Computing power, memory and Alpha

Intelligence is also storage-driven.

1. It's not computing power, it's storage.

Where does the difference between people really come from? We are increasingly convinced that the real difference lies not in "computing power", but in "storage".

The difference in brain processing speed and reaction ability between people isn't that significant. The average person's IQ is roughly between 85 and 115, and even so-called geniuses rarely exceed two standard deviations. Furthermore, basic processing speed essentially stops improving after the early twenties, and may even begin to decline slowly. If intelligence depended solely on hardware (intelligence), a person's ability curve should peak in their twenties—but this is clearly not the case. Most true masters in various industries reach their peak in their forties or fifties.

Where does the difference come from? From storage (resources). The experience, knowledge, and patterns accumulated layer by layer from childhood—these are the real reasons why a person's ability curve can continue to rise for decades after hardware growth stops. To put it more directly: the difference between people is primarily not a difference in hardware, but a difference in storage (resources).

This judgment is not unfamiliar to us. On another level, the AI ​​system itself is undergoing a shift in focus—from “computation-driven” to “storage + connectivity-driven” (we wrote about this in “Deep Research on AI Architecture: The Era of Storage and Connectivity?” [1]). Over the past two decades, GPU computing power has increased by about 59,000 times, but DRAM bandwidth has only increased by about 110 times and interconnects by about 29 times—the growth rate of computing has far outpaced storage and data transfer. As a result, computing power itself is still scarce, but the bottleneck at the system level has shifted from “how fast is the computation” to “how timely is the data delivery and how much context can be stored”. In the end, the capex of storage + connectivity will surpass that of computing. What we are talking about today is essentially the reuse of the same pattern in another dimension—not only is the bottleneck of silicon-based systems shifting towards storage, but human intelligence itself is also storage-driven.

The "memory" in the title refers to this very thing—the structured, repeatedly accessed long-term storage in the human brain. However, this observation is too weak to simply state that "experience and knowledge are important"—it doesn't explain why some people spend their entire lives in the same industry and remain mediocre practitioners, while others, having seen the same amount of information, can see structures that others miss. To answer this question, we need to refine the concept of "storage" further. The key is not how much you store, but whether what you store has structure. A more precise statement is:

Individual ability ≈ Encoding and abstraction ability × Structured long-term storage

Here, "encoding and abstraction ability" doesn't refer to innate brain speed, but rather the ability to transform input into a usable structure—the same information entering the brain can be distilled into three variables and framed in a framework by some, while others are left with only fragments; the difference lies here.

Why must it be broken down into a product? Because storage is never passively accumulated. Two people read the same 10-K. One can extract three orthogonal key variables and integrate them into an existing framework; the other only remembers a bunch of scattered numbers. Over time, the former's mind develops a clearly structured network with interconnected nodes, while the latter's mind is filled with unrelated fragments. Their "storage capacity" may seem similar, but the quality of their stored structure is on completely different levels—the latter is what truly determines the output.

A classic psychological experiment illustrates the power of this structure. Show a chess grandmaster a 5-second chessboard, and they can reconstruct over 93% of the piece positions; however, if the pieces are randomly placed and don't conform to any real game logic, the grandmaster's advantage almost completely disappears, approaching the level of an average person. The grandmaster's "super memory" is mostly not innate—what they truly possess is the ability to rapidly compress input into a structure. When the input conforms to this structure, the memory is instantly retrieved; when the input doesn't, they regress to the level of an average person.

Therefore, the key difference between experts and novices is not the hardware aspect of "being quick-witted" or the quantity of knowledge they possess, but rather the ability to perceive the same world through a more refined and compressed framework. Most of this framework is developed through gradual training.

2. Mental Model: Personalized Compressed Structure

This framework is particularly explanatory when applied to the investment industry.

Someone who has researched 1000 companies will almost certainly outperform a smart person who has only researched 50—this is common knowledge in our industry. But upon closer examination, this common knowledge is incomplete. True alpha doesn't come from simply "reading a lot," but from those who, despite reading the same amount of material, can extract different structures from it.

Everyone who has worked in investment research for more than five years has encountered two types of colleagues: one type meticulously prepares every report, their Excel spreadsheets filled with pages, but if you ask them what the core variables of the industry are, they'll give you a textbook answer; the other type seems to have less workload, but when discussing a company, they can pinpoint the real bottleneck, the hidden misalignments between management's incentive structure and stock price performance, and why the industry's cycle over the past decade might be different this time. Two people reading the same material and attending the same conference call, but what ultimately settles in their minds is on completely different levels.

In the investment world, the term "mental model" is often used to describe this difference: different mental structures for organizing information result in completely different tradable propositions from the same input. A good mental model is not just about memorizing facts, but about how you encode the relationships between those facts—which are the main variables, which are noise, and which signals have hidden causal chains. The quality of this mental model is the truly effective term in the product formula.

What's truly scarce in investment research is never diligence or intelligence, but the ability to compress a bunch of seemingly unrelated signals into a single tradable proposition.

Here's something counterintuitive: the research nuances of any company are endless. You could read its annual reports for a lifetime, meet with the CEO weekly, and obtain all its operational data—you still wouldn't finish. Even more problematic is that "exhaustive research" itself is an illusion—repeatedly proven, even company insiders, and even management themselves, often make mistakes in their own judgments about the company's future. If you believe that "as long as you look enough and ask enough detailed questions, you'll get the right answer," you'll likely get bogged down in endless details and make decisions based on a false conviction.

The real solution is the opposite: instead of pursuing exhaustiveness, structure complex problems into a few fixed, compressed dimensions. Internally at Bedrock, any company's research is compressed into three interlocking frameworks—

• Competitive Advantage (CA) Framework

  This framework answers the question, "Why does this company win?" It breaks down easily generalized concepts like moats, business models, management quality, and industry position into a set of structured dimensions that can be scored and compared across companies and industries. The goal is not to provide a qualitative "good/bad" judgment, but to make convictions quantifiable, trackable, and questionable.

• Runway Research Framework

  This framework answers the question, "How long can this company win?" It compresses intertwined variables such as growth potential, market penetration, product cycle, and technology iteration pace into a quantifiable "runway length." A company's current strength is entirely different from a company with a two-year runway compared to one with a ten-year runway.

• Pricing Model Framework

  This framework answers the question, "How much is this company worth?" Instead of mechanically applying a DCF, we translate the CA score and the runway length into specific valuation ranges, allowing the judgments of the first two frameworks to be directly mapped to position and buy/sell point decisions—ensuring that we ultimately bet on "undervalued high-quality runways," rather than "good companies at any price."

The significance of these three frameworks isn't about "thinking more comprehensively than others," but rather about forcing you to retain only the variables that truly influence the final trading decision from an infinite amount of detail. Research shifts from "gathering as much information as possible" to "continuous calibration on a finite set of variables"—we no longer need to "know everything about the company" to make a convincing judgment; we only need information that moves any one of the variables in these three frameworks. All other nuances, no matter how interesting, can be temporarily set aside. This is where mental models truly function at the P&L level.

And this framework itself isn't copied from any investment book. It's the result of layers of cognitive correction accumulated through repeated mistakes, market lessons, and drawdowns. In other words, the framework itself is the most typical form of a personalized mental model—it's the intersection of general investment methodologies (DCF, Five Forces, Moats) and our own judgments (which mistakes we've made, which variables we believe are more important).

3. The Law of Drift: Why Mental Models Are More Important Now

One thing this framework has recently made us rethink is at what level LLM actually changes the structure of human capabilities.

A common misconception is that as large models become more powerful, human capabilities are being raised across the board. But looking at it using the product formula above, things are much more subtle.

The progress of large models is constantly raising a standardized "intellectual foundation"—everyone who obtains it starts with the same computing power. This has a two-sided impact on capability structure: it does raise the lower limit for each individual, but it also makes the "lower limit" itself cheap. When everyone can have models write a decent industry analysis, create a clean financial model, or translate a foreign language annual report, the marginal value of these tasks is compressed to near zero. This is the same principle as the transformation of the industry by the widespread adoption of Excel and Bloomberg twenty years ago—the more standardized the tools, the less valuable the capabilities dependent on the tools themselves become.

But the real key issue isn't here. The key issue is that the models themselves are also learning increasingly powerful compression capabilities.

Over the past two years, the reasoning ability, context length, and structure recognition capabilities of cutting-edge models have all improved significantly. Many things that were originally industry-specific "unique frameworks"—decision trees for medical diagnosis, the structure of legal documents, and code review patterns—are being trained into vertical models or specialized agent products. The financial investment research field will inevitably reach this point as well; it's just a matter of time.

Here, we need to acknowledge a contrasting optimistic narrative: some might argue that after models absorb standardized work, human attention is freed up to focus on higher-order judgments—this isn't a zero-sum game, but rather a reorganization of division of labor. This argument isn't without merit. However, looking back at every tool revolution—from Excel to Bloomberg to quantitative finance—the reality is closer to an "arms race" than a "reorganization of division of labor": each advancement in tools turns previously profitable skills into entry requirements, raising the overall skill requirements and reducing the number of people actually earning excess returns. There's no reason to believe this time will be different.

Therefore, the conclusion is clear: alpha pricing will continue to drift upwards. Those who made money ten years ago by "knowing how to use Bloomberg, build DCF, and read English annual reports" can no longer do so today—these skills have been completely absorbed by tools, becoming mere entry requirements. In the next decade, skills like "being able to structurally read company data, quickly extract core industry contradictions, and conduct comparative analysis using standard frameworks" will likely also be gradually absorbed by agents, becoming new entry requirements.

Returning to the isomorphism mentioned at the beginning: these are two versions of the same drifting pattern—in AI architecture, capex drifts from GPUs to HBMs and high-speed interconnects; in investment research, value drifts from "knowing how to use tools" to "having one's own mental model." When a certain level of capability is standardized, value inevitably shifts to the level above it—this pattern outlives any specific technology cycle.

4. The part of you that depreciates the slowest is the part that's part of your body.

If value inevitably drifts upwards, then the truly sustainable alpha becomes a concrete question: at what level will it stop? What is the basis for the statement mentioned earlier that "personalized mental models are the most typical compressed structure"? Why does it depreciate more slowly than general models?

Our answer is: from those mental models with the slowest depreciation rate. We intentionally avoid the term "moat" here—there are no permanent moats, only differences in the rate of depreciation.

Here, we need to distinguish between two types of mental models. One is the general model—abstracting the basic logic of an industry into a reusable framework, such as Porter's Five Forces, DCF, and value chain analysis. The stronger the public nature of this type of model, the easier it is for LLM to absorb it, because they are themselves public corpus, and their depreciation rate is the fastest. The other is the personalized model—a structure built layer by layer with your own unique experiences, preferences, judgments, and lessons learned from failures. BEDROCK's CA × Runway × Pricing framework is a specific form of this personalized model: it uses publicly available investment methodologies (moat, growth, valuation) as its framework, but the scoring weight of each dimension, the threshold for each calibration, and the judgment of "why this time is different" are all things we have learned from our own mistakes.

The reason these personalized models depreciate more slowly isn't because "they lack publicly available data, so the model can't learn from it"—with the advancement of technologies like personal memory, RAG, and vertical agents, personal experience is gradually being digitized and used by the model; it's just a matter of time. The real reason is that it's rooted in your own history of mistakes. Every pitfall you've fallen into, every painful lesson learned from the market, every cognitive correction preserved in drawdowns—this is a structure that must be accumulated over time and at a cost. The model can learn your output, but it's difficult to replicate the process by which you generated that output—this gap will be a real problem in the foreseeable future.

More importantly, these personalized models determine the questions you ask. The same large model can be tuned to completely different depths by a seasoned investor, not because they "write good prompts," but because they know what questions are worth asking, which parts of the answers are worth pursuing further, which types of model outputs are reliable, and which require further self-verification. This "ability to ask questions" is itself an outward manifestation of a personalized mindset.

The real differentiation in the LLM era isn't between "those who can use AI and those who can't," but rather between those who use AI with rich, personalized mental models and those who use AI as a general-purpose search engine. The former essentially grafts their years of accumulated knowledge onto a new colleague with standardized, high intelligence, significantly raising their skill ceiling; the latter merely consumes the standardized output of the model, their skill ceiling capped by the model's average level.

Future alpha will increasingly concentrate on those with both deep, personalized experience and a willingness to frequently offload standardized tasks using AI.

5. Frame clarity = radius of the commanding agent

There's one more layer missing here—and this layer's implications are perhaps deeper than what's been discussed so far.

The "personalized mental model × AI collaboration" mentioned earlier still defaults to a scenario of "one person + one large model," essentially an extension of individual capabilities. But this is merely a transitional form. The true next generation of work is not like this—it's about one person directing a team of dozens of agents working simultaneously in parallel. This may seem like simply "using more AI," but it represents a fundamental shift in the demands on individual capabilities.

Where does this shift lie? In the past, an analyst's output was limited by their own mental bandwidth—how many documents they could read, how many models they could run, how many questions they could clarify in a day. With the advent of AI, this bottleneck was partially opened up, but only partially, because the analyst was still interacting with a tool. However, when agents truly mature, the bottleneck will shift completely from "individual computing power" to something entirely different—can you clearly break down a complex task into 20 parallel subtasks, assign each subtask to the appropriate agent, accurately verify their outputs, and then integrate the results into an organic whole?

This is a very different capability. It actually reduces the demand on individual raw computing power—you no longer need to do everything yourself; but it significantly increases the requirements for clarity of framework and abstraction ability. The reason is simple: a chaotic mind can't manage 10 agents. If you yourself can't clearly explain how a research should be broken down into steps, what the input and output of each step are, and what counts as "doing it right"—then you simply can't deliver the task. You can only continue to do it all yourself, continuing to be limited by your own computing power limits.

In other words, the clarity of your framework directly determines your command radius. An investor with a clear CA × Runway × Pricing framework can have 5 agents simultaneously handle preliminary analyses of 5 companies because they know what each agent should output, which dimensions of output are reliable, and which require recalibration. An investor without a framework, even if given the same 5 agents, can only have them do some scattered research and summary writing work—because they cannot break down "researching a company" into deliverable sub-tasks.

Therefore, the statement that "personalized mental models depreciate the slowest" is only half true. A more complete statement is: personalized mental models not only depreciate the slowest, but they are also a prerequisite for your future ability to lead an agent team—the framework is also the language you use to direct agents; the clearer and more precise it is, the larger your collaboration radius. The truly scarce people in the future will not be those who "know the most," nor those who are "best at using AI," but rather those who can clearly articulate their research methodologies to deliver to an agent team.

This applies not only to investment research. The software engineering field has already experienced the same transformation much earlier—PingCAP's CTO, Huang Dongxu, mentioned in an interview that a product line he leads, with only one or two developers, accomplished in three months what 100 people in a traditional software company would do in a year. 90% of the company's code is automatically generated by AI; human engineers no longer read the code details, focusing only on higher-level things like engineering architecture. "Traditional software engineering is over," he said. He named this underlying capability "Harness Engineering"—the ability to manage a digital team composed of different roles. This is the same concept in investment research as "using a clear framework to direct agents."

When the marginal cost of execution approaches zero, the scarce resource is no longer your execution ability, but rather "what you want to do." The bottleneck quickly shifts upstream—towards the person who can continuously raise new questions, new perspectives, and new hypotheses. Raising demands itself becomes the new core competency.

Here's an easily overlooked corollary: this shift is particularly unfriendly to one type of person—those whose methodologies heavily rely on "tacit experience" and intuitive practitioners who think, "I know at a glance." These methodologies were a significant source of alpha in the past, but in the agent era, they become a bottleneck—because tacit knowledge cannot be directed, parallelized, or scaled. Conversely, those willing to spend time making their tacit experience explicit, structured, and transformed into a framework that can be delivered to agents will be significantly amplified in this era. This has little to do with "diligence" or "lack of experience"; what truly determines your fate is whether you are willing to put your intuition into words.

6. Short feedback loops = Amplifiers of creativity

But the "command radius" isn't the only amplification of capabilities in the AI ​​era. Besides "being able to do more simultaneously," something else is quietly happening—the number of ideas you can try and the number of times you can experiment are also being greatly amplified. I've recently felt this very strongly. The feedback loop in the AI ​​era has been compressed to an unprecedented degree—you propose an idea, and you see results within minutes; a flawed hypothesis can be disproven within hours; a research angle that previously required weeks to verify can now be completed in an afternoon. This short feedback loop itself is an independent amplifier: it greatly amplifies the creativity of those who already have a constant stream of ideas, dare to try and fail, and are willing to iterate quickly. I must admit, this way of working has a near-addictive feel—every new idea can yield results in the time it takes to drink a cup of coffee, every failure comes so quickly that there's no real pain, and you can't stop thinking, "Let's try this angle again?" At the end of the day, you'll find yourself having traversed a cognitive path longer than a whole week in the past.

So a more complete picture is this: Harness Engineering allows you to direct more agents to work in parallel at the same time (expanding the execution radius), and short feedback loops allow you to try more ideas at the same time (expanding the exploration radius). These two expansions are not substitutes, but multiplicative—the strongest individuals have both a clear framework to direct agent teams and a continuous stream of ideas that dare to be immediately falsified in feedback loops. Returning to the formula at the beginning, individual output in the AI ​​era can actually be broken down into another layer: execution radius × exploration radius—the former relies on the clarity of the framework, and the latter relies on the shortness of the feedback loop and how many ideas you are willing to cram into it.

7. The Silicon-Based World's Mirror Image of This Event

The above discussion focuses on the "human" side—how mental models are structured, directed, and amplified. However, it's worth highlighting that "structured personalized storage" is also occurring in the silicon-based world, redefining what true storage needs really are.

Today's storage needs are mostly write-once-read-rarely—files, logs, and web snapshots are stored, with 99% of the data accessed zero times, making the philosophy of "never deleting data" valid. But "personalized memory" in the agent era is entirely different: it's write-once-read-many-times. Every piece of raw information is processed into multiple representations—vectors (embedding, translating each piece of text into a string of numerical coordinates that can be retrieved by "meaning"), knowledge graph nodes, timestamp metadata, semantic clustering, and user profile fingerprints—which are then repeatedly queried and reassembled by the agent in every interaction.

The biggest misconception this leads to is the belief that "storage needs" will continue to grow primarily in terms of capacity, as in the past. In fact, the more crucial aspect is the fundamental shift in access patterns: the same data is amplified into multiple structured representations, repeatedly read, and cross-referenced—capacity only increases moderately, but bandwidth and I/O intensity experience a much steeper expansion. Capacity is linear, while I/O is a much steeper curve; the difference between these two is what truly defines the bottleneck of next-generation storage.

Furthermore, there are two easily overlooked second-order judgments here.

First, technological advancements will not reduce this demand; they will only increase it. A common intuition is that "future embedding technologies will be more efficient, so storage requirements will decrease." The opposite is true—the more powerful the embedding technology, the richer, more dimensional, and denser the relational graph translated from each piece of raw data becomes. A 768-dimensional embedding can express hundreds of response modes; an 8192-dimensional embedding can express thousands. Each upgrade in "understanding ability" generates more and more complex derived representations from the same raw data. Storage requirements will not decrease with technological advancements; they will only continue to rise as "each piece of data is understood more deeply."

Second, the core differences between different agents will increasingly focus on the agent memory layer. Two agents, even running the same large model and possessing the same raw dialogue history, might have vastly different relationships—one using rudimentary embeddings and a sparse graph, the other using sophisticated embeddings and a dense graph. The latter can only "think of" a fraction of the relationships the former can. The difference in agent capabilities will gradually shift from "whose underlying model is stronger" to "whose memory architecture is more sophisticated." The sophistication of the memory architecture depends almost entirely on the representation density and access frequency that the underlying storage system can support—this is another force that pushes value from "computing power" to "structured storage," and it operates at the layer closest to the user.

8. Mental models are an asset that needs reinvestment.

So, what does this mean specifically for you personally?

Your true competitive advantage lies not in how fast your mind works, but in the mental models you've developed and how clear those models are to the point that they can be utilized by external "computing power."

This leads to a concrete and more substantial deduction for investment research:

Don't outsource your judgment to models. This isn't a soft excuse like "you'll become a more mediocre version," but a reality about asset depreciation: your personalized mental model is a cash flow asset that requires continuous reinvestment. Every time you encounter a judgment problem, think carefully, and make a mistake or get it right, you're updating this structure; every time you directly throw a problem at a model and accept its answer, you're skipping a reinvestment. In the short term, you're more efficient; in the long term, your differentiation depreciates at a rate you can't even perceive. A year later, the distance between you and "newcomers who only know how to use AI" will be narrowed, not widened.

The correct approach is the opposite: let the model offload the standardization work for you, and invest the saved time entirely in what only you can truly internalize—visit more companies, talk to more managers, experience more cycles, make more of your own mistakes, and make more of your own corrections. Then, write down these insights, structure them, and turn them into a framework that you can teach to agents.

In an era where computing power is becoming increasingly cheaper and general-purpose mental models are becoming increasingly standardized, only those structures that are ingrained in you, bearing your unique imprint, are the true alpha that can transcend cycles.

And that is the true, genuine you.

Notes

Further Reading: "In-Depth Research on AI Architecture: The Era of Storage and Connectivity?", published on this WeChat official account by BEDROCK Research. This article demonstrates, from the perspective of AI hardware architecture, the scissors gap between the growth rate of GPU computing power and storage bandwidth/interconnectivity, the structural scarcity of HBM, and the trend that storage + connectivity accounts for a larger share of AI capex than computing.