zlumer.eth

@thdxr imagine if you could prompt with a dance instead of speaking

some of you have asked for more examples of how i voice prompt here's an example - you can see how it's rambling, not very coherent and yet it has 10x more information than what i would be willing to type out

@zeeg @stalmico what you're saying is that we will set up (llm-based) pipelines to extract important events from session replays instead of frontend requests which sounds undoubtedly cool but tbh is pretty close to event based analytics

@zlumer @stalmico in a few years you'll see what im talking about. you dont need to instrument these event-driven analytics anymore as llms can pattern extract on top of complex heavier data sources like replays, and sampling should become enough

Rent free baby Actually what you do is hold a lifelong grudge and use every ounce of power you have to make them regret their decisions

@zeeg @stalmico maybe for b2b you can sometimes get away with just session replays or even speaking directly to your customers but b2c without event-based analytics is driving the mountain bike with your eyes closed

@stalmico We follow the law of focus otherwise I’d have already shipped product analytics Not out of the question in the future tho… surprisingly you dont need analytics most of the time when you have traces and replays and well functioning agents…

@robinhanson it's really much more simple than that. people living paycheck to paycheck can't afford kids (or at least believe they can't afford kids) people living paycheck to paycheck go where higher-paid jobs are -> jobs are in big cities -> people who can't afford kids go to big cities

I agree that this is a striking patten.

@KentonVarda interesting! I've been sending every prompt in the same harness to different models at once and noticed claude to be the worst offender here: almost 100% of prompts start with a subagent gemini launches subagents maybe 10% of the time this is non-scientific obv

@zlumer This was GPT 5.5 (in OpenCode).

My agents are always spawning subagents and telling them to do exactly the thing I told the main agent to do, except with a less-clear prompt. I feel like this isn't helping.

Some people see a racehorse running in a field, and they want to strap plow to it. They think an amazing racehorse is going to be an amazing plow horse. Then they are angry when the racehorse snaps its legs pulling a heavy plow behind it. They send the horse to the butcher and take the loss on the chin. "Horses are so difficult. It's impossible to find a good one!"

@thdxr Couldn't be more true. Previously you needed like 3-5 juniors destroying the codebase for years to make it rot, now you just let an agentic bro ship slop for a month and the code is irrecoverable

1 year old codebase is the new 5 year old codebase

@FUCORY it actually made me more curious about smithers

Choosing React for smithers was a difficult choice. I knew choosing it would make people go "wtf" and never give it a chance. It also makes it look a bit more like a gimmick rather than new shiny tech What makes me confident I made the right choice is how strong llms are at using smithers. Especially with our new docs llms one shot complex smithers workflows so well it make my head spin Drafting off of patterns the llm already has great training data on like file systems, clis, and frameworks like React is the best way to get great LLM-experience out the box

@antirez fast mobile typing requires swipe-keyboards and their UX negatively correlates with size. you want the keyboard to be as small as possible to not move your finger too much

Important features: 1. It has *two* screens, it is not a foldable screen. So there is no problem with folding angle and when closed it will have the two screens facing in the opposite direction. 2. The aspect ratio is different, it is a taller phone, so when open you can type with two hands. 3. The software and keyboard are designed for this use case. It has a native SSH client, a Linux subsystem, ...

I want something like that (made with AI image generation, not a real thing).

@sanxiaozhizi tsx is 25% more efficient tho it's 3 letters "tsx" vs 4 letters "node" — imagine the time saved over years

Node.js 20 is EOL. Say goodbye to: - tinyglobby: use `fs.glob` instead - picomatch: use `path.matchesGlob` instead - tsx: run TS files directly with type stripping only

@zeddotdev I'd really love to use it but gh issue 46320 is blocking 100% of my work

We launched Zed 1.0 this week! 🎉🎊🥳 As per usual with a new version, we shipped goodies: Stop using breakpoints as makeshift bookmarks... because bookmarks are here! 🔖 Right click on the gutter to add a bookmark and see them all with `editor: view bookmarks`. Thanks austincummings!

@rough__sea I find it frustrating that non-js apps abuse the js registry as a free cdn

Ironic coming from me - but I find it frustrating to install Node.js just to run `npm i -g @openai/codex` - codex is completely programmed in rust

@vkrajacic that's assuming that faster coding leads to better products, which is easily provable to be if not completely false then at least non-linearly correlated

X is flooded with people advocating for AI and vibe coding, but I see almost no cool demos or anything I can download to improve my daily work. I don't care how many tokens you burn, how many LOC you generate, which IDE you use. Can I use your application, please?

being a founder is reaching status lows i have never before seen my entire time in tech

agents in arcnext.dev can now drive the web tabs - open, click, type, scroll on any page - one shared space with you - sky-blue glow to know where agents works here's claude filling my yc s26 application

reasonable moderate take running agents in parallel is often performative but not always, sometimes it's genuinly useful

@FUCORY I still use it today whenever I need a surgical code editing tool (as opposed to an agentic general task solver)

@zlumer I mostly used aider throughout 2024

Fucory harness ranking update S-Tier Amp, Pi, Smithers A-Tier Conductor, OpenCode TUI, Warp B-Tier Codex, Antigravity, T3 Code C-Tier Claude Code, Gemini CLI, Cursor, Cline, Opencode GUI D-Tier Avante, Copilot, Windsurf, Devin F-Tier Amazon Q, Claude Desktop, Aider Biggest winner: Opencode for fixing it's agent perf a bit and competitors getting worse Biggest loser: Claude Code TUI for becoming unreliable and performing worse than claude in other harnesses

@kylebrussell what if your code is a landing page website tho? surely AI could write 75% of it even two years ago

I think this is false memory? The code coming out of the tools was not very good two years ago