Author|@flirting with models

Translation|Aki Wu on Blockchain

The guest of this episode is Jeff Yan, the founder of Hyperliquid. Jeff's career began in high-frequency trading at Hudson River, and he later transitioned to the crypto world, where he built one of the largest market makers in the space. He delves into the infrastructure of centralized crypto exchanges, adversarial algorithms, and why the profits and losses of HFT (high-frequency trading) may actually be predictive of mid-term price trends. He explains the current issues he sees with decentralized exchanges and introduces the early concepts of Hyperliquid. This episode was released on May 8, 2023, showcasing many of Jeff Yan's early thoughts.

How to Transition from Harvard to Crypto Trading

Jeff Yan: My experience is probably similar to many HFT practitioners: I graduated from Harvard University with a major in computer science and mathematics, and then directly joined Hudson River Trading, which is a large market-making firm in traditional finance. I was working on U.S. stocks, and it was a great experience. When I joined, the company had about 150 people, and it has grown significantly since then. I benefited immensely from being exposed to the most interesting problems, where engineering and mathematics could perfectly combine, which was almost "heaven" for quant. By 2018, with the surge of Ethereum smart contracts, I read the (Ethereum) Yellow Paper and instantly had an epiphany; I was convinced it would be the future, so I left to work on a trading protocol in the L2 direction.

At that time, we chose to enter the prediction market space because Augur was already showing strong signs of product-market fit (PMF), and we were more skilled and focused on the underlying technology of exchanges. After completing our funding, we moved to San Francisco to build a team. However, a few months later, I decided to shut down the project because the timing was not right: on one hand, regulatory uncertainty was extremely high; on the other hand, user acquisition was very difficult. Most people were not familiar with smart contracts, and their interest was more focused on token speculation, so the real demand for DeFi had not yet formed, leading to the project's eventual suspension.

I then spent some time reflecting and traveling, ultimately choosing to return to trading. Compared to continuously "searching for PMF" in the market, trading itself is more direct and interesting. Initially, I considered joining a mature company, but thinking about my experience with crypto products and my familiarity with industry mechanisms, I started with proprietary crypto trading. At first, it was just a side job, but soon I saw significant opportunities, and the business expanded much faster than expected. I was surprised by the inefficiency of the market. After that, I almost immersed myself in it for nearly three years: the real systematic launch began in early 2020, coinciding with the market cycle. As the market size and trading volume grew by 10 times or even 100 times, we also scaled up, ultimately entering the top tier of market makers in centralized exchanges (CEX).

About a year ago, we began systematically evaluating DeFi trading opportunities. This is similar to the observations made when we first entered CEX trading—inefficiencies were widespread. However, the difference is that some DeFi protocols have inherent design flaws that limit trading experience and capital efficiency. Meanwhile, after the FTX incident, the market's awareness of "not your keys, not your coins" and counterparty risk significantly strengthened, leading to a continuous rise in demand for truly decentralized products. Based on these changes, we judged that the window for building decentralized exchanges had arrived. Over the past one to two quarters, we have continuously invested resources to advance this direction; the high-frequency trading (HFT) business is more in a relatively stable operation and maintenance state, while our current main investment and focus are on solidly implementing this decentralized trading technology stack and completing its systematic construction.

Market Making vs. Taking Orders: What’s the Difference?

Jeff Yan: In my view, this is indeed a significant decision that needs to be made when entering high-frequency trading. From a macro perspective, there are many similarities between the two: both have extremely high infrastructure requirements and are highly sensitive to latency. However, in many key aspects, they exhibit opposite emphases: market making relies more on infrastructure capabilities; taking orders relies more on statistics and mathematical modeling.

I believe the choice of which path to take mainly depends on what type of work and research you prefer to engage in. For example, in market making, you are somewhat constrained by the counterparties that "hit your quotes," leaving very little room for error. Typically, leveraging and placing orders across multiple products and price levels can create significant implicit risk exposure; once an error occurs, the cost of tail risk is often very high. In contrast, taking order strategies can trigger only once a day—still potentially an effective high-frequency strategy, which may be based on news or certain segmented signals.

Because triggers are less frequent, you have the space to refine your models: most of the time, not triggering does not matter, as long as the performance is good enough when it does trigger. Conversely, market making does not have this flexibility—even if it runs well 99% of the time, if it is slightly slow or fails to follow the data in that 1% of the time, the related losses could be enough to wipe out the entire PnL of the other 99%.

Corey Hoffstein: Can we understand it more intuitively this way: the party choosing to "take orders" is willing to cross the bid-ask spread because they expect the price to continue moving in their direction, thus willing to bear the spread cost; while the party "making the market" hopes the price remains as stable as possible within their trading time window—when someone crosses the spread and executes with them, they then complete a hedge or reverse transaction on the other side to capture the spread profit. Is this distinction reasonable? One party prefers the market to be sideways within their time window, while the other prefers directional movement.

Jeff Yan: Yes, that’s basically how to understand it. In high-frequency trading, we usually assess markout (lookback returns) over very short time windows, but this judgment also holds for more general trading frequencies: as long as you choose to "take orders," at the moment measured at the mid-price, you will immediately incur a certain loss (spread and fees). Only if, subsequently, within your set prediction time window, the average price movement can cover this immediate loss and further compensate for the fees, does your strategy have a positive expectation.

Market making, on the other hand, is exactly the opposite: at the moment of execution, your "initial PnL" is often at the highest level that the trade could achieve—because you just captured a spread. What you are betting on is that this portion of profit will not be completely eroded by "adverse selection" on average.

Therefore, in market making scenarios, if you observe all transactions by time dimension for markout, it is usually more common for PnL to decrease over time; your expectation is simply that the rate of decrease does not turn negative.

Corey Hoffstein: Before our call, you mentioned that scaling the business is most challenging not in research but in infrastructure. I also saw a similar statement from you on X: "Knowing how to normalize data does not guarantee you can make money, but not knowing how to do it guarantees you won't." Can you talk about your biggest lessons learned in infrastructure and why it is so critical?

Jeff Yan: This question can roughly be divided into two parts, which are closely related: one is "trading infrastructure," and the other is "research infrastructure." Data cleaning leans more towards the latter, being part of statistical practice; the former is the narrow definition of high-frequency trading systems. Both are extremely important.

While the research aspect is more well-known, it is important to emphasize that the "signal-to-noise ratio" and noise patterns in high-frequency trading are several orders of magnitude worse than most subjects in academic research, making the handling of outliers much more critical.

If there is a lack of a proper handling framework for these issues and you simply ignore outliers, then once a black swan tail event occurs, the model may be directly breached; but if you fail to standardize or filter properly, extreme samples will dominate model training and parameter selection. In practice, using quantiles in many tasks is often more robust than directly using raw values; even when using raw values, you need to make a clear trade-off between "discarding outliers" and "trimming outliers," and these choices often have a significant impact on the final results.

The biggest lesson sounds very simple: you must personally review the data. Don’t assume you are smart enough or that your pipeline is "clean" enough for model inputs to automatically meet expectations. The time spent checking raw data can hardly be considered "too much"—because each review almost always brings new discoveries. In the early stages, the team should fully record all raw data streams provided by the exchange, check them one by one, actively identify anomalies, and perform consistency checks.

There was a seemingly absurd but real case: at one point, a trading firm had a defect in its market data feed, swapping the "price" and "quantity" fields. For example, Bitcoin's 20,000 / 0.1 was recorded as 0.1 / 20,000, leading to a complete distortion of our internal statistics and counting logic. Many teams had to urgently shut down or switch to backup data sources. Such events illustrate that no matter how "robust" your logic is designed, it cannot cover all exceptional cases, so you should maintain as close a relationship with the raw data as possible and ensure traceability.

Additionally, you must pay close attention to timestamps. Exchanges often provide multiple timestamps in the data, and their true meanings need to be unpacked and aligned. This is particularly crucial for understanding "black box latency"—what exactly are you measuring? Are you truly "keeping up" with the market, or is the other party pushing lower-quality data? By breaking down and comparing timestamps, you can better distinguish these situations, thus assessing whether the link is healthy and whether the latency is within a controllable range.

What is "Fair Price"? How to Measure It, and Why Does High-Frequency Market Making Trade Around It?

Jeff Yan: Different trading firms do have varying definitions of fair, often depending on their trading styles. However, the commonality is that fair essentially condenses your modeling results into a "predicted price." This abstraction is very valuable because it breaks down "how to construct a profitable strategy" into two equally challenging tasks: price prediction and order execution.

This also echoes your earlier question about market making versus taking orders: market making leans more towards the execution side, while taking orders leans more towards the modeling side. For taking order strategies, research and decision-making almost revolve around "fair price." As for what information should be included in fair price, it depends on where you believe you have advantages in data processing and where the market's efficiency gaps specifically exist.

Moreover, fair price does not necessarily have to be singular. Within a more machine learning-oriented framework, you can maintain fair prices for different prediction horizons simultaneously, such as 1-second predictions and 1-day predictions; execution strategies will utilize them in different ways, and the corresponding optimization objectives may vary in terms of PnL.

For beginners, a relatively effective "rough-cut method" is to first provide a single value around which you are willing to quote or cross the spread, treating it as your "oracle"; then, with a historical price series in hand, further think about how to achieve optimal execution around that value.

Corey Hoffstein: Can we simplify this understanding: if we observe a particular exchange, assuming Binance aggregates almost all liquidity, we can consider Binance's price as the fair price; if other exchanges (like OKX) have a lag of milliseconds to seconds, we can trade across the spread based on Binance's fair price and wait for it to "catch up." Of course, there are more statistical approaches, which do not take a single exchange as the "true value," but rather estimate the fair price by combining order book-related signals. Does this explanation hold? I'm not entirely sure.

Jeff Yan: Yes, the thought process is correct. Using the most liquid trading venue as the fair price is indeed a good first approximation. In the early days, there were often price discrepancies of around 10% between exchanges, and the main challenge at that time was not price prediction, but how to efficiently transfer funds between exchanges; thus, this method was very effective back then. In recent years, the market has undergone an evolution: liquidity was first dispersed and then flowed back and concentrated on Binance (especially recently). Therefore, as you said, using Binance's price as the fair price is a reasonable starting point.

However, it is important to emphasize that directly equating a single external source with the fair price should be approached with caution. For example, the lag at OKX might only be a few milliseconds, and actual trading may not be as straightforward as described. Further, suppose there is an opportunity: whenever Binance's price changes and no one is taking orders at OKX, you follow that price movement and attempt to arbitrage—this may be effective in most cases, but this is still the crypto market, which carries discontinuity risks: for instance, if OKX suddenly goes into wallet maintenance, causing a temporary disruption in deposits and withdrawals between Binance and OKX, the arbitrage chain cannot close, and prices may diverge. If your fair price relies solely on Binance's price, you may face passive risk exposure.

Therefore, there are many detailed factors. Even within this seemingly intuitive framework, it is far from as simple as "taking a number from a data source as the fair price"—it can only serve as a good first approximation.

Corey Hoffstein: This leads to my next question: the various technical characteristics and "traps" of crypto exchanges. Historically, their technical reputation has not been stable: you previously mentioned examples of "dirty data" (such as swapping price and quantity fields), API crashes, poor documentation, hidden endpoints, and even undisclosed parameters. I remember you recently gave an example on X: being able to bypass risk control engines or run them in parallel—these completely undocumented details constitute "orthogonal alpha" that is independent of price prediction. My question is: how much alpha can work like deeply understanding API details and accurately measuring endpoint latency contribute? Compared to that, how do traditional statistical alpha (like using order book signals to judge pressure and direction) weigh against each other?

Jeff Yan: I remember that tweet you mentioned received quite a good response.

Corey Hoffstein: By the way, I still can't tell if that was an April Fool's joke.

Jeff Yan: April Fool's Day has passed, and I admit that was a joke. However, it is closer to reality than people think. The real "punchline" is that part of it is true. I've always wanted to write a follow-up; this is a good reminder—I'll post it after recording this episode.

Returning to your intuition, I believe your judgment is correct. When someone has worked at a company for a long time, they often develop preferences; or they enter with a preference—such as "I studied mathematics, so I should work on cooler machine learning models, mining signals, and generating alpha; that's the key because it's the hardest." This "only doing models" mindset may work in large companies because the division of labor is detailed enough; but if you need to run the entire business yourself, relying solely on that won't get you far.

The "dirty work" you mentioned—mastering the API, filling in documentation gaps, measuring endpoint latencies—is very critical. My understanding of high-frequency trading (and many things) is that it resembles a product of multiple factors rather than a simple sum. Your investments in different "buckets" may seem additive, but the output often reflects a multiplicative relationship. For example:

Overall Efficiency ≈ Infrastructure × Model.

If the "infrastructure" factor is only 1 and the "modeling" factor is 10, then when you invest one unit of effort, the rational choice is often to first address the shortest board. The difficulty in high-frequency trading lies in accurately assessing what level each of these factors is at. Therefore, in practice, continuous "meta-analysis" is needed—am I currently working on what truly matters? You will quickly find that the answer is not obvious. Many competitive advantages lie precisely in the ability to judge priorities.

In this sense, those seemingly "dirty jobs" are often crucial. One should pragmatically pick the low-hanging fruit and follow the 80/20 principle. When the market is smooth, the most common pitfall is: "The foundation has been laid, so we can do some cooler machine learning research and pursue innovation." We have also paid a price for this. It does not mean that this direction lacks alpha, but rather that the investment scale is large, and the marginal returns often diminish quickly.

When your team is small, existing strategies are still effective, and market opportunities are abundant, it is even more important to repeatedly ask yourself and face honestly: what should be prioritized right now? Don't be "lured" by superficial data to chase directions that should not be prioritized at the moment.

Corey Hoffstein: For those looking to engage in high-frequency trading in the crypto space, you have suggested two paths: one is to trade directly on Binance and focus on generating alpha (I understand this leans more towards "actively taking orders" rather than "placing orders to make a market"); the other is to choose an exchange with long-tail characteristics, deeply understand its infrastructure "features," and seek advantages accordingly. Can you elaborate on why you believe these are the two optimal paths? What are the methodological differences between the two?

Jeff Yan: This can be likened to an intuitive conclusion of a "bell curve"—do not stay in the middle. If we understand the horizontal axis of the bell curve as different exchanges, the most prominent issues often lie in the middle range, roughly corresponding to platforms in the 2nd to 7th tier.

Their trading volumes are far smaller than Binance, but the competitive intensity and "toxic flow" are similar, or even the quality of the flow may be worse. At least on Binance, we know that its retail flow accounts for a very high proportion, which brings a "buffer effect"—the mix of toxic and retail flow is more friendly. Top HFT firms have basically fully integrated with the leading exchanges (you can roughly understand this as the top 15), trading at full capacity with larger and more mature strategies; it is difficult to "squeeze" much profit from these middle-tier platforms. If you are willing to challenge highly scalable large CEX strategies, start directly with Binance, and generalize as much as possible—there is no reason to start from the "middle."

The other path you mentioned is also valid: going to the far left of the bell curve. Look for small and overlooked opportunities—either too small to be worth the time of big players or too niche for them to cover. The infrastructure of niche exchanges is a great example.

Exchange systems are developed and implemented by people. Just as many DEX protocol designs may have obvious shortcomings, some small centralized exchanges may also have clearly identifiable flaws in their technical implementations. If you truly understand the "peculiarities" of how they operate, that alone may translate into a strategic advantage. Infrastructure is often also an important source of alpha, and there is no absolute clear boundary between "model vs. infrastructure."

You might worry about "lack of generalizability": for example, you have mastered a specific "exploitation method" at a small exchange, but it does not directly help at Binance. I believe the outside world generally underestimates the value of "running an effective strategy." For most teams, this should be the primary goal; as for the scale of the strategy, it does not necessarily need to be overly concerned about in the initial stages.

Of course, there is a basic premise: if the platform is small enough to have almost no trading volume, research and deployment lack meaning. But as long as there is a certain trading scale, it is usually possible to achieve some returns. More importantly, if the strategy has a high Sharpe ratio and sufficient robustness to retail events, then the skills and experience you gain will be something that most participants do not possess.

Even if the specific strategy may not be directly generalizable, my experience is that as long as you complete the "research → launch → production" closed loop, the insights gained in this process often exceed expectations; even if you later pivot to top platforms like Binance, the overall difficulty will significantly decrease. Additionally, while many detailed differences may not transfer one-to-one, you will begin to extract common principles from "proven effective things" and continuously generate new ideas; these ideas are usually significantly better than those conceived from scratch.

Therefore, both paths have their value. If you find it difficult to choose one, you can start small and gradually move to larger ones; frankly, trying both paths is also not a bad idea.

Corey Hoffstein: You mentioned "toxic flow." Can you provide a definition for those who have never heard of this concept?

Jeff Yan: Essentially, it refers to "information-advantaged flow." I have a framework for understanding the growth of the crypto market: I entered the market when it was not too early, and I can only imagine the earlier state by looking back. Even at the stage when I entered, the scale of retail funds was already considerable, and there were large participants, but the core contradiction in the supply-demand relationship at that time was still that the available liquidity was insufficient to meet the trading needs of the retail side. Therefore, retail flow is the most direct and worthwhile target to capture. The most intuitive approach is to write a relatively generic market-making strategy to provide liquidity through placing orders. As long as retail investors transact with your orders, you can largely retain the portion of profit contributed by crossing the spread; at that time, this model itself could sustain profitability. This, in turn, constituted a strong signal: the dominant flow in the market at that time still mainly came from the retail side.

However, as time passed, market participants gradually became aware of this and began to deploy market-making strategies on a large scale. As liquidity on the market-making side continued to increase, the significance of taking orders rose, and the bid-ask spread was continuously compressed. To continue capturing quality retail flow, takers began to emerge and more selectively "filter" the poor orders from the market-making side, taking them one by one. This is a relatively common path in market evolution. It should be added that taking orders also provides important value to the market; simply dividing "market making = market makers, taking orders = counterparties" is not accurate, as both roles often intertwine in practice. In my view, a more ideal market structure allows participants to trade freely in their own ways.

However, from the perspective of market makers, this type of taker flow significantly increases the difficulty of strategies: the previously relatively easy model—continuously placing orders and earning small spreads each time you are filled—may be "pierced" by a few trades. For example, you might accumulate about 1 basis point of profit from approximately 99% of retail transactions, but in about 1% of trades, you could incur a one-time loss of 10 basis points (this is just a mental model, not an exact figure). In this structure, tail losses can consume most of the regular profits.

Therefore, "toxic flow" largely refers to this type of flow represented by takers, which has an information advantage. Of course, whether it constitutes "toxic" depends on the specific strategy you are running; but in most contexts, it can usually be relatively intuitively distinguished between "retail flow" and "institutional/high-level flow."

How Common Are "Adversarial Algorithms" Luring HFT in the Crypto Market?

Jeff Yan: Crypto indeed has a "Wild West" vibe. From a more positive perspective, Crypto is also an experiment, where stance and perspective are particularly important. Regulators often latch onto one point—"they haven't followed our carefully crafted securities laws." Meanwhile, supporters of DeFi argue that these securities laws themselves likely bear the marks of lobbying and human judgment; crypto may provide a more libertarian experimental space: what truly needs regulation? I'm not sure, and the reality is likely somewhere in between. I'm neither a regulator nor a policymaker; I'm just sharing some philosophical observations. Back to practical matters, if you don't pay attention to strategies that involve manipulation and extraction, trading in the crypto market will be very challenging.

Another reality is that it's not that exchanges are unwilling to regulate, but rather that many times it's unclear who should regulate which exchange—at least to me, this point is not clear. Many legal frameworks differ significantly between countries, which may be one of the important reasons for the long-standing issues. Moreover, operating an exchange is itself extremely challenging, and they also need to handle a multitude of other matters simultaneously.

To give a more specific example: spoofing is a very common behavior. I don't intend to get bogged down in its strict technical definition under U.S. securities and futures law; the spoofing I'm referring to is more in a broad sense: from the order book and the subsequent price trajectory, you can often clearly observe that someone is placing large orders but clearly has no real intention to execute—indeed, if they were to execute, they would find it disadvantageous. Although it's difficult to prove "intent" legally, these orders are clearly not meant to be filled but rather to create the illusion of abundant orders on one side. The result is that if certain algorithms view order book liquidity as a signal for price direction, they may be misled and subsequently place orders in that direction. Once the "induction" takes effect, the spoofing algorithm will either place maker orders that are easier to hit or actively take those passive orders that have been exposed under the induction.

Such situations are very common. Another more straightforward example is various forms of market manipulation, such as organized groups engaging in "pump and dump."

Out of observation, I have infiltrated several such groups, never participating in trading, only observing. The scale of such phenomena is not small. Recently, related behaviors have indeed been cleaned up significantly, which is a good thing; but in earlier years, they could even create exaggerated trading volumes: a certain "insider" announces a token, and then retail users rush in (I don't fully understand their organizational methods), while insiders use the traffic to complete their sell-off. For high-frequency trading, such scenarios may seem manageable on the surface, but the actual handling difficulty is very high because the strong mean-reversion effect often reverses and "kills" strategies.

As for coping strategies, it goes back to the trade-offs you previously mentioned between infrastructure, models, and strategies—where should energy be invested? For me, these types of issues fall into the category of "miscellaneous/special scenarios" that also belong to risk management and special situation handling.

In short, if this part of the work is not completed, even if other aspects are nearly perfect, this area may still become a key factor determining the long-term average PnL's success or failure across different market states and different assets.

Jeff Yan: When we first encountered such situations, we were indeed shocked. Looking back, we were somewhat lucky: the assets we initially traded were either hard to manipulate or the other party hadn't had the chance to act yet. We had no foresight of this issue and built our system under the premise of "ignorance," and at one point, PnL progressed smoothly. But once caught, the impact can be very severe—if strategies are not constrained, one could lose a day's PnL in a minute. Sometimes automated trading is the most "foolish" trading because it is essentially just a limited state machine lacking human discretion, executing only along preset paths. Our response was quite pragmatic: of course, you can sit down and delve into modeling to predict whether manipulation exists; but one of our advantages at that time was our quick response, data-driven approach, and not being fixated on the "most standardized" path. For us, the approach was—once a specific loss pattern appeared, we would directly shut down the related logic;

Jeff Yan: Such rules could often be written within an hour and directly deployed to the production environment. At that time, we strictly followed the 80/20 principle: we would indeed miss some opportunities because of this, but it freed up time and energy to expand and push forward those key matters that could amplify PnL by ten times, rather than being continuously constrained by these issues. Perhaps about 5% of the time, we would forgo potential gains due to shutdowns, but this is essentially a trade-off and judgment—investing resources into the most valuable work.

As subsequent resources and time became more abundant, we gradually deepened this area: we now have more complex models to predict related market states and identify ongoing behaviors; compared to the earlier more "discrete" on/off handling methods, we now adopt more continuous parameter and weight adjustments to dynamically constrain and adaptively configure strategies.

As of now, we have formed a relatively deep understanding of how these manipulative behaviors operate and their identifiable characteristics. But it should still be emphasized: for newcomers, the 80/20 principle remains the most important action guideline.

Does Market Manipulation Occur More in Long-Tail Coins and Small Exchanges?

Jeff Yan: Such situations are relatively rare for Bitcoin and Ethereum on any exchange because their liquidity is more abundant. I believe this depends more on the assets themselves rather than the exchanges. I have seen (manipulation/spoofing) behavior on almost all exchanges; the methods vary across different platforms, and you can feel that the participants are not entirely the same, but the overall patterns are generally consistent.

There exists a "sweet spot": if a token has almost no trading volume, it usually isn't worth investing in; but for some altcoins with a certain trading scale, it's different—algorithms may expect a certain level of trading and liquidity on them, thus creating "inducible" space, allowing manipulators to profit from it.

Corey Hoffstein: I have always believed that the way we observe the market is often constrained by our own trading cycles. As a high-frequency trader, your intuition about microstructure may differ significantly from mine, as I have a longer holding period and lean more towards fundamentals. You once tweeted comparing the market to a viscous fluid, where external shocks manifest as damped oscillations in the price discovery process. I find this metaphor very interesting; could you elaborate further?

Jeff Yan: I also value understanding the essence of things. This probably relates to my background in mathematics and physics—if I can't understand the underlying mechanisms, it's hard for me to innovate within a "black box" system. Therefore, I tend to construct some mental analogies and metaphors to help understand how the market operates.

Using the "viscous fluid" model as an example, we can return to a more fundamental question: why can high-frequency trading make money? Many retail investors view it as a form of "predation," believing we are "front-running" or "hunting stop losses." I'm not claiming that high-frequency trading is "doing good," but I do believe it provides necessary services to the market to some extent.

We can abstract the external factors influencing prices as "shocks" applied to the system (which for us largely have randomness): for example, someone urgently needs to execute a trade and must obtain liquidity immediately; or a news event changes the "fair value" of an asset. Although some may attempt to interpret the event itself, such demands are often sudden and typically "exit after execution." The order book is essentially a strong PvP (participant vs. participant) arena, where many participants enter with a clear urgency to execute; and it creates a feedback loop: momentum trading triggers more trades, leading to various unstable equilibria.

In this structure, prices often first experience a significant initial shock, and then market participants gradually "enter" and engage in a game around the true fair price. The first jump is usually the largest; thereafter, some will judge that "overshooting has occurred" and engage in mean-reversion trading—this could come from both mid-frequency and high-frequency participants, who might believe "the mean price will revert in the next 5 seconds." Meanwhile, others may believe the event has far-reaching effects and choose to push prices continuously upward until larger increases occur; for instance, events like "Elon incorporating Doge into Twitter" may be seen as "having a real impact" within their narrative framework, thus breaking through positions on the mean-reversion side.

Overall, this resembles a "price voting" and continuous game played with real money. Its key feature is that the amplitude of fluctuations gradually converges. As participants gradually establish their target positions, funds continuously complete a weighted average, and prices ultimately converge to a more stable fair price range.

In this process, the core function of high-frequency trading remains to provide liquidity by buying low and selling high. If we view the price path as a curve oscillating up and down, high-frequency trading buys when the curve is low and sells when it is high, and its trading impact, on average, smooths that curve—prompting prices to approach the fair price more quickly and operating as closely as possible around the fair price during the price formation process.

Thus, within this analogy framework, the stronger the high-frequency capability and the more abundant the market liquidity, the more this "fluid" behaves with higher viscosity (stronger damping). This mental model may not be rigorous, but it generally conveys the meaning I intended to express in that tweet.

Why Can High-Frequency Trading's P&L Predict Mid-Frequency Price Movements?

Jeff Yan: These are some of our internal "exploratory ideas." I previously mentioned that iterating around validated effective directions is almost always better: it has a higher hit rate and is easier to scale. However, we also reserve space for a few bolder explorations, as occasionally they do yield results. This time is a relatively successful "interest project," and at the project's inception, we didn't have a strong prior judgment.

The motivation mainly lies in: our available capital scale has exceeded the capacity that high-frequency strategies can effectively support; although we have connected to multiple exchanges, this is more about constant-level expansion, and marginal returns are continuously diminishing—because the platforms we subsequently connect to are getting smaller. Thus, we began to consider: could we extend our reach into the mid-frequency domain—ideally, that would be an "ideal asset" with a Sharpe ratio of 3-4 and a capacity hundreds of times that of high-frequency strategies. This idea sounds very appealing.

However, we generally acknowledge the basic framework of efficient markets. Yes, we have advantages in high-frequency trading, but if given a set of daily data and asked to predict daily returns, we would also find it difficult to directly identify reliable entry points. Based on this cautious attitude, this "brainstorm" provided us with a relatively feasible path: in mid-frequency trading, if we can obtain data sources that are valuable to others but inaccessible to them, that alone could form a strategic advantage. We cannot acquire "alternative data" like satellite images or statistics on parking lot traffic like some institutions do. So, what do we truly possess? We possess our own HFT PnL—this is proprietary data, and it is clearly not random noise; its time series form reveals certain structural characteristics, making it worthy of further study.

Further probing, what factors is it related to? Returning to the previous discussion about "toxic flow and retail flow," it is highly correlated with retail flow. A relatively straightforward prior is: if you can distinguish between types of market participants and understand their behavioral patterns, you often gain better signals. The overall prior is still that "most signals lack stable predictability," but the direction is not necessarily clear. Therefore, our thinking is: since we have this indicator, and it is related to retail flow, which in a probabilistic sense is related to price formation—then we should truly delve into this path and conduct serious analytical validation.

Jeff Yan: We indeed conducted this analysis. The overall approach was to incorporate a series of P&L-centric features (such as changes in P&L, the "derivative" of P&L, etc.) into a regression framework to predict price performance at a mid-frequency scale over different time windows. Initially, we were also uncertain about how to implement mid-frequency research, so we adopted a relatively "broad coverage" approach: starting with 5-minute returns and gradually expanding the time scale to several hours.

Jeff Yan: The research mainly relied on our internal dashboard data system, which can aggregate the P&L of different strategies across different exchanges and assets, supporting segmentation by exchange/strategy/type, etc. Due to significant data noise, more robust processing is required; clearly, we wouldn't directly regress the P&L of a single cryptocurrency against its mid-frequency trends—noise is too high, and interpretability is limited. We basically followed the 80/20 principle, using bucketing and grouping methods to avoid obvious overfitting while adhering to existing priors, resulting in a rather interesting and counterintuitive conclusion:

Whether making markets or taking orders, the PnL on the high-frequency side is significantly negatively correlated with subsequent returns of crypto assets, and the effect strength is not weak. We were quite excited when trying to capture this in real trading: within a 1-2 hour prediction window, the magnitude of this effect was roughly several dozen basis points, and the capacity was relatively high.

But the problem is: this signal almost only suggests shorting, lacking a symmetrical reverse effect (theoretically it may exist, but our strategy iterations would avoid being in a continuous loss state for long). In other words, when we are making money, the model's implication is closer to: one should short.

Jeff Yan: So, what exactly should we short? Intuitively, it should be shorting perpetual contracts or futures. However, in practical implementation, two real constraints arise.

The first is funding rates. When such situations occur, many mature participants are often also shorting; even if the underlying signals they focus on differ, the alphas may be highly correlated, leading to market behavior trending in the same direction, and funding rates will reflect and absorb this level of crowding.

The second is that the individual assets where the signal is most "significantly effective" often belong to extreme samples, and these assets are practically difficult to short due to reasons such as insufficient liquidity, limited borrowable securities, and incomplete contract tools.

Nevertheless, the overall effect remains usable. High-frequency trading naturally creates inventory, allowing for internal hedging or internalization between strategies; even without internalization, one can impose a bias on inventory targets when the signal is strongest (for example, by trying to reduce positions or avoid holding positions), thus contributing positively to overall P&L.

Regarding "abstracting a reusable mid-frequency short strategy," we believe its persuasiveness is insufficient, so we did not package it as an independent strategy. This also belongs to the type of alpha that is closest to being publicly shareable; however, whether it is executable still depends on your strategy combination and trading process design, and it could very well transform into actionable alpha under a specific system.

Corey Hoffstein: I really like this idea: directly shorting futures may not be feasible because funding rates may have already priced part of the signal; but adjusting inventory bias to capture this alpha is an alternative path and can have a substantial impact on P&L.

This reminds me of some practices in my frequency band—take DFA as an example, they do not explicitly trade momentum, but when buying value stocks, they will exclude stocks with significantly weak momentum: they do not directly build positions based on momentum as a factor, but wait for the negative momentum phase to dissipate before intervening in value. This is similar to the logic here: treating theoretically "orthogonal" alpha not as an explicit position expression, but integrating it into the trading process, continuously improving results through marginal advantages and subtle improvements. This concept is quite enlightening.

Jeff Yan: Let me add something. Your earlier example is very interesting; I hadn't encountered it before, but I have indeed heard some "manual traders" who primarily trade large positions mention similar practices: in the crypto market, whenever the 50-day moving average crosses above or below a certain moving average, they trigger corresponding actions. The implication is that their core decision does not rely on technical analysis, but when a certain technical condition arises, they view it as an execution trigger. I haven't specifically studied the case you mentioned, but it reminds me of similar methods—waiting for a relatively reliable "conditional signal" to change before executing the established trading process.

Corey Hoffstein: Yes, the essence is to wait for a conditional signal to change. Very interesting. We have talked a lot about centralized exchanges, but not much about on-chain strategies/decentralized exchanges. You mentioned that one of your favorite but later discontinued on-chain strategies was RFQ. Could you explain what it is, why you liked it so much and it worked well at the time, and why you later stopped?

Jeff Yan: About six months ago, we began to increase our investment in the DeFi direction. At that time, the industry generally believed that the best opportunities were migrating on-chain, while centralized exchanges had entered a stage of diminishing marginal returns (overall trading activity was relatively low). Therefore, we decided to invest more time in researching DeFi. During that period, RFQ (Request for Quote) formed a wave of enthusiasm. Douglas from CrocSwap recently posted some interesting tweets, and I largely agree with his viewpoint: this design is not ideal—it attempts to directly transplant effective mechanisms from traditional finance (TradFi) into DeFi, but may not fit the on-chain environment.

To help unfamiliar listeners understand, let me provide some background: the starting point of RFQ is relatively clear—helping market makers filter "toxic flow" and allowing retail users to connect directly with market makers. Retail users initiate requests: "I am a retail user, please give me a quote." Market makers return a quote (usually better than the market spread, or at least able to meet the larger scale that retail users wish to execute). After receiving the market maker's signed quote, the retail user broadcasts that signed payload to the smart contract; the contract verifies the signature's validity and completes the asset settlement between both parties. Its essence is closer to a "protocolized OTC" mechanism.

The concept sounds reasonable and is also quite common in traditional finance (TradFi): users can obtain larger trades that are not easily "front-run" by high-frequency trading, which is a better service for the retail side. However, in the DeFi context, this is almost an obvious design flaw because you cannot prove that the other party is indeed a "retail user"—on-chain anonymity is the default, and there is no KYC layer of identity verification.

To validate this judgment, we wrote a very simple Python script to batch initiate quote requests. The result was that market makers indeed provided extremely favorable quotes: the spread was about only 5 basis points, and the quote was valid for 60-90 seconds. In most cases, from the market maker's perspective, obtaining such trades should be attractive; the scale they were willing to offer was also quite considerable (in the hundred thousand dollar range). But this directly exposed the mechanism's flaw: in a system where identity cannot be verified and anonymity is the default, anyone can masquerade as a "retail user" and exploit the quote mechanism for an advantage. This was also the fundamental reason why we initially liked but quickly stopped using such on-chain RFQ strategies.

Jeff Yan: Our operational approach was actually quite simple: first, wait for price fluctuations. The crypto market is inherently volatile, and once the price changes, we would broadcast that already signed transaction on-chain. The counterparty has limited options. The Sharpe ratio of this strategy was very high. Furthermore, it didn't even have to wait for price fluctuations to trigger—it is essentially closer to a "free option," with clear time value: you can wait until the last moment before the quote expires to decide whether to submit the trade.

Jeff Yan: This also made the returns more stable. That's how we executed it at the time. Clearly, we were not the only participants doing this (we might have been one of them), and market makers quickly reacted: they began to stop providing us with normal quotes, claiming "you are making us lose money, you are clearly not retail," and thus either offered extremely wide spreads or simply stopped quoting altogether. We could also respond to this, for example, by changing addresses or wallets to continue requesting.

From a principled perspective, I don't believe there is an inherent problem with this strategy. It feels more like a real-world concern: the maximum value of executing this strategy may lie in proving to the market that the microstructural design of RFQ has fundamental flaws—capital and intellectual resources should shift towards more reasonable mechanisms. Perhaps in this sense, our "experiment" has already fulfilled its function.

I understand that many RFQ mechanisms have now introduced market makers' "last look" rights, rather than allowing "retail" to retain final decision-making power. As you said, we later also stopped these types of strategies. I do believe this is an evolution; however, once market makers are given the last look, the core advantages of RFQ will essentially be weakened or even disappear. This point can also be seen in related discussions on Twitter: it is very difficult to outperform centralized limit order books (CLOB) at the mechanism level, and I do not believe RFQ can achieve this stably in DeFi. The above experience further illustrates that after our repeated trial and error, we increasingly feel that this track is still immature, and many protocol designs have not been fully thought through.

In this context, we made a strategic judgment at the time: rather than engaging in "arbitrage-style adaptation" within existing mechanisms, we should build a truly decentralized price discovery platform aimed at retail users.

Why Shift to Hyperliquid

Jeff Yan: The reason we decided to get involved personally is that we experienced strong confusion during the DeFi trading process: even during the low period of DeFi in mid-2022, retail flow was still considerable, but users were using protocols with extremely poor experiences. They were paying high Gas fees under poor performance of the underlying public chain while using mechanisms that were not ideally designed (such as RFQ). Even more surprisingly, users were indeed willing to continue using them; data clearly showed that demand was always present. Based on this judgment, we further delved into research.

I don't quite remember the exact position of the FTX incident on this timeline, but it should have been shortly before its collapse. After the FTX crash, the market narrative quickly shifted to counterparty risk: phrases like "Not your keys, not your coins," which had previously sounded more like slogans, suddenly became the most concerning issue for most people. This further reinforced our belief that "we should build some kind of infrastructure." But what exactly to build, we also weighed repeatedly: first, we needed to clarify the real needs of users and the unmet gaps in the market.

At that time, there were already a large number of clone versions of swaps in the market, with various minor innovations and aggregators emerging one after another, leading to a vast parameter space due to different curves and formulas. However, we were not optimistic about the AMM route: a significant portion of the so-called "market-making" liquidity was driven more by erroneous or misleading narratives, resulting in low-quality liquidity (for example, the way "impermanent loss" is presented, as well as the residual effects of liquidity mining). Even if AMMs do represent the main demand of the market, this track is already highly crowded, and launching another similar product does not clearly provide incremental value.

Therefore, we turned back to centralized exchanges to observe: what do users really need? Where does price discovery mainly occur? Where is effective liquidity concentrated? The answers were highly consistent—perpetual contracts. Perpetuals themselves are an extremely clever innovation (the idea can be traced back to traditional markets, but has been fully developed in the crypto market). However, in the decentralized space, there are almost no projects that provide this capability in a truly decentralized manner. Although dYdX adopts an order book format, its matching is still relatively centralized; it is the closest solution, but only to that extent. Our conclusion was: since the gap is clear, we should fill it ourselves.

The value proposition for traders is also very direct: if you appreciate the centralized trading experience of Binance or Bybit but do not want to bear custodial risks—then Hyperliquid aims to provide just that choice.

Hyperliquid has recently launched a closed beta test, aiming to provide a trading experience consistent with centralized exchanges (CEX): narrow spreads, near-instant trade confirmations, and almost zero Gas costs (only used for DoS defense). Under non-congested conditions, its blockchain can process tens of thousands of orders per second. All actions are fully transparent and on-chain, with all operations presented as on-chain transaction records—this is the vision we pursue.

Our primary target users are in the DeFi circle, as educating a broader audience on "you can trust on-chain contracts without custodians" is costly and not our specialty; DeFi users are willing to use it today. What we need to do is prove to them that among many protocols, most are not serious, some are merely stopgap solutions or temporary mechanisms based on local prices—suitable for gamblers but not for serious traders who need real liquidity and reliable price discovery. What we aim to provide is usable liquidity and a decentralized price discovery mechanism.

Blockchains and smart contracts can inherently take on custodial and settlement functions and establish verifiable trust at the mechanism level. However, promoting this concept is not easy and is not our strength. Our strategy is more direct: demonstrate differences through products and facts, showing users that among the complex choices of protocols, most solutions are not rigorous, and many are just short-term patches (some even based solely on local prices). Such mechanisms may be more suitable for "Degen-style" participants, rather than professional traders who require stable trading experiences and real liquidity.

What sets Hyperliquid apart is that we have centered our design around the aforementioned needs from the very beginning. To this end, we have made significant technical innovations and invested most of a quarter in concentrated development. Initially, we were also attracted by dYdX's path: off-chain matching, on-chain non-custodial settlement. However, after further analysis, we believe that this model has structural flaws— the degree of decentralization of a system depends on its most centralized component. Based on this judgment, we could not accept this solution, as it is difficult to scale to the size and vision we envision.

Therefore, we returned to the starting point: complete decentralization must be achieved. Under our constraints, this almost means having no choice— we can only develop our own public chain. We are not inclined to take detours, nor do we blindly follow existing conclusions. The outside world generally believes that building an L1 is extremely difficult; our approach is to first focus on solving the consensus problem. Tendermint is not perfect, but it is mature and has been tested in many practical scenarios. We chose to build on its foundation and have progressed to the current stage.

Why Hyperliquid Chooses to Build Its Own L1 and Views It as a Key "Vertical Integration" Decision

Jeff Yan: In recent years, L1 has become an important narrative in the industry, with many large investments centered around it, such as Solana, Avalanche, and other so-called L1 projects. The concept itself is not complicated: L1 refers to the blockchain itself. In contrast, there is the "smart contract-based implementation path"— that is, implementing exchange logic on top of an existing L1 (such as Ethereum or Solana) through smart contracts, with that L1 responsible for execution and settlement.

This is important because there is a subtle incentive structure involved. Many teams are willing to "build on a certain L1," partly because it is easier to gain support and promotional resources from VCs/funds holding large amounts of tokens; the value of a general-purpose smart contract L1 also relies on the application ecosystem it supports, so they naturally tend to attract developers to "deploy contracts based on me." In contrast, the Cosmos SDK chain based on Tendermint is closer to a self-sovereign model— it lacks strong external incentives to drive its expansion, and its value does not directly flow back to a single entity.

From my personal experience (we have tried both paths), it is hard to imagine that simply using an existing L1 as a general contract platform can build a truly high-quality exchange, especially in the derivatives space, particularly in an order book model. A kind of "corroboration" is dYdX: as a recognized pioneer, it also chose to shift to building its own blockchain after five years of operation. Their motivations may include legal pressures (I can only speculate), but in any case, the architecture they currently operate under is clearly not fully decentralized; once the new chain is ready, the old architecture will gradually phase out. For us, if the goal is to create a truly high-quality exchange, pursuing the L1 path is the more orthodox and scalable direction.

To give a more specific example: if the exchange is fully realized as a "smart contract," it will inevitably be subject to the strong constraints of the underlying contract platform's rules. Taking Ethereum as an example, transactions and state updates typically need to be triggered by user transactions. Therefore, for one of the most basic operational actions of a perpetual contract exchange— for example, settling funding fees every 8 hours— if there are 100,000 "sessions/positions" in the system, the number of storage slots that need to be updated cannot be accommodated within a single block. You would have to design an additional mechanism for "who triggers the funding fee settlement": it may require auctioning off trigger permissions, designing incentives and fee distributions, subsidizing Gas costs for the triggerers, etc. More critically, this process often cannot be completed atomically, and the actual effect will end up being "settled approximately every 8 hours," but the specific execution timing depends on the activity level of participants at that time, which may be delayed by several minutes. For strategy traders, it is difficult to establish a stable execution and risk control framework around such uncertainty.

Such operations are considered "basic actions" for all perpetual contract exchanges. If implemented on a self-developed chain, it can be significantly simplified: the funding fee settlement logic can be written into the consensus protocol. For example, it can be stipulated that when a new block is produced and its timestamp exactly meets the integer multiples of every 8 hours since genesis, the system automatically triggers the funding fee settlement and executes the corresponding logic, making the overall implementation more direct and controllable. In other words, operating a perpetual contract exchange is, in essence, closer to building an L1 rather than just writing a few smart contracts.

Why Hyperliquid Believes Order Book DEXs Are Superior to "Rate Layering" Liquidity Pool Models?

Jeff Yan: The concept of rate layering is indeed quite representative. You will see that many AMMs are slowly "evolving" towards an order book format. For many DeFi practitioners, this inevitably brings a sense of frustration— as if they are still reinventing the wheel. There may be local innovations during the process, but from a more fundamental structural perspective, the liquidity pool model has its "clever" aspects, but also carries a sense of being "over-packaged."

This is because AMMs largely originated from the computational and storage constraints of the past. Going back to 2018 (when Uniswap first appeared), the on-chain arithmetic operations that could be supported were extremely limited, and a single transaction could often only update a very small amount of storage state, with users unable to accept excessively high Gas costs. AMMs were formed as a compromise to be "barely usable" in an environment where computational power and storage were highly constrained.

Its operation, in a sense, relies on persuading funds to enter the pool to provide liquidity. At the same time, packaging impermanent loss as a marketing narrative, in my view, is a very clever but also questionable approach: emphasizing to retail users that "putting in is not trading, but gaining returns"; even if losses may occur, they are explained as "impermanent," negligible costs. Whether this narrative fully reveals the risks is at least open to discussion.

From the trader's perspective, the arbitrage opportunities in AMM pools are relatively explicit in mechanism: arbitraging around these pools can yield profits. Nowadays, such trades are quite crowded, but in the early stages, it was indeed a strategy with a clear positive expectation. Unlike order book markets, liquidity providers in pools are often not professional market makers, but more retail LPs. Many people put funds into the pool and then do not manage them for a long time, relying on so-called "mining returns" to maintain participation. If there is a lack of effective management and risk awareness, their expected returns over the long term may not be ideal, and they may even continuously suffer losses from adverse choices.

"Liquidity mining" forcibly pulled in liquidity in the early stages through incentives; when incentives wane, some funds may still remain in the pool, and there may even be participants who lack continuous attention to their own exposure. This structure itself is hard to call sustainable. Some may counter that "trading volume is still high," but within your narrative framework, this is more likely related to incentives and marketing mechanisms rather than a natural result under long-term equilibrium. In long-term equilibrium, the real sustainable liquidity may gradually decline until LPs are forced to raise rates to cover the implicit costs of adverse choices; once rates rise, retail flow will be further suppressed, creating a negative feedback loop.

At that "actually sustainable" liquidity level, if you account for returns and risk costs, the results are often not ideal. This is also one of the fundamental reasons I believe the liquidity pool model is difficult to sustain in the long term.

The so-called "upgraded version" can be understood as GMX and a series of "GMX clones": they no longer rely on constant function curves but instead turn to using oracle prices. To ensure that the oracle readings at the time of trading are as close to "real prices" as possible, these protocols often set many restrictions and makeshift designs. However, even so, related issues still frequently arise— for example, someone may manipulate prices on centralized exchanges and then trade on GMX based on the manipulated oracle quotes for profit. Mechanically, these methods still resemble "band-aid" patches rather than fundamentally solving the issues of price discovery and counterparty information advantages.

In my view, with recent advancements in underlying technologies such as L1 consensus, we no longer need to make excessive compromises between "decentralization" and "usable trading forms": we can maintain decentralization while also supporting order book-style trading. Based on experience, this is almost the only fully validated path that can achieve real price discovery and form a "real market."

Will Building an L1 Slow Down Price Discovery Due to Cross-Chain and Fiat Inflows and Outflows?

Jeff Yan: This is indeed a common issue across the entire crypto industry, not limited to DeFi. Even when arbitraging on centralized exchanges, the funding deposit and withdrawal routes still operate on public chains, and once there is congestion on-chain, the transfer efficiency will significantly decline. In our initial phase, we chose to focus on perpetual contracts, which is essentially an 80/20 trade-off: the vast majority of trading volume and price discovery is concentrated in perpetuals. Based on this, we made another 80/20 decision by initially standardizing the use of USDC as collateral, thereby streamlining the core path; subsequently, we can gradually introduce various stablecoins to diversify risk, which is not difficult. For most users, this model is smoother: after depositing USDC into the bridge/chain/contract system, they can express opinions and trade a large number of crypto assets in one place.

In terms of expressing views and price discovery for highly volatile assets, as long as there is collateral, positions can be established and arbitrage can be implemented. A typical scenario is spot-perpetual arbitrage: earning funding fees and trading the spot-futures price difference. In this structure, the perpetual side can be fully completed on Hyperliquid without the need to frequently move spot or USDC.

Of course, the cross-chain issue itself still requires ongoing attention. There are already many noteworthy full-chain technology directions; we have integrated some of these solutions and will continue to support relevant evolutions. Due to resource and priority constraints, we will not personally invest in "original innovations" for multi-chain infrastructure, but the ultimate goal is clear: assets can come from any source chain and be collateralized through trusted minimal or decentralized bridging methods, thus directly serving as collateral for Hyperliquid.

Currently, whether through the UI, Python SDK, or directly calling the raw API to try Hyperliquid, the end-to-end latency is approximately between 100–300 milliseconds; due to the randomness of block production, this value is not completely deterministic. You might think this is an order of magnitude slower than Binance's order placement latency. However, the impact of latency does not accumulate linearly like transaction fees; for our core user group—retail users—humans find it nearly impossible to reliably distinguish differences between 100 milliseconds and several tens of milliseconds.

Even if they can distinguish, in most cases, it does not constitute a critical factor; they care more about receiving "instant" interactive feedback. In typical market conditions, prices do not undergo substantial changes between 100 milliseconds and 10 milliseconds.

For the vast majority of trading scenarios, this portion of latency introduced by block time can be approximated as zero, so by running our self-developed L1, we can basically control it within an acceptable range. In contrast, confirmation times exceeding 10 seconds on chains like Ethereum would significantly impair the experience—prices may fluctuate greatly within 10 seconds. Moreover, from the user's perspective, the benefits of latency improvements exhibit clear diminishing returns. In terms of order book speed, a more critical metric is TPS; for decentralized exchanges, this specifically refers to the number of order placements, cancellations, and other operations that can be processed per second.

Indeed, compared to centralized exchanges like Binance, the throughput of decentralized exchanges often lags by an order of magnitude.

However, in my view, this gap does not necessarily constitute a substantive problem—computing power continues to improve, and current performance is "good enough" in an engineering sense. I do not know the exact metrics of Binance's matching engine, but assuming it can handle 1,000,000 transactions per second, while our self-developed L1 aims for 100,000 transactions per second, this does not simply lead to the conclusion that "Binance is 10 times better." The protocol can consciously design the upper limit at 100,000 TPS, which is still sufficient to support price discovery for the covered assets and most user needs. Indeed, during periods of extreme volatility, some frequent reorders may be delayed by several blocks, or even ten blocks, before being processed on-chain.

But similar phenomena can also occur in centralized exchanges. Therefore, although there is a numerical gap of "one order of magnitude," it does not necessarily correspond to a "one order of magnitude" cost consequence.

On the contrary, if we change the comparison object to a general-purpose smart contract chain, which may only have 10 TPS, then the gap between 10 and 100,000 becomes decisive. Furthermore, this is not just a TPS issue; it also involves engineering trade-offs. dYdX has consistently emphasized the off-chain order book path— to my knowledge, even in version 4, their plan is still for validators to run their own order books, only putting settlements on-chain.

Theoretically, this could bring an order of magnitude increase in TPS, but the cost is also quite high. It would significantly amplify the MEV space and also make "what constitutes a fact" more ambiguous—in my view, the order book should be seen as part of the system state; placing it off-chain would make the overall system harder to reason about and verify. Therefore, I prefer to accept a few orders of magnitude of performance degradation in exchange for significant improvements in robustness, resilience, and transparency; in my view, these benefits far outweigh the costs.

Additionally, we have conducted systematic research on the latest consensus studies. It is foreseeable that consensus will become the main bottleneck of the system, but there have been many new results in recent years, and the directions are very valuable. Tendermint is indeed relatively "old," with its core ideas being at least a decade old. The academic community has accumulated deep knowledge on related issues, but many newer consensus protocols have not yet reached production-level maturity. Therefore, we currently choose Tendermint as a phased solution, but almost all parts outside of consensus are self-developed: we do not rely on the Cosmos SDK but instead use Rust to implement high-performance components from scratch.

Meanwhile, we have completed relevant research and will continue to follow up. For us, once a better and production-level consensus protocol emerges, the migration cost of replacing Tendermint with a new solution is not high; when conditions mature, we expect to achieve at least a 10-fold performance improvement. We remain optimistic about our technical path: the self-developed parts and PoC (proof of concept) are ready, and the benchmark testing data is performing well. If the confirmation platform could not support the target load, we would not be advancing the current market promotion and user growth.

What I focus on is control, will, success, action, and determination—this aligns closely with our approach of "establishing ourselves through action" and sets us apart from many teams. The goals we set are often not "conservative": for example, "Can we create a Binance that does not sacrifice experience under complete decentralization?" Most people might think this would take at least five years. But we do not preset conclusions; instead, we study from first principles and truly engineer them into reality. This will and execution are equally critical for trading: you need both the will to win and the motivation to profit; lacking either makes it hard to reach the end. Today, we are building a larger system, and this "war chariot" approach to advancement is even more important— the market indeed needs it, but few are willing to take it on, partly because it is indeed very difficult. Our choice is straightforward: we will do it ourselves.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。