micah

Lilian Weng @ OpenAI on Harness Engineering a deep breakdown of harness design patterns, workflow automation, persistent memory via file systems, and how modern agents like claude code actually work. the best blog post I've read all month. link: lilianweng.github.io/posts/2…

Waterloo makes you write your exam here btw..

just learn from elite company blogs: - google for systems (gfs, mapreduce, etc.) - anthropic/openai for agents (lilian weng's blog) - stripe's blog for high-stakes engineering (zero downtime, idempotency, payments, etc.) it's absurd how much knowledge is free

software is entirely about doing the *right* work

if you're 18-25, in software/cs, and relatively ambitious, rn is the best moment in history to be alive

coded by hand today just to feel something

AI widened the skill gap in the most skill gapped industry there is lol 10x engineers are now 100x engineers 1x engineers are now 0.1x engineers

As someone who grew up in Canada, SF is easily the most beautiful city I've been to, by far

the levels of modern AI engineering: L1: basic prompt engineering, RAG L2: context window arch., dynamic pruning L3: harness engineering & loops L4: deterministic evals, synthetic data L5: "hey, are you conscious?"

CMU's new AI agents course is as good as it gets, and totally free it goes all the way from multi-agent coordination to computer-use agents, deep research, and advanced reinforcement learning algorithms incredible for anyone building agent systems

POV: a day as a human in 2027 7am: wake up 8am: AI solves 91 year old unsolved problem 8:07am: "it's fine, my career requires taste." 8:09am: open favourite model, start prompting 8:10am: notification, "taste v1.0" added in new model 8:15am: log in for UBI, go enjoy rest of day

The modern engineer's dilemma: - optimize for output = use AI - optimize for learning = don't use AI What's the most efficient way to optimize for both learning and output at the same time?

POV: You open LinkedIn, only to see a 12 year old Waterloo CS major is outcompeting you

ngl for me ai's productivity boosts are half using ai for stuff, and half pressure from "you have one year left before AGI to escape the permanent underclass"

Stanford has a new course on modern software engineering... and man do I wish I had access - coding agent internals - advanced context engineering - structuring repos for agents, self-improving software, etc. every university should be teaching this stuff rn

it's interesting how computer science is shifting to involve more math and physics glueing together APIs is becoming automatable, while deep mathematical insight, ml theory, and systems reasoning are more valuable than ever is anyone else noticing this?

The #1 reason why knowledge is going to remain valuable: someone has to do the prompting, and knowing things makes you way better at it link: seangoedecke.com/llms-reward…

I miss when Karpathy used to casually drop a 4 hour long session that condenses 1000 hours of ML learning into a short, beautiful video

- linear algebra - probability & stats - combinatorial optimization the most important areas of math for cs. a lot of what feels like magic (ML, AI, advanced algos) is just these core areas in disguise.

with LLMs you can self-teach literally anything. you should learn math, learn physics, learn computer science, learn the startup ecosystem. literally all you have to do is ask "what's the best resource to learn this topic" and then when you get stuck ask "why am I stuck?"

Stanford just released CS329A, a graduate seminar course on self-improving AI agents a goldmine for going from zero to hero on multi-step reasoning, planning, agentic evals, and scaling RL: a must-watch playlist. youtube link: youtube.com/playlist?list=PL…