10 thoughts on today's AI: It won't take your job, it will only create qualified "cows and horses"

Author: BALAJI

Compiled by Tim, PANews

Here are 10 thoughts on AI that I personally believe are practical on an economic level. Let’s get started.

1. First, there is no single general artificial intelligence (AGI), but rather multiple AGIs. In other words, what we're actually observing is a multimodal AI landscape (multiple powerful models coexisting), rather than a single dominant AI (a single, all-powerful model monopolizing). The current trend is that numerous models from different camps have reached similar capabilities, rather than a single, top-tier model creating an absolute generational gap. Therefore, we can foresee that the future will be one of multiple forms of human-AI integration, balancing each other, rather than a single dominant AGI turning all humans into paperclips (implying destructive control or punishment).

2. Current AI shifts costs to the input prompt and verification stages. Essentially, today's AI only handles tasks between intermediate steps, not the entire end-to-end process. Therefore, even if AI accelerates intermediate steps, all business costs will still shift to the input prompt and result verification stages.

3. AI stands for augmented intelligence, not artificial intelligence. Today's AI lacks true subjective consciousness because it doesn't exist completely independently of humans. Existing AI systems are unable to set complex goals or effectively verify their outputs. Humans still expend considerable effort on setting goals, verifying outputs, constructing prompts, and integrating systems. This means that the smarter the user, the greater the AI's intelligent amplification capabilities. Therefore, it should essentially be called augmented intelligence, not artificial intelligence.

4. AI won't take your job away, but will enable you to do any job. It can make you a barely qualified UX designer, a decent special effects animator, and so on. But this doesn't mean you can really do the job well, because final polish often requires professionals to do it.

5. AI isn't replacing human jobs, but rather jobs performed by the previous generation of AI. For example, Midjourney replaced Stable Diffusion, and GPT-4 replaced GPT-3. Once you offload a task (such as image generation or code writing) to AI, you simply invest your budget in the latest model. Therefore, it's always the new generation of AI that replaces jobs.

6. AI is better at visual expression than text. This means AI has a greater advantage in front-end development than back-end development, and is better at image and video processing than text processing. This is because user interfaces and images can be easily verified by the human eye, while large blocks of AI-generated text or code require significant human effort to verify.

7. Lethal AI has already arrived, in the form of "killer species" like drones. Every country is pursuing this technology. So, the real concern isn't image generators or chatbots.

8. AI is probabilistic, while cryptography is deterministic. Therefore, cryptography can act as a check and balance for AI. For example, AI can crack CAPTCHAs, but it cannot forge on-chain balances. Furthermore, it can solve some equations, but not cryptographic ones. Therefore, cryptography generally represents what AI cannot do.

9. From an empirical perspective, AI is driving decentralization, not centralization. The decentralization effect currently exhibited by AI is undeniable: this is driven by the coexistence and development of numerous AI companies, the ability of small teams to achieve significant improvements with the right tools, and the continued emergence of high-quality open source models.

10. The optimal AI penetration rate isn't 100%. After all, 0% AI is slow, while 100% AI is garbage. Therefore, the ideal AI penetration rate is somewhere between 0% and 100%. While the specific value varies depending on the situation, it's important to understand that neither 0% nor 100% is optimal. This is the Laffer Curve for AI:

Today’s AI is limited

In short, fundamentally, this is a restricted AI model, not an omnipotent AI model.

The economics of AI are limited because each API call is expensive and competing models are constantly emerging.

AI is mathematically limited because it cannot (provably) solve chaos, turbulence, or encryption problems.

AI has practical limitations because it requires human prompting and verification, and performs tasks by operating between intermediate layers rather than through a complete end-to-end process.

AI is physically limited because humans are still required to perceive the environment and input this information through prompts, rather than AI collecting environmental information on its own.

What is clear is that these limitations may be overcome in the future. Someone may be able to unify the probabilistic thinking of artificial intelligence with the deterministic, logical thinking of traditional computers, but this remains an open research question.