The leaky hole of abstraction gets smaller.

Discussions about letting AI do the programming are gaining momentum in 2025.

This is not the first time a system in which concrete, lower-order linguistic expressions are generated by higher-order linguistic expressions of a higher level of abstraction.

Compiler generates machine language description from C description

LLM generates programming language descriptions from natural language descriptions

https://gyazo.com/0d8bb71bc34d857c71a39472647f0fe2

A:

It starts out in a "generated, but useless" state.

B:

Next, "useless as generated," "but usable if reworked."

If the cost of rework is less than the cost of creating it from scratch, it makes economic sense.

Up to this point, knowledge of lower level languages is required

C:

Opinions differ on whether "useful in its generated form" comes

Argument that "abstraction has leaks, so it never completely eliminates the need for lower layers of knowledge."

Software engineers who disagree with the opinion that "C will be realized" as of 2025 often use this as a basis for their position

On the other hand, how many people are not satisfied with the quality of machine language generated by C compilers as of 2025?

You'll need it in cases where you want to build the fastest thing in the world.

What percentage of people who call themselves software engineers today are in the business of creating the world's fastest implementations?

https://gyazo.com/b60bbde37eecadae97ec03e0824344ef

Lateral deviation is bias, spread of distribution is variance

https://gyazo.com/cd488ea841e660fcf52b201f87b1d812

The probability of getting into range R of the expected output is different for A with a large variance and B with a small variance.

As technology advances, the variance gets smaller and smaller, increasing the probability of getting into Range R.

No matter how much it increases, the probability of the overhanging E is not zero.

When it is sufficiently high, the importance of high-level linguistic expressions that appropriately produce output in the expected range of output increases relatively more.

The probability of a human putting a bug in a C description is higher than the probability of a compiler making a mistake.

The expected return on "checking the compiler's output machine language with the suspicion that the compiler is wrong" and "checking the C description with the suspicion that the human is wrong" is higher for the latter.

If we compare "checking the source code output by the LLM with the suspicion that the LLM is wrong" and "checking the natural language description given to the LLM with the suspicion that humans are wrong", the former is still likely to have a higher return as of March 2025.

I think the importance of the correctness of setting the "expected output range" in the first place increases when looking at the activity as a whole.

[It is futile to make something that no one will use.

The ability to understand the needs of users in the world and set an appropriate "expected output range" will account for a higher percentage of the overall value creation process

relevance

---