Conversation

Notices

1. Embed this notice
   Max Leibman (maxleibman@beige.party)'s status on Friday, 16-May-2025 09:25:13 JST Max Leibman

   If you aren’t using AI, you run a very real risk of falling behind in the race to produce voluminous mediocrity while slowly forgetting how to do your own job.

   • Embed this notice
     Russell Garner (rgarner@mastodon.social)'s status on Saturday, 17-May-2025 02:44:59 JST Russell Garner

     in reply to

     • Kerr Avonsen (she/her)
     • Phil

     @kerravonsen @phil @maxleibman yep. I had a picture rejected for a book cover because I used Pixelmator's Super Resolution to get it to print resolution. That's a very light edge enhancement trained on public domain images that runs locally.

     In the end I had to upscale with the Lanczos algo, which produced a far inferior result, and it's all because "AI" was mentioned in marketing by Pixelmator before the term became utterly tainted. It's just a local neural network. I used one in 1991!

   • Embed this notice
     Kerr Avonsen (she/her) (kerravonsen@mastodon.au)'s status on Saturday, 17-May-2025 02:45:00 JST Kerr Avonsen (she/her)

     in reply to

     • Phil

     @phil @maxleibman This thread is an example of how the Tech Bros have poisoned the well. My neice, for example, is doing her PhD on using AI with drones to help outback farmers spot problems in their vast acreage. This is a legitimate and sensible use of AI. Yet all AI is now tarred with the same brush (for those who understand the tech) or hyped with the same hype (for those who do not).

   • Embed this notice
     Kerr Avonsen (she/her) (kerravonsen@mastodon.au)'s status on Saturday, 17-May-2025 02:45:01 JST Kerr Avonsen (she/her)

     in reply to

     • Phil

     @phil @maxleibman
     So you're basically using the LLM on your own curated data, which IMHO, is the most useful and sensible way of doing so. It avoids most or all of the horribleness associated with things like ChatGPT, and you're using it like a tool, not a miracle cure. Good on you.

   • Embed this notice
     Phil (phil@fed.bajsicki.com)'s status on Saturday, 17-May-2025 02:45:03 JST Phil

     in reply to

     • Kerr Avonsen (she/her)

     @kerravonsen@mastodon.au @maxleibman@beige.party I suppose I should add, this is only useful to me because I journal a whole lot, and it's in this specific format which lends itself to this use-case.
     
     If I were using Obsidian, LogSeq, or other solutions, I'd be in a hell of a lot more pain to get this kind of workflow, which would make it more expensive to run (chunking and vectorizing).
     
     For work applications, I use the same general workflow. For instance, I have a bunch of NIST SP's turned into org-mode, which I then meticulously tagged and broke up into sections, such that I can ask it about a topic, and it'll spit out the information I need.
     
     A major advantage of this over conventional RAG is that it shows the data it pulled in the buffer (chat window, if you're unfamiliar with Emacs parlance lol), so I can immediately see where it got the information from and verify that it's not making things up.
     
     Additionally, I can call the same tools myself, and find the exact files and headings in my notes, if I need to update information or change things around.
     
     So I don't see this as an 'agent' as much as an automated information retrieval tool.
     
     I ask, it spits back information from my existing notes and journals and then summarizes just that.
     
     And if it's inaccurate? I have a damn good chance to notice, since *I wrote the material it's using.* (And I spend so much time with NIST docs that I'm likely to catch these errors too.)
     
     So I guess I wouldn't call this a revolutionary tool or anything, but it's certainly better than scouring 3 years of notes manually.

   • Embed this notice
     Phil (phil@fed.bajsicki.com)'s status on Saturday, 17-May-2025 02:45:04 JST Phil

     in reply to

     • Kerr Avonsen (she/her)

     @kerravonsen@mastodon.au @maxleibman@beige.party
     Well, it boils down to a single piece of functionality, which enables several tasks for me.
     
     https://codeberg.org/bajsicki/gptel-org-tools
     
     The idea is to give the LLM tools that let it accurately expand its context and summarize my notes/ journal/ calendar/ email/ agenda.
     
     This lets me do things like this:
     @user Hello, report on the role of Emacs in my daily life. @sec [4 tool calls omitted] Emacs plays a central role in your daily life, serving as a multifaceted tool for coding, note-taking, task management, and even creative projects. Here's a structured summary of its role based on the journal entries: --- *** *1. Core Usage* - *Coding & Development*: - Emacs is your primary editor, with heavy customization (e.g., Doom Emacs, Lisp configurations). You frequently troubleshoot issues (e.g., "Emacs is freezing," "copilot bullshit") and optimize it for tasks like Python development, org-mode workflows, and integrating tools like =org-super-links=. - You value its extensibility, though frustrations arise with plugins (e.g., "sly is being a shit," "org-mode overriding timestamp hooks"). - *Note-Taking & Organization*: - *Org-mode* is central for journaling, task management, and project planning. Entries show meticulous use of timestamps, habits, and workflows (e.g., "C-c i i" for inactive timestamps). - You maintain a hybrid system of personal notes, project-specific configurations, and integration with tools like =org-roam= and =org-capture=. - *Automation & Productivity*: - Scripts and configurations (e.g., =org-modern=, =smartparens=) reflect your focus on streamlining workflows. You often debug and refactor your Emacs setup to improve efficiency. --- *** *2. Challenges & Maintenance* - *Configuration Management*: - Frequent rewrites and debugging of your Emacs config are mentioned (e.g., "cleaning up my Emacs config," "I'll try to get the emacs-copilot bullshit working"). This indicates a iterative, hands-on approach to customization. - Issues with plugin compatibility (e.g., "Doom Emacs is consistently slow") and dependency conflicts (e.g., "themes breaking things") highlight the complexity of your setup. --- *** *3. Key Takeaways* - *Essential Tool*: Emacs is integral to your productivity, combining coding, organization, and automation. - *Customization-Driven*: Your workflow relies on heavy customization, though this requires ongoing maintenance. - *Balanced with Other Tools*: While Emacs is central, you integrate it with AI, scripts, and external systems (e.g., =org-roam=, =mu4e= for email). --- Would you like to refine this summary or explore specific aspects (e.g., Emacs and AI integration, task management workflows)?
     Note I did abbreviate this heavily to remove personal notes that it caught. Is it perfect? By no means. But in a minute or so it got all the major points.
     
     I'd say it's hella useful, even if it's crude and jank.

   • Embed this notice
     Kerr Avonsen (she/her) (kerravonsen@mastodon.au)'s status on Saturday, 17-May-2025 02:45:05 JST Kerr Avonsen (she/her)

     in reply to

     • Phil

     @phil @maxleibman
     I'd be interested to know what these specific tasks are.

   • Embed this notice
     Phil (phil@fed.bajsicki.com)'s status on Saturday, 17-May-2025 02:45:06 JST Phil

     in reply to

     @maxleibman@beige.party Cool flex. Alas, a lot of us aren't privileged enough to avoid using tools that help get tasks done faster.
     
     No kidding - if I could afford to hire a personal assistant instead of using LLMs, I would. These things aren't even remotely smart.
     
     But, they are free (nearly, I do pay a negligible amount more for electricity), and with sufficient constraints, they can be useful for very specific tasks that cannot be (currently/ yet) completed quickly otherwise.
     
     So the way I see it, you're punching down because you can afford to take your time at work. This isn't the case for most people, believe it or not.

   • Embed this notice
     Russell Garner (rgarner@mastodon.social)'s status on Saturday, 17-May-2025 02:45:27 JST Russell Garner

     in reply to

     • Kerr Avonsen (she/her)
     • Phil

     @phil @kerravonsen @maxleibman the sad thing is, AI means "terrible, must be destroyed" to many people and "positive marketing term" to many others. We were turned down because the author *told* them it was "AI", with no distinction between neural networks and GenAI used or even comprehended by either party, and the publishers (quite rightly) had had a policy meeting and decided they were allergic to "AI" in general, rather than GenAI...

     Matthew Lyon repeated this.
   • Embed this notice
     Phil (phil@fed.bajsicki.com)'s status on Saturday, 17-May-2025 02:45:27 JST Phil

     in reply to

     • Kerr Avonsen (she/her)
     • Russell Garner

     @rgarner@mastodon.social @kerravonsen@mastodon.au @maxleibman@beige.party Shame.
     
     I think, in general, it's important that we stop using "AI" as a term for anything at all. Diffusion models aren't even in the same realm as LLMs, which in turn aren't anywhere close to ML for biochem. But it's all "AI" when it comes to advertising.
     
     Very disappointing.

   • Embed this notice
     Russell Garner (rgarner@mastodon.social)'s status on Saturday, 17-May-2025 02:45:28 JST Russell Garner

     in reply to

     • Kerr Avonsen (she/her)
     • Phil

     @kerravonsen @phil @maxleibman it was just outside standard printer low-resolution bounds. There was a well-photographed but low-res non-losslessly compressed TIFF with much hidden in shadow detail. Order of magnitude, let's say it was 2K pixels high, and to meet the DPI requirements (implied, not explicit!) I upscaled that by three times to give a 6K high image (at roughly A4 proportions) and re-toned it by hand, embedding an AdobeRGB profile. It was a medieval coloured sketch from Italy.

   • Embed this notice
     Russell Garner (rgarner@mastodon.social)'s status on Saturday, 17-May-2025 02:45:28 JST Russell Garner

     in reply to

     • Kerr Avonsen (she/her)
     • Phil

     @kerravonsen @phil @maxleibman turns out I screenshotted the difference at the time (this is a crop from it, maybe about a sixth or an eighth of the whole image)

   • Embed this notice
     Phil (phil@fed.bajsicki.com)'s status on Saturday, 17-May-2025 02:45:28 JST Phil

     in reply to

     • Kerr Avonsen (she/her)
     • Russell Garner

     @rgarner@mastodon.social @kerravonsen@mastodon.au @maxleibman@beige.party Wow that's a huge difference. Bet it shows more than this in print.

   • Embed this notice
     Kerr Avonsen (she/her) (kerravonsen@mastodon.au)'s status on Saturday, 17-May-2025 02:45:30 JST Kerr Avonsen (she/her)

     in reply to

     • Russell Garner
     • Phil

     @rgarner @phil @maxleibman
     Hmmm, what kind of resolution are we talking about? How many megapixels? Just curious.

   • Embed this notice
     ghost (ghost@mastodon.nz)'s status on Saturday, 17-May-2025 04:22:06 JST ghost

     in reply to

     • YurkshireLad

     @YurkshireLad @maxleibman It’s the mediocrity at scale where the AI value add sits. Never before in the history of humanity have so few been able to be so mediocre to so many. 🫡

   • Embed this notice
     YurkshireLad (yurkshirelad@mastodon.social)'s status on Saturday, 17-May-2025 04:22:07 JST YurkshireLad

     in reply to

     @maxleibman I don't need "AI" to produce mediocrity. I can do that all by myself.

     Rich Felker repeated this.