Weekly notes

• How to Scale Your Impact at the Staff-Plus Level: Things to avoid and get better at in here. Delegation, cross team communication, and making something nebulous and ambiguous more concrete. Emphasis on helping to identify points of leverage for the team and organization becomes important here too.

Quote post

I do not have confidence in “our technology is totally profitable on any reasonable timescale” when the answer to “how do we afford it” is “we turn the solar system into computronium and use the entire sun as an energy source for it” even if it were possible without breaking the laws of physics. - Liz Fong-Jones

A corollary is that all of the over-hyped/speculative use cases of GenAI are not going to take off, because the price is going to have to rise. only the cases that it’s worth spending thousands of dollars a month for are going to survive. making crappy derivative art? not one of those use cases.

(The things that already are possible with a local LLM are probably fine as is, those have a reasonable power budget. but in general, what I said stands about LLMs being a tool for trained professionals only, not for consumer use)

Nice little thread about value vs promise vs cost of llms. It’s not clear yet that we’re going to be able to get out of the current llm craze what certain people have put in.

• AWS in 2025: The Stuff You Think You Know That’s Now Wrong: Stuff that has changed about aws over the years. Nice list.
• Under the hood: how AWS Lambda SnapStart optimizes function startup latency: A snapshot is taken of a lambda function in a way that makes it useful as a starting point for future invocations that make the bootstrap process more efficient.
• Easy will always trump simple: Easy wins because simple is hard to reason about and implement sometimes.
• Can LLMs replace on call SREs today?: Clickhouse runs a test using well known failure scenarios in a demo application to determine how well various llms will be able to figure out what’s going on. At this point llms are a replacement for people.
• Cloudflare incident on August 21, 2025: We saw this in our pre-prod environments the other day. Traffic management controls and fairness in multitenant systems usually come in after you have a successful product as far as I have seen in my work life. Often takes a moment like this to get visibility on the need for a project like this. (Non-functional requirements ftw!)
• Dynamo, DynamoDB, and Aurora DSQL: Some great tech built by amazon over the years between dynamo (voldemort?), dynamodb, and dsql.
• Designing Software in the Large: Notes from the book A philosophy of software design. I also enjoyed this book a lot. Concise with many good ideas to think about.