Why Has “Verification” Become the Most Expensive Luxury When AI Begins to “Mass-Produce” Answers?
Hello everyone, I’m your financial journalist and economist. Today, we’re going to discuss a very insightful article from Havenlon Labs. This article doesn’t boast about how intelligent AI is; instead, it reveals a profound economic structural change that’s underway: in the age of AI, the processes of finding answers and proving answers are becoming completely separated.
To make this easier to understand, I’ve broken down this article, which is full of mathematical and philosophical concepts, into five key points that are easy to grasp.
---
Key Point 1: From “All-Round Experts” to “Assembly Line Division of Labor”
In the past, we believed that a great scientist or engineer had to be both the one who came up with ideas (the “questioner”) and the one who verified those ideas (the “judge”). For example, Newton not only discovered gravity but also derived the corresponding mathematical formulas.
But now, AI is separating these two roles:
1. AI is responsible for “frantically trial and erroring”: It acts like an endless printer, generating a vast amount of hypotheses, code, solutions, and even rough drafts of mathematical proofs at extremely low costs.
2. Specialized systems are responsible for “strict verification”: Tools like Lean, which are used for formal proofing, or future risk management systems, don’t require intelligence; they just need to be rigid, strict, and precise to check whether what AI produces is correct.
In one sentence: AI has made generation extremely inexpensive, turning verification into the new bottleneck and a scarce resource. The competition in the future won’t be about who produces the most answers, but about who has a more reliable mechanism to sift through the garbage and find the gold.
---
Deep Dive: Five Easy-to-Understand Explainers
1. The First Thing AI Changes Isn’t “Intelligence,” but “Physical Capacity”
Title: AI Isn’t a Genius; It’s a Machine for Exploring Infinite Possibilities
Many people think that AI’s breakthroughs in mathematics are due to its intelligence. That’s not the case.
There’s a natural limit to human research: our energy is limited, and we can’t split our attention. A mathematician can only think of a few ideas a day; if one path doesn’t work, they have to stop and try another. This is called “low concurrency.”
AI is different. AI agents can work thousands of times at once:
- Agent A is trying to prove Strategy 1;
- Agent B is constructing counterexamples;
- Agent C is searching for relevant theorems;
- Agent D is adjusting parameters…
Most of these attempts will fail, but it doesn’t matter because the cost of failure is very low. In the past, humans wouldn’t dare to try absurd ideas easily because time was precious; now, AI can afford to make many mistakes because it can process things extremely fast.
Economic Insight: The first step in AI’s industrialization isn’t about creating truth but about generating hypotheses. It has turned the once-expensive process of exploration into a cheap form of large-scale searching. It’s like how diamonds used to be found by miners digging one by one; now, AI uses thousands of excavators simultaneously. Although most of what’s dug up is just rock, the speed is overwhelming.
2. A New Bottleneck Has Emerged: We’re Being Overwhelmed by Answers
**Title: Easy to Generate, Hard to Verify; Systems Face a “Verification Debt”
When AI can generate ten thousand research paths in a day, the question changes:
- In the past, the question was, “Can we find the answer?”
- Now, the question is, “Out of these ten thousand answers, how do I know which one is true?”
This creates a huge capacity asymmetry:
- Generation capacity has skyrocketed, with costs approaching zero.
- Verification capacity is growing slowly and remains expensive and complex.
This has already happened in software engineering. AI writes code quickly, but testing, reviewing, integrating, and security checks can’t keep up. If code is written faster than tests, the system crashes. Now, this asymmetry has expanded from coding to “knowledge production.”
Economic Insight: A system that can produce ten thousand answers but can only reliably verify a hundred doesn’t actually have ten thousand usable answers; it only has ten thousand “candidates.” The gap between “candidate answers” and “credible answers” will be the biggest economic friction in the AI era. Whoever solves this verification bottleneck will gain a new competitive advantage.
3. Why Do We Need “Stupid” Tools? The Value of Lean Lies in Its Rigidity
Title: Smart People Need a “Stupid” Butler; the Narrower the Verification System, the More Reliable It Is
The article mentions a tool called Lean, a formal proofing system. Many people wonder: If AI is so intelligent, why use a seemingly “stupid” tool like Lean?
Because AI and Lean perform completely different tasks:
- AI (the generator): Needs to be powerful, flexible, and intuitive, responsible for randomly exploring a vast space of possibilities and coming up with various ideas.
- Lean (the verifier): Needs to be narrow, precise, and strict. It doesn’t understand mathematical intuition; it only follows logical rules. It’s like a precise ruler that will report an error if there’s any logical flaw in the proof chain.
We often make the mistake of thinking that to supervise an intelligent system, we need another intelligent system.
The truth is: The generation system must be powerful (to expand possibilities), while the verification system must be narrow and precise (to narrow down the credible options).
Economic Insight: In business systems, the generation side needs to be innovative and flexible, while the verification side needs to be compliant and certain. Don’t expect AI to verify itself; that would lead to a vicious cycle of “who supervises whom.” The best structure is to let AI create and let the rule engine ensure quality.
4. The Underlying Logic of Trust Mechanisms Has Changed: From “Trusting the Person” to “Trusting the Structure”
**Title: No Longer “Who You Are,” but “Where the Evidence Is”
In the past, we built trust based on the identity of the subject:
- I trust this doctor because he graduated from Peking Union Medical College.
- I trust this bank because it’s a large institution.
- I trust this developer because they come from a big company.
But AI is a “strange” entity: it can be extremely powerful, but it can also make basic mistakes; it might be flawless in one reasoning process and completely wrong the next. You can’t guarantee that AI will always be trustworthy.
The mathematical community has long had this tradition: **the validity of a theorem doesn’t depend on who proves it.” A unknown person’s perfect proof makes the theorem valid; a Fields Medalist’s flawed logic makes it invalid.
AI has spread this idea to all fields. Future systems won’t require the generator to be perfect; they’ll require the results to be independently verifiable.
Economic Insight: This is the beginning of a decentralized form of trust. We don’t need to issue “trust certificates” for every AI agent; instead, we need to establish a system where any AI-generated result must pass independent and strict verification. The goal is not to eliminate errors but to prevent them from being considered “facts.”
5. The Scarce Resource of the Future: Not Answers, but “Right of Confirmation”
Title: Answers Are Cheap, but “Verification” Is Expensive
If this trend continues, the economic structure will change dramatically:
- What becomes cheap: Code, solutions, hypotheses, mathematical conjectures, creative copywriting—everything that’s generated—will have near-zero marginal costs.
- What becomes expensive: Determining what is worthy of acceptance.
When the world has a million candidate answers, the truly scarce capability will no longer be “finding answers” but establishing a reliable mechanism to reduce “a million possibilities” to “a few credible results.”
This is similar to the information explosion era, where attention and trust have become the most scarce resources. In the AI era, verification capability will be the new scarce resource.
Economic Insight:
1. For companies: Your core competitiveness may no longer lie in how powerful your AI model is (because everyone can buy one); it’ll lie in how rigorous your verification system is. Can you quickly and cheaply filter out high-quality, low-risk results from the vast amount of AI-generated content?
2. For individuals: People with critical thinking and verification skills will be more valuable than those who can simply use AI to generate content. You need to be the one with the “ruler,” not just the one with the “paintbrush.”
---
Conclusion: Tips for Everyone
Although this article discusses advanced mathematics and AI architecture, it reveals a very real workplace and economic trend:
1. Don’t Just Focus on Generation: If you only learn to use AI to write code or articles, you’ll quickly be obsolete because AI can do these tasks too cheaply.
2. Emphasize Verification: Learn how to check AI’s output, how to design processes to prevent AI’s mistakes from entering the production environment, and how to establish standardized evaluation systems. These will be the “hard skills” of the future.
3. Understand the Division of Labor: The future workflow will be “AI for divergence, humans/rules for convergence.” Your task is to design the “convergence” mechanism.
To conclude, with the most brilliant sentence from the article:
> “A truly mature system doesn’t require the subject to be always correct; it requires that a result cannot automatically gain the status of a fact, enter a system, or change the state just because it seems intelligent enough.”
In this era of AI proliferation, suspicion and verification are the highest forms of productivity.