runesleo

A Polymarket player with about a 96% win rate on the main battleground: I split BipBop’s 1,632 trades into separate positions
This account on the leaderboard: BipBop
1,632 settled positions, main battleground win rate about 96%, positive for multiple consecutive months.
Most people’s first reaction is: copy him.
We actually backtested that layer that looks “like him” — turning “buy 0.85–0.95 high-probability markets, and get to settlement” into a mechanical strategy, backtesting 5,378 markets (high-probability bucket compared group): win rate 88.9%, net profit −2.5%.
Win rate is an illusion. A

Codex is down—congratulations, everyone is about to welcome a new round of credit limit resets!

When you’re not at home and the elderly are not by your side, and they also have very specific needs, with AI now, it really becomes convenient to create highly customized small tools.
For example, in my family, an elderly person needs to do muscle rehabilitation training for a period of time. I turned the movements and intensities that have already been confirmed into a web page specifically for him: when you open it, you can immediately see what to train today, with one item per screen that you can follow directly. If the plan changes later, I can also update the page for them right away.
In

Everyone is asking AI to write code, but today I did something even more thrilling—getting it to write a piece of code that can automatically spend real money and place orders with zero human supervision.
First, I laid out the decision logic card: I thought through the boundaries, the worst-case scenarios, and the kill conditions before I dared to touch it. I initially wanted Codex to set up the actuators by itself, but the dispatch channel was unstable all day, so I switched to building it incrementally in my own main thread.
Before going live, I ran 7 rounds of adversarial code reviews, and

Five days ago, I locked all four lanes—Claude, Codex, Cursor, and Grok—into the strongest model with the highest reasoning setting. The idea was: since I paid for it, I should use the best fully.
Five days later, I realized I can’t handle it at all. The quota drops too fast. The strongest model ends up doing redundant checks and formatting work—pure waste.
Now I’ve changed it: I split by judgment density. For tasks where mistakes would cost money or contaminate data—reviews, cross-validation, funding strategy—I keep the strongest seat. For day-to-day implementation and development, I use a mid

I gave the same UI requirement to the same model and ran it 9 times: 3 runs with no Skill, 3 with apple-design, and 3 with ui-ux-pro-max.
The results forced me to change my evaluation approach: the same set of Skill can produce one great output and another failure.
After anonymizing the 9 sets of results and handing them to two people for independent scoring, the average scores across the three groups are:
apple-design 91.3
ui-ux-pro-max 89.0
no Skill 85.8
But none of the three groups achieved 3/3 with zero defects.
apple-design and no Skill each had a case where the primary button on the ph

The most dangerous failure in AI workflows isn’t a direct error.
It’s when the main model goes down and the fallback also fails, yet the system still outputs a seemingly complete result—making people think it’s done.
I hit this once: during daily information processing, the main model got cut off at 60 seconds, the fallback didn’t succeed either, but in the end I still received an extraction text that looked respectable. The problem isn’t “switch to another route and keep trying,” but that after switching routes, nobody clearly explains what happened.
Later, I added three receipt checks

How to build in public?
Today, I had AI work from early morning to late night—pushing tasks, research, scripts, and queues forward a lot. But on X, there’s still nothing worth posting. The content system feels like it has no topic selection at all.
On the 16th, I published 8 pieces of content; on the 17th, I did more actual work, yet the supply clearly dropped.
So I made Work → Surface v1.
It passed 19/19 tests, and it also reads and verifies 99 live task cards. Currently it only has one mature reusable object, which can automatically derive 2 preview-ready packets.
The system won’t automatica