Resilience Coffee/2026/04/02

- Preparing for interviews using coding assistants - Eating schedule as a barometer for resilience - Artemis II - Is Hardware the future? - Recycling - dogfood bags - Working with LLMs when you don't have the expertise

  - Nate Jones "Models amplify confidence but not experience"
  - You own the output
  - LLMs sound similar to grad students at the edge of knowledge boundaries, unable to verify hypotheses of understanding, or mirror user input speech
  
 article quote: "In trials where the software performed suboptimally and automatically recommended a very poor solution (e.g., failed to consider uncertainty in weather predictions), many of the participants selected this poor flight plan despite the fact that they looked at all of the relevant data and explored (but rejected) much better alternative plans. By contrast, in conditions with intermediate levels of automation, where pilots were required to identify a priori aspects of a solution that they wanted to see included and "delegate" them to automation to realize if possible, these types of errors were more likely to be avoided."

- Constructing workflows for really flakey LLM outputs (how do you protect against temporary LLM insanity/ incompetence; alternatively: sudden ramping of token prices/ dropping of usage limits?) (Seems like there's an analogue to "building a service with 4 9s off of primitives that only offer 2 9s")

- feeding LLM output to another instance or model as adversarial balance to sycophantic tendencies

- curriculum design for standardized leet-code gatekeeping test to be studied for, and pattern-matched.

when you set an LLM against another (or itself), what do they do? (I'm curious as to how they perform ... maybe I should try it out.)

- https://www.usenix.org/system/files/login/articles/login_june_07_jones.pdf Sysadmin: Hiring Site Reliability Engineers (Google)


- in the same way that certain types of dynamic websites can be "compiled" to be static websites and lose the need for a compute backend and just serve static assets: how do we use LLMs to boostrap the creation of deterministic software (that doesn't have an llm in the workflow)