2025/06 - blog.lmorchard.com

Month: 2025/06

2025 June 04
- Hello world!
- The Verge, How to move a smart home
  - We've moved a lot. Mostly, I distrust smart home gadgets and don't have many. But, several of the houses we've owned had lingering smart devices. Many of them ended up useless. Occasionally a Nest thermostat could be coaxed to betray its former owner and work for me. For the most part, it's a mess.
- Once upon a time in college, I got a dial-up network connection working to my Commodore Amiga 1200 in my dorm room. I sprinted across campus to a computer lab to telnet back into my A1200. It was so neat. And pointless. But neat.
  - This, of course, was before it occurred to me that anyone with my temporary IP address could have also telnetted into my A1200. 🤷‍♂️
- Had some adventures in vibe coding, last night. Maybe I'll write about it? I keep reading folks saying this stuff doesn't work, but... it does?
# 11:59 pm
- miscellanea
Adventures in Vibe Coding with Grafana and Claude

Since re-launching my blog, I wanted to monitor traffic and logs more closely. Nothing groundbreaking, but it had been a while since I'd run Grafana, Prometheus, and Loki on my own hardware.

Turns out there's this handy all-in-one docker-compose setup that runs on Synology NAS. It fired up with minimal fuss, and soon I had metrics machinery humming in my basement—except the package didn't include Loki. A quick docs consultation got it running alongside the rest.

My blog is a static site hosted via AWS S3 and CloudFront. Both services dump logs into an S3 bucket, but I'd never bothered reading them before—and didn't want to start now. Instead, I loaded up Claude.ai and described my problem:

I want to get logs out of CloudFront. I have enabled new-style log delivery that stores gzipped JSON logs in an S3 bucket at s3://lmorchard-logs/blog.lmorchard.com/ with names like E5YXU82LZHZCM.2025-06-04-04.d024d283.gz

Can you help me write a script for my home Loki server to download only new log files and push them into Loki?

Claude stepped right up:

I'll help you create a script to process CloudFront logs and push them to Loki. Let me write a Python script that tracks processed files and handles the gzipped JSON format.

After some vibey iteration, we landed at this artifact:

It's quite verbose and could use some tightening up. But, I really don't care—it does the quick & dirty needful.

I wrote zero Python. I just henpecked Claude to add features until the script did what I needed. I wasn't even in an IDE, just the Claude.ai interface in a browser. An interesting thing to note is that Claude didn't have access to my AWS resources—I didn't even give it a sample of my logs. But, still, what I told it about JSON, S3, and CloudFront was enough for it to be off to the races.

Anyway, after a quick review and a satisfactory dry run, I dropped it into a cronjob to grab new logs every 5 minutes. Then I pestered Claude with Grafana dashboard questions I could have figured out myself. But why read docs when you can just ask? (Which I realize is ironic, since I wrote Too long? Read anyway. but I think I make an exception for LLMs.)

Total time from idea to working dashboard: about an hour.

Not revolutionary, but pretty satisfying for barely having to think about it.
# 3:26 pm
2025 June 02
- Hello world!
- Jotted down a couple posts today on AI stuff that aren't particularly revelatory.
  - If anything, they're just me trying to think out loud and clarify.
  - I'm probably going to try writing more stuff like this, if only to be Wrong on the Internet and lure someone in to correct me. 😅
- Dang it, I don't wanna go to bed, I just discovered strudel.cc
# 11:59 pm
- miscellanea
Quoting W. David Marx on Gen AI

W. David Marx, GenAI is Our Polyester:

Everyone knows happened next: There was a massive cultural backlash against polyester, which led to the triumphant revaluation of natural fibers such as cotton and linen. The stigma against polyester persists even now. The backlash is often explained as a rejection of its weaknesses as a fiber: polyester's poor aeration makes it feel sticky. ... While polyester took a few decades to lose its appeal, GenAI is already feeling a bit cheesy. We're only a few years into the AI Revolution, and Facebook and X are filled to the brim with “AI slop.” Everyone around the world has near-equal access to these tools, and low-skilled South and Southeast Asian content farmers are the most active creators because their wages are low enough for the platforms' economic incentives to be attractive.

This along with remembering that some professors are going back to handwritten essays (and also that handwriting is better for memory and learning) had me wondering if there's going to be a handcrafted backlash in the next few years?

I write journal entries nearly every day by hand—albeit these days on an e-ink tablet. I think that helps me focus on what I want to dredge out of my head. I keep meaning to get back to that handwriting recognition project I started a few weeks ago, since no product I've tried yet has been able to turn my writing into clean machine-readable text.

But, then again, maybe producing machine-illegible works by hand will be the next big trend?
# 4:58 pm
- genai
- ai
- llms
My New Rube Goldberg Blogging Machine

According to the count in my archives, I've published over 50 blog posts in the past few weeks. That's roughly 50 more than I managed in the previous 10 years! These aren't masterpieces—mostly just random thoughts and half-baked ideas. But as I mentioned before, I'd rather throw a bunch of stuff at the wall and see what sticks than spend another decade crafting the perfect post that never gets published. So, here's how I tinkered my way into a writing setup that seems to actually be working. [ ... 873 words ... ]
# 4:04 pm
The Bomb Still Works: On LLM Denial and Magical Thinking

I found myself in a frustrating argument with someone convinced that LLMs are pure vaporware—incapable of real work. Their reasoning? Since LLMs were trained on stolen material, the results they produce can't actually exist.

Not that the results should be considered illegitimate or tainted—but that they're literally impossible. That the training data's questionable origins somehow prevents the technology from functioning at all.

I couldn't convince them otherwise. But, life isn't fair and both things can be true simultaneously: the origin of something can be problematic and the results can be real.

This analogy kept coming to mind: If someone steals materials to build a bomb and successfully builds it, they have a functioning bomb. The theft doesn't retroactively prevent the bomb from existing or reduce its explosive capability. Proving the theft might help with future bombs or justify going after the bomb-maker, but it doesn't cause the current bomb to magically self-dismantle.

This seems obvious to me—embarrassingly so. Yet I keep encountering this form of reasoning about LLMs, and it strikes me as a particular kind of denial.

There's something almost magical in the thinking: that moral illegitimacy can somehow negate physical reality. That if we disapprove strongly enough of how something was created, we can wish away its actual capabilities.

The ethical questions around LLM training data are important and deserve serious discussion. But pretending the technology doesn't work because we don't like how it was built isn't engaging with reality—it's a form of wishful thinking that prevents us from dealing effectively with the situation we actually face.

Whether we like it or not, the bomb has been built. Now we need to figure out what to do about it.
# 12:20 pm
- llms
- ai
- ml
Why Prompt Engineering Isn't Just Good Writing

Someone told me that prompt engineering isn't real—that it's just techbros rebranding "good writing" and "using words well." I disagree, and here's why:

Prompt engineering fundamentally differs from writing for human audiences because LLMs aren't people. When done rigorously, prompt engineering relies on automated evaluations and measurable metrics at a scale impossible with human communication. While we do test human-facing content through focus groups and A/B testing, the scale and precision (such as it is) here are entirely different.

The "engineering" aspect involves systematic tinkering—sometimes by humans tweaking language, sometimes by LLMs themselves—to activate specific emergent behaviors in models. Some of these techniques come from formal research; others are educated hunches that prove effective through testing.

Effective prompts often resemble terrible writing. The ritual forms, repetitions, and structural patterns that improve LLM performance would make a professional editor cringe. Yet they produce measurable improvements in evaluation metrics.

Consider adversarial prompts: they're often stuffed with tokens that are nonsense to humans but exploit specific model quirks. Here, the goal is explicitly to use language in ways that aren't human-legible, making attacks harder to detect during review.

Good writing skills can help someone pick up prompt engineering faster, but mastering it requires learning to use words and grammar in weird, counterintuitive ways that are frankly sometimes horrifying.

All-in-all, prompt engineering may still be somewhat hand-wavy as a discipline, but it's definitely real—and definitely not just rebranded writing advice.
# 12:12 pm

Month: 2025/06

2025 June 04

Adventures in Vibe Coding with Grafana and Claude

2025 June 02

Quoting W. David Marx on Gen AI

My New Rube Goldberg Blogging Machine

The Bomb Still Works: On LLM Denial and Magical Thinking

Why Prompt Engineering Isn't Just Good Writing