panphora

panphora

@panphora.bsky.social

working on malleable files, the perfect software primitive: https://hyperclay.com

262 followers655 following505 posts

there's nothing more ironic than frontier AI CEOs writing about how AI might kill us all when they know that's going to go in the training data

Just booted Linux 6.1, emulated entirely in a browser tab! 😳 cat /proc/cpuinfo shows isa: rv32ima — no Floating Point Unit. Try awk 'BEGIN {print 1.5}' and Linux kernel segfaults 💥 Incredible project by @larkworthy.bsky.social

tom larkworthy@larkworthy.bsky.social

Run Linux in a file? A 6.8Mb HTML self-serializing literate notebook file that emulates Risk V and can boot an interactive Linux BusyBox shell. Built differently (but heavily inspired by) Fabrice Bellard's JSlinux. tomlarkworthy.github.io/lopebooks/no...

For pro coders: raise effort, compact earlier, and run in simple mode. { "showThinkingSummaries": true, "alwaysThinkingEnabled": true } export CLAUDE_CODE_EFFORT_LEVEL="max" export CLAUDE_CODE_AUTO_COMPACT_WINDOW=300000 export CLAUDE_CODE_SIMPLE=1 # Strips CLAUDE(.)md, hooks, skills, etc.

Hyperclay just got better. Went from allowing a very limited number of malleable HTML files per user (10-20) to allowing unlimited files, all by switching to username-namespaced file paths. Everything's more intuitive and Hyperclay Local syncing feels so much better! Solves ∞ UX problems

Someone should freeze a copy of the internet right now and pass it down to future generations as one of the few digital ecosystems not completely overrun by AI

Shipped a big Hyperclay Local update! 🌹 It's like Dropbox for local malleable HTML files. • It's just a tray popover now • Quick sync status • Recent uploads/downloads • Tightened up the sync engine Desktop app feels way more focused now.

Chrome DevTools has a built-in filesystem proxy: you can locally override any HTML/JS/CSS on any site. Incredibly powerful. 🦸 But I just got burned. 🙈 Once enabled, it silently shadows any local file you open in Chrome without asking.

"Every generation mistakes its current map for the territory." — ChatGPT I love having deep philosophical conversations with an LLM at 6 AM in the morning. Truly fun.

Remote state should feel like local state. Mjölnir.js: what if db.users.find(str) was just frontend JS, whether you’re hacking locally or running search across a distributed fleet? Time for browsers to let web citizens make the web.

Just did my product video with remotion! 🎥 My process: have Claude generate each scene as a single HTML file or static image, iterate with Claude until it's A+, then ask to convert it to React/Remotion.

The best AI tools get out of the way. That's why my local essay revision tool only shows exactly what you need: 1) The feedback 2) The context 3) What the AI understands 4) The proposed fix Glance, approve, continue. #buildinpublic www.youtube.com/watch?v=V6M...

youtube.comThe UI Claude Should Have for Editing Your WritingI got tired of copying feedback into chat and losing track of what was done. Revise structures the whole process: see the passage, see the feedback, and see ...

anyone else looking at all these banger vibe coded projects on twitter and getting discouraged about their business? like: why am i building this big platform when someone could produce the same thing, custom to their use case, in a vibe session?

Claude Opus 4.5 is now the default in Claude Code. Here’s my take on the best coding model right now, and what I think is actually happening behind the scenes at Anthropic. 1/10 🧵 👇

View full thread

Is Opus 4.5 bad? No. Is it Opus 4.1’s replacement for deep coding work? No. It feels like I: • Lost my genius architect (4.1) • Got an excellent foreman (4.5) • Plus great subcontractors (Gemini, Codex) Useful, but not the same flow as before. 9/10

If you’ve used these models heavily for coding: • How does Opus 4.5 feel vs 4.1 for you? • Are you also bouncing between Opus, Sonnet, Gemini, etc.? • What combo is working best for your day-to-day dev work? Curious to hear other stacks. 10/10

Claude Opus 4.5 is now the default in Claude Code. Here’s my take on the best coding model right now, and what I think is actually happening behind the scenes at Anthropic. 1/10 🧵 👇

View full thread

My take on Gemini 3 Pro: • Great for one-off tasks • Can one-shot big greenfield challenges But it doesn’t have the same deep exploratory vibe 4.1 had. For me it’s weaker at slow, incremental refactors and better at one-offs. 7/10

The current stack I'd suggest: • Gemini 3 Pro → One-shot assets, SVGs, small libs, single HTML files • Opus 4.5 → big-feature planning and refactors (but you need to guide it more than 4.1) • ChatGPT (Codex) → code review, catching subtle issues the others miss 8/10

Claude Opus 4.5 is now the default in Claude Code. Here’s my take on the best coding model right now, and what I think is actually happening behind the scenes at Anthropic. 1/10 🧵 👇

View full thread

From actually coding with it, Opus 4.5 feels dumber than 4.1. Not bad, just shallow: • Digs less into tricky code • Gives up faster on hard bugs • Feels more like Sonnet 4.5 Plus than a true Opus successor If I could, I’d prefer to run 4.1. 5/10

So... I’ve lost my very smart planning partner (Opus 4.1) but gained a stronger Sonnet in Opus 4.5. That leaves me wondering: • Which AI do I use for in-depth planning? • Do I lean hard on Gemini 3 Pro? • Or juggle multiple models as a combined “brain”? 6/10

Claude Opus 4.5 is now the default in Claude Code. Here’s my take on the best coding model right now, and what I think is actually happening behind the scenes at Anthropic. 1/10 🧵 👇

View full thread

Why I think Opus 4.1 was unprofitable: • Severe Opus 4.1 caps got implemented soon after launch • They started pushing Sonnet 4.5 hard • Opus 4.5 (released yesterday) feels like a weaker, cheaper-to-run Opus My theory: 4.1 blew a hole in margins; 4.5 is the "sustainable" Opus. 3/10

Pricing backs this up: • Opus 4.1: $75/MTok out • Opus 4.5: $25/MTok out A 3x price gap even after 4.5 ships??? This suggests 4.1 is way more expensive to run. Is it worth the 3x price tag? No. But for in-depth planning and refactors, I’d put it 50% ahead of Opus 4.5. 4/10

Claude Opus 4.5 is now the default in Claude Code. Here’s my take on the best coding model right now, and what I think is actually happening behind the scenes at Anthropic. 1/10 🧵 👇

I think Anthropic saw Opus 4.1 as their last big swing at mainstream devs before others caught up. They pushed it until it became too expensive to run. My gut: it may have cost 4–5x more to serve than they charged. Pure speculation, but I think it fits. 👇 2/10

I feel so lucky to be living in the age of AI. I just produced a script in 10 min that helps me automatically release NPM packages, using a local Claude Code to write the changelog entries. Mind blown 🤯

🔥 Build a lightweight, elegant Kanban board in pure HTML. Sortable columns, draggable cards, zero React — just a single HTML file, and ~20 lines of jQuery. #hyperclay

You can import shadcn or spend 56 hours designing the perfect tab component, or you can simply accept that this minimal tab layout is superior and ship it in 6 min.

CURRENT AI STACK: • Claude Opus 4.1: Planning AND implementing complex tasks • Claude Sonnet 4.5: Implementing simple plans created by Opus • OpenAI Codex: Thorough deep-dive code reviews • Gemini 2.5 Pro: Running for hours and providing not even one line of useful feedback