The challenge in building agent applications is that the all of product state is related to the agent, so every feature in your product is coupled with the agent! So every team member now needs to know how most of the agent functionality works. Requires structuring team roles and products differently.
1
1
187
The most valuable semiconductor company in the world has only 2 members out of 10 who have ever worked in a semiconductor company
What?! So journalists with sketchy motives portrayed rumors from god-knows-who (lobbyists?) as facts sourced from intelligence officials and gaslit the NVIDIA board to push out the founder, CEO, and once-in-generation visionary?
This raises serious corporate governance issues.
175
Applies to any knowledge worker. Wake up and increase the breadth of your work, and go higher up the stack before your role is obsolete.
162
We found that every single open-source document format converter sucks. If you want to convert btw formats: markdown, html, docx, xlsx, you have to write your own! Which we have to painstakingly do
1
157
Imagine telling someone 5 years ago you would spend $10 bln to train a model that predicts next characters in a sentence
Replying to @jukan05
Correction: The $10 billion was reportedly spent on training the entire Claude 5.5 model family, including Opus 5.5.
3
320
We sold Ribbon last week for life changing money. Built here in Toronto. I’m not sharing this as a flex. I just want founders/builders to know that you can do it and you can make it from here. You can sell to the world from our city. You can do it. I will help you do it.
Anuar retweeted
Someone I know asked Astra to play Factorio. It wins, but it takes forever and guzzles tokens. If you ask Astra instead to “design a system to win Factorio,” it builds software that runs 1000x faster. Mechanistic software gives you cheap leverage, and LLMs can write mechanistic software on demand.
Coding agents seem to be mainly capable of adding or editing existing code and NOT removing it. Complexity and size of the codebase grow much faster now. I have to constantly force agents to “simplify” or “check for redundancy” to remove things
2
3
210
Imagine all world knowledge is a giant repo that you can run Claude Code on with tool access to every API in the world! 🤩🤯
2
2
265
Most human jobs will be commoditized and see their wages drop, but the most skilful humans will see their wages skyrocket as AI scales outputs exponentially.
The dominant model people use to think about AI's effect on the labor market is automation. But perhaps commoditization is a better model. Automation implies job apocalypse. Commoditization implies plentiful jobs with low wages. 1/
191
The best way to start with evals on your agent harness is to write deterministic unit tests, not custom metrics. Your coding agent would write the cases and checks on output yourself. This way tests are deterministic and you don't have to maintain golden set or review metrics
19
11
584
The most important for me is “Do we even need this” and “Is there a simpler way to do this” 😆
The anti-slop ai coding prompt I’ve found the most success with is definitely “deepen modules and simplify interfaces”
It does such a good job with deduplication and centralization of core functionality, and gets the ai implementing much better levels of abstraction
It’s led to tons of consistency improvements, bug fixes, and, most rewardingly, reduced LOC. I run a deepening pass on every PR, and a nightly automation to tackle overdue maintenance
Comes from @JohnOusterhout’s Philosophy of Software Design (highly recommend), which I had read but didn’t fully leverage until @mattpocockuk put me onto it
2
3
180
You would upgrade to god-level founder once you acquire a skill of going through walls under level of pain
1
2
215









