Recent Notes
View all"The last week was funny and also tiring, I worked 14 hours per day on average. My normal average is 4/6 since early Red…
If I had to ask myself the question — had I the right to gain power? — then I certainly hadn’t the right
### The two-minute rule - Take whatever habit you're trying to build and scale it down to something that takes 2 minut…
- Every action you take is casting a vote towards being the type of person you want to be - Act of doing it is proof you…
Change in the environment such that the behaviour that you want is obvious. You want to be able to walk into your rooms…
In a lot of ways, I actually think the most powerful form of mental toughness, the most powerful or resilient form of pr…
No amount of information is going to allay the fact that all of your knowledge is about the past and all of your decisio…
People talk themselves out of things long before the world actually shuts the door.
### Work backwards from magic "One of my little sayings internally is I try to work backwards from magic. So what would…
### Reputation "I guess the main way that I think about it is I want to be known as someone who is useful. Useful is a…
The financial behaviors are downstream of the psychological experience: When the future feels foreclosed, your brain sta…
The political scientist Peter Turchin has a framework for understanding this group historically. He calls it “elite over…
# Do as much as necessary and as little as possible - Nvidia's job: turn electrons into tokens. Do as much as necessar…
# AI is a five-layer cake - AI isn't just a model. It's a five-layer stack. - The bottom layer is energy. Above that si…
# Conceding China is a loser's premise - Nvidia's share is growing, not shrinking. The assumption that they'd lose Chin…
"This is one of the concerns that I have about the doomers describing the end of work and killing of jobs. If we discour…
- Moore's Law now delivers about 25% per year. That's it. - Between Hopper and Blackwell, the transistors themselves got…
When designing a consumer product, you should consider every tap by a user to be a miracle. The motivation to stop using…
They also don’t realize that yea I could keep going through thier onboarding flow, but timewise it’s competing with me l…
There’s some weird cognitive bias where product creators vastly overestimate how likely people are to do or pay attentio…
May 2026
A top-tier CTF competitor explains how frontier AI models (GPT-5.5, Claude Opus 4.5) have turned security competitions from skill ladders into pay-to-win orchestration races, breaking the primary pipeline for developing elite security talent.
GGUF files contain chat templates, sampler configs, and special tokens—but missing metadata for tool calling formats, think tokens, and multimodal projections still forces developers to write model-specific code.
Why building small, "unimportant" software projects for personal curiosity is more valuable than the tech industry wants you to believe—from the creator of Redis on his post-retirement work.
Your codebase is a Marble Madness level and your AI agent is the marble—your job is removing hazards so it rolls smoothly, plus a vision of "Barbapapa software" that reshapes itself while running to fit your exact needs.
A simple two-step technique for generating AI images with accurate text and numbers: create a precise SVG "underdrawing" with correct layout, then have the AI model paint over it to make it beautiful.
James Clear reveals why ambitious people fail by optimizing the wrong variables - and his exact systems for life positioning, from deleting all social media to sequencing advantages across decades.
DeepSeek just released the largest open weights model ever (1.6T parameters) that performs within 3-6 months of GPT-5.4 and Gemini-3.1-Pro but costs 10-20x less, with efficiency breakthroughs that use only 10-27% of the compute for long context windows.
AI is about to trigger the biggest SaaS churn wave yet—not in 2024 or 2025, but in 2026 when contracts renew and companies finally have time to rebuild their stacks around agents, internal AI tools, and new workflows.
AI isn't just changing what software does—it's breaking the SaaS business model by enabling internal builds, forcing vendor consolidation, and making 18+ month CAC payback periods mathematically impossible when customer retention collapses.
When you hit the afternoon focus wall, the solution isn't more coffee or a break—it's literally staring at a wall for 5-10 minutes to reset your dopamine-fried brain.
Cursor trained their AI coding agent to learn what information matters by making self-summarization part of the RL training loop, enabling it to solve problems requiring hundreds of actions that exceed its context window.
OpenAI's experiment creating a deliberately chaotic "goblin mode" for ChatGPT revealed that intentionally misbehaving AI teaches more about alignment and human-AI dynamics than always trying to be helpful.
April 2026
Cursor reveals how they make the same AI models work noticeably better through obsessive harness engineering—measuring everything from "Keep Rate" of agent code to fixing "context anxiety" where models refuse work as their context fills up.
A financial planner earning good money explains why she still feels crushing money anxiety—and why the gap between what your income "should" buy and what it actually buys isn't a personal failing but a feature of how the economy has been redesigned.
Linear doesn't use OKRs or A/B tests—they hire for taste and judgment through paid work trials, prioritizing "slope over credentials" and autonomous builders over polished interviewers.
Tarantino reveals how working at a video store nearly killed his filmmaking dreams by being "dream-adjacent"—comfortable enough to put his ambitions to sleep, but not actually his dream—and the all-night self-interrogation ritual that finally broke him out.
Jensen Huang argues Nvidia's real moat isn't chip specs but ecosystem lock-in through CUDA, and makes the contrarian case that export controls are backfiring by forcing China to build a competing AI stack that will become the global standard.
Waymo adapted Google DeepMind's Genie 3 to simulate driving scenarios that are nearly impossible to capture in reality—from tornadoes to elephants—by transferring vast world knowledge from 2D video into multimodal 3D simulations with camera and lidar outputs.
Waymo's new world model can simulate tornadoes, elephants on roads, and other scenarios their fleet has never encountered by leveraging Google DeepMind's Genie 3—fundamentally changing how autonomous vehicles prepare for edge cases.
Claude inverts the typical AI orchestration pattern: instead of a smart model delegating to dumb workers, a cheap model does all the work and only escalates to the expensive genius when it gets stuck—delivering near-Opus intelligence at near-Sonnet costs.