Four levels of AI

Author: To Visa.

Programming, Agent, judgement and panorama- How AI deals with certainty and uncertainty.

One, many of the arguments about AI today are actually different from the beginning.

Let AI write a well-defined program and let AI judge whether there is an AI bubble now, not one problem.

It is not a question of having Agent carry out his mission in accordance with the established rules and to discuss who will benefit from AI in the future.

One important reason why there is now much debate about AI is that these different levels have not been clearly distinguished.

In my view, AI deals with the real world and can generally be divided into four levels:

The first level is programming.

The problem is clear, and it's clear how to solve it. Entering what rules are, what rules are, and then how to judge right or wrong is clearer. What AI does is turn people's intentions into procedures, and then make the process rigorous.

The second floor is Agent.

The external environment will change, but “whatever happens” can be defined. Agent can observe the environment himself, call tools, and carry out his mission, but its actions should remain within a pre-defined border.

The third level is judgment.

The question is clear, but the answer is uncertain. “Is there a bubble in AI now?” is one of these. We can find evidence, we can find proof against it or we can interpret it differently, but no procedure can prove what the future must be today. Such questions are uncertain: they can lead to probabilities or conditional conclusions, but they cannot be considered as established facts.

The fourth floor is the panorama.

This floor is more troublesome.

We don’t even know what the real problem is. For example, “Who will benefit from AAI in the future?” It seems to be asking which companies will make money at first. The study will only find that it involves workers, consumers, governments, the state, capital owners, and even people’s cognitive capacity and autonomy.

The first and second tiers have developed a more defined structure. The third and fourth layers are still in uncertainty.

One big problem today in the AI field is that these four layers are often mixed.

Particularly dangerous is the fact that a judgement that has not yet been formed on the third level is directly handed over to Agent on the second level for execution; or that questions that have not yet been clarified on the fourth level are left to AI to demand an answer.

This is what this article really wants to discuss:What exactly should AI do in the face of varying degrees of certainty and uncertainty?

II. True division depends on certainty

The real differences at the four levels are:The extent of this problem has been determined.

The four tiers here are not labeling products, occupations or industries, but describe the state of certainty in which a problem is at the moment. The same job can contain different levels of problems simultaneously and move between levels as awareness increases or rules fail.

The first layer is clear. We already know how to do it, and we just need to repeat it precisely.

The second level, while the outside world will change, we already know what is happening and what actions are being taken.

The third level is different. We know what we're asking, but we don't know what to conclude.

The fourth floor goes one step further: we don't even know what to really ask ourselves.

Four layers can be said in four words:

• First floor: How to do it.

• Second floor: when to do it.

• Third floor: What should we say?

• The fourth level: What is the real problem?

The more you move up, the greater the uncertainty.

First tier: process structure - already known how to proceed

The program structure solves the most defined category of problems. For example: “Consolidate two tables by security code, repeat the latest one with a retention time, and make mistakes when fields conflict.”

This mission may be annoying, but it's already clear. AI can help people write programs, and when they're done, it'll give them to the machine for steady execution. Here's one thing that matters:

The certainty of the procedure is not that the code itself was created, but that before it was written, the issue had been made clear.

If you ask me to replace it with, "Do me a good stock analysis system." Things are going to be different. What's "good"? What's the data?

If AI starts writing code at this point, it can only guess by itself. The program may run up, but not necessarily what the user really wants. So we often see when actually using AI programming:The clearer the intention, the better the procedure; the broader the title, the more loopholes.The reason is not mysterious. AI is not eliminating uncertainty, but is secretly supplementing the rules that are not spoken out.

So the most important sentence on the first floor is:AI programming can only translate established rules into definitive implementation.

Second tier: operational structure — the world is uncertain, but rules of conduct have been established

The focus of Agent in this paper is not on which technology is used, but rather on the fact that it has already been granted some autonomy: to observe the environment within its stated objectives, competences and operational boundaries, to call upon tools and to carry out its tasks.

Agent faces more complex situations. The outside world does not repeat itself by a set of fixed inputs. For example, an unmanned car does not know if the next person in front will suddenly cross the road.

But it can be done in advance: when to slow down; when to stop; where to stop; where to stop; and when to return control. That's the structure of action. So what Agent really does, not all the uncertainty, is it?While the external environment is uncertain, the rules of operation have identified problems.

If Agent is going to enter a scenario that will have real consequences, I think there are at least a few things to do:

• Know what the target is;

• Capacity has been tested;

• Know when to go on;

• (a) Restrictions on the exercise of their powers;

• (b) Be able to stop after failure;

• The course of action can be replayed.

And the "returnable" here is not just a log left by the system. It should be done:

A third party that was not involved in the operation of the original system could also re-evaluate, on the basis of the record, what it had seen at the time, under what rules, why it had acted and whether it had exceeded its authority.

It is only then that it can be established that Agent was wrongly executed; the data were wrong; or the original rules of operation were wrong.

So the real question on the second level is not "Agent will do it?"“To what extent does Agent qualify for it?”

Third floor: structure of judgement — the question is clear, but the answer is not yet known

The third level is the one that is now most susceptible to confusion in AI applications.

For example, “Does the AI industry have bubbles now?” The question we know is investment, demand, utilization, finance, prices, income, cash flows. But it will not happen in the future, and today it will not be possible to get a definitive answer like running procedures.

It's not Agent, it's a set.Structure of judgement。

A complete judgement should at least answer:

• What are we judging?

• What are the facts supporting the current judgement;

• (a) The facts against it;

• (b) Whether there is any other explanation for the same set of facts;

• What are the boundaries of the evidence;

• Are there any important variables missing;

• What is most reasonable in the light of these evidences now;

• This judgement must be revised as new facts emerge.

So the third level can be institutionalized. But institutionalization is not the answer.

Institutionalized.How do we form judgment?

The method can be fixed, the conclusions cannot be fixed. This is the fundamental difference between the third and second levels.

The second tier can already say: “If this happens, so.”

The third level can only say: “We will judge this for the time being, based on current evidence.”

This also explains why simply bringing in a few more Agents does not solve the third level of problems.

A person who sees more, a person who can't watch, and then a few votes will not suddenly make the future certain. Five can be wrong together.

Different viewpoints can compete at the judgement stage, but unsolved differences cannot be directly granted the right to act.

So what the third level really needs is a judgement structure, not more Agent.

Level 4: Panorama Structure - Not even a real problem has been found

The fourth floor is the hardest. It's often not a clear question at first, but just a feeling:"It doesn't seem right here."

Or:"This explanation is too simple for you." We have only one vague concept. We don't know what the real problem is, we don't know which variables are most important, and we don't even know by what scale to look at.

For example, we study: "Who will benefit from the future?" If you follow normal AI usage, you can get a list in a few minutes:

Chips, Cloud Calculator, Software, Robotics. It looks like it's complete. But when we really do this, it's not just a company.

Those who benefit may include: business; capital owners; workers; consumers; Governments; and the State.

The “benefit” itself may also be: profit; wages; consumer surplus; taxation; national competitiveness; individual capacity; and even cognitive autonomy.

What does that mean? It means that the problem at first has not really developed.

So the most important thing on the fourth level is not to define questions immediately, let alone answer them immediately.Let's keep the issue open.Let different perspectives come in; let different facts come in; let reverse examples come in; let different scales come in.

The normal search is: I know what the question is, and then I find the answer. That's not the whole picture.

The whole picture is that I know that there may be problems here, but first try to open up the world as possible and then see what the real problem is.

When can we get to the panorama? A practical criterion is that objects, scales and key variables are basically stable, that new information changes conclusions and does not change “what are we asking” from the point of view to the third level.

Our research processes are usually:

Sweeping through the gaps, building structures, attacking structures, forming judgement, compressing documents.

The most important step here is actually “attack structure.” Because once people and AI have a beautiful explanation, they're becoming more and more fond of it. It's easy to find all the information that you can find to prove right. So we have to change the key words deliberately; we need to look for the opposite view.Looking for the hardest to explain; even asking: Are we asking wrong questions from the outset?

This is not about gathering more information. It's really about preventing:The first framework, which seemed reasonable, closed the issue too early.

So the fourth level uses AI, and I think it's very important:Let's use AI to expand the world, not let it define the world first.

vii. Why is it difficult to scale the landscape into a way without AI?

There were, of course, people who had a great sense of holisticism in the past. But that was often the result of long-term accumulation, talent and interdisciplinary training, which was difficult to repeat.

People have limited reading speed; limited memory; limited search scope; and limited cross-cutting comparative ability. When it comes to the fourth tier, it is easy for a person to capture his/her best-known explanation and dig down. It may end up digging deep, but just studying like legs or noses. LLM and search change not only speed. More importantly, they decrease dramatically.Keep it open.And the cost of it. AI can quickly launch dozens of different perspectives; search across different domains; compare different theories together; change keywords; look for counter-examples; and attack the structures that have just emerged.

The cost of this was extremely high in the past. So AI's real new research value is not just to answer a question more quickly.Turning a vague feeling into a real question worth answering more quickly.

In this sense, for the first time, LLM and search significantly reduced the cost of expanding the cognitive vision of a common human system. But it does not guarantee that the vision becomes complete: if there is no active change in frameworks or an alternative, AI will also bring people into the first beautiful answer.

VIII. STEPS DOWN THE QUESTION: FROM A SIDE TO PROCEDURAL

When the four levels are linked, there is an interesting process:The panorama structure, the judgement structure, the operational structure, the program structure.

We call it:Question down.

The fourth level is to find the real problem through a landscape. When the problem is more stable, it goes into the third level and begins to form a judgment. The parts of the judgement that are already stable and can become rules continue to sink into the second layer of operational structure.

And this process has been going on. It's probably the past: experts have been doing it for years, slowly building up experience, forming systems, creating processes, and finally writing software.

Today AI is accelerating the process dramatically. It may turn into a blurry problem AI is rapidly spreading, and people find gaps AIS continues to search for and form judgement, turning stabilization into operational rule AAI programming program execution. So AI's real impact on the economy is probably not just "doing one's job many times faster."

A deeper layer is that AI is accelerating those parts that can be identified and validated, from uncertainty to certainty about the process.

Nine, but not as much as possible.

It's easy to go to another extreme. Since it can be downgraded, should all the questions become Agent and then the program? No.

The principles that should be truly justified should be:The problem can only be reduced to the extent that evidence permits.

Many of the third-level issues can be kept on the third level for a long time. Many of the fourth-level issues may also remain open for a long time. Only those that have really achieved certainty should continue to go down.

Otherwise there's a very dangerous thing:Hypocrisy.

For example, it would have been a complicated question of how well a business was operating. To manage itself, it could then be put down as a KPI. The program could then do the best that could happen to KPI. But what if KPI did not represent the real goal of the company?

The more accurate the machine is, the more stable the error is. The LLM era must be particularly vigilant about this.

Because one of the things that AI's best at doing is to make up for what is missing is very complete. So today there is a new risk, which was not so obvious:

In the early computer age, there was a fear that machines would not solve unstructured problems; in the generation of AI, we had to prevent machines from disguiseing unstructured issues as already resolved.

The machine didn't do it in the past, the questions were easily exposed. Today AI can answer almost anything.

The real danger is:It's done so much like real.

Ten, it must be able to rise again.

Realism does not always work by yesterday's rules. Markets change; technology changes; participants learn; regulation changes; indicators that are valid can fail. Stock markets are particularly obvious. A rule that, if everyone knows, trades ahead of schedule may be useless in itself. So a truly reliable system cannot just be the fourth floor, the third level, the second level, the first level.

It must also be reciprocated. If the program or Agent continues to experience anomalies, it should go back to level three: is the original judgment wrong? If it is further discovered that not even the ruler was wrong, then it has to go back to level four: Are we asking the wrong question? That's “upgrade.” A truly mature AI system, not just knowing when to do it automatically, must know:When did it become ineligible to continue doing so automatically?

Eleven, before giving orders to AI, make a judgment.

The four layers are different, but there is one common denominator. Unless it is a very simple and specially defined first layer of tasks, none of them should skip a process:We'll discuss it, we'll make a judgment and then we'll give instructions.

• The first tier is the main confirmation: the specifications are clear.

• The second tier is primarily confirmed: whether rules of operation, competencies and conditions for cessation have been established.

• The third level of the main discussion: where exactly can evidence and judgement go?

• And the fourth floor is ahead: first find out what's going on.

So AI uses a really universal process, and it shouldn't just be: people write Prompt for AI execution.

More rationally:Fuzzy intent, human discussion, judgment, clear instructions, AI, Agent or program execution validation.• Reopen the discussion as necessary.

A lot of AI failed, not because Prompt was bad.Before the judgment was formed, it was already urgent for AI to start working.

It's very dangerous because AI will automatically fill you in with the missing “judgements”, and we don't necessarily know if they are valid.

Conclusion: AI is not a capability, but four different uses

Now you're asking: "What can AI do?"

The problem itself may be too thick. AI should be in different positions in the face of different problems.

• If you know what to do, you go into the program structure.

• If you know what action should be taken in the event of a situation, enter Agent.

• If the question is clear, but the answer remains uncertain, it will remain at the judgement level.

• If the real problem is not even found, make the whole picture first.

So it's really important not to give more and more things to AI.

It's about asking:Which floor is this question on?

I am increasingly thinking that this may be the basic principle of the AI era in dealing with certainty and uncertainty:

Not all questions require answers.

Some issues require judgement first.

There are still a number of issues that need to be addressed first and foremost.