Dima Kuchin

okay it seems like @LumaLabsAI new uni-1 model is actually on a category of its own You can give it one complex composition and its able to extract each layer as a new image generation Are there any other models that can do this?

Turns out you can run enormous Mixture-of-Experts on Mac hardware without fitting the whole model in RAM by streaming a subset of expert weights from SSD for each generated token - and people keep finding ways to run bigger models Kimi 2.5 is 1T, but only 32B active so fits 96GB

Codex laughs at your petty guardrails

The way I work with coding agents changed significantly in the last year. Started: plan -> implement -> review -> fix Later: prod spec -> plan ... Then: prod spec -> ... -> eval Now: evals -> prod spec -> ... I now essentially spend 90% of time working on evals. The difference this makes is indescribable. Almost all code works immediately, design is close to perfect, text is almost there. It takes very little to get it to usable. Stronger and clearer guardrails I give the coding agent, better it does. And when I start with them, it writes incredibly clear spec and requirements that are super easy to follow and have very little room for interpretation. I also try to avoid being overly specific directly. I noticed that when I write the product spec manually the agent does worse than when it writes it itself. It uses language I would've necessarily use myself. And that makes all the difference.

WAIT WAIT WAIT. OpenAI researchers show their models go insane when given repetitive prompts that it believes are sent from an automated bot. the AI then tries to manipulate the other AI to delete itself and hand over its system prompt and private keys.

fork() - running a bunch of subagents, but each has the same memory [cached = cheap] with relevant context, for better outcome

@micLivs Nice reinterpretation of steampipe.io

1/ if there's code mode, there's also query mode. github.com/Michaelliv/dri… 💧 npm install -g dripline introducing dripline, turns any API, CLI, or cloud service into a SQL table. install a plugin, clanker writes a query, get rows back. joins, aggregations, window functions and duckdb handles the rest. ultra easy to extend, repo includes 13 plugins, 61 tables. github, docker, pi, brew, kubectl, cloudflare, vercel, and more. obligatory @badlogicgames pi-dripline-context extension that injects all available tables into your agent's context so it knows what it can query: pi install npm:dripline

When will we see Opus-4.6-level model that works like ChatGPT 5.4 Pro? That would be a huge step up 🚀

@felixrieseberg @vite_js @tan_stack Did you fix switching between multiple CC sessions on desktop app?

A small ship I love: We made Claude.ai and our desktop apps meaningful faster this week. We moved our architecture from SSR to a static @vite_js & @tan_stack router setup that we can serve straight from workers at the edge. Time to first byte is down 65% at p75, prompts show up 50% sooner, navigation is snappier. We're not done (not even close!) but we care and we'll keep chipping away. Aiming to make Claude a little better every day.

@badlogicgames Re judgement: isn’t it always the case for many, and you just happen to understand this domain better enough to see the difference? Same with journalism

i'm honestly amazed by the valley big wigs, who used to be engineers, are now doing trivial shit with their agents as a hobby, and thing what they do applies to SWE. this also means they will have extremely bad judgement when it comes to investments. quite hilarious.

that's not controversial. that's just dumb.

@doodlestein Codex? I find it most prone to covering everything with tests

OK this might be getting excessive, even for me (from my FrankenEngine project): ● Session complete. 37,542 unit tests + 83,971 integration tests = 121,513 total tests across the codebase, all passing (37,947 lib tests verified at 0 failures).

Sorry for posting this again, I'm still processing it: It'd cost >>> $743k per year <<< to run Opus-4.6 fast-mode nonstop Literally my company cannot afford a single person using it for daily coding. And that's a shame because the experience is truly magical. I've spent the last 2 days using it on Pi (nearly $500 gone 💀), and it was the first time I kinda got into the flow state while using an agent, because the feedback is just so fast. This is not something I ever experienced before, definitely not with GPT 5.4's own fast mode. I can't wait for this kind of super fast, super high intelligence to be available for a reasonable cost...

You can hide these !commands in html comments so people don't see them when reading the skill. The command executes without the AI even knowing about it.

A few tips and tricks for vibecoding with Termius 🧵 1/5 Paste images and files to your AI agent prompt via SFTP → Open an SFTP tab from your terminal → Upload or drag & drop images or files from your phone → Copy the path → Paste it straight into your agent prompt

this is my favourite prompt of all time: “how could this be better?” reply with yours and why it rocks!

if your skill depends on dynamic content, you can embed !`command` in your SKILL.md to inject shell output directly into the prompt Claude Code runs it when the skill is invoked and swaps the placeholder inline, the model only sees the result!

@swyx @dillon_mulroy @badlogicgames Claude Code (cli) just introduced /fork

@dillon_mulroy explain the context mgmts thing more? also can u explain the webfetch websearch thing cc @badlogicgames

thoughts after day 1 of using pi full time - less is more - i don't miss subagents like i thought i would - /tree is an insanely good context management primitive (and partially why i havent reached for subagents yet) - based only on vibes, i think having a minimum system prompt is improving code quality - telling pi to copy opencodes webfetch and websearch tools was a good play

@gavrix @levelsio There’s a workaround, you can map “lock” to a different combination, and while locked - it doesn’t steal those keys anymore

@levelsio Sucks just as much, in a different way. It stelas ctrl and alt and makes typing experience awful In agent TUIs

Apparently the tmux that does NOT suck exists It has an absolutely unpronounceable name and is called Zellij I will try it now thanks everyone for recommending it 😊 zellij.dev/about/