Re: Just take me out back and shoot me...

The key I've found was to work in incremental steps. 

Basically, my Emacs+Claude integration looks like this: put the point
on the start of an s-exp, call M-x claude-rewrite-sexp and give it a
prompt. So at any point of time, it's only rewriting one small S-exp.
If I ask it to rewrite an Emacs region, if frequently gets the
parenthesis wrong, but if I ask it to rewrite an S-exp, never hits the
parenthesis issue.

If it struggles, I give it a few other carefully chosen sample files to
show it the patterns I'm using. (For instance, special macros,
functions and what not.) 

Also, ask it to write unit tests, it does a pretty good job and you'll
be validating its work as you go. (Btw, I don't even use Opus, I just
use Claude Sonnet, and it's still great.)

Claude Code is a different beast. I haven't had success with production
code with it yet, but it was useful in order to test out complex
refactorings to see where I might get stuck if I had to do it myself.

--Arnold

On Fri, 2025-08-08 at 08:15 +1200, Eben Bruyns (as eben at sdk dot co
dot nz) wrote:
> Oh I want to pipe up here. I'll share a little bit about what I've
> been doing in lisp while I'm at it.
> 
> Some of you might have seen me pop up in here asking some strange
> looking questions at times, time to show and tell. 
> 
> The main project I've been working on is an extensible launcher
> platform for a Gaming handheld machine called the Steam Deck. I
> started off writing this in python as a FOSS project, but since the
> community turned out to be basically me, I ditched the efforts there
> in favour of finally learning Common Lisp properly.
> 
> I started off using these AI tools (co-pilot, chat gpt etc), as a
> test. With python it was yielding somewhat ok results initially.
> Here's the open source repo for those
> interested: https://github.com/ebenbruyns/junkstore
> 
> I ran into problems with it pretty quickly. I can say with reasonable
> certainty that every single hard to find bug was casually introduced
> by these AI systems being "helpful". They've created many subtle race
> conditions that took hours or days to debug, and once I looked at the
> code I realised that I would never have written it that way to start
> with.
> 
> I'll give you guys some hours spent on this just so you can have
> context and put it all in perspective. The FOSS code took around 1000
> hours to develop, test debug etc. I have another component written in
> python too that adds functionality to the FOSS version that's paid
> and took another 400 hours or so.
> 
> When I decided to stop writing open source and in Python I switched
> to Lispworks (best decision ever). I've spent another 4500-5000 hours
> working on this new version, but AI can't code lisp. Instead what I
> found was that all the tasks I used AI for I could just abstract away
> in macros, clos and other lisp specific goodness that made the AI
> obsolete really.
> 
> You can see the application here: https://www.junkstore.xyz/
> 
> Then a few weeks ago I talked to a friend and they put me onto Claude
> code. The results have been a mixed bag. I'm doing some pretty heavy
> reverse engineering stuff (as you can imagine with a project like
> this), and it's helped me build some tools to aid in this very
> quickly, but it destroys things as quickly as it can create them.
> It's been a wild roller coaster of  2 weeks with this thing and I
> think if I just coded stuff in lisp I could potentially have done a
> lot more in that time.
> 
> But it's not all been wasted. My wife, who is not a programmer, has
> been using Claude code to crunch, twist and bend the website. I think
> the results are pretty decent. As an experiment I thought I'd throw
> her the secure portal UI (since all the real code is in services the
> security aspect is abstracted away from it and it's just user
> experience that can get affected). I have to say the results and the
> speed at which she was able to do things is incredible. I did have to
> review the code, and I actually need to review the code much closer
> than I would with human devs. The problem is that you can't tell them
> why something is bad and not to do it again, they don't remember. So
> you have to check for the same stuff over and over again.
> 
> I think there's certainly value in these tools, but I don't think the
> value is as much as the industry wants you to believe. The real trick
> would be to find a way to put these tools on rails and keep them on
> rails.
> 
> I've done some experimentation with setting these tools up with
> testing feedback loops and you can get better results if you do that,
> but it's still an exercise in frustration most of the time. If I had
> the resources I'd get my hands on some serious hardware that can run
> a "large" small language model or something and specialise that on
> the stuff I'm doing, to see if that yielded better results. Living on
> a sailboat with solar power and using a 15w power budget on my
> machine doesn't really leave much room for this type of thing though.
> 
> Anyways, I just wanted to share this as I think it might be
> interesting to some of you.
> 
> Regards,
> 
> Eben Bruyns
> 
> 
> On Fri, 8 Aug 2025 at 06:24, Alexey Veretennikov (as alexey dot
> veretennikov at protonmail dot com) <[email protected]> wrote:
> > Yes indeed Claude could write Lisp code now. I've tried and it
> > works for some primitive functions/algorithms. I successfully used
> > it with Aider for CL code.
> > I would love to see some MCP agents with integration to popular
> > APIs in LW, so we can try to use it at least in chat mode with
> > ability to modify open files.
> > Would be also nice to have a feature in Editor for auto-reverting
> > modified buffers, so integration with command line tools like Aider
> > could be easier. I've submitted the feature request for that and
> > hope it is not a huge work to do, it would help a lot.
> > Would be nice as well to have a terminal with ansi colors support
> > so we can launch i.e. aider directly from LW.
> > 
> > BR,
> > /Alexey
> > 
> > 
> > On Thursday, August 7th, 2025 at 20:14, Arnold Noronha (as arnold
> > at tdrhq dot com) <[email protected]> wrote:
> > 
> > > Claude (not Claude Code, not Claude Chat, but the Anthropic API)
> > > actually writes Lisp really well, as long as you give it the
> > > right
> > > context because everyone's Lisp code looks different.
> > > 
> > > It even handled CAPI quite nicely, even though it hallucinated a
> > > bit..
> > > but close enough to guide me through my first CAPI app.
> > > 
> > > It's also remarkably good at writing Elisp, which makes it easier
> > > to
> > > write dev tools. I used that to bootstrap a little Cursor like
> > > Emacs
> > > integration, which I use almost daily now.
> > > 
> > > Thanks,
> > > Arnold
> > > 
> > > On Thu, 2025-08-07 at 09:59 -0700, David McClain (as dbm at
> > > refined-
> > > audiometrics dot com) wrote:
> > > 
> > > > I finally caved in this morning and tried ChatGPT. I am long
> > > > familiar
> > > > with “AI” in its present state, having directly used
> > > > Perception,
> > > > Kohonen, and Probabilistic networks back in the 1990’s for
> > > > control of
> > > > “rubber mirrors”, target tracking for missile guidance, etc,
> > > > etc.
> > > > 
> > > > I was holding off because this is not AI, just a recognizer of
> > > > patterns with an ability to generalize. Komogorov wrote a
> > > > theorem a
> > > > while back that a feedback perceptron network with N inputs +
> > > > feedback can emulate an Nth degree attractor in high
> > > > dimensional
> > > > fractal systems. That’s why we could use the rubber mirror to
> > > > cancel
> > > > out atmospheric scintillation of star images on the Kitt Peak
> > > > 4m
> > > > telescope back in the early 90’s.
> > > > 
> > > > At any rate, I saw a video expressing the virtues of ChatGPT
> > > > for web
> > > > search, rather than Google search, since it doesn’t offer up
> > > > tons of
> > > > extraneous finds based on whomever paid the most to get in
> > > > front of
> > > > your eyeballs. And so I tried it, and it does work better.
> > > > 
> > > > But, as long as you already know a great deal about some topic,
> > > > you
> > > > are prepared for its hallucinations (which I have yet to see),
> > > > and
> > > > its tendency to offer up a consensus opinion as authoritative
> > > > fact.
> > > > You have to push back and challenge its answers, and that
> > > > generally
> > > > provokes an even better response.
> > > > 
> > > > So, along the lines of a discussion about Transactional
> > > > Conventional
> > > > Hewitt Actors, it offered up the statement at one point about
> > > > parallel concurrent activation of the same Actor by two
> > > > different
> > > > dispatch threads sending two different messages at the same
> > > > time:
> > > > 
> > > > "You retry only if the actor behavior was concurrently
> > > > changed, which is rare in practice if messages are processed
> > > > quickly.”
> > > > 
> > > > I challenged that as an opinion, not fact, since I had the
> > > > experience
> > > > of a debug monitor Actor becoming overwhelmed by parallel
> > > > message
> > > > collisions in my Ionosphere Watch System:
> > > > 
> > > > "Your statement "You retry only if the actor behavior was
> > > > concurrently changed, which is rare in practice if messages are
> > > > processed quickly." is just a (hopeful) opinion. It could very
> > > > well
> > > > happen that collisions in some situations could happen very
> > > > frequently.”
> > > > 
> > > > To which it acknowledged my criticism as valid, and offered an
> > > > even
> > > > better answer, mentioning an exponential back-off, like we use
> > > > in
> > > > TCP/IP and Transactional Memory systems.
> > > > 
> > > > That was an interesting suggestion, but I can probably get by
> > > > just
> > > > using a “put the collision message back on the queue for later
> > > > delivery attempts”, instead of a direct immediate retry with an
> > > > exponential back-off.
> > > > 
> > > > At any rate, I am now willing to concede that ChatGPT may be
> > > > useful,
> > > > especially if you are alone in a room staring at 4 walls. But
> > > > I’m
> > > > still not willing to consider it true AI. And I haven’t allowed
> > > > it to
> > > > write any Lisp for me. Maybe someday...
> > > > 
> > > > - DM
> > > 
> > > 
> > > 
> > > _______________________________________________
> > > Lisp Hug - the mailing list for LispWorks users
> > > [email protected]
> > > http://www.lispworks.com/support/lisp-hug.html
> > 
> > _______________________________________________
> > Lisp Hug - the mailing list for LispWorks users
> > [email protected]
> > http://www.lispworks.com/support/lisp-hug.html


_______________________________________________
Lisp Hug - the mailing list for LispWorks users
[email protected]
http://www.lispworks.com/support/lisp-hug.html