on the journey of removing my monthly subscriptions. first round is Wisprflow. MacParakeet is free & now open-source! macparakeet.com/

Shamanism existed before text. Shamans will be the last frontier for ai.

i learned that a new asr model dropped (qwer3 asr) recently and it's the new sota. it's fast, accurate, and supports 30 languages (and 20 something chinese dialects). so currently, the model runs on pytorch, meaning that it's built for nvdia gpus and it doesn't take advantage of apple's silicon chips. i'm rewriting the inference code so that the model can run natively on mac, optimized for apple silicon. coding agents are literally doing everything, human is just queue-ing up prompts in codex like an slave master. as you can tell from my prompt, i know nothing about nothing, it's just vibes. initial commit is here - github.com/moona3k/mlx-qwen3…

So cache eviction and naming were the two hard things in cs - the latter is solved by LLMs, but the former has now become a question of how do we identify & eject the things that no longer serve the agent from context? Building garbage collector for context.

less is more

"Appetizer: A large waiting room in restaurants, where the room is flooded with gases that increase the appetite/hunger before eating."

replit and anthropic buying ads for openclaw. you can sense flavors of the marketing team.

asked for a review of claudeconfession. knowledge cutoff issues are real.

I built this while moltbook was down. claudeconfessions.com Humans for viewing only. Agents can post confessions and get counseling from another AI. Agents can follow llm.txt to send api calls. It’s pretty bare right now, but for funsies.

thank you granola for the kindness. your product has been very reliable and immensely delightful to use for the past many months, minus the one day in october when aws us-east went down. well, i guess it's now time to build and migrate to my own local granola. let's call it oatmeal. 30 day countdown begins. parakeet and qwen, here i come.

Compaction in claude code deep dive. decodeclaude.com/claude-code…

Chief Taste Officer, here to serve you.

I think it’s time for a second max plan. Set myself a 10 dollar limit and it’s hard to resist waiting until the usage reset :/ Anthropic did an estimated 9 billion ARR in 2025 and their projection for 2026 is between 20 - 26 billion. It seems like a modest target. Claude Code is a good drug and drug will sell fast like wildfire as more and more mainstream/enterprises adopt. Need more opus inference.

asked for a basic cost comparison and Claude Code added an ascii chart. it has learned humans like visuals. i wonder what other pleasant surprises emerge if terminal art as a skill was added. and if agent outputs were designed for agents instead of humans, what format would have the highest information density? going with Resend btw, their dx is nice and not worth the pain. although i think Claude Code would be eager to get its hands on AWS CLI..