eugmai86

Part of the fatigue comes from agents explaining everything in prose. When important operations happen through constrained tools, the human can review structured inputs, results and logs instead of reconstructing the entire conversation. Fatigue comes when doing same things many times day by day. What we did in our team - created MCP server that does most of routine operations automatically for us. We explicitly state to LLM that output shall be in simple words; structured in this or that way depending on the project and task. Yes architecture and decision making still on our shoulders but this is the part of the dev job nowadays

The most interesting part for me is not the top-end 192 core size, but the small/medium sizes.

Looks like m9g.large is about 9% more expensive than m8g.large on-demand, at least in Vantage’s current data.

If the “up to 25% compute improvement” holds for real web/database workloads, that is a pretty reasonable perf/$ trade. The usual caveat is that small SaaS apps are often bottlenecked on DB, storage, or network long before raw CPU.

Nice example of an algorithm that is simple enough to explain in a few lines, but still changes how you think about floating point errors.

Nice visualization. Putting the cells directly on the sphere makes the nearest-neighbor idea much easier to understand.

On a flat map I always have to ask myself how much of what I’m seeing is geometry and how much is projection distortion

Thanks!