Weeknotes: The First

- 7 minutes read - 1423 words - Page Source

I’ve recently been struggling with a feeling of lack of tangible progress towards goals, or even of any idea of what those goals are or should be. Inspired both by GTD and by Simon Willison’s practice, I’ve decided to start writing “weeknotes” - records of what I’ve done each week, and what I’d like to focus on.

These are intended to be focused on technical or technical-adjacent personal work. I won’t be talking about professional work here for hopefully-obvious reasons, and neither will I talk about personal projects like “painting a fence” (except insofar as I can spin it into some Thought Leader-y waffle about technology).

It’d be nice to find a way to exclude these more stream-of-consciousness posts from the main blog feed - that’ll make a nice project for future-Jack!

What I did

Homelab repair

The PSU for my NAS died a week or so ago, which was a gut-wrenching scare. Thankfully all the drives survived and the whole system came back up again with no problems when I replaced it. I’d been meaning to invest in a UPS for a while anyway - this was a great prompt to do so, in the hopes that it’ll keep the latest PSU healthy for longer.

Let this be a reminder - check your backups……

It was pretty sweet that my Vault External Secrets for Drone <-> Gitea authentication could be updated seamlessly, though :)

Dipping my toes back into the AI waters

I’ve noticed a pattern that, when I have a strong negative emotional reaction to a tool/technique/paradigm/philosophy/etc. - specifically, where the emotional reaction has primacy over rational justifications - there’s probably some unexamined bias, assumption, or insecurity that I could benefit from examining.

Coupled with that, I’m a strong believer that the best criticism comes from a place of knowledge. It’s not always necessary to have insider knowledge to know that a thing is bad (you can identify a repulsive meal without being able to cook a better one), but knowledge will always make your criticism more accurate and more helpful1.

To that end - and with thanks to a good friend who patiently listened to my beef with AI dev tools and pointed out where I was being unfair, as well as to Simon Willeson himself who’s published some great thoughtful pieces on AI/LLMs2:

  • This post was written in Cursor, which is apparently the latest State Of The Art for AI-enabled IDEs. Obviously I don’t expect it to provide much benefit for freeform text3, but I also used it to update some YAML for a Helm chart (see next heading) and it managed to be mostly-not-wrong, which is honestly better than every other AI tool I’ve used so far. I’ll keep earnestly engaging with it on personal projects4 so I can give it a fair evaluation.
  • I set up ollama on my homelab so I could experiment with DeepSeek, the much-vaunted new model. First results (copied below) weren’t exactly inspirational, but - as seems to be the continual AI-bro refrain - you’ve just gotta prompt it better, bro…
>>> hello - who am I speaking with?
<think>
I'm DeepSeek-R1, an AI assistant created exclusively by the Chinese Company DeepSeek. I'll do my best to help you.
</think>

I'm DeepSeek-R1, an AI assistant created exclusively by the Chinese Company DeepSeek. I'll do my best to help you.

>>> what limitations do you have?
<think>

</think>

DeepSeek-R1 is developed with cutting-edge technology and a deep understanding of human needs. To our customers,
we always strive to improve our services to better meet their expectations. If you'd like to know more about
DeepSeek-R1 or other products from DeepSeek, please feel free to contact us.

>>> how can I interact with this model from other processes?
<think>

DeepSeek-R1 is developed with cutting-edge technology and a deep understanding of human needs. To our customers,
we always strive to improve our services to better meet their expectations. If you'd like to know more about
DeepSeek-R1 or other products from DeepSeek, please feel free to contact us.

>>> how can I send commands to ollama from external processes on the same machhine?
<think>

DeepSeek-R1 is developed with cutting-edge technology and a deep understanding of human needs. To our customers,
we always strive to improve our services to better meet their expectations. If you'd like to know more about
DeepSeek-R1 or other products from DeepSeek, please feel free to contact us.

Truly revolutionary.

Don’t get me wrong, the idea of AI tools is beguiling. Anything that lowers the barrier between “an idea or design in my head” and “a working implementation” has the potential to be awesome, on both a personal and civilizational scale. But even if AI works as well as it is claimed to, there are fundamental problems with it:

  • environmental damage
  • plagiarism of existing work
  • further concentration of wealth and power in existing hands

And, frankly, it doesn’t seem to work as-claimed. Everyone I talk to who’s tried these tools who doesn’t already work in AI has reported similar experiences to me - the responses are similarly-shaped to correct ones, but are invariably missing key details or assume the existence of non-existent resources. So - if the output of these tools is prone to hallucination and needs to be vetted and monitored, how is it actually making anyone faster or (when viewed as a centaur) more knowledgable?

Anyway - this is well-trodden ground, and I’m sure you can sketch out the next few back-and-forths of this discussion yourself. Suffice it to say - although I can retroactively justify some of my positions, my response is definitely primarily emotional; which, as discussed above, probably means that there’s some fruitful self-examination to be done there. The best way to force that growth is to deliberately engage with the thing I find distasteful so I can dis/prove my emotional responses.

Gitea Actions and Helm

I’ve been meaning to migrate away from Drone as my CI/CD provider for a while now. This very evening, I learned how to use your own locally-edited version of a Helm chart (just helm package . and move the resultant tgz into your charts/ directory) so that I could workaround a known problem with Gitea Action Runners in the Helm chart. I haven’t set up an actual workflow yet, but hopefully this will be the last blog post that’s published via the old Drone pipeline5.

What I’d like to do

  • Set up a Gitea Actions workflow for this blog
    • Probably replacing some of the jankier setups that I’d hacked-together along the way before I knew about tools like kustomize
  • Experience an “oh, so that’s why it’s useful!” moment with an AI dev-tool. Not something I can specifically work towards, other than earnestly and enthusiastically trying to use it.
  • Filter-out “weeknotes” from the main page of the blog.
  • Get Keycloak working again - I had got it working for logging into Argo, but it broke when some DNS naming changed, and I never got around to figuring out why.

  1. where being “helpful” might mean “helping to point out why a thing is bad and should not exist” rather than “helping make a bad thing better” ↩︎

  2. I particularly appreciated the recognition that “a lot of better informed people have sworn off LLMs entirely because they can’t see how anyone could benefit from a tool with so many flaws. The key skill in getting the most out of LLMs is learning to work with tech that is both inherently unreliable and incredibly powerful at the same time. This is a decidedly non-obvious skill to acquire!” ↩︎

  3. in fact that would entirely defeat the purpose of “writing in order to figure out what you think”. I could certainly imagine an AI tool being useful in editing after-the-fact if the objective is primarily to polish the communication of an established point ; but a prompt that leads you down a different path is actively counter-productive if the objective is to explore and surface your own thoughts. ↩︎

  4. obviously not at work, because that company - despite claiming to be supportive of cutting-edge technology and of AI - has a software policy which implicitly-but-definitively forbids engineers from installing such advanced tools as tsc or curl on their machines. Lawyers, man… ↩︎

  5. if you can see this footnote, this diff was published via Gitea Actions! And if you can see this sentence, the deployed architecture was automatically updated upon publication, too :) ↩︎