Why is AI not working yet?

Visual China.

Text |Li Zhiyong.

Clearly, AI is an infrastructure that redefines what computing is and how it works

If compared with the past, thenProgramming in the past solidified the intelligence of the programmer, and the intelligence of the programmer handled problems within defined boundaries through the programTherefore, the rise of pan-IT is accompanied by the rise of the programmer group, whether it is the number of programmers or the income. AI folds this intermediate link to a large extent, and the dialogue is computation, and at the same time makes this computation more generalized and borderless。From this point of view, the rise of AI is destined to be accompanied by the decline of the programmer community (first in numbers, but not in the sense that the profession is gone).

User - programmer - program - computer - computing mode of the Internet).

User - the computing model of artificial intelligence).

The ultimate success or failure of infrastructure must depend on the outside, not on its own characteristics, such as whether it is good or not, and how big the model isWait. In the past, microkernel operating systems were once expected, but in fact, neither Linux nor Windows is a microkernel, and pure microkernel systems such as minix are just teaching aids.

The same is true for AI as infrastructure, which must move beyond mere content generation tools to become a general-purpose computing platform that provides new computing methods for a variety of occasions if it is to succeed.

In the past, both Windows and Linux provided this kind of versatility, from ATMs to large screens at airports, to set-top boxes at home, and even a little smart alarm clock, they were providing basic computing power. (Sometimes these systems crash, and it's surprising to see that many of them are actually XP instead of crashing.)

When AI breaks down the boundaries of content generation tools, it becomes the new computing base (the universal counterpart of general artificial intelligence is in this context). AlsoOnly after becoming this kind of universal computing base can AI truly usher in its own sea of stars.

InThere is now an invisible genetic chain between today's content-generative AI and this general-purpose computing base.

This genetic chain is the boundary of the content generation tool.

The GC (Content Generation) of AIGC is both the fulcrum and the chain.

In the field of tools, the AI has actually done well enough this time, but the pool is too small, and if it does this, it will suffocate everyone.

Note 1:If you are really interested in the technology of the new computing model under the large model, you can scan the end of the article*** to aim,This course is done by a buddy of mine and Zhihu.,It's a more serious and non-nonsense course.。 We had a lot of contacts when Zhuo Ran was a triangular beast in his early years, and we also followed this model together in GPT2, and I still have a spectrum of level, but the course still feels a little difficult, and it also costs at your own discretion.

Note 2: For more information about the computing mode of AI, see "Will the open-source large model Llama 2 play a role similar to Android?".Let's take a concrete example to see why this pool is too small.

Starting point Chinese online there is a web writer pen name called I eat tomatoes, this 1987 classmate was originally a college student majoring in mathematics at Soochow University, according to the normal trajectory after graduation, there is a high probability that he cannot continue to do mathematics-related work, at that time employment is likely to be a programmer and other related directions. But he didn't take the usual path, and started to create online articles during college, and achieved good results, and in November 2012, he ranked second on the "China Internet Writers Rich List" with a royalty income of 21 million.

Suppose he writes a book of **3 million words a year, and enlarges and counts 10 million tokens. Now I don't write this part myself, I use AI. Just choose the ** of a large domestic model as a reference, according to 1500 yuan 50 million tokens, then this part of the income that can be created by artificial intelligence companies is 300 yuan, accounting for a little more than 100,000 in the 21 million income. Zoom in againIf there are 10,000 tomatoes that I eat, then AIGC can earn a total of 3 million in the online literature industry。This is not enough for a team's salary for a year, especially in the case of high-end talents, which is not even enough for one person.

If the large model only does content generation, the value created is roughly the same as the existing value of the industry.

And a lot of people have rushed in, it's very much like making a small pond with a bunch of sharks, and when they are so hungry, they can only desperately roll up and kill each other, and then there is a high probability that they will all die, and there is no one left.

If AI can't go one step further than GC, it must be this ending: a high degree of involution with happy expectations.

This kind of involution is a complete negative feedback and a dead end for AI as a whole.

Everyone expects a new general-purpose computing platform and application, but it's really just a content generation tool that creates a little bit of new value. How can it be possible to live up to the feedback for a long time!

So how can AI come out?The answer is that you need to pass the Turing Test 20。

Note 3: For a summary of AI's business model, see Can AI Make Money?The original Turing test looks like this:

This is a purely intelligent test, and the essence is to pursue logical self-consistency in a closed system.

Now let's add a similar concept of agent to this test:

This is Turing Test 20。and 1What is the core difference between 0 and 0?

To hallucinate, there are boundaries.

1.0 is a volley system, and the illusion of plausibility actually helps to pass the test, but 20 no,The tester receives feedback from both the real-world scenario and the testeeThe second is that the limitation of the test boundary requires a higher depth of intelligence. This is very similar to what Zhao Kuo can say when he learns the art of war, but he may not be able to fight;Being able to fight a war doesn't necessarily mean that you know everything in the art of war, but you have to know one of the water stations, land battles, and horse stations.

The ability to break the boundaries of content generation and become a modern general-purpose computing platform for every occasion depends critically on whether intelligence can keep up. Whether intelligence can keep up depends on whether it can pass the Turing test in each scenario20。

Linux and Windows are old-fashioned ways of exporting intelligence, but they provide enough certainty that they plus programmers have reached what used to be called software devouring the world. This is a good enough computing model, but the current AI is not.

Now large models and the like do provide better forms of computation, but the key is that it is not intelligent (it can't pass the Turing test 2.).0) Result in the combination of the past system and programmer cannot be replaced. The boundaries of intelligence limit the boundaries of an application.

Note 4:Turing Test 2For the expansion of 0, see "The progress of AI is not too fast, but too slow" At that time, not only customer service and outbound calls will be built based on AI, but every existing application (office and so on have begun, and the game is likely to explode with a real multi-dimensional narrative is highly random, focusing on intelligent modern games), advertising screens, smart speakers, TVs and even mobile phones will be reorganized. BecauseThe basic computing paradigm has changed, and its interaction carrier will inevitably change, and the magnitude of this change may be greater than the change from PC Internet to mobile Internet. From this point of view, it can be entered into the robot must be the next general-purpose computing platform type product.

Extremely, except for extremely mechanical products such as neon lights and calculators, everything else will change.

This perspective can describe the density of intelligence that becomes a sceneObviously, the intelligence density of screws is lower than that of math problems.

The higher the intelligence density, the more the calculation method and the corresponding product will change, because the value is greater. Then match the perspective from digital to physical, with or without illusion. Take the Turing test 20 is the foundation, and the combination of these three perspectives also restricts the development path of future intelligent applications.

The embodiment of the implementation of this route is what we often call the agent.

If we draw a coordinate system of the density of intelligence (the origin is 0), the degree of physics (the origin is 0, which represents the pure digital application), and the harmfulness of the illusion (the origin is 0, which means the illusion is harmless), and arrange this diagram in the center position, then in the following diagram, what is the most head?

The answer is likely to be:Games, multi-dimensional narrative games.

Note 5For the form of landing application of large models, please refer to "AI Agent: The Value Bridge between Large Models and Scenes, but Not Suitable for Pure Technology" to pass the Turing Test 20 to have a real agent. However, it should be noted that the agent is not an extension of the larger model, but a new species. Making a car engine is not the same thing as making a car, although a car can't run without an engine at all.

Only agents can start the AI wave, and the agents that can start the AI wave are not other applications that simply integrate AI features, but intelligent native agents。In this case, the agent is not only a conduit for delivering intelligence to specific scenarios.

Composition of a smart native app:

In this mode of thinking, AI is natively destined to be placed at the center of a structure:

Here, the large model plays the mode of the engine, and it cannot pass the Turing test 20, then the smart native application will be much like a car pulled by a mule.

After passing this test, and then adding the above-mentioned perception, goal, feedback, and enablement links, the agent can truly become the implementation carrier of new general computing. Only by passing this test can the engine be replaced with a steam engine, an internal combustion engine, a turbocharger, etc., step by step. The expansion of the scope of the agent is destined to be a process of step-by-step improvement of intelligence.

Note 6: For more information about intelligent native, see From Mobile App to AI Native AppThe answer is no. That's why the current AI is not good.

We could lower the standard and make it more vertical, and as long as the range is narrowed enough, all the tests will pass, but that doesn't make sense. Let's take a look at the previous specific example to understand the entire logical chain above:Can't pass the Turing test 20, it can't become the base of intelligent native applications, it's just stuck in a small pool like AIGC, so the current AI is not good.

If you want to start broadcasting, but you don't want to be on it yourself, but want to be your own number** or doppelganger, then what do you have to do if this number** person really wants to achieve results?(Achieving an effect means that someone is willing to watch it, there are fans, etc.).

The first is the most basic production and research partFirst build your own shell, that is, the image should look like that, and then match it with the ability to see, hear, speak, and think(Input/output, storage, and CPU for computers.)Seeing, listening, and speaking are basically technologies that have been honed repeatedly over the past decade, such as image recognition, language recognition, speech synthesis, etc., and the thinking part is based on a large model, which is responsible for synthesizing various inputs to produce its own output. When the programmer connects all of this, there is basically a digital thorn that can give feedback based on various inputs from the audience. But the product basically completes the hand-eye part here, and the brain part belongs to it, but it is not easy to use. At this time, even if the best large model is imported, it is still a very stupid bot, let alone achieving results, basically no one will finish watching any live broadcast paragraphs。At this time, it is not promising to make efforts to involute in a simple single-point technology (including large models), so you can't get fans and you can't retain them, and the return is roughly 0.

The first step to improve is definitely to hope to add personality traits, so that its personality traits are more similar to yours, such as whether you are friendly to others, whether your expression is sharp, and you need to be more social: Able to speak, able to communicate with feelings, etc. At this time, try to remember what you have said to someone in the past. This part is not purely technical, but the technical relevance is still very high, and it is usually necessary to find an old driver who has done it in the past, and it is estimated that it will not be possible to do it purely (Note 1That's the value of the lesson.) This step is a threshold, and it is calculated to pass the Turing test 10, others can't tell if it's you or not, but now it's actually impossible to completely handle this, it's okay to chat without borders, and it's not as good as you think when it comes to personality characteristics. If you can't figure it out, what will happen?Will look a little bit smart and a bit like you, chattering there, but uncharacteristically and interestingly. Will you be able to attract fans?It depends on what you're broadcasting. I guess it's not possible to broadcast the animal world, and the entertainment is probably enough. This is the topic below, and the key factor is further shifting from technology to product.

Pass the Turing Test 10 smart products are already useful, before this is a pure tool, after this is a bit of an agent, but the value is not as big as thought.

Pass the Turing Test 1What is the use of a digital clone like 0?

Its advantages are large information throughput, tireless, and human-like;The disadvantage is that I am not smart enough, I can't do a good character, talent, outstanding opinions, interesting improvisation, etc. So what is the right thing to do?It's suitable for doing things where the content itself is interesting, and the anchor is a supporting role.

What are those things?For example, broadcasting the animal world, telling stories, broadcasting news, and occasionally interspersed with some interaction. This is essentially a better smart speaker.

What is this doing?It is to reduce the need for intelligence in the scenario. If the supply of intelligence is insufficient, it can only be downgraded.

So what does the ideal situation look like?

Ideally, this digital clone should also be able to access real-time hotspots, dynamically generate content to be output, such as **, and then be the anchor。This kind of hot spot should match everyone's concerns, be novel, and match the rules of the platform, not only the positive rules, but also the reverse scale, otherwise it will be carried away or **. This part will derive a lot of detailed work, such as the theme is now the main promotion, which has to follow the platform, otherwise if you have bad weight, it will not push you, and it will be in vain. This is an intelligent process for the platform, a process of comprehensive analysis for the audience, and a creative and innovative process for creation. This thing is capable, and it can be regarded as passing the Turing test 20, once it passes, it can at least be divided into two worlds with humans. can't pass, for example, regardless of the timeliness of the content or the capture of hot spots on the platform, it's still half a job!It's the lack of smart supply. If this part is successful, then it can basically have fans. It's only here that my brain has grown up, and I've developed my own style.

Assuming this can be done, is it over?

Not yet. These are all done, and the relationship between silicon-based intelligence and silicon-based intelligence is mainly solved, which is equivalent to being able to compare the rules of the platform and the hot spots of reality.

Guo Degang's main method of praising people is to repeatedly mention this person. If you are a streamer, it will obviously work better if someone pulls it. Then who to cooperate with and how to cooperate still needs people to do it. Only when this scene is fully covered can you be considered the real **.

What is said above can be summed up in a diagram that is analogous to autonomous driving:

From this point of view, less than 10 percent of what can be done is now. What's more, this is just a relatively simple C-end scenario, and the B-side scenario is much more complex than this.

Note 7For a more detailed description of this part, see "For example: the brain, hand, and mind of intelligent native applications" AI has always been in such a state, once there is a little breakthrough, everyone will rejoice, and then the expectation will go up, what can match this expectation immediately?It's capital and marketing heat. So it's going to get all over the place very quickly and see a lot of high valuations. ButIt is much more difficult to match the business with this expectation, but this is the first nature of the industry. The speed difference between these companies in different industries is not the same, the matching speed of the Internet is actually the fastest, and the matching speed of AI is likely to be more like traditional software, second to the Internet, but faster than consumer products.

Note 8:The *** original link to the previously mentioned large model course is also it.