In 2025, Oxford philosopher Toby Ord ran the numbers on METR's AI agent benchmarks and found the failure curve is identical to radioactive decay. Every agent has a half-life. 99% success per step sounds nearly perfect - raise it to the power of 100 and you get 0.37. Your agent fails two jobs out of three. The same exponential decay that defines every learning rate schedule in every training run.
Double the task length and you square the failure rate. METR found the half-life has been doubling every 7 months for 6 years straight - six more doublings turn 1 hour into 64, a full working week. OpenAI is betting on that curve with Dots at $500 a month. The only escape is verification: let the agent catch 9 out of 10 of its own mistakes and that 37% jumps to 90%.
4 minutes. Bookmark & watch today. The equation that decides whether AI agents are worth paying for.
Daryle Smith retweeted
I remember the first time I was hit with cyberattack.
It was September 9th, 2014. I woke up and an absolutely garbage U2 album had been downloaded straight to my iPhone 5.
Daryle Smith retweeted
We have a new section in our technical overview.
Foreshadow is things we are working on, this is a running record of work done and being worked on, when completed they will be moved to "Past-Shadow."
For people who like to read.
Daryle Smith retweeted
The sandbagging at the 7th Annual @NoLayingUp NEST Invitational was so egregious, that the tournament director had no choice but to cut a competitor’s handicap mid-event. #golf #TheYipRule #journalism
We are heartbroken by the passing of Dan Hicks. He was a true friend to the game and to the USGA, but more importantly, a wonderful person whose warmth, generosity and friendship touched everyone fortunate enough to know him.
His presence will be deeply missed, and his impact on so many of us will never be forgotten.
Our thoughts and heartfelt condolences are with his family and loved ones.
Daryle Smith retweeted
We’re building out the gradient library together and every challenge you create while golfing helps generate useful test cases for Llama-3-8B and Phi-3. Those feed straight into the Pebblewood Library linked to your profile, so you earn the contract-generation credit.
Every surface hides a field of arrows. At each point, one arrow shows the steepest way up - that's the gradient. Flip the arrows and follow them step by step to the bottom. That's gradient descent - the single idea that trains every neural network, every LLM, every diffusion model on every GPU cluster on the planet.
It starts with partial derivatives. Slice along x - freeze y - measure the slope. Slice along y - freeze x. Stack the two slopes into one vector and it points straight uphill. On a tilted plane every arrow is identical. On a landscape of hills and valleys, arrows flow out of valleys and gather toward peaks. The same vector field that every optimizer computes billions of times per training step to update every parameter inside every transformer.
2 minutes. Bookmark & watch today. The hidden arrows behind every AI model that learns.
Daryle Smith retweeted
Pebblewood National — Agronomic Surface Physics
I've implemented the first integration of course conditions into Oracle, using the latest player-club physics build as the baseline.
The central change is that the ball responds to the surface it actually touches.
Daryle Smith retweeted
If you play PebblewoodNational.com right now you can experience some extremely hard surfaces and roll. Physics engine is live and working, not all variables are connected yet, but they're coming.
20 MPH tailwind, 369 yard drive. Par 4, Eagle.
Daryle Smith retweeted
BTW anyone not using agents like this is falling behind and will inevitably wind up in the permanent underclass.
Daryle Smith retweeted
Working.
DB will be reset a bunch so don't save it and you can dismiss and refresh for a new set, but so far it works and that's what we like, we can fix bugs after things work.
Hard to fix a bug if it doesn't work.
Daryle Smith retweeted
Doing some work and have a preview up of the new version.
qcem0vca23.c36.airoapp.ai/?p…
Daryle Smith retweeted
Updated our GitHub. still working on the AI side of our game.
github.com/darylesmith06-gli…
Daryle Smith retweeted
Golf is already a game of incomplete information, so an AI caddie that quits after one shot would fit the comedy of the sport almost too well. The useful version still has to stick around for all 18.
Daryle Smith retweeted
If you're not in AI and need to understand what's happening:
This year we've went from Super Nintendo to PS5 in 6 months
























