runesleo

Karpathy 4/30 at Sequoia Ascent compresses this year's most useful AI explanations into three key points. After reading, your perspective on AI will change.
1. AI is not just "faster," it’s a new paradigm
In the past two years, everyone has been talking about AI making things faster.
Karpathy says this is a misinterpretation.
Here are three examples of AI redefining tasks:
- menugen: image input and output, no traditional code, the entire app is swallowed by LLM
- .md skills: installing software without writing .sh scripts, just write a Chinese/English explanation, letting the LLM

AI has helped me increase my workload by 10 times,
The current bottleneck is my own brain.
Prefrontal cortex is severely damaged 😂

Vision Pro connects to Mac with a ultrawide monitor, Polymarket monitors odds fluctuations, Codex runs strategy code, and opens another window to watch live streams, leading to new ways of watching sports, preparing early for the World Cup.

The Codex quota that was supposed to refresh tomorrow was suddenly reset again ahead of schedule.
Codex is really interesting in this regard.
The weekly quota is supposed to refresh every 7 days, but in reality, it secretly resets every few days.
Is this really a marketing strategy or some kind of sneaky move?

A tool that monitors itself over 5 days, triggered 306 times, 100% false alarms.
Turns out I’ve been inhaling its own exhaust 🤦
Added a hook to Claude Code, to pop up a banner reminder before SSH cross-machine or writing critical files, worried I might accidentally violate the SSOT rule.
Paired with a stats script to count trigger counts, and today I looked at the sample, all false alarms.
The bug isn’t in the hook, it’s in the stats script: it greps for the string "⚠️ cross-machine" in the logs.
But there are at least three types of echoes in the logs: the banner output by the hook

Leo Labs Big Group 4/26 Group Member Discussion Highlights (Recent 380 Messages)
━━━━━━━━━━━━
1️⃣ Kelly Criterion was collectively rejected in the group — unanimously switching back to fixed positions
Group discussion on position management for over 30 minutes:
• "Kelly needs to know the true probability, retail traders simply can't calculate it"
• "Kelly is an optimization method under stable profit conditions, it doesn't determine profit and loss itself; claiming it as an essential system component is nonsense"
• Practical consensus: fixed position size + increase principal after profit + do

Cursor is so generous, giving away a token quota of 10,000 USD all at once, expiring at the end of May.
When I first started using vibe coding, I heavily used Cursor for a period of time, then gradually shifted to mainly cc and codex;
I didn't expect to have the chance to revisit Cursor and see what fun and useful products and tools I can develop, now I have something to play with!
Thanks to @cursor_ai @edwinarbus 🙏

Thank you, Elon
Even though it has decreased, it’s still more than expected.
About $225 in revenue corresponds to every 1 million impressions.

Creating automated strategies for Polymarket, everyone is very cautious when deploying, but when disabling, hardly anyone talks about it — today I learned this lesson the hard way after 14 hours.
Three weeks ago, I stopped a strategy bot running on Polymarket. pm2 stop, saw the status turn to stopped, and thought the issue was resolved.
Today, I casually did a 10-minute server cleanup and restarted the process manager. That "dead" strategy from three weeks ago was revived and quietly ran for 14 hours in real money mode.
I checked the root cause — pm2 stop only marks the status as "stoppe

Check out @predictionindex’s latest weekly market forecast report for this round: Polymarket + Kalshi together account for approximately 75% of trading volume, with all other platforms combined under <$30B。
Two details not mentioned in the report are actually more critical for prediction market players：
Kalshi is catching up to and even surpassing Polymarket
DeFiRate’s latest week：Kalshi $2.9B（60%），Polymarket $2.0B（40%）。 The two sides are locked in a tug-of-war in the weekly data—last week Poly just overtook Kalshi, and this week it was pulled back again.
The U.S. market is even more e

Another thing that’s not really useful but very satisfying: today I used Claude to reduce my Gmail inbox from 37k to 11k. I didn’t read a single email the whole time, just clicked twice: one to send App Password, and one to import the filter.
It’s all done with Python + IMAP, in four steps:
1. Export all senders categorized by frequency
2. Archive 555 spam domains to Archive-Junk with one click
3. Proactively stop and say, “I accidentally marked 317 emails from banks / brokerages / Coinbase / Claude as important. Do you want to recall them?”
4. Generate Gmail Filter to let these doma

Recently, I conducted a multi-model competition, and I made an interesting discovery.
For the same problem, I asked Claude and Codex to independently come up with solutions.
Codex missed one thing: my strategy is distributed across more than 20 independent processes, and its default assumption that "all components run in the same place" made the solution invalid. Claude immediately spotted this issue.
On the other hand, Claude missed another thing: it was fixated on creating new independent modules, but I already had a complete framework ready to be integrated—just needed to add one fiel

A common mistake in quantitative strategies: trusting "seems reliable" data without running upstream checks.
My weather strategy previously covered 34 cities, gradually expanded over several past expansions, and I haven't reviewed it in a while. A few days ago, during a routine audit, I casually pulled the full market data from the official Polymarket API and found there are actually 44 cities — I missed 10 myself.
Even more interesting, I compared it with the weather tool site that everyone in the industry references. It also missed its biggest city: Panama City. On Polymarket, there are 66 e