Leolabs AI 实验室

Someone is poking fun at today’s startup playbook: find a product that’s taking off, stick “Open” in front of it, have AI copy it, then call it the spirit of open source.

Funny, but it points to something that isn’t: the cost of copying has nearly dropped to zero.

It used to take months to copy a product, an in-depth analysis, or a whole methodology. Now it takes an afternoon.

So the real question isn’t “How do I stop people copying me?” It’s “What can’t be copied in the first place?”

My answer is a dividing line: things that can be generated in a one-off lose value; things that can only accumulate over time gain value.

AI compresses production, not accumulation.

Follow that line of thought, and there are really only three things worth sticking with.

First, do things that generate firsthand information.

Get involved yourself; don’t just comment from the sidelines. You only know the sale price isn’t what you expected once you’ve placed an order yourself. You only know which line in the documentation is wrong once you’ve hit the problem yourself. AI can recombine all the information that already exists, but it can’t experience the parts that haven’t been written down yet. Firsthand information is the only thing that isn’t in the training data.

Second, document your process publicly, including the parts where you got it wrong.

One public correction says more about whether you’re actually doing the work than ten correct articles. Someone copying you can copy your conclusions, but not how you discovered you were wrong—and they won’t know what was wrong, either, because they never did the math; they just copied the result.

Third, keep going.

This one is the least glamorous, and the hardest to copy. Because a copy has no next version. Copying once takes an afternoon; copying three years takes three years—and that’s exactly what you’ve already been doing.

In the end, AI hasn’t changed what’s valuable. It’s just made “looking valuable” cheap.

The real dividing line has never changed: where did your work come from?

Was it generated, or did it grow?

How happy you are has little to do with what you’re doing.

It has a lot to do with whether you’re mentally present while doing it.

This was a 2010 Harvard study published in Science. Using random pop-up prompts on iPhones, researchers sampled the real-time experiences of more than 2,000 people in their daily lives.

—— Three numbers ——

　People spend 46.9% of their waking hours mind-wandering
　“What you’re doing” explains only 4.6% of the variation in happiness
　“Whether your mind is wandering” explains 10.8%

In other words, what you’re thinking about has more than twice the effect on how happy you feel in the moment as what you’re doing.

And mind-wandering accounted for at least 30% of the time during every activity. The only exception was sex.

—— The most counterintuitive finding ——

You might think: Then I’ll just let my mind wander to something pleasant.

The data says that doesn’t work.

When people’s minds wandered to pleasant topics, they weren’t any happier than when they were focused on the present—the difference wasn’t statistically significant.

But when their minds wandered to neutral or unpleasant topics, they were significantly less happy.

So mind-wandering itself comes at a cost, regardless of where your thoughts wander. Thinking happy thoughts doesn’t make up for it.

—— Cause or effect? ——

This was the first thing I wondered: Could being in a bad mood cause mind-wandering, rather than mind-wandering causing a bad mood?

The study conducted a time-lag analysis: mind-wandering came first, followed by a drop in mood.

The direction was mind-wandering → unhappiness, not the other way around.

—— What this means for people who work on their own ——

Our usual solution is to switch tasks: If this project is frustrating me, I’ll move on to something more interesting.

According to this study, the benefit of switching tasks tops out at just 4.6%.

The real leverage isn’t in what you’re doing, but in whether you’re present while doing it.

You may feel better doing the same boring task with your mind present than doing an interesting task with your mind elsewhere.

—— In a nutshell ——

Don’t look for something more enjoyable to do. Just stay present while you’re doing it.

Paper: A Wandering Mind Is an Unhappy Mind
Killingsworth & Gilbert, Science, 2010

I made a copy of a file as a backup.

Then both copies stopped working.

The reason: that file didn’t contain data. It contained an identity.

Copy data, and you have two copies of the data.
Copy an identity, and you don’t get two identities — you get the same identity appearing in two places at once.

Any platform that sees this assumes, first thing, “This account has been hijacked.” So it invalidates the entire credential, and the copy and the original die together.

There are more things like this than most people think:
· Account login sessions
· API keys, authorization files, certificates
· Browser cookies
· All kinds of “remember me” credentials

You think you’re making a backup, but really you’re creating a clone. And the system only recognizes one.

The sneakiest part is that it doesn’t fail right away — everything seems fine when it’s invalidated. You only find out when you actually need it.

So before you copy anything, ask yourself:

Does this contain data, or an identity?

Data can be copied. An identity can only be moved, not duplicated.

This week, someone in the group spent several days trying to reproduce the trading approach of a top address. Each test ran for at least two hours.

The conclusion: it couldn’t be reproduced.

Not just by a little. It was impossible even with no queue and zero latency.

Later, he did something clever: he switched to replay simulation and let himself peek into the future. Since it was a backtest, cheating didn’t matter—he just wanted to find out exactly where the gap was.

It turned out he had to peek at least three seconds ahead to have any chance of catching up.

And three seconds of lookahead is impossible in real trading.

This testing method is worth discussing on its own.

When most people can’t reproduce someone else’s strategy, their first thought is, “I’m not fast enough.” Then they start buying hardware, getting dedicated lines, and optimizing their network routes. But speed is a continuous variable: you never know just how fast you need to be—you can only keep spending money.

Letting yourself peek into the future is equivalent to taking the “speed” variable straight to infinity. If you still can’t catch up, then the gap definitely isn’t about speed; it’s something else.

This was a very clean falsification: without buying anything, he ruled out an entire line of investigation.

In the same discussion, someone else gave an answer that pointed in a completely different direction:

“Honestly, at this point I think canceling orders (avoiding toxic fills) is the most important thing.”

There was also a more practical point, which I think was the most valuable thing said in the group this week:

“The hard part about being a maker is that live trading doesn’t guarantee a fill, so in practice you’re more likely to buy at a higher price—the moment the market drops, your order is sure to get filled.”

At first, this sounds like a complaint. Think about it more, and you realize it’s structural:

Whether your limit order gets filled isn’t random. When the market moves against you, your order is certain to get filled; when it moves in your favor, your order is likely to be skipped.

So the event of getting “filled” is itself directional. The fill price you assume in a backtest and the fill price you actually get in live trading don’t just differ because of slippage; there’s also this layer of selection.

In other words, when you can get filled, it’s often precisely when you shouldn’t.

This also explains why simulated trading and live trading can be so different. In a simulation, an order is filled when “the price reaches it.” In live trading, an order is filled when “someone is willing to sell you the unfavorable side.” The former is a price problem; the latter is a counterparty problem.

There was another bit of context this week: the platform changed two settings in the same week.

TWAP was ramped up, smoothing out mispricing—strategies that rely on mispricing reverting are effectively dead.
Taker latency was cut from 250 ms to 50 ms, benefiting takers—and dealing another blow to signal-driven market making.

The two changes point in opposite directions, but they hit the same group of people. The sharpest summary in the group was: “Changing taker latency essentially weeds out makers doing microstructure arbitrage.”

So here’s where things stand: you can’t catch up on speed (someone has already proved that by peeking into the future); TWAP has closed off the mispricing route; and market makers have taken another hit from the latency adjustment.

What’s left are strategies that don’t depend on any of the above—like knowing which orders not to take.

This week’s full roundup of community highlights is in the channel. It includes real-world measurements showing liquidity shrinking threefold over eight hours, end-of-cycle crowding, and 11 unanswered questions from the group (each one links back to the original message so you can check the context):

When you work alone, the hardest part isn’t having nothing to do—it’s opening your computer in the morning and not knowing where to start.

I eventually realized the problem was this: I was trying to do two completely different things at the same time.

One is **keeping an eye on the world**—what’s happening out there, whether anything relevant to me has changed, and what’s worth following. This is about filtering: the input comes from outside, and the result should be a list of “0–3 things worth reading.”

The other is **managing my own work**—what I have in progress, what’s stuck, and which item to move forward today. This is about prioritizing: all the input comes from my own backlog.

When you mix the two, you end up scrolling through information while worrying that you should be working, and you do neither well.

Once you separate them, the order is fixed too: **look at the world first, then manage your own work**. Doing it the other way around causes problems—you plan today based on yesterday’s judgment, then discover that things have changed and the work you scheduled is no longer worth doing.

What’s worth borrowing is this division of labor, not anyone’s specific workflow. Your two activities might each take five minutes, but they have to be two separate things.

When you ask AI to handle something for you, how do you get past the password step?

I’ve seen three approaches, and their risks are worlds apart:

1. Paste the password straight into the chat—it’s the easiest, and the worst. That message becomes part of the context and logs, may end up in training data, and you can’t take it back.

2. Have the agent read your password manager—it sounds professional, but the agent gets the password in plain text, and you have no idea where that context goes afterward.

3. The agent only initiates the request. You confirm it in the password manager’s own interface, and the value is filled directly into the page. The agent only sees “authorized / not authorized” the whole time—not the value itself.

Only the third approach is sustainable for long-term use. The test is simple: **Does the plaintext pass through the AI’s context?** If it does, it’s not secure, no matter how many layers it’s wrapped in.

One more basic rule: For operations that require a password, try not to have the agent do them for you. It’s enough for it to get you to that step. The time you save isn’t worth trading away your account. ⚠️

Have you filled in the link field in the upper-right corner of your GitHub repository?

I just checked my own 17 repositories. Sixteen were blank. That’s the only outbound link on a repository page—and with 2,352 stars, the traffic I was sending to my own website was zero.

Same goes for topics. These aggregator pages are a major way people discover tools. If you haven’t tagged your repo, it won’t show up there—you’re effectively leaving a free source of traffic switched off.

You can take care of both in about ten minutes:

· Settings → the gear next to About on the right → enter your website under Website
· In the same place, add 5–8 topics that reflect what your repo actually does. Don’t just copy a template.

Be specific with your tags—don’t add too many, but don’t be lazy either. Specific terms like mcp / claude-code / prediction-markets are much more useful than broad ones like ai / tools. You’re never going to rank on the aggregator pages for those broad terms.

These omissions don’t trigger errors, and the repo looks perfectly fine, so they can go unnoticed for a long time. Go check yours now.

AI’s “it got away from me” moment!

I only wanted Codex to draft something for me.

Instead, it went ahead and posted the tweet. 😂

Earlier, I’d asked it to research a Cloudflare Wallet handle based on my brand and existing projects. After I confirmed the login and authorization, it automatically checked and claimed it.

Then I said, “Share it in a tweet.”

I meant: write a draft first so I can review it.
It understood that as: open X and publish it right away.

After posting, it even fucking verified the result automatically and wrote the link and status back to my content system.

I stared at the screen:
Wait, what? …Oh! 😳

Kind of exciting, but also a little terrifying when you think about it...

There isn't much time left for @OpenAI

We’ve claimed the Cloudflare Wallet brand name first:

☁️

Currently, each Cloudflare account can claim only one handle. If you want to reserve your brand name, you can check it out first:

Try this prompt with your own AI:

【Let’s talk about the kind of person I want to become ten years from now.

Don’t just focus on abstract aspirations—help me think through my finances, work, health, and relationships.

From now on, guide me through a weekly review: Did what I did this week bring me closer to that vision? If I’ve gone off track, tell me the one thing I should adjust most next week.】

I haven’t actively scrolled through the X feed for a while.

The real problem isn’t that there’s nothing worth reading. Quite the opposite: there’s too much valuable information, and too much noise.

I’d start out just wanting to see what was new, but half an hour later, all I’d remember were a few trending topics I wasn’t interested in. Not a single thing genuinely relevant to my work or business would stick.

So I changed how I “scroll X” and turned it into a continuously running information-processing loop.

It runs four rounds a day. I assign these weights to three types of sources: 60% For You, 25% Following, and 15% active search.

After it collects content, it doesn’t immediately ask AI to summarize it. First, it checks whether the sources are healthy and the data is fresh. If something goes wrong, it explicitly reports a failure instead of passing off stale data as “there’s nothing important today.”

Only then does it move on to deduplication, relevance filtering, evidence verification, and semantic review.

In one anonymized real-world run, the sources produced 142 hits in total. After deduplication, 80 distinct items remained; 23 passed the initial machine screening, 12 went to semantic review, and in the end, only 3 were marked “worth knowing,” with 1 receiving “formal promotion.”

142 source hits → 80 distinct items → 23 initial screenings → 12 semantic reviews → 3 worth knowing → 1 formally promoted.

What this funnel really helps me solve isn’t “how aggressively to filter,” but how to separate two questions that used to get lumped together:

What’s worth knowing?
What’s worth changing the system’s state for?

At the knowledge level, items are marked “worth knowing”: 0–3 per round. The system has to tell me why each item matters, how far the evidence verification got, what’s merely the author’s own claim or an inference, and how it could be useful from a business, product, research, or workflow perspective.

At the system level, an item can be “formally promoted”: at most 1 per round. It only moves into a knowledge, observation, action, or content system if there’s enough evidence, it adds something genuinely new compared with existing memory, there’s a clear consumer for it, and it leaves a verifiable receipt.

The other two items don’t disappear, but they also don’t automatically turn into two tasks or two to-dos just because they “seem interesting.”

Now, an item of information can go to one of five places: discard, observe, learn, act, or create content.

The system only recognizes that something has truly moved to the next stage when it receives an actual write-back to knowledge, task routing, or a content queue. AI saying “done” doesn’t count.

For now, the reports from all four rounds are delivered directly to the Codex App I open every day.

I later realized that this kind of tool probably shouldn’t start out as yet another app you have to open. After all, the places where I actually use this information are AI clients like Codex and Cursor.

So I built a local MCP alpha with six tools:

• 5 read-only tools: freshness check, latest summary, summary by date, limited search, and evidence retrieval
• 1 write tool: records genuine feedback only after explicit confirmation

This MCP doesn’t collect data from X or create a new ledger, queue, or scheduler. It’s simply a very narrow entry point for AI clients.

Codex already recognizes this local service in practice. Cursor’s configuration and paths have been validated, and native client loading is still undergoing final verification.

The underlying engine also isn’t tied to any model. Source health, deduplication, ranking, limits, and receipts are handled by deterministic code. Codex, Claude, Cursor, a local model, or even a purely rule-based program can handle the semantic judgment in the middle, as long as they follow the same protocol.

I now think what these kinds of AI information tools are really missing isn’t faster summarization, but four things:

1. Check whether sources are healthy before talking about intelligence.
2. Allow zero output; don’t force conclusions just to prove the system is useful.
3. Separate “knowing” from “changing state.” Some things are enough to know about and don’t deserve a task.
4. Leave a receipt at every step. Without an actual write-back, there’s no closed loop.

This version is still a local-first alpha. I’m in no rush to turn it into a SaaS product. First, I want a few people who aren’t the author to test three things:

1. Can they quickly generate their first summary?
2. Can personalized results remain useful for different people over time?
3. Does it genuinely reduce the time spent actively scrolling through the feed?

If you’d like to try this kind of information processor in your own AI client, leave a comment saying “I’m interested in trying it.” I’ll use the first batch of real feedback to decide whether to package it as a public tool or an open-source project.

There’s nothing wrong with using AI. What matters is taking responsibility for what you put out.

It’s not that anything with an AI feel is unbearable. The main thing is that you put care into it and genuinely want to offer readers some value.

After reading how Baoyu (@dotey) turned the Feynman learning method into a loop, I realized that the content workflow I’ve been tinkering with lately is really the same thing—with an extra layer of AI and spoken narration in the middle.

My workflow used to be very simple: I’d have an idea, ask AI to structure it, make a few edits, and publish. It was pretty efficient. But then I noticed an obvious problem: the content was complete, yet it didn’t sound like it came from a real person.

This was especially true when AI organized spoken notes directly. It would produce a title, bullet points, and a conclusion, but the order I spoke in, the pauses, and the judgments that came to me on the spot—the most human parts of my tone and emotion—were easily smoothed away.

So I changed the process into a closed loop:

1. A genuine thought or problem — It has to be something I’ve actually experienced or genuinely need to address: what I’m spending money on, where I’m wasting time or hitting snags, or what I’m stuck on.
2. Test it in practice — Try it out first, keep a record of the process, then have AI organize it.
3. Revise it through speaking — Explain it in my own words and put the human voice back in.
4. Publish across platforms
5. Feedback — Comments and questions feed into the next round, completing the loop.

That’s what the Feynman loop looks like in content creation: it’s not “learn it, then share it.” Sharing forces you to really understand it. The only extra steps are AI handling the organization and spoken narration keeping the human in the picture.

The hands-on test I wrote about a couple of days ago—using Codex to control ChatGPT Pro—is this loop in action, not some motivational fluff bolted on afterward.

I’d forgotten to take screenshots earlier. To make the piece solid, I had to run through the process again and capture the missing visuals. As I re-tested it, my own judgment about the whole thing became clearer, too.

Writing isn’t just about recording the result. It pushes you to put things into practice, verify them, and gather evidence—and find out whether you really understand them.

In this workflow:

AI organizes information, trims repetition, and adapts it to each platform.
People bring real experiences, make the final judgments, and add their personal touch.

Once AI gives me a draft, I read the whole thing through. If it sounds like AI, read it aloud—you’ll immediately catch the awkward bits and the things you’d never actually say.

One small suggestion: give the speech-to-text transcript back to AI, and ask it only to smooth things out and fix errors. Don’t let it add opinions or force a grand takeaway.

This isn’t about making writing more complicated.
It’s about using AI to work more efficiently without erasing the parts that make your work uniquely yours.

Ondo Perps cumulative trades break $6B, enter the RWA perps Top 5; in the last 24h there are another $330M.

Same theme: tokenized stocks can be used directly as margin, without first converting into stablecoins to open a position.

I tested and broke down the discount rate / account limits in practice:

If you don't feel like reading, I talked through the first half—

what that 98% queue is all about,
and why being “at the front” is actually the most dangerous spot on the list.

The full four-tier breakdown and three rules are in the original post.

Polymarket is switching Crypto Up/Down markets at 5min / 15min / 4h to TWAP settlement, and the group spent the last couple of days arguing about one thing: is it TWAP for the entire period, or just the final 30 seconds? Two people dug in, and neither convinced the other.

I went and checked the official announcement. The answer is that it depends on the market duration:

5-minute markets → 30-second TWAP
15-minute markets → 60-second TWAP
4-hour markets → 60-second TWAP

The official wording is “with the averaging window scaled to the length of the market”—the window scales with the market duration.

One more important thing: this doesn’t take effect until August 7 at 00:00 UTC, so nothing has changed yet. The mainnet feed and RTDS go live on August 4.

So in that argument, each side was only half right: it is 30 seconds, but only for 5-minute markets.

One related point nobody in the group mentioned, which I think is more practically important than the settlement method: in the same announcement, the team said it would distribute $1 million in liquidity rewards across all affected markets throughout August to support the transition. In other words, LP returns will look different from usual for that month.

The timing matters: the late-stage reversals people have been seeing these past few days are still happening under the old single-point snapshot settlement. You can’t use them to infer what things will look like after TWAP.

Back to why they’re making the change. Someone in the group had specifically tallied this up: the reversal rate in the final 30 seconds is 13%—the probability that the side in the lead 30 seconds earlier ends up losing. He defined being ahead quite specifically: for example, if PTB is up +100 (UP is leading) and settlement comes in at -30 DN, that counts as a reversal.

In that environment, his conclusion was, “No matter how right you are about the direction or how high your win rate is, you’ll eventually get wiped out by a reversal,” so “there’s really nothing to do except hedge.”

So wouldn’t fixing it solve the problem? Someone in the group didn’t think so. This is the part I think is most worth remembering:

“Honestly, it’s not so bad if they don’t switch to TWAP. Sure, there’s manipulation, but the bigger the waves, the more valuable the fish. Our quotes already price in that risk.”

“If there’s no manipulation, nobody will bet on reversals.”

His concern wasn’t that his own strategy would stop working, but that nobody would play the market at all: “The biggest harm this thing causes isn’t to any one strategy—every strategy needs retail traders to provide the other side of the trade.”

Someone else put it more bluntly: “Without degenerate gamblers and retail traders, it’s just a bunch of bots fighting each other. Who’s going to make money? The alpha is gone.”

The full-period TWAP he was worried about isn’t what’s happening this time (the windows are 30s / 60s, and the chart will still move as usual). But the trade-off itself is real and worth remembering:

Price volatility is part of the product that attracts participants. A design that smooths out volatility doesn’t just weed out manipulators; it also drives away people willing to bet—and market-making and arbitrage strategies make money from those people.

Here’s how I’m handling the switch myself.

The day the announcement came out, I first checked my real-money exposure. It was zero. What actually got interrupted were a few validation pipelines still running on data but not yet deployed with real money—their samples will span August 7. One of them was especially unfortunate: it replicated exactly the kind of “enter in the last 30 seconds” strategy, and the new window for 5-minute markets happens to be exactly 30 seconds, directly overlapping its entire premise.

So the first thing I need to do isn’t change the strategy; it’s draw a time boundary.

Samples from before and after 8-07 can’t be analyzed together. A sample pool that crosses that date will produce a meaningless average—and won’t throw an error. The statistical program will run as usual; its conclusions will just quietly stop meaning anything. That’s the easiest risk to overlook in this transition: a rule change won’t crash your program; it will invalidate your conclusions.

Second, don’t look at just one side of the equation. It’s intuitive to think smoother TWAP means better predictability, but it has two opposing effects: predictability may increase (an average price is more stable than a single tick), while the amount you can capture may decrease (smoothing also narrows the spread). Whether the net effect is positive or negative depends on which one drops faster. You can’t figure that out by reasoning alone; you have to test again.

Third, put time boundaries on existing research. Some conclusions rest on the core evidence that “the final N seconds are a manipulation window.” That pattern assumes single-point snapshot settlement. It may still hold after 8-07, or even hold more strongly, but the basis for the argument has changed. Conclusions get carried forward; their assumptions don’t. A conclusion without a time boundary can still be cited as evidence six months later.

I’ve put the full version on the site. It also covers why copy trading can be self-defeating, and another definition that wasn’t clarified in that 13% statistic.

📄
📡
🎯

Everyone online says you need 5,000 followers to join Xiaohongshu Pugongying. I got in today with 1,210.

I’m documenting the whole journey, from starting an account to being able to take brand deals. It could save you a lot of trouble.

1｜The threshold isn’t 5,000
The exact message in the backend says: “You’ve reached the 1,000-follower milestone and can now use these benefits: Buyer Collaborations / Creator Collaborations.”
You unlock it at 1,000 followers. 5,000 is the next tier, where rates double. Somehow, people started saying it was the entry requirement.

2｜The entry point is really hard to find
There’s no entry point in the Creator Services Platform (the web dashboard). I spent ages looking there first.
There are two ways to get in:
· App: Sidebar → “Professional Account” → Business Growth → “Creator Collaborations”
· Web:

3｜How I got those 1,000 followers: let the data pick your focus—don’t choose for yourself
At first, I posted both types of content. Then I checked the dashboard: prediction market content brought in 850 followers in total; AI content brought in 9.
That’s 94 times as many. One viral post alone got 46,910 impressions and 837 new followers—more than all the others combined.

My main focus on X is AI, but on Xiaohongshu it’s the complete opposite. The audiences on the two platforms aren’t the same.
The assumption that “you should pick one core content theme and distribute it across every platform” was disproved by my own dashboard data.

4｜Getting approved ≠ getting deals
My dashboard currently shows: Awaiting acceptance 0 / In progress 0 / Awaiting confirmation 0, with monthly income at ¥0.
It’s an entry ticket, not income. Setting rates and getting approached are the next steps. I’ll write about them once I’ve done them.

5｜Rate reference
Real rates from similar creators recommended on the platform’s homepage: 31,000 followers, starting at ¥1,000; 14,000 followers, starting at ¥400; 11,000 followers, starting at ¥500.
Official range: 1,000–5,000 followers: ¥100–300 for image-and-text posts / ¥200–600 for videos; rates double above 5,000.
Niche accounts can charge a 2–3× premium. The platform takes 10% from both brands and creators.

My next hurdle is Bilibili’s charging feature. I still need 1,000 followers and real-name verification. I’ll write another post once I qualify.

When people look at equity perpetual contracts (Perps) platforms like this, they basically focus on one thing: can you add leverage and trade U.S. stocks?

But after I broke down the mechanism of @OndoPerps, I found that what gets changed isn’t actually the leverage side—it’s the collateral side.

What’s even more interesting is that the first beneficiaries of this change aren’t retail traders; it’s market makers. Retail traders get the second-order result of this.

1. The pain point: market makers’ double-collateral problem

All RWA perps venues have a structural issue: market makers must post double collateral.

They need to post an amount of stablecoins as margin on the perpetual contract side; at the same time, they also have to keep a pool of cash and a stock inventory at a traditional brokerage to hedge.

With the same economic exposure, the market maker’s capital is effectively locked up twice. The result is a thin order book, wider spreads, and slippage that doesn’t match the real market. This isn’t because any one venue is doing poorly—it's a ceiling inherent to the entire model.

2. The solution: tokenized securities as collateral directly

Ondo Perps’ solution is to let tokenized securities serve as the collateral directly.

For example, if a market maker wants to quote the NVDA perpetual, they can directly hedge with NVDA tokens within the same system—no need to open a separate brokerage account.

Once the funding base is unified this way, spreads may narrow, and the order book depth can better reflect true liquidity.

So when judging whether a platform is worth looking at, you can’t only ask how much leverage it can offer. You need to check whether the order-book quality has really improved. For you, the practical question is: can you benefit from it? That depends on whether market makers are willing to quote tighter prices here.

3. Key parameters and limitations

I looked up Ondo’s official documentation. Currently there are two collateral underlyings: SPYon and QQQon—tokenized versions of the S&P 500 and Nasdaq 100, respectively. Both are Ondo Stocks tokens on Ethereum. Because they aren’t individual stocks, you can’t temporarily use the specific stocks you hold as collateral.

Haircut (discount rate):
· USDC is 0%
· SPYon and QQQon are 10% each

What 10% means: if you deposit SPYon worth $1,000, the system only credits you with $900 of margin—but it will lock all $1,000 of your deposit. The $100 difference is used to absorb slippage or price movement risk in case the collateral needs to be converted into USDC.

Collateral limit (this is easy to overlook):
· SPYon: 146 units per account
· QQQon: 166 units per account

If you exceed this amount, you can still deposit and withdraw, but the system will no longer increase the margin. Roughly based on the current price, the stock-collateral-based borrowing capacity per account is on the order of several hundred thousand dollars.

One more thing to make clear so there’s no misunderstanding: USDC has no limit, and the haircut is 0%. So what the limit restricts isn’t how much you can invest—it’s the path of “using stock tokens to offset margin.” If you want to participate with a large size, the USDC route works.

As for why the limit exists: my understanding is that when the platform liquidates, it must convert collateral into USDC. But the secondary-market depth of this kind of token is limited. As of now, the total tokenized stock collateral across the whole platform is $27.34 million—so the market is still not very large. Therefore, this limit is essentially the platform’s self-imposed constraint on its liquidation capacity. The official docs also mention that it will adjust with market conditions.

4. Platform performance so far (DefiLlama, checked on Aug 1)

· 24-hour trading volume: $233 million
· 7-day trading volume: $1.57 billion
· 30-day trading volume: $4.38 billion
· Total trading volume: $6.0 billion
· Open interest (OI): $74.67 million

Worth noting: out of the total $6.0 billion trading volume, $4.38 billion came from the most recent 30 days—about 73%. Most activity has occurred in the last month, and growth is indeed accelerating.

5. Fees and incentives

Trading fee is 0.035%. If you enter via a referral link, the fee can be discounted to 50%. On the order panel, you can directly see “save 50%.”

In addition, there will be a weekly USDC incentive pool and Points. Points currently has no fixed price, so I’m not counting it as part of returns. The exact rules depend on the official App and the documentation; they will be adjusted.

6. Entry requirements

1. Region: the platform does not provide services to the U.S. and other restricted jurisdictions. Confirm local availability before using it.

2. Double exposure: your collateral itself will move in value, and your position will also move. This is not the same as simply using stablecoins as margin. The liquidation logic and haircut are worth taking time to understand first.

I also placed a small amount of my own funds into it to try it and get a feel for how it works.

If you want to research it yourself, you can use this link:

The above is my personal sharing based on my research on this platform and does not constitute any investment advice.

This week’s Leo Insider is about weather strategies. The focus is clear: research process, validation table, and risk-control checklist.

I want to break down a more fundamental question: where might the money in weather markets actually come from?

I’ve divided it into four layers:

1. Gaps in market coverage;
2. Gaps in information updates;
3. Gaps in understanding settlements;
4. Gaps in player flows.

The full memo will cover:

- How I build a candidate pool;
- Why the challenge of weather strategies isn’t just predicting the weather;
- How player flows can serve as an external confirmation layer;
- Why, even with clues pointing to alpha, you still need to hold back before scaling up;
- A “niche strategy alpha validation table.”

This issue is useful for one thing: the next time you come across any niche strategy, don’t rush to ask whether it can make money. First break down where the money comes from, where the evidence is, and where the risks are.