some things that took way too long to figure out:
• LLM inference in a Web Worker so the UI stays 60fps
• hybrid RAG: BM25 + vector search → RRF fusion
• agent loop detection (catches the LLM calling the same tool 3x)
open source: https://t.co/oowKlYLl3P
i was paying for ChatGPT, Grok, and Gemini for stuff a 1.5B model handles fine. so i built this.
40+ models run on your GPU via WebGPU or plug in a free Groq key and run qwen3-32b at 229 t/s like in the screenshot. document Q&A, web search, Python sandbox agent all in the browser
Built PairLane because online coding collaboration still feels broken
Realtime collaborative engineering workspace with
collaborative coding
live cursors
collaborative drawing
shared workspaces
Built with Next.js, https://t.co/n5oxgiXIIB, Monaco, WebRTC
https://t.co/vI17W79eId
google I/O is tomorrow and they already
spoiled half of it.
here's what's actually coming:
→ Gemini Intelligence — AI that moves across
apps, understands your screen, books things
for you without switching apps
→ Gemini in Chrome with "auto browse"
→ Googlebook — new laptop hardware, Gemini at the core
→ Android XR glasses going consumer
→ Android Auto fully rebuilt around Gemini
→ Aluminium OS — Android + Chrome = one OS for laptops
→ new Gemini model dropping (GPT-5.5 class)
this isn't a chatbot update.
this is Google trying to make Gemini the OS layer.
https://t.co/zmc3yrEF6L
keynote: May 19, 1PM ET....watch it.
@sama the real variable isn't age it's context volume vs precision required.
voice wins when you have more to say than patience to type it.
typing wins when every word needs to be exact.
the age correlation is just a proxy for use case.
granola broke an a16z partner's local agent workflow when they changed how they store data.
instead of reverting, they shipped a full API.
that's the correct call.
local caches aren't contracts. if you're building agents on top of another app's undocumented internals, you assumed the risk.
the API is better infrastructure anyway.
@cjpedregal
Algorithms by Jeff Erickson - one of the best algorithm books out there.
The illustrations make complex concepts surprisingly easy to follow. Highly recommend this.
https://t.co/8G06RjGnMA
shipped something today
every AI tool agrees with you. tells you your idea is great. maybe needs some refinement. very cool very exciting
i got tired of it so i built mindfuck — it tells you where your idea breaks, what problem you're actually solving, and what to do about it. monday morning. concrete. no vibes
pip install mindfuck
(yes that's the real name yes it's on pypi...yes i checked)
@jacobeffron@swyx this. i run 1,200+ agent executions/day and
"stability" just means you've seen every failure
mode enough times to have a handler for it
@katedeyneka@ycombinator Reelful's camera-roll-to-reel flow is exactly what agentic pipelines should feel like.
👏 congrats on YC app .... rooting for the W
Name one thing that sounds fake but is actually true about building AI agents in production.
I'll go first: the hardest part isn't the AI, it's the retry logic.
Just dug into @EntelligenceAI adversarial verification loop...most AI code reviewers: flag everything suspicious → noise → engineers stop readingEllie: generate finding → actively try to disprove it → only surfaces if it survives
That's not a feature. That's a product philosophy.
Rare.
@sorenrood@runcomputing the hardest part of this role isn't the agent architecture it's that ops workflows were never designed to be observed. no logs, no states, no handoffs.
you're basically building the observability layer before you can build the automation.