2026W10
- Nobody Gets Promoted for Simplicity — The Engineer A vs Engineer B comparison is painfully recognizable. A ships a 50-line solution in days, B builds an event-driven architecture over three weeks, and B gets the promotion because “designed scalable architecture” reads better in a review doc than “solved it simply.” The line that stuck: “anyone can add complexity. It takes experience and confidence to leave it out.” I’ve been on both sides of this — writing the simple solution and feeling like I had to apologize for it, and reviewing someone else’s overengineered system while thinking “this didn’t need to exist.” The interview culture point is sharp too: system design rounds actively punish straightforward answers by pushing “what about ten million users?” until you cave.
- Don’t become an Engineering Manager — Zaides used to argue that management experience was universally valuable. Now he’s reversed that position, and the reasons are hard to dismiss: Amazon increased its IC-to-manager ratio by 15%, Staff engineers out-earn EMs across the industry, and the pace of technical change makes stepping away from code genuinely risky. The detail about his friend getting Staff offers paying 20-30% more than his internal EM promotion says a lot about how the market has shifted. What I respect is that Zaides stays an EM anyway because he enjoys it — he’s not telling you to optimize for comp, he’s telling you to stop assuming the management track is the default next step.
- The stranger secret: how to talk to anyone – and why you should — The opening anecdote is what stays with me: an elderly woman on a nearly empty train asks to sit nearby, and they end up talking for 50 minutes. The author senses, without being told, that this woman is probably heading back to an empty house and needs to process her day out loud. Sometimes the most useful thing you can do is just not say no to proximity. Remote work has made me worse at this — I’ve optimized so hard for uninterrupted focus that I’ve forgotten what it’s like to let a random conversation happen.
- LLMs Are Good at SQL. We Gave Ours Terabytes of CI Logs. — The counterintuitive move here is denormalization: they stamp 48 metadata columns onto every single log line instead of normalizing into separate tables. It sounds wasteful until you realize ClickHouse compresses repeated values so aggressively that
commit_messagehits a 301:1 compression ratio. The agent behavior patterns are interesting too — across 8,534 sessions, they found agents don’t run one clever query, they investigate iteratively like a human would, starting broad then drilling in. Average investigation: 4.4 queries. The freshness point is the real takeaway: “did I break this, or was it already broken?” is the question that actually matters, and you can only answer it with current data. - Leaving Google has actively improved my life — The most telling detail isn’t about Google’s products — it’s about distribution. Google pays Apple $20B annually to stay the default on iOS, and Chrome holds ~70% browser market share. People don’t choose Google, they just don’t un-choose it. The author’s point about search is the one I keep coming back to: switching to alternatives turned “Googling” back into “surfing the web,” reconnecting with actual discovery instead of algorithm-filtered answers. I’ve made similar switches and the difference is real, though I’ll admit YouTube is the one I can’t quit either.
- The Hunt for Dark Breakfast — This is the kind of thing the internet was made for. The author maps breakfast foods onto a simplex using milk, eggs, and flour ratios, identifies known clusters (the Pancake Local Group, the Egg Singularity), and discovers a gap — a theoretical breakfast that should exist but doesn’t. The breakthrough: IHOP adds pancake batter to omelettes, which means the “Dark Breakfast Abyss” is occupiable. There’s even a recipe: ¼ cup milk, 4 eggs, ½ cup flour, instructions unknown. Ending the whole thing with a Lovecraft quote about humanity’s inability to correlate all its contents is the perfect touch.
- The path to ubiquitous AI — The ENIAC parallel is the framing that makes the technical claims click: room-sized and impractical became ubiquitous through specialization, and Taalas is betting the same applies to inference. Hard-wiring a single model (Llama 3.1 8B) into custom silicon to get 17K tokens/second at 20x lower cost and 10x lower power is a bold trade-off — you lose flexibility entirely but gain numbers that would change what’s economically viable. The aggressive quantization (3-bit and 6-bit) introduces quality degradation, so the real question is whether “good enough but everywhere” beats “better but expensive.” They spent $30M of $200M raised with a 24-person team, which at least shows discipline.