Jeonlees

认认真真撸毛

981Following

1.5Kfollowers

Feed

Holding ETH or BTC and don't want to sell for now, but need some stablecoins, what to do? Selling your assets directly at this time feels awkward. Your original holding plan hasn't changed; you just temporarily need liquidity. Once you sell, you'll have to reconsider when to buy back later. f(x)'s fxMINT offers another approach: Collateralize assets, mint fxUSD, and repay the debt and retrieve the collateral once your funds turnover is complete. The official underlying collateral currently includes wstETH and WBTC; WETH entries will be converted into wstETH. ✰ ꙳ ☾.˖✰ ꙳ ☾.˖ Currently, the opening fee for ETH and BTC positions is 0.5% of the debt, and the closing fee is 0.2% of the repaid debt. Under normal circumstances, there is no continuously accumulating annual borrowing interest. Assuming you mint 5,000 fxUSD, the opening fee is about 25 fxUSD, and then repaying the same amount of debt incurs a closing fee of about 10 fxUSD. In other words, just looking at these two fees, it's roughly 35 fxUSD. So borrowing for a few days versus a few months feels completely different. Short-term turnover may not be cost-effective, but if the usage period is longer, the fact that there is no ongoing annual interest accumulation makes it worth comparing. Of course, actual costs also include Gas, exchange fees, and possible temporary borrowing costs under special market conditions. ✰ ꙳ ☾.˖✰ ꙳ ☾.˖ Regarding position risk. For example, if the collateral asset is worth $10,000 and you borrow $3,000, the initial LTV is 30%. If the collateral asset drops 30% in value to $7,000, and the debt remains unchanged, the LTV will rise to about 42.9%. You haven't borrowed an extra cent, but the position risk has already changed. f(x) has designed a rebalancing mechanism for such situations. Once a specified threshold is reached, the system may liquidate part of the collateral and reduce the debt to lower leverage. This mechanism can indeed help adjust position risk, but it also means that after rebalancing, the amount of collateral you hold may decrease. So it cannot be simply understood as "no liquidation risk." Another easily overlooked cost: wstETH itself is an interest-bearing asset, but according to official documentation, the yields generated by fxMINT collateral are mainly distributed to stablecoin stakers. The borrowing position retains the price exposure of the corresponding asset but does not continue to receive this portion of the yield. If you originally hold wstETH long-term, the yield you give up here should also be counted as part of the borrowing cost. ✰ ꙳ ☾.˖✰ ꙳ ☾. Regarding fxMINT's "0% annual interest," it's recommended to calculate whether this suits you. How long to borrow, how much to borrow, and how you plan to repay are more important than just looking at the interest rate. Because over a longer period, the one-time opening and closing fees do have an advantage; but if you only borrow for a few days, considering Gas and other costs, it may not really be cost-effective. Also, position space matters. After borrowing, ETH and BTC prices will continue to fluctuate, and LTV will change accordingly. Rebalancing can handle some risks, but the cost may be a reduction in collateral quantity, which should not be ignored. So I think the most interesting aspect of fxMINT @protocol_fx is not simply lowering "borrowing interest," but giving long-term ETH and BTC holders another option: they don't have to rush to sell assets and can still access some liquidity first. Whether it's worth using ultimately depends on each person's holding cost, borrowing duration, and tolerance for price volatility. This is only a personal research and user experience sharing, not any investment advice. Remember to DYOR.

Many people see Backpack and still think of it as "an exchange issuing a token." But after BP went live, Backpack actually faces even more challenges. What Armani needs to deliver is not just new features for the exchange, but also integration with brokerages, securities access, and a complete connection between traditional assets and on-chain assets. The Founder aspect is worth discussing, and here’s why: After issuing the token, how much interest do founders still have that depends on running the company well to realize value? In Backpack’s announced mechanism, founders, the team, and investors have no initial BP allocation, nor a separate team token share. The founders’ company equity still relies on business growth to realize value. This gives the concept of "long-term interest alignment" something concrete, not just statements in interviews. Official BP mechanism: Looking at its current product direction, you can also see how far this company wants to go. Exchange, brokerage, wallet, real US stocks, tokenized stocks, perpetuals—each alone could be a business. Backpack wants to connect them all, aiming to cover the entire financial needs within a user’s account. The commercial value of this is much greater than "just launching another product." Users trade crypto today, buy US stocks tomorrow, then transfer assets on-chain for use. If the experience is smooth enough, they have reason to keep their account with Backpack. Company growth can come from more assets and more use cases, without having to find a new batch of users every time. The so-called "Silicon Valley company" story, I think, should be told like this: Continuous product iteration, continuous accumulation of user relationships, and the earlier business providing a foundation for the later business. The longer the company operates, the more valuable these accumulations become. Of course, at this stage, the most important is still Exchange + Brokerage. First, get trading, liquidity, and fund flows right, so users want to use it and keep assets there, then talk about a bigger map. This priority is pragmatic. "Whatever you want to trade, however you want to trade, you can use Backpack" sounds like a slogan but actually demands a lot from the company. Behind it requires long-term investment in licenses, custody, clearing, risk control, and technology. Whether the founder has patience and will continue investing will be reflected in the product over the next few years. The market is starting to evaluate Backpack over a longer timeframe, and the founder must continue proving the company is worth holding and using long-term. Making Backpack a pillar of the global economic system is where Armani wants to go. And the current path must start with every delivery after the token issuance.🎒

A robot works, fails eight times out of ten, what would you do? Axis recently conducted an experiment: let it keep working, keep the few successful operations, and then use them to train the next version of the model. In the end, the success rate increased from 22% to 52%. I saw @axisrobotics and found it interesting at this point. Previously, the most notable thing was that it collected nearly 7 million robot trajectories. But what I want to know more is whether this robot with a still not high success rate can help people figure out: what data exactly should be collected next to enable it to accomplish more tasks? This is also why I recently revisited Axis. The initial Axis V1 approach was straightforward: the system issues tasks, humans remotely operate, after data validation, the data is used to train the model. The advantage of this method is practical. It collects data quickly, at low cost, and participants don’t need to have a robotic arm at home. Axis Hub has already accumulated about 6.91 million trajectories and 222,000 users. But the more data collected, the more a problem arises: how much can the robot still learn from the data collected later? For example, "putting a cup into a basket" has already collected 10,000 trajectories. If the next one is just a similar environment and similar action, continuing to collect will certainly make the numbers look better, but how much the model can learn from it is another matter. You can’t just assume that more trajectories mean the robot is definitely more capable. Axis V2 started trying a different collection method. They added human-gated DAgger, a somewhat complicated name, but the method is easy to understand: The robot tries by itself first, with a human watching nearby. If it can do well, let it continue; if it’s about to fail, the human takes over and corrects the action. Previously, humans demonstrated the entire task from start to finish; now, they watch where the robot gets stuck and help just at that point. I think this change is quite practical. For parts the model already knows, there’s no need for humans to repeat the work. What’s more worth keeping are the moments it almost messed up and humans stepped in to save it. However, human intervention doesn’t necessarily mean the data is useful. Axis published an experiment: out of 660 human-corrected data points, only 161 passed closed-loop validation and entered training. This result actually made me want to look more closely. They didn’t just stuff all the painstakingly collected data in. As for whether the filtering was good, in the end, it depends on whether the model’s previously error-prone areas improved after using this data. Their proposed Compounding Data Engine, I also started to understand from here. First, let the model run and see where it fails; then arrange tasks around these problems, have humans fill in the parts it can’t do; after training, put it back to test. Sounds simple, but the hard part is making each round effective. You can’t just be busy collecting data and still have the robot make the same mistakes. The Grounded RSI experiment, where the success rate rose from 22% to 52%, is also an attempt along this line: first let the policy run successful rollouts in the real environment, then use this data to train the successor policy. This improvement is worth noting, but I want to know what happens next. With a different set of tasks or environment, can it still achieve this? How much more can the success rate increase with continued training? These questions tell more than a single experiment’s numbers. Looking at their recent collaboration with OpenRoboto on the Open Axis Benchmark, you can see the evaluation is moving in this direction. Using a fixed question bank for a long time, everyone gets more familiar with the tasks, so scores naturally might get higher. But whether the robot can handle unseen tasks is sometimes hard to judge just by the score. Open Axis Benchmark rotates test tasks from the ever-expanding Axis task library. If the evaluation finds the robot consistently fails a certain type of task, the next batch can collect more data of that type, train, and then test again. This at least gives a chance to clearly see whether the previous problems are solved after adding this data. Of course, changing tasks isn’t a cure-all. How to control task difficulty and compare scores across rounds still need to be watched closely. So now, looking at @axisrobotics, my interest in the "Web3 + Robotics" label isn’t as big as at first. I’m more concerned about one very concrete thing: when a robot team brings a not-so-good model to them, can Axis help find the problems, collect the right data, and make the model a bit more usable? If they can do this reliably, customers have reason to keep paying. This reminds me of Scale AI. But robotics is more complicated: just because the data itself is good doesn’t mean the trained model can perform well in the real world. Changing the robotic arm, environment, or slightly shifting object positions can all cause issues. Axis Dataset V1 has already shown some early results. π0.5 on LIBERO-Plus improved from 83.9 to 88.8. In Booster Robotics’ case, using a large amount of simulation priors plus about 30 real machine demonstrations, the success rate went from the out-of-the-box π0.5 of 37.5% to 87.5%. These results are worth following, but they don’t yet prove the business model is fully viable. How much of what’s learned in simulation transfers to the real machine? Can the original data be reused if switching to a different robot? How to ensure quality when crowdsourcing grows? After the model gets stronger, can it still find valuable data to supplement? There’s still a lot to watch going forward.