The Lake They Couldn't See
Every fifty years a new kind of gold appears, and we ride past the thing worth keeping to go dig it up. We are doing it again.
Cyrus Radfar
Posts and experiments from Cyrus Radfar
Latest posts
The Afternoon of Life
I'm a builder, and I'm about to stop building to drive my family through every national park in the country. Not a midlife crisis. A crossing — the one an old Swiss doctor said we all have to make, and that nobody in my world is taught how to survive.
Isolation Is the New Runtime
Sandboxing is thirty years old and has barely changed. What changed is who's in the box: an agent that reads its next instruction off the open internet and can't tell a command from a trap, which is why the OS is reorganizing around it. A line from chroot in 1979 to Microsoft's MXC in 2026, and the trade we made along the way.
Where's the political center, if money didn't move it right?
The federal minimum wage has been $7.25 since 2009. The senators who kept it there call themselves moderate. I built a dataset of 85 polls and the funder profiles of ten think tanks and ran the analysis. The 'moderate' position polls at the size of the far right.
Why cars charge 5x faster in China when the research is shared?
BYD's claim of 500 kilometers of range in five minutes is real, independently observed, and on a $27,400 sedan. The chemistry is public knowledge. So why can't you do this in Bentonville? This is the consumer companion to the AI infrastructure story: five thousand megachargers, four V4 sites, one fired team, one frozen program.
What AI Needs That $700 Billion Can't Buy
A breakdown of the supply chain behind the AI buildout: a 38-year-old transformer, a single steel mill in Butler, Pennsylvania, a 5-year wait on turbines, half a million missing electricians, and a lengthening list of people who said no to the check. What AI needs to run at scale, and where those needs collide with ours.
Looms Taught Us to Store, Share, and Run Logic
The loom is a computer. It has been one for 30,000 years. The Inca had 7-bit ASCII in string. Lacemakers ran a two-opcode instruction set. NASA called the women who wove the Apollo computer 'Little Old Ladies,' then erased their names.
Should capitalism be a choice?
My kids can refuse to play each other's game. Adults in America can't. A long argument for the right to not compete, and for protecting the alternatives to capitalism the way we protect Yellowstone.
ADHD is not a deficit. It's a mismatch.
The history, the stigma, the neuroscience, and what it's like to have a brain that doesn't run on importance. A personal post about ADHD from someone who has it.
Stop reading walls of text from your AI agent
A simple slash command that forces your agent to walk you through decisions one at a time instead of dumping a plan you won't read.
I built a tech debt simulation. At 10x coding speed, feature output approaches zero.
Defect rate per line rises with codebase size. At 10x speed you hit the sprint where rework consumes more capacity than new work. I built a model to find that sprint and ran it through four scenarios.
Pretext is a text measurement library. The most interesting use cases have no DOM at all.
Pretext measures text 300x faster than the browser by replacing layout reflows with arithmetic. Cursor already shipped it. The use cases that matter most are in 3D, on servers, and in CI.
AI agents keep failing. The fix is 40 years old.
AI agents fail in production because of mutable state, hidden dependencies, and side effects the agent can't see. The fix is functional programming. SUPER and SPIRALS are the frameworks I use.
I built a filesystem flight recorder because an AI agent ate my work
An AI agent overwrote hours of uncommitted edits. Git couldn't help. So I built Unfudged, then Hacker News told me to open-source it.
Understanding Google's TurboQuant Insight
KV cache quantization explained: why compressing LLM memory below 4 bits breaks attention, and how Google's TurboQuant (PolarQuant + QJL) fixes it. Compares KIVI, KVQuant, QuaRot, MLA, and H2O. Updated April 2026 with framework support for vLLM, SGLang, llama.cpp, and TensorRT-LLM.
Models are optimizing their own tooling now
For sixty years, self-improving AI was the apocalypse scenario. It arrived as a sampling temperature tweak. The gains are real anyway.
On resilience over optimization
A response to the idea that the best engineers optimize systems — and why I look for something different.
ADHD Reader
A speed reading tool that flashes one word at a time against a 3D canvas. Paste markdown or load a blog post, set your pace, and read without moving your eyes.
Humanity trumps politics: we must stop the escalation
Gilens and Page found US citizens have near-zero impact on policy. If Americans can't control their government, why are we punishing 90 million Iranians for theirs? An Iranian American developer on sanctions, the Versailles playbook, and why the tech community needs to speak up.
The thieves are upset about theft
AI labs trained on everyone's data without asking. Now they're furious that other AI labs are training on their outputs. The pattern is older than electricity.
Zen Bloom
An exploration of how ideas form, find each other, and grow — rendered as a particle flow field.
Configuring this Blog with a Delightful Example
A kitchen-sink post that exercises every typographic element — headings, blockquotes, images, code, lists, and more — so the design configurator has something real to chew on.
Sandman
Zero-blue night filter for macOS, Linux, and Windows. Zeros out blue and green at the hardware gamma level. No CPU overhead. Screenshots stay full color.
A single data center drinks five million gallons of water a day
America's energy and water infrastructure cannot sustain AI buildout without radical change.
Productivity is up, but is your time worth more?
Coding productivity metrics are rising, but the real question is whether human judgment is becoming more valuable.
Toggle your inner critic
High taste is a superpower in review, but a liability during creation—separating judgment from making is the skill creators need most.
The GPU accumulation era is ending
Raw compute scaling is hitting limits; the next breakthroughs require hardware-software co-design and specialization.
How do you understand code you didn't write?
Code generation speed isn't the bottleneck—understanding and verifying AI-generated code is, and current tools optimize the wrong 5% of development.
Will it work out?
Uncertainty has shaped human decision-making for millennia—from ancient dice to modern language models—teaching us to work with randomness, not against it.
I could've been an awesome Dungeon Master
Reflecting on D&D world-building skills and what generative world models mean for the future of storytelling and creative expression.
Building for myself first
Using agentic development to prototype Swift apps that solve personal problems with privacy, ownership, and connection at their core.
A fundamental shift on AI opportunities
The real opportunity in AI is disrupting incumbents and creating new products, not building more AI platforms.
The cave you fear to enter
As AI handles optimization, human superpower shifts to ideation and the courage to create badly—making and learning without measuring.
You didn't need to predict the App Store
Mobile blindsided many companies that moved slowly; AI will repeat this cycle—start small, measure, and upgrade every quarter.
AI autonomy level framework
Framework for understanding AI autonomy levels outside self-driving, from copy-paste to full autonomous systems.
What the heck is Cy up to?
Learning at speed to build AI-enabled products that return people to the real world, moving beyond screens.
You down with A.P.P.?
A rap parody celebrating AI Paired Programming—the synergy between humans and AI in software development.
Willpower is not unlimited
Willpower and resilience are finite resources that must be managed strategically, not glorified into burnout.
AI won't magically replace companies
Economic pressure and frozen job markets will drive a new generation to build, sparking another wave of founder-led disruption.
Hot take on the future of coding
Most individual contributor developers will need to re-skill within five years as AI agents take over execution work.
No team votes itself out of existence
Conway's Law shows why organizations keep building what they did yesterday—and why Tesla's system reimagining beat a hundred years of optimization.
Hot take on Sora
Sora's outputs remain largely unusable at current pricing; the model needs dramatic improvements in consistency to justify its value.
Will there be 10 billion humanoid robots?
The raw material constraints for 10B humanoid robots by 2040 are severe—especially lithium and rare earth elements.
Excuse the nostalgic rant
San Francisco's tech community gathering places have vanished, taking with them the energy that built the startup ecosystem.
1-bit models and the future of on-device AI
Binary neural networks achieve comparable performance to 16-bit models while using 3x less memory, 4x less GPU resources, and 100x less energy.
AI is a new entity
Understanding that AI agents represent a fundamental shift beyond smarter tools—a genuinely new form of intelligence with exponential capability.
Google's sloppy Gemma launch
Google's open-source Gemma launch failed to account for community infrastructure, creating a poor first impression with key stakeholders.
On a personal note
Reflecting on a transformative period of recovery, exploration, and the decision to build something new in the age of AI agents.
Projections on OpenAI after Altman
When risk-averse leadership replaces visionary founders, the focus shifts from innovation to asset protection.
Stop romanticizing entrepreneurship
People aren't inspired by a hero's journey of suffering—they're inspired by movement and momentum. Entrepreneurship is just another path.