Harness 套利期 从 SaaS 边缘抢救 DeFi

Looking back 500 years, the labor-capital conflicts under the capitalist system have always been marked by the continual victory of capital.

On the production side, the level of labor participation has gradually shrunk to the machine operation level; on the consumption side, user value lies in producing usage data for the platform.

Together, they support the company’s valuation in the capital markets.

But human organizational models have long been impossible to fully quantify; white-collar KPI/OKR still follow a hierarchical system, and annual salaries of millions or piecework wages are variants of Taylorism.

Without a clear formula, capital cannot value them, which in turn affects capital efficiency. Whether algorithmic stablecoins are the Holy Grail of DeFi remains uncertain; the computability of organizations is indeed a measure of financial leverage.

Large models decide to brute-force with token quantities; the collapse of secure SaaS is just a surface phenomenon. Product design is also on the way—replacing niche professional skills and scaling them up is the key, as innovation ventures into uncharted territory.

This offers us endless insights, especially as DeFi’s DAO model is gradually collapsing and tokenomics is increasingly bankrupt.

In one sentence, why are AI organizational models and token models more efficient than DeFi?

How did all this begin?

Token cheapening, Agent practicalization.

For 300% profit, capitalists can sell their own nooses;

To keep their current job, workers can write skills for Agents.

At the capital level, Agents empowered by skills hold a status as sacred as profit.

Agents represent “human ability” refined into skills; moreover, human organizations are transforming into interaction ritual chains centered on Agents.

The so-called prompts, contexts, and now Harness engineering are all about turning human organizational models into no-man’s land—at least reducing human involvement.

Your next colleague might not be a robot, but could be an “ability” instinct.

This is not a fantasy; the scaling laws at the data level are gradually failing. But data collection and production are no longer critical. Before AGI succeeds, new valuation targets are needed.

Illustration: Content no longer valuable

Overall info: @ARKInvest

Starting from Claude’s selection of programming fields to realize AGI, AI is surpassing entertainment chat modes and entering the existing markets of programming, security, and the newly released design.

Will this disruptive innovation ultimately create new economic increments, or will it push the economy into a permanently low-employment mode where work is replaced by tokens? We are witnessing this process.

But the current cheapening of tokens will decentralize capabilities once monopolized by a few large enterprises to small and micro businesses, thereby shaping super-individuals—this is not a fantasy.

Take China as an example: token call volume from 2024’s 100 billion/day → end of 2025’s 100 trillion/day → now 140 trillion/day. Content and data production are entering a zero-cost era.

It’s important to note that the scarcity of computing power is a relative state. Large enterprises no longer monopolize “ability,” but still seek to maintain advantages through monopolizing “computing power,” though they cannot stop the inevitable trend of overall token cheapening.

The paradigm comparison of foundational large models varies widely, but the evolution of “how AI helps humans” has long been overlooked.

In my view, Harness is a spatial form that allows Agents to focus on tasks within boundaries for the first time, employing a depth-first strategy, distinct from the breadth-first question-answering approach.

Illustration: Evolution of Agents

Illustration source: @zuoyeweb3

Since the Tab key was first used for code completion, humans have become merely an input layer for AI—only a matter of time.

The cost of trial and error has exponentially decreased, enabling more interesting attempts at human collaboration:

• Software: SaaS, human ability no longer comes from humans but from emergent Agents

• Hardware: compute cards + HBM, data centers now directly serve AI needs

• Space: Harness, not a physical space for human collaboration but a digital space for Agent interaction

• Interaction: the demise of Doubao phones, Google supporting GUI Agents at the Android system level

The ability of AI to generate text has limited commercial value; the cost of text generation is very low for humans. But “what to do” will cause token consumption to surpass that of image and video generation—similar to AWS selling not just servers but usage time.

AI sells not tokens but “work capability,” which is the root of fear in the SaaS industry. Unfortunately, DeFi has already become SaaS, not large models.

DeFi Protocol SaaSification

DeFi is not outdated but overly mature.

AI is reinventing software engineering. It’s not just SaaS being replaced, but SaaS is undoubtedly the most typical.

Even the Bloomberg Terminal’s most valuable aspect isn’t technological sophistication but authoritative information, which is built on decades of industry contacts, networks, and non-standard data.

Agents offer a choice: infer future trends from data, even risking stepping ahead of peers to earn small profits.

Illustration: SaaS collapsing

Illustration source: @zuoyeweb3

You can think of Agents as cleverly exploiting capital’s profit-seeking nature. Of course, you can wait for complete Bloomberg Terminal data, or use pieced-together, inaccurate data to gamble for gains.

This isn’t new; IBKR founder Thomas Peterffy first “invented” or assembled physical trading terminals in finance, all originating from an idle P101.

If a data-utilization method can generate more profit, then you can get more data, and the flywheel starts.

SaaS monopolized the past; AI sales will dominate the future.

Unfortunately, we need to approach DeFi from this angle. Remember Dune/DeFiLlama’s API paywalls, begging for golden data, or Arkham Exchange’s eventual shutdown.

Data in crypto is never valuable.

But crypto is an open financial system, and its generated data can be repeatedly learned. Even before AI, the speed of forked projects has slowed to monthly, with Meme-like PumpFun copies compressible to seconds.

An counterintuitive inference: DeFi is the testing ground for financial systems. The AI+DeFi approach we’re exploring today may become the template for future financial evolution.

• For example, before the 2008 financial crisis, uncollateralized LIBOR “triggered” the financial tsunami; it was replaced by the SOFR index derived from US Treasury trades, but the over-collateralization mechanism guaranteed DeFi’s finality in liquidation.

• For example, large model providers don’t want to sell tokens based on consumption; they must tier marketing, capability customization, and professional transformation. Token economics have turned “use value” into a twisted game.

Crypto tokens obsess over utility; AI tokens obsess over economic value.

From this perspective, DeFi hacks are just routine stress tests—open systems cannot self-repair external entropy bugs.

Like the dark humor of Article 22, without external signaling, crypto defaults to assuming the current environment is safe. When a security crisis occurs, it collapses into centralized handling.

For example, in the Drift incident, the blame shifted to the slow-frozen Circle.

Illustration: Code cannot solve security issues

Illustration source: @zuoyeweb3

It can be said that before AI capabilities leap, DeFi has already undergone SaaSification—only charging per transaction, unable to directly migrate “finance” onto the chain.

RWA on-chain lacks liquidity, and DeFi has no good solution.

But the evolution of Agent capabilities suggests a glimmer of an unclear dawn for rewriting DeFi rules:

1. Token economics: deploy usage across channels based on “capital efficiency”;

2. Rule setting: Mythos provides secure finality, AI defense walls battle zero-day crises;

3. Human organization: great—DeFi has long been managed by just a few people with hundreds of billions.

Revival of engineering narratives

Where does security come from? The determinism of Turing machines. Where does danger come from? Infinite possibilities.

Garry Tan of YC said, “Fat Skill, Thin Harness,” which resonates deeply. Essentially, it’s about setting fundamental rules—“order-based freedom.”

Turing machines can combine infinitely; von Neumann architecture always has a time lag in computation; large models cannot produce truly random numbers.

In a future where data is no longer valuable, only human behavior can generate value through money flow.

But human behavior also needs time to be fully learned by AI, then internalized into engineering and coding expressions.

Endless pursuit of the infinite is ultimately impossible. LLMs cannot eliminate hallucinations entirely; they must approach a state of “beyond AI and human reach” to allow market mechanisms to price them. Only then can we truly believe in smart contracts.

Current smart contracts are hardly successful: DAO forks, Curve language bugs, even Drift multisig—all prove that “humans retain ultimate control over code.”

Moral questioning has no economic value. The reason DeFi’s collaboration models have shrunk from DAOs to foundations and “teams” is rooted in the practical needs for contract upgrades and business cooperation.

But humans simply cannot write forever safe and dynamically upgradable code—remember, it’s impossible.

If it’s never upgraded, Curve’s own history shows that dependency on technology stacks can also cause issues.

Decisions made now shape the past, which in turn shapes the future.

From the Simmons Grand Prize Fund to Numerai’s AI strategies, AI in finance is not rare. Another counterintuitive case: trading signals can actually help AI evolve.

Illustration: 10 Years of AI and DeFi

Illustration source: @zuoyeweb3

AI models still follow the computational paradigm—state machines processing signals. Without external signals, they lack the ability to simulate the outside world. Yang Le Kun and Li Fei-Fei’s focus on world models is meaningful here.

But from a DeFi perspective, enabling AI autonomous trading depends on human intent being learned through Agent behavior. This underscores human importance: even if Agents replace human labor, they are mimicking and summarizing human actions.

Even humans cannot be intentionally random; tiny deliberate variations follow statistical patterns. Human physiological traits also introduce randomness—e.g., “I naturally prefer Ethena’s market-making strategies and dislike XX’s arbitrage strategies,” which carries fuzzy preferences.

It’s very certain that making blockchain/DeFi the infrastructure for AI has faced tragic failures over the past decade: deAI/deAgent/deOpenclaw all encounter similar fates.

Directly applying the latest large models to reshape DeFi structures—like Mythos testing with default security, real-time detection of any changes—can improve safety levels.

In human organizations, AI’s choice is “no humans,” only “capabilities.” DeFi is the most suitable industry for this—no doubt. After rule design, DeFi will only enhance capital efficiency under safety constraints, following a tiered process similar to autonomous driving levels (L1/2/3/4): from information authorization → limited fund access → full fund access.

If Agents continuously learn engineering trader skills and Curator management abilities, they will inevitably surpass humans in trading and profit. But unfortunately, the DeFi data accumulated so far has not been systematically learned or trained by AI; current crypto AI is still in the money-raising stage.

But I am very confident that the actual use of funds will be the next major wave of AI-driven DeFi transformation—inevitable.

So, after safety (contracts) and organization (humans) are re-upgraded, what will token economics look like?

• In PoW era, tokens are proof of computational power consumption, similar to current AI tokens;

• In PoS era, tokens are discounted claims on expected yields, and AI tokens are evolving toward this (providing capabilities to replace humans as a form of economic value);

The crypto tokens of the AI era have already transcended our engineering scope, and only rely on theoretical, irresponsible predictions.

Referring to Sky’s token distribution controlling APY across channels, and Claude’s token consumption-based pricing of model capabilities, future crypto tokens are likely to be a kind of “return on capital” certificate.

Note the distinction: PoS tokens like $ETH ’s expected yields are based on economic hypotheses—prior-based experiential reasoning. But AI’s engineering design and DeFi parameters will approach real-world data infinitely, with highly credible return and risk rates, verified in real time.

Even users can determine the current token price based on the large models and Agents used in DeFi protocols, and Harness optimization scores—buy if optimistic, sell if pessimistic.

Conclusion

Countless frustrations and the unpredictable future of humanity.

The future of DeFi divides into economic and technological aspects. Tokenomics currently has no good solutions, but security shows a glimmer of hope. Claude Mythos can threaten the world; conversely, that means controlling money.

AlphaGo solved Go completely; Claude will solve programming entirely. Such scenarios will only increase in the future. Theoretically, DeFi contracts, human organizations, and even economic units all have room for optimization.

At least, humans need not worry about being completely replaced. In an era where data is no longer valuable, behavior still has its meaning. For now, Agent takeover of humans remains in “micro-tasks,” “micro-payments,” and other small details. Repeating and copying these behaviors to generate value—AI is driving down the value of data and content toward zero cost. Meanwhile, the unit economic value (cost) of AI tokens and crypto tokens is continuously declining—this is the trend.

It can even be said that this is the first time money truly opens its doors to individuals—whether for AI work or crypto consumption.