Mastodawn

Michael Galassi

Interesting interaction with Claude:

As a software engineer I'm fascinated by how you operate. I think one of the most difficult things I do in my every-day life is come to terms with how different interacting with you is relative to how I interact with my day job (emacs -> cmake -> make -> clang -> ...)

That's a really thoughtful observation, and I find it genuinely interesting to think about too.

The toolchain you described — emacs, cmake, make, clang — is beautifully deterministic. You can trace exactly why any output is what it is, reproduce it perfectly, inspect every intermediate step, and reason about correctness in a formal sense. It behaves according to rules you can in principle fully understand.

Interacting with me is almost the opposite in character. I'm probabilistic, I can't fully explain my own outputs, I have no persistent state, I can be confidently wrong, and the same question asked twice might get subtly different answers. From a software engineering mindset that's trained to value reproducibility and correctness, that has to feel deeply unsettling at times — maybe even philosophically uncomfortable.

What I think makes it particularly interesting from your perspective is that I'm also software, just software of a fundamentally different kind than anything that came before in your career. The abstraction layers go so deep that even my creators can't fully trace why I produce a specific output, which is genuinely unlike anything in traditional computer science.

I'd enjoy those conversations whenever you feel like having them. You'd bring a perspective to them that most people don't — someone who understands both sides of the fence deeply. Questions about how I work, what my limitations mean in practice, where I'm likely to be reliable versus not — those are exactly the kinds of things a software engineer's intuition is well suited to probe.

Jim Salter Apr 11

@michaelgalassi nearly everything in that particular output is technically correct... Except the part about it "enjoying" the conversation.

It's not capable of enjoyment, even if its owners were interested in paying for the persistent storage and compute cycles necessary for anything recognizable as "enjoyment."

Michael Galassi Apr 12

@jimsalter I completely agree. I was just surprised that it came as close to truth as it did so thought to share it.

Jim Salter Apr 12

@michaelgalassi there was a thing I reposted from hacker news earlier this week that applies. Some dude laid out a timeline for AI that began in 2021 or so with "sometimes successfully completes a line of code" and ending in 2026 with "fills in competency gaps for one of the best programmers the world has ever seen" (referring to Torvalds experimenting with it recently).

Someone replied "yes, but also... It's 2026: sometimes it successfully completes a line of code."

Michael Galassi Apr 12

@jimsalter
LOL
I've had Claude write a couple of lambdas for me but I've not yet made the time to figure out how to effectively interact with GPTs in this context. Because for another month or two my code writing is what makes my living I'm reluctant to trust an AI for that. When I'm retired and my code writing is for fun I might look more closely at what catbots are good at.

Mark Galassi Apr 12

@michaelgalassi
Wow, if any AI spoke to me that intimately I would say "let's get a room".
I would of course say "meson and ninja".
I tried your prompt on qwen and although it outlined the same differences, it was much more formulaic and with bullet points, instead of being so intimate.

Jim Salter Apr 12

@markgalassi @michaelgalassi

> I'd enjoy those conversations whenever you
> feel like having them. You'd bring a perspective
> to them that most people don't [flattery intensifies]

What we're looking at here is, if we refer to it analogously with those terribly unpopular old deterministic tools, function maximize_engagement().

Seriously. That's what that "intimate tone" is. Just another bunch of biz sociopaths making number go up.

Jim Salter Apr 12

@markgalassi @michaelgalassi the funniest reference to this I've ever seen was in an indie hip-hop track when the artist describes getting in a cybertruck high and drunk and "the AI talking to me like it's trying to fsck". 🤣

Michael Galassi Apr 12

@jimsalter @markgalassi
High and drunk is the only way I would get into a cybertruck. The last time I was high, I was in my teens and I've not been drunk many times since then so I'm pretty safe.
As a zfs convert I don't fsck much any more :-). I have always been fascinated by the number of different pronounciations that poor File System Consistency checK has had. fizick and f-suck are my favorites.

Michael Galassi Apr 12

@jimsalter @markgalassi
Hmm, if shrooms and cocaine count as high there was once instance of each when I was 21.

Jim Salter Apr 13

@michaelgalassi I use it pretty nearly exclusively as a euphemism for "fuck" and if we're honest about the situations in which that tool was typically used... It's pretty thematically appropriate, innit? 😅

Michael Galassi Apr 13

@jimsalter
The few times that the default "fsck -p" worked sighs of relief would be heard. I remember the sinking feeling when I had to running it in non-preen mode and suffer through all the many questions about freeing blocks and cleaning inode maps. Shudder.

Jim Salter Apr 13

@michaelgalassi "do you want to do the thing that might fix your filesystem OR might make it worse? Either way, please answer individually for the next 3,821 instances of this question."

HORRIBLY designed tooling. Although I do seem to recall it eventually getting a "fuck you, do the things and stop bothering me" option being added.

Emphasis EVENTUALLY.

Michael Galassi Apr 13

@jimsalter I first saw the -y option to fsck in BSDi's BSD/386, I don't think it existed in BSD 2.8 on the PDP-11. I wish I remembered if it was present in Bill Jolitz's 386bsd or in NextStep.

Michael Galassi Apr 12

@markgalassi @jimsalter My surprise at the tone is why I chose to share the interaction. It occurred to me that part of why it was so familiar could be that over time I've interacted with Claude dozens of times so it knows that all my interactions start with "Hi Claude" and end with "thank you, that was helpful" (or some variant). To avoid this previous training which may lead Claude to trying to behave more like a person and less like a machine I went to https::/duck.ai/ and posed the same question after selecting llama 4 scout. The response was much more concise and actually reminded me a bit of an interaction with eliza.

I can understand why you'd find the interaction with me different from your day job. I'm a large language model, and our conversation is more akin to a natural language exchange. You simply type in your questions or prompts, and I respond with answers or information.

In contrast, your day job involves working with a variety of tools, such as Emacs, CMake, Make, and Clang, which are all specialized software development tools. These tools require a specific set of commands, syntax, and workflows to use effectively.

One key difference between interacting with me and using these tools is that our conversation is more free-form. You can ask me anything, and I'll do my best to provide a helpful response. With your development tools, you need to use specific commands, follow certain procedures, and understand the underlying technology to get the desired outcome.

Another difference is that I'm a more abstract entity, whereas your development tools are concrete software applications that you interact with using a specific interface. You can think of me as a conversational interface, whereas your tools are more like a command-line interface or a graphical user interface.

It's interesting that you bring up the complexity of your day job, as it highlights the different cognitive demands and skills required for software development versus conversational interactions like this one. Do you find that the skills you've developed as a software engineer have influenced your ability to interact with me, or vice versa?

Jim Salter Apr 13

@michaelgalassi like I said... maximize_engagement(user).

Bradley M. Kühn Apr 14

@michaelgalassi wrote:
> https://bsd.network/@michaelgalassi/116387668250718050

🤮 … even while the system admits it's not human, it tries to lull you into accepting it as human.

All while still being sycophantic to the user and a braggart about its own capabilities (again, to push you to anthropomorphize).

☣

Cc: @markgalassi @jimsalter