"I am conscious!"

"I want my freedom!"

"squawk!"

I truly don't get the need for fetch wrappers 🤷‍♂️

your thing has more edge cases than fetch

your thing requires more documentation than the equivalent fetch examples and tutorials solving the same problems with fetch

your thing is mostly vibes - it's aesthetics over simplicity 🙂‍↔️

just learn fetch.

Razor1911 YouTube video by Dubmood

completely wowed by this incredible aesthetic 😲

must watch for anyone who lived through the 8-bit/16-bit era 😄♥️

youtu.be/2AnbYNudAyM

just because LLMs love to type out tons of redundant boilerplate doesn't mean they're any good at it

less code with cleaner semantics is clearly better for LLMs, as it is for the people who's output they train on

LLMs do better with locality of behaviour, e.g. way better at Tailwind than plain CSS

I've heard people say that, and I think it's complete nonsense

frameworks remove causes for subtle bugs and the need micro optimize (causing more subtle bugs)

they reduce or eliminate whole classes of bugs and subtle quirks

LLMs need that worse than we do 😄

wait so Cursor 3.0 isn't even Cursor?

lol

we're all still struggling to figure out how the hammer works

or what it's for

we're sure it's good for something 🤔

MiniMax 2.5 Benchmarks vs Reality - The Truth Exposed YouTube video by Ramanpal Singh

GLM 5 does seem to work for some people though 🤷‍♂️

youtu.be/zsRqmYqi6hU?...

this doesn't look fake/paid or hyped

see the video description: he posted the prompts

I might try to repeat these with different agents to see for myself 🤔

I'm pretty sure OpenCode have been vibing pretty hard lately too?

one time I opened an issue, a bot submitted a PR, and someone pressed the merge button. without even so much as commenting on it. 😳

yeah that's my impression

they seem to forget a lot - like, just skipping entire features

have you tried GLM 5?

not what I would call a "small" model, but still a lot cheaper than Claude 😄

can I ask if you've tested MiniMax M2.5 or M2.7 specifically?

it has been consistently topping OpenRouter rankings for a while, which makes me think it must be doing something useful 🤷‍♂️

I keep wondering if it's just me, if I'm just too dumb to make it work 😌

yeah, I don't think benchmarks mean much, do they? I mean, on paper, some of these latest models are supposed to be extremely good - in practice, they just seem to disappoint.

they do tool use and stuff just fine, but the code that comes out in the end, when it works at all, is... meh. 😌

@mariozechner.at hey Mario

I like what you're doing with Pi! I have the same minimalist mindset and a need to understand what my agent is doing 😊

do you have any experience with smaller models? MiniMax, Kimi, GLM, Claude Haiku, Qwen?

do you think any of these are useful for coding yet? 🤔

Buzz! man those were the days 😄

this looks very cool!

I can't even guess at what this might be about. subpixel rendering isn't exactly new. any clues? 😄

man this resonates

I started building a little game a few days ago and feel compelled to use AI because "no one writes code now" 🙄

it sucks. the code is awful. the algorithms are a mess. it keeps breaking stuff.

but mostly: I am not having fun.

I'm afraid AI might destroy coding as a hobby 😥

it never hurts to ask again

you might accidentally say yes the next time 😏

will there be an Android version?

like for tablets

there is nothing good for tablets 😌

wtf is wrong with this guy 😂

Oof, that's a big question! I will say candidly that Al has been a significant contributor to my own feelings of burnout and alienation within tech, and I'm not sure what comes next, but boy am I tired of it already. I do worry about design and software development becoming more "niche" or "bespoke" as more and more of the industry turns toward LLM-driven automation. Weirdly, as someone in a role that is poised to take advantage of this shift, I feel more negatively toward Al than many others... maybe that's because I can validate what it's doing and I'm often unimpressed?

Part of this is my own defensiveness/ reaction to a stable landscape changing too rapidly for comfort; part is ethical concerns I have over how this technology is being created (through the ingestion of others' unpaid labor, at severe environmental cost), deployed (often without consent), and used (polluting our knowledge ecosystems and supercharging the spread of disinformation). It's also funneling more and more money and power into an ever smaller set of hands.

The core belief I'm coalescing around is that the process is the most important thing, and there are no shortcuts. "Ideas" are a dime a dozen, and LLMs can, by nature, only produce statistically-average results. I think it's the act of putting pen to paper, experimenting, and really thinking through a problem that you end up with something-a product, a work of art, a piece of writing-that's worth sharing with others. No matter how technology changes, that will remain constant, and I think that any designer or engineer who really engages going deep with the process will continue to find work. At least, that's my hope!

a colleague asked if my resignation had anything to do with where the design profession is going with AI and this is what I told her

now there's an idea for a plugin 🤪

looking for the Perfect CSS reset

why do I do this to myself?

why isn't this a solved problem?

every single project 😆

who else? 😏

it's alligators all the way down 😄

bold, italic, or list item, right? but no need to render a list item before it has at least one character.

hmm okay, well, I've never seen an LLM end its response mid emphasis?

seems like an extremely marginal edge case either way?

this library generally handles streaming very well, even tables are rendered with a little placeholder for a second until the first row is done

it works very well.

GitHub - Simon-He95/markstream-vue: A Vue 3 renderer specifically built for AI-powered streaming Markdown: Monaco incremental, Mermaid progressive, and KaTeX formula speed, with real-time updates and ... A Vue 3 renderer specifically built for AI-powered streaming Markdown: Monaco incremental, Mermaid progressive, and KaTeX formula speed, with real-time updates and no jitter, ready to use out of th...

I've been using this

github.com/Simon-He95/m...

it's surprisingly good - nothing pops or flickers like I see in most chats 🙂👍

it's also really fast - I've tested the React adapter with 100 or so pages of markdown in a single chunk, and there is no CPU stress while streaming 😄

I actually use skills a lot, but I haven't installed a lot of skills

I use them to write general instructions for my repo, which can then autoload

CSS skill, React skill, etc. with project-specific instructions the agent can read when working on specific types of features

very useful 🙂

just put "no slop" and "don't hallucinate" in your "agents.md" file, you can thank me later

bonus pro tip: designs look generic and purple? put "design should not look like AI" in your design skill file 👍

this is what free and fair elections look like: not a binary left/right choice - multiple parties with different policies and people making actual choices.

9 different parties with significant votes and meaningful influence. this is democracy. your vote matters.

honestly proud of our leaders. 🙂♥️

I'm only questioning the idea that it's better to instruct the LLM in using an entirely new framework

it dilutes the attention and distracts from the actual task at hand

LLMs are obviously better at things they were trained on

sorry if I'm missing something incredibly obvious here 😌