The state of agent permissions is a mess. I can set MCP permissions for claude web, but those can't carry over to claude code unless I go back to full manual approval of everything. And ChatGPT won't let me connect Google drive at all unless I give full access.

A second brain only helps with understanding the first brain. If you're trying to do something in the real world, you need a second real world. That's the "digital twin" concept Palantir and other ontology systems realized.

I'm liking Muse a lot. It just gets stuff done and I don't have to worry about MCPs or connectors. I can do all that, but agents won't change the world if people need to understand MCP servers to use them.

The Axiom Reason ethos demands we update our strategy when the load-bearing premises change. Our new strategy enables us to evolve Axiom without the constraints that come from it being the primary revenue generator.

This was me until I met my wife. I stayed wrong about so much because I was better at arguing. It isn't purely about being smart though, because many smart people recognize it's not worth arguing and let people be wrong. It takes an intelligent person who is willing to fight your epistemic firewalls because they care enough about you to not let you stay wrong.

The first one I found out about was anchoring bias. Nothing else comes close in importance.

We've not been reading the code long before LLMs. No one on web search at Google understands the entire code base. We operated probabilistically, had robust deploy procedures, and tested everything multiple ways.

This is directionally correct.

Running Qwen image 2.1 locally is a lot of fun. Something feels a lot more satisfying about crafting the perfect image when I own the whole stack.

I love ntm with beads because I can so easily take advantage of reset announcements to chew through some of the backlog in my beads graph.

Okay so we spent 6 months building a higher order ontology (what we call a metaontology) on top of Lean 4. But what we built is NOT what people need. The debate has been settled: proofs are programs, programs are proofs. You can't escape a graph because it's always a graph at one layer of analysis. That's NOT what needs solving. The substrate doesn't actually matter much. The hard part is getting people to articulate, and in many cases extricate, the core ontological concepts they actually use to guide their actions—the most important of which formalize into dreadfully simple structures. Until that happens, formal ontology will remain a dreary planning and optimization engine grinding away to shave $0.00003 off a bottle cap's material cost.

Running nine omp agents locally on Qwen3.8-27B with Opus 5.5 coordinating. I'm using ntm to let Opus read and send keystrokes to the worker panes.

This is such a smart idea. And it validates what a lot of us have been seeing: If you let an AI have too much freedom to manage its own harness, it devolves into slop or ineffective token churn.