@sudoingX I switched to Forgejo running on my homelab and felt more unchained than I have in years, wild how big a shift it unlocked for me
English
Tobias Davis
1.6K posts
Tobias Davis
@saibotsivad
Bespoke Software Artisan & System Architect. DMs always open 👀 #javascript #security #serverless
Midwest USA Katılım Eylül 2012
151 Takip Edilen216 Takipçiler
if your whole dev stack still lives on microsoft's servers in 2026, you don't own your work, you're a tenant. and you're one pricing change away from finding out.
github proved it this year, they floated per-minute charges on self hosted runners, then walked it back after the backlash. but they floated it. the meter is always one quarter away, and your code, your ci, your issues, your releases all sit on a platform whose terms you don't write.
forgejo is the way out. the community run fork of gitea, nonprofit governed, and it moves the whole stack, git, ci, issues, packages, onto your own metal. nobody's terms but yours.
if github locked you out tonight, could you still ship tomorrow? if not, you don't own your stack, you rent it. build on rented land if you want, plenty do, just don't be shocked when the landlord changes the locks.
Sudo su@sudoingX
here is how i'm using my dgx spark apart from running local ai models. i quit using github actions. my ci runs on self hosted runners on the dgx spark now. the whole pipeline on my own metal for free, no per minute meter from microsoft. the stuff you reflexively push to the cloud, all of it in local now. the hardware is capable, it is already paid for, and it was only ever doing one job. this is the part of the local movement nobody talks about. it was never only about running the model locally. it is about owning the compute, and once you own the metal, you stop renting everything you built on top of it. if you have the hardware and the will, you can walk away from most of your monthly bills. i'm doing it one runner at a time.
English
give me a limited tool I can own over one that "mysteriously" gets nerfed every time the more expensive upgrade arrives (yes this is about local llm)
English
@puruvjdev I haven't actually used it yet, but if I understand correctly @radicle is aiming at that and looks very promising. Sometime this year I'm hoping to test it out.
English
imagine if every single thing about a repo on github was a project of .git. you could nuke gihtub and have it rebuilt the same way next day
English
@awesomekling Most of the Studio Ghibli movies are very heartwarming and family friendly, we love them. Make sure you're emotionally steady before watching Grave of the Fireflies, it's good but sad.
English
saw Spirited Away for the first time last night and I really liked the radish spirit. need to check out more films like this :)
English
@nickreese I just watched a video of @mattpocockuk using his "grill me" skill which sounds very much like what you're doing. Haven't tried it yet but the idea seems solid. github.com/mattpocock/ski…
English
As AI gets smarter I'm noticing that I'm prompting less and doing a lot more steering and often that is harder to manage than a long well thought out document... but producing those well thought out documents is the bottleneck.
Recently I've been having AI "interview me" of what my goals are for the thing that is going to be built, then I dictate it and have it transcribe it and log it verbatium to what I said.
Incredibly this starts snowballing as it will find related concepts and how things link together.
Really going to stick with this process. Blown away by the results in just the past couple days.
I am WAAAY MORE verbose in audio than typing, so it is helpful and I explore tangents more often which often help identify possible sharp edges of my initial ideas around implementation.
English
@awesomekling @valigo This really bothers me, I've noticed language shift in authors who I'm like 80% sure aren't writing with AI but I know they read lots of AI, the voice is shifting towards AI 😢
English
@valigo Even if we don't use LLMs for writing, we're still reading more and more LLM output every day! The influence is unavoidable
English
Once upon a time I was writing some kinda public post for the WebKit project at Apple, and I thought I'd be a little funny and throw in some light-hearted jokes.
I forget what the jokes were (100% sure they were cringe by today's standards) but I remember asking Darin Adler to review the post and he said it was nice I was writing in my own voice, and that I don't need to change it just to sound professional.
Really made an impression on me. As all of our writing styles now converge in the LLM age, what will remain of my voice? Of yours?
English
@mjasay Same as cameras were: people going wild about 1.5 megapixels being so much better than 1.2, but eventually it slowed down and 18mp could be better than 20 if the software was right. I think that's what is next. Hoping.
English
This is *exhausting*. Watching folks trumpet how great 5.6 is this week (a few weeks ago it was Claude Sonnet 5, and in a few more weeks it just might be a Gemini model), it's unclear to mere mortals what one gets by moving to a new model (or, rather, what you give up by using an earlier model). When I use ChatGPT or Claude to help with a project, I never have a clue what the different versions of models represent (fewer tokens? But at what cost in utility? What's the tradeoff?).
I suspect I'm the norm, the not exception.
The model providers are in a drag race and, in general, LLMs keep getting better/more capable. But if I'm an enterprise, it's bewildering to know which to use, where, and for how long. The whole AI arms race has an air of near-immediate obsolescence. That makes it hard to depend upon.
Gergely Orosz@GergelyOrosz
Its not even a question that Google / Gemini is no longer in the top 3 coding model providers Which is surprising given the spend and resources Google has, and how they seemed to want to play to win Claude, GPT, and probably Grok and GLM seem to be the top coding LLMs
English
@ShaanVP Definitely need to connect with @ProfCalNewport he talks about this idea quite a lot (at least ~recently) in his podcast, trying to play it out a bit more to see what might the future look like. I haven't done daily writing yet but it's on my list!
English
Daily writing is the new weightlifting.
Back in the day, 90%+ of the population worked manual labor on farms. We didn’t have “gyms” because everyone was breaking their back working all day.
Once we got machines, we needed a new way to keep our body in shape - lifting heavy objects (3 sets of 10), voluntarily, at the gym.
The same thing is happening for writing. Everyone is outsourcing their thinking to AI.
The brain is a muscle, and it will atrophy if you don’t use it.
English
@inkdrop_app This is an interesting idea. A well written issue is better than a PR, but I wonder if a "half baked" PR (more like a sketch than a fully working change) would be useful? I've made them at work and they are helpful. Good thought to chew on!
English
The cost of sending PRs is nearly zero now. I don't feel like reviewing them (I hate AI's redundant inline comments). Instead, I'd only take the ideas and implement them myself if they sound good. PRs are now more about sharing ideas than contributing hard work.
English
@sudoingX What do you use Cursor for if you're all-in on local?
English
you rent intelligence for $20 a month, forever, and every prompt is a window into your mind that opens on someone else's servers.
the guy with a used $900 card breaks even at month 45, rides free for years, and his thoughts never leave the room. renting only looks cheap because nobody counts forever.
> the card i mean: a used rtx 3090, 900 to 1200 usd, 24gb of vram.
> the model: qwen 3.6 27b dense at q4, the current king
of a single 24gb card, 35 tok/s flat no matter how deep
the context runs. chats, codes, runs agents, one shotted a full game build on my bench.
fully offline, no dario or altman snooping on your prompts.
English
@sudoingX Where does cursor fit in the local LLM paradigm? I'm using local as my main starting point, Claude for dev work, haven't explored the local-drives-Claude idea yet
English
people keep asking what i use for software development.
anon. if you're serious about building software, or about agentic coding at all, the answer is cursor. use it and thank me later.
claude code and codex still feel experimental to me. i've said that before and i'll say it again. the cursor is incomparable. that's the entire reason.
English
Graphed my AI spend since 2024
lately it's mostly Cursor (main) + Claude, with a bit of OpenAI for ChatGPT and Gemini for specialized OCR
English
@nickreese Ugh it's absolutely terrible. I've often wondered if I could pipe LLM output through a secondary transformer to strip out the word vomit 🤮
English
GLM 5.2: "That is a load-bearing observation."
I absolutely despise this language.
English
@wesbos I was thinking of this yesterday 🤯 you know how you look at search queries to your site to help with SEO? (been out the game a while, may not be applicable now?) I was thinking that about OSS agent searches, it could help me shore up docs, if I could get summaries
English
I'm working on a paper "Autonomous Skill Acquisition in Adversarial Networks" to show a better path forward for AI Agents. Very inspired by the GEPA paper (2507.19457) and some original research on memory categorization across disciplines.
English
@danielkhunter @whoiskatrin I built a chat tree app as a TUI with "blessed" because I didn't know Tauri enough 😞 is there a good template to start with?
English
English
Love this post from @jeremy_daly on how prompt engineering has quickly evolved to become systems design. datacamp.com/blog/context-e…
English
@lucasmeijer It doesn't feel great, TBH, like there are surely going to be better tools (skill / memory recall) within a year, I would think.
English
@lucasmeijer Then it deletes planning docs and other task docs, capturing all lessons into ADR. Then I manually review and curate (often fully delete) the final ADR. Then feature branch merges into main nice and clean, but I had the benefit of the extra docs while it was going.
English
I'm slowly migrating from "its a good idea to have lots of .md files for plans/bugs/tasks in the repo" to "yeah lets not do that at all".
English