface robotics to AI— never has the mismatch between attitudes and value been so great and intention. It hurts people this upcoming year.
But wait to read what I send and make up your own mind. I stay away from most tech companies for now. There is a fundamental error over there.
Warm regard,
Bernat
David Bernat, Ph. D.
Property of Starlight LLC.
On Sun, Jan 4, 2026 at 5:07 PM gene heskett via Discuss discuss@lists.openscad.org wrote:
On 1/4/26 14:19, Jon Bondy via Discuss wrote:
I think that any statement about how AI performs for a given task must
end with "... and it will improve over time". The key choice seems
when to get involved with AI.
For me, I want to know that I did it, I figured it out. I'm less
invested in efficient success and more on growth and competence. I
doubt that I will ever find a good reason to use AI.
Jon
Makes 2 of us Jon. Same reasoning. AI is only as good as the training.
On 1/4/2026 2:05 PM, Torsten Paul via Discuss wrote:
On 1/3/26 17:33, David Bernat via Discuss wrote:
To create OpenSCAD code the modern way is to use an LLM for natural
language, an LLM which is specialized in understanding OpenSCAD
code, and an OpenSCAD MCP which has a set of direct APIs for
generation, documentation lookup, testing, and debugging etc. The
user uses an interface specific to understanding how to code, test,
etc., of which the commercial ones are Claude, Windsurf, Cursor,
etc., but the open source OpenCode is by far my favorite.
I certainly like the Open Source part of it. I don't mind if there's
the option for users to connect to something they pay for. But to
be worth maintaining the code in OpenSCAD, (talking about my personal
time and choice) there has to be an option that is free and open in
the actual meaning, not as in OpenAI by just name.
While the discussion of the software side is interesting and all,
my question are on a different part of that whole topic.
-
Who is going to train an LLM on this and on what code? And is
code even enough to solve this? I'm not sure a text-only LLM will be
able to really work for CAD as it has likely no training data that
would relate to 3D concepts.
-
Where's an example that actually provides reasonable results?
I've seen quite a number of claims of "I used AI to create this
OpenSCAD model", but my tests so far are not very promising. While
the coding syntax seems to be getting better, there are actually
results that can be run with no or only minor syntactic errors. I
have not seen any case yet where an even slightly more complicated
design actually worked.
Example: Design a dice with rounded corners, place the numbers on
the sides in correct orientation and order for a dice.
Earlier LLMs did not even get the general dice correct, mostly just
tossing the text into the X-Y plane instead of correctly aligning
them. Interesting is also to see that some choose digits, some go
for the classic dots.
The newer ones at least tried to get the orientation but still
failed to make it work in a way that it's complete even with
multiple corrections.
ciao,
Torsten.
OpenSCAD mailing list
To unsubscribe send an email to discuss-leave@lists.openscad.org
[Test Info: listname=discuss@lists.openscad.org
user_email=jon@jonbondy.com deliver_to=jon@jonbondy.com]
[member=jon@jonbondy.com]
That does "Property of Starlight LC." mean. Is Bernat, the author of these posts a living individual or an agent?
Michael Marx,
OpenSCAD Administration.
_____
From: David Bernat via Discuss [mailto:discuss@lists.openscad.org]
Sent: Monday, January 05, 2026 9:24 AM
To: OpenSCAD general discussion Mailing-list
Cc: gene heskett; David Bernat
Subject: [OpenSCAD] Re: the OpenSCAD MCP server for LLMs [a discussion of first results for beginners]
Jon & Gene,
I think your mindsets of ML are rapidly out of date, and you’ll be pleasantly surprised about what processes AI actually supports in modern (2026) hands of chief leadership of the technology. When I clear internally to share my deck & response to TP later tonight or tomorrow I will. In the meantime though it is very clear OpenSCAD possesses more than enough materials.
I’ll agree with you on one other frontier too— there is something negligent and criminal about the rollout of these tools, even in the open source community. I’ve been an AI leader for more than ten years now — the tools in the open source miss critical gaps which hurt long term value of the open source and wage economies.
Different tech companies and we’d be better. I’ve created original research on everything from general relativity to cosmic constellations to brain machine interface robotics to AI— never has the mismatch between attitudes and value been so great and intention. It hurts people this upcoming year.
But wait to read what I send and make up your own mind. I stay away from most tech companies for now. There is a fundamental error over there.
Warm regard,
Bernat
David Bernat, Ph. D.
Property of Starlight LLC.
On Sun, Jan 4, 2026 at 5:07 PM gene heskett via Discuss <discuss@lists.openscad.org> wrote:
On 1/4/26 14:19, Jon Bondy via Discuss wrote:
> I think that any statement about how AI performs for a given task must
> end with "... and it will improve over time". The key choice seems
> when to get involved with AI.
>
> For me, I want to know that __I__ did it, I figured it out. I'm less
> invested in efficient success and more on growth and competence. I
> doubt that I will ever find a good reason to use AI.
>
> Jon
Makes 2 of us Jon. Same reasoning. AI is only as good as the training.
>
> On 1/4/2026 2:05 PM, Torsten Paul via Discuss wrote:
>> On 1/3/26 17:33, David Bernat via Discuss wrote:
>>> To create OpenSCAD code the modern way is to use an LLM for natural
>>> language, an LLM which is specialized in understanding OpenSCAD
>>> code, and an OpenSCAD MCP which has a set of direct APIs for
>>> generation, documentation lookup, testing, and debugging etc. The
>>> user uses an interface specific to understanding how to code, test,
>>> etc., of which the commercial ones are Claude, Windsurf, Cursor,
>>> etc., but the open source OpenCode is by far my favorite.
>>
>> I certainly like the Open Source part of it. I don't mind if there's
>> the option for users to connect to something they pay for. But to
>> be worth maintaining the code in OpenSCAD, (talking about my personal
>> time and choice) there has to be an option that is free and open in
>> the actual meaning, not as in OpenAI by just name.
>>
>> While the discussion of the software side is interesting and all,
>> my question are on a different part of that whole topic.
>>
>> 1) Who is going to train an LLM on this and on what code? And is
>> code even enough to solve this? I'm not sure a text-only LLM will be
>> able to really work for CAD as it has likely no training data that
>> would relate to 3D concepts.
>>
>> 2) Where's an example that actually provides reasonable results?
>>
>> I've seen quite a number of claims of "I used AI to create this
>> OpenSCAD model", but my tests so far are not very promising. While
>> the coding syntax seems to be getting better, there are actually
>> results that can be run with no or only minor syntactic errors. I
>> have not seen any case yet where an even slightly more complicated
>> design actually worked.
>>
>> Example: Design a dice with rounded corners, place the numbers on
>> the sides in correct orientation and order for a dice.
>>
>> Earlier LLMs did not even get the general dice correct, mostly just
>> tossing the text into the X-Y plane instead of correctly aligning
>> them. Interesting is also to see that some choose digits, some go
>> for the classic dots.
>>
>> The newer ones at least tried to get the orientation but still
>> failed to make it work in a way that it's complete even with
>> multiple corrections.
>>
>> ciao,
>> Torsten.
>> _______________________________________________
>> OpenSCAD mailing list
>> To unsubscribe send an email to discuss-leave@lists.openscad.org
>>
>> [Test Info: listname=discuss@lists.openscad.org
>> user_email=jon@jonbondy.com deliver_to=jon@jonbondy.com]
>> [member=jon@jonbondy.com]
>
Cheers, Gene Heskett, CET.
--
"There are four boxes to be used in defense of liberty:
soap, ballot, jury, and ammo. Please use in that order."
-Ed Howdershelt (Author, 1940)
If we desire respect for the law, we must first make the law respectable.
- Louis D. Brandeis
Don't poison our oceans, interdict drugs at the src.
_______________________________________________
OpenSCAD mailing list
To unsubscribe send an email to discuss-leave@lists.openscad.org
[Test Info: listname=discuss@lists.openscad.org user_email=david@starlight.ai deliver_to=david@starlight.ai]
[member=david@starlight.ai]
