Secta

This guy built a home AI supercomputer out of Mac Studios. Five machines, wired together, running as one local inference cluster through Exo. The point is not that it looks insane on a desk. The point is what it replaces. Most people rent intelligence one message at a time. ChatGPT Pro. Claude Code Max. Cursor. Transcription tools. Research tools. API credits. The stack quietly turns into a $400/month bill just to use compute sitting in someone else’s datacenter. This setup flips the model. Macs use unified memory, so the GPU can use the same memory pool as the system. Exo then splits large open-source models across multiple machines, letting the cluster run models a single computer could not handle alone. DeepSeek. GLM. Kimi. Qwen. The kind of models people usually access through cloud APIs can now sit on a desk, running privately, overnight, with no rate limits and no token anxiety. The old flex was paying for every frontier model. The new flex is owning the box that keeps thinking after you close the tab. Bookmark this and read the full setup below.

A guy makes ~$10,000/month testing security systems for major grocery chains. He walks into a supermarket with the director’s approval and blends in like any other customer. Except he is not looking at the shelves. He is looking for the small failures everyone assumes are already handled. Emergency communication. Alert workflows. Response chains. Tiny broken processes customers never see. One report. One payment. A few weeks later, the problems are gone. I built the software version of that mindset with Claude Fable 5. Not “ask one question, get one answer.” I mean giving Claude a full goal, a workspace, files, tests, and a definition of done — then letting it inspect, fix, verify, and keep going until the job is actually finished. That is the real shift. The new skill is not prompting. It is delegation: clear brief, clear finish line, enough context, and a hard stop condition. Read the full guide below.

Gavin Baker accurately predicted Micron's 10x run in 2025 Now he's making his next big call: "DRAM is probably going to be 30 to 40% of all hyperscaler CapEx next year, every hundreds of billions of dollars spent goes straight to DRAM" He calls it the single most important bottleneck in all of AI, above lasers, power chips, NAND or HDDs, Elon is even turning the TeraFab toward memory because he sees the same thing Only three firms on earth can make the memory these AI servers need "This is as close to magic as science can get" And Micron just locked in supply contracts covering half its revenue, with floor prices already above the margin peaks of past cycles, the economics of the stock are not what they were "If you still hold $MU at a commodity discount to $ASML and $LRCX, that discount is no longer earned" bookmark it, one of the strongest memory theses for 2026-2027 ↓

Fable 5 is back.