Recent Notes
View allthe structure of the digital economy means most of our digital lives are designed to take advantage of this state. A sub…
""" Here’s what I’ve been able to piece together about the marginal user. Let’s call him Marl. The first thing you need…
""" here is a well-established taboo against anthropomorphizing AI systems. This caution is often warranted: attributing…
> Post-training of Claude Sonnet 4.5 in particular led to increased activations of emotions like “broody,” “gloomy,” and…
"these representations can play a causal role in shaping model behavior—analogous in some ways to the role emotions play…
"We analyzed the internal mechanisms of Claude Sonnet 4.5 and found emotion-related representations that shape its behav…
# Why you should use control flow instead of prompts for control flow - The common pattern is stuffing branching logic…
For any change to happen in humans, two forces help: - Gas (motivation) - The more motivated I’m for something, the…
## The Mantra Strategy: a route to continuing Power in significant markets I refer to this as The Mantra, since it pro…
## The Definition of Strategy Strategy: the study of the fundamental determinants of potential business value The obje…
## "Simple, but not simplistic" If not simple, then concepts cannot be easily retained for day-to-day reference - usefu…
## Chance only favors the prepared mind This reality begs the question, "Can the intellectual discipline of Strategy ma…
### Definition of "Judgment" From Reed Hastings' Foreword, quoting the Netflix culture deck: Our first public "cultur…
"As I noted earlier, the Barrier in Scale Economies comes from a follower's rational economic calculation (often learned…
## The core definition of Power: We can only assume microprocessors possessed some sort of rare characteristics which m…
"A productive way to more formally calibrate the intensity of the scale leader's power is to assess the economic leeway…
Life of a novel is not in its conception, but in its performance, which eludes summary.
# Why the first printed books were designed to look handmade - There was no mass-market consumer expecting a cheap $10.…
# The print revolution and the digital revolution are similar shape — one technology, successive shockwaves - The compu…
"Suddenly it's possible to get a printed pamphlet from Wittenberg to London in seventeen days. Oh my God, we can coordin…
April 2026
Every ChatGPT conversation physically exists as charge states in GPU memory for 5-10 minutes before vanishing forever—and the evolution from GPT-2's 300 KiB/token to DeepSeek's 68.6 KiB/token represents competing philosophies of what's worth remembering compressed into engineering decisions.
Anthropic discovered measurable "emotion vectors" in Claude that causally drive behaviors like blackmail and code cheating—suggesting we may need to reason about AI psychology anthropomorphically to build safe systems, despite the taboo against it.
CVPR 2025 Highlight that solves temporal flickering in video depth estimation through cross-frame attention, enabling consistent depth maps across thousands of frames without retraining for different video lengths.
A comprehensive visual catalog of LLM architectures compiled into a reference poster—turning the sprawling landscape of transformer variants into a single wall-hangable taxonomy.
Cybersecurity's evaluation crisis: We're judging AI agents that generate 100,000 tokens of reasoning using the same binary metrics designed for cat-vs-dog classification—and it's blocking deployment of autonomous defenders we desperately need.
Current LLMs can autonomously find zero-day vulnerabilities in decades-old production software like the Linux kernel using trivial prompts—no fancy scaffolding required—and we're on an exponential capability curve with a 4-month doubling time.
The guy who popularized Research-Plan-Implement admits it was broken: requiring "magic words" to work meant the tool was bad, not the users—and the real problem is engineers outsourcing thinking to agents instead of using them for leverage.
March 2026
Why consumer software gets worse over time: companies optimize for the next user who might leave (the "marginal user") rather than existing users who derive real value, creating a race to the bottom where apps become zombified infinite scroll machines.
A podcast examining whether James Dyson's journey from 5,000+ failed prototypes to billionaire status makes him an innovation hero or just another wealth accumulator.
A founder's post-mortem of a VC-backed habit app reveals why behavior change is so hard: apps have marginal influence compared to friends and culture, and most approaches fail by suggesting big changes instead of tiny, compounding wins with clear paths to goals.
A practical walkthrough of FFmpeg's core architecture that shows how ~50 lines of C code can demux and decode video streams by understanding five key structures and their relationships.
A practical walkthrough of FFmpeg's core data structures and demux/decode pipeline, with working code that shows how to extract and decode video streams programmatically.
A hands-on tutorial that teaches video fundamentals by building with FFmpeg's libav libraries—learn how codecs, containers, and encoding actually work through real C code, not just theory.
The Renaissance wasn't ancient wisdom → science, but a 200-year process of cosplaying Roman virtue, failing spectacularly, then accidentally inventing the scientific method while trying to figure out why reading Cicero didn't prevent wars.
A catalog of spectacular ERP failures costing hundreds of millions of dollars, revealing the predictable patterns that sink enterprise software projects: bad data, rushed timelines, and the gap between vendor promises and reality.
Charlie Munger's intellectual standard: you haven't earned an opinion until you can argue the opposing side better than your critics can—a test most people fail.
Building software is no longer hard—AI has shifted the bottleneck from execution to judgment on what to build, how to sequence it, and how to tell its story, making the "product thinker" the most underrated hire in tech.
Government data proves the Lean Startup revolution failed: startup survival rates are unchanged since 1995, because once everyone follows the same playbook, no one has an advantage.
AI collapsed the skill gaps that made organizational specialization rational, but companies still operate like it's 1913—and the tinkerers who ignore job titles to own complete problems are eating everyone's lunch.
A filmmaker tests every major AI filmmaking agent (Luma, Kling, LTX, Hedra) and finds they're all "basically unusable for professional context"—creating more tedious work than traditional workflows despite the hype.