Using AI to really master those techniques I never quite mastered in college

We are likely to see “select a model/reasoning effort” UX go away. Not to simplify for the customer but to let companies manage their costs better. Any competent user selects the beefiest model for every task given the choice but the compute supply is not keeping up.

It should be pretty obvious at this point that AI is a "force multiplier" not a "labor substitute". It helps experts be better at things they are already good at. It doesn't let beginners match experts. If you can't write, anything you write with AI will be unmitigated slop.

There is a real art to finding the perfect quantity of context to give your agents You have to provide enough info to enable pattern matching to your specific tech (the same way a human does) But not so much you eliminate the ability to focus clearly on singular tasks

“hey claude buddy just so you know every time you ask me ‘should I investigate the crash?’ it physically wounds me. A tiny robot comes and stabs me in the leg with a fork :(“

If an LLM gave you a cake recipe, would you bake it? Investing your resources (ingredients and time) without validating precise amounts and ratios against an external source?

It turns out that LLMs are least useful when building something new, where patterns still need to emerge and direction needs to be set And incredibly useful in a mature engineering org with clear best practices and precedent for problems

My favorite agent I've built is Lydia. She has full access to my Home Assistant APIs, available via telegram bot. I've transitioned my wife to bugging her with asks to turn things on/off + create routines. Chat is an amazing interface for regular people and complex systems.

When starting on interview prep I just assumed there would be some high quality guided + adaptive learning tool to get you up to speed on leetcode Doesn’t exist though believe it or not Should we build this?

The most basic skill I’ve been teaching to get more out of agents is feedback loops First task: write code however you want, then ask the agent “build this until checkstyle passes” Has converted many skeptics once they realize it can meaningfully simplify their work

None of the big 5 companies seem to care about quality anymore. There was a time when eg Amazon greatly cared about customer obsession (when it was still Day 1 and Bezos was there), Apple about quality. Now it's chasing whatever they think will make the most money to sustain

guys I’ve been playing Engineering Manager Simulator and its super fun but $1k/day seems pretty expensive to play

I recently gave a talk to my engineering org on how I am regularly getting 2hr+ autonomous execution loops via agent orchestration. This is an excellent summation of the same principles I've learned through spending a HUGE amount of my employer's tokens.

I always found “personified agents” a cringe AI-psychosis approach but unfortunately its a really effective model to build around So now I’m often repeating a mantra of “yeah I named my agents and gave them personality quirks BUT I really am sane”

This tenet is clear: As your agentic system matures it should become more deterministic worn paths convert from reasoning to deterministic code

Is the best agent interface on a phone still telegram? Has anyone made an actual high quality mobile interface to work with agents?

Anytime you touch an LLM your first task is to constrain its world view to roughly match yours

openclaw is gonna be so pissed when it falls for an internal prompt injection test and has to do a 4-hour company mandated training