AI on kenji.blog

The Wall of AI Development

Sun, 20 Jul 2025 21:52:42 +0900

The Wall of AI Development

〜Made it to the 8th station, but from there it was hell〜

Hello, I’m kenji.

Recently, I’ve been hearing a ton of people saying, “I asked AI and it made an app for me!” We’ve moved past the era of no-code and low-code, and are entering a phase of “vibe coding” (writing code by feel).

For example, if you say, “Make an app that processes images and posts them to SNS,” AI will spit out the code and even the UI. It’s amazing, you’d think we don’t need humans anymore, right?

But you know, that’s like landing at the 8th station of Mount Fuji in sandals.

Getting to the 8th station looks “easy”

Development with AI is really easy at first. File I/O? It runs for now. Network? Well, it connects. Database? JSON is fine. UI? ChatGPT spit out React code. Billing? Just copy-paste the Stripe API.

Here, you get the illusion of “Am I an engineer now?”. But the real hell begins from here.

Why you get stuck

The reason is simple: “AI makes it look like it’s done, but leaves the final adjustments entirely to humans.”

Data gets lost in File I/O
DB normalization is weak so searches are slow
UI is not intuitive
Crashes from access concentration because no load testing was done
Rejected by Apple/Google review
(Almost) causes an uproar by violating legal requirements

In other words, **AI is good at creating a prototype that looks like a finished product , but ** finishing it into a product that actually works in the real world is still a human’s job today.

But there’s a paradox

Here you notice one paradox.

What if AI becomes able to truly do everything in the future? In other words, what if “AI can perfectly handle security, billing, and design all by itself”?

At that time, the need for humanity to create apps itself might disappear.

Because if a user directly tells the AI, “I want to do this kind of thing,” the task will be executed immediately without an app.

🌀For example

In the past, we “launched a calculator app to calculate”, but now it’s done just by saying, “Hey Siri, what’s 12 x 32?”.

In the same way, if you say “AI, process this photo and share it”, the process might be completed without going through any app UI or API.

In other words, when the goal of “making apps with AI” is realized, a world where apps themselves are unnecessary might arrive.

In the end, what we can do now

So, what should we do?

Right now, “above the 8th station” is still the human domain, so polish that
Focus on essential user understanding and service design
Think about “what’s valuable” rather than “what can be made”
Rather, explore ways of providing value that are “not bound by the format of an app”

AI is a tool, a lift, and sometimes a competitor. But, “what to make” and “why to make it” remain questions for us humans.

Conclusion:

The wall of AI development is not a paradox of technology, but of structure

An era where anyone can get to the 8th station of Mount Fuji. But the real thing starts from there.

And by the time we reach the summit, a future might be waiting where we ask ourselves, “Did we even need to climb the mountain?”.

Steps to call TinyLLaMA from C++ (using llama.cpp)

Sat, 19 Jul 2025 09:40:53 +0900

✅ TinyLLaMA × C++ Setup Steps (using `llama.cpp`)

🔧 Step 1: Prepare llama.cpp

1-1. Required Environment (Minimum)

OS: Windows / Linux / macOS
Development Environment: g++ / clang / MSVC
Git / CMake

1-2. Get and Build llama.cpp

1
2
3
4
5
6


git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
mkdir build
cd build
cmake ..
cmake --build . --config Release

On Windows, it is easier to use cmake --build . --config Release in the Visual Studio Developer Command Prompt.

📦 Step 2: Download and Convert the TinyLLaMA Model

2-1. Get the Original Model from HuggingFace

Example: TinyLLaMA-1.1B

1
2
3
4
5
6
7


# Download using transformers if necessary
pip install transformers huggingface_hub

python3 -m transformers.models.llama.convert_llama_weights_to_hf \
 --input_dir ./TinyLlama-1.1B-Chat \
 --model_size 1B \
 --output_dir ./hf_model

This is a step to convert into Hugging Face format.

2-2. Convert to GGUF format (for `llama.cpp`)

1
2


cd llama.cpp
python3 convert.py ./hf_model --outfile tinyllama.gguf

2-3. Model Quantization (Size Reduction)

1

./quantize ./tinyllama.gguf ./tinyllama-q4.gguf q4_0

q4_0 is 4-bit quantization. The model size will be reduced to around 350MB.

🧪 Step 3: Call the Model from C++ (Code Example)

3-1. Simple C++ Code (Inference)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41


#include "llama.h"
#include <iostream>

int main() {
 llama_model_params model_params = llama_model_default_params();
 llama_context_params ctx_params = llama_context_default_params();

 llama_model *model = llama_load_model_from_file("tinyllama-q4.gguf", model_params);
 llama_context *ctx = llama_new_context_with_model(model, ctx_params);

 std::string prompt = "The user says they want to read Excel data, filter it, and save it. What is the node configuration?";
 llama_batch batch = llama_batch_init(512, 0, 1);
 llama_token BOS = llama_token_bos(model);
 batch.token[0] = BOS;

 // Tokenize
 std::vector<llama_token> tokens(prompt.size() + 8);
 int n = llama_tokenize(model, prompt.c_str(), tokens.data(), tokens.size(), true);
 tokens.resize(n);

 for (size_t i = 0; i < tokens.size(); ++i) {
 batch.token[i + 1] = tokens[i];
 }

 batch.n_tokens = tokens.size() + 1;
 llama_decode(ctx, batch);

 // Get Inference Result
 for (int i = 0; i < 50; ++i) {
 llama_token next = llama_sample_token(ctx, nullptr);
 std::cout << llama_token_to_str(model, next);
 llama_batch next_batch = llama_batch_init(1, 0, 1);
 next_batch.token[0] = next;
 next_batch.n_tokens = 1;
 llama_decode(ctx, next_batch);
 }

 llama_free(ctx);
 llama_free_model(model);
 return 0;
}

🧱 Step 4: Compilation Method (Example)

1

g++ -I./llama.cpp main.cpp ./llama.cpp/build/libllama.a -o tiny_infer -pthread -std=c++11

libllama.a will be created in the build/ directory after building.

✅ Example Structure of Output (Organized)

1
2
3
4
5
6


my_app/
├── tinyllama-q4.gguf # Quantized model (~350MB)
├── main.cpp # C++ code above
├── llama.cpp/ # llama.cpp core
└── build/
 └── libllama.a # Compiled library

🧠 Additional Notes for Applying to Use Cases

Include code in C++ to match and select node templates based on the output
Example: If “Excel”, “filter”, and “save” are included -> Generate corresponding nodes
A simple structure like if statements + JSON template loading is fine for this part

📌 Summary

Item	Content
Recommended Model	TinyLLaMA-1.1B-Chat v1.0 (GGUF + Quantization)
Size	~350-450MB (4-bit quantization)
C++ Integration	Possible using `llama.cpp`, almost no external dependencies
Processing Capability	Sufficient for basic intent understanding and output generation (Natural Language -> Structure)
Scalability	Can be made into a node generation AI by combining slot filling and template calling

How to generate illustration images using AI (Stable Diffusion)

Thu, 06 Apr 2023 00:43:19 +0900

What is Stable Diffusion?

Stable diffusion is an AI developed by a research team at the University of Munich in Germany that generates images from inputted text information. By training on various images, it is possible to generate a wide variety of images, from photorealistic ones to illustrations.

This time, I will introduce how to generate illustration images using the pre-trained data of Stable diffusion.

What you need

Google Account

That’s all.

Generation steps

Open https://colab.research.google.com
Select File on the top left and choose New notebook
Select Edit and choose Notebook settings
Change the Hardware accelerator to GPU
Paste and execute the following code

1

!pip install diffusers==0.8.0 transformers

Paste and execute the following code

1

from diffusers import StableDiffusionPipeline

Paste and execute the following code

1
2


pipe = StableDiffusionPipeline.from_pretrained("gsdf/Counterfeit-V2.5")
pipe.to("cuda")

Paste and execute the following code

1
2
3
4
5


prompt = "((masterpiece,best quality)),1girl, solo, animal ears, rabbit, barefoot, knees up, dress, sitting, rabbit ears, short sleeves, looking at viewer, grass, short hair, smile, white hair, puffy sleeves, outdoors, puffy short sleeves, bangs, on ground, full body, animal, white dress, sunlight, brown eyes, dappled sunlight, day, depth of field"
n_prompt = "EasyNegative, extra fingers,fewer fingers"

image = pipe(prompt, negative_prompt = n_prompt).images[0]
image

The Prompt used here is based on the Prompt at https://huggingface.co/gsdf/Counterfeit-V2.5.

AI on kenji.blog

The Wall of AI Development

The Wall of AI Development

Getting to the 8th station looks “easy”

Why you get stuck

But there’s a paradox

🌀For example

In the end, what we can do now

Conclusion:

Steps to call TinyLLaMA from C++ (using llama.cpp)

✅ TinyLLaMA × C++ Setup Steps (using `llama.cpp`)

🔧 Step 1: Prepare llama.cpp

1-1. Required Environment (Minimum)

1-2. Get and Build llama.cpp

📦 Step 2: Download and Convert the TinyLLaMA Model

2-1. Get the Original Model from HuggingFace

2-2. Convert to GGUF format (for `llama.cpp`)

2-3. Model Quantization (Size Reduction)

🧪 Step 3: Call the Model from C++ (Code Example)

3-1. Simple C++ Code (Inference)

🧱 Step 4: Compilation Method (Example)

✅ Example Structure of Output (Organized)

🧠 Additional Notes for Applying to Use Cases

📌 Summary

How to generate illustration images using AI (Stable Diffusion)

What is Stable Diffusion?

What you need

Generation steps

Generation results (some)

References

AI on kenji.blog

The Wall of AI Development

The Wall of AI Development

Getting to the 8th station looks “easy”

Why you get stuck

But there’s a paradox

🌀For example

In the end, what we can do now

Conclusion:

Steps to call TinyLLaMA from C++ (using llama.cpp)

✅ TinyLLaMA × C++ Setup Steps (using llama.cpp)

🔧 Step 1: Prepare llama.cpp

1-1. Required Environment (Minimum)

1-2. Get and Build llama.cpp

📦 Step 2: Download and Convert the TinyLLaMA Model

2-1. Get the Original Model from HuggingFace

2-2. Convert to GGUF format (for llama.cpp)

2-3. Model Quantization (Size Reduction)

🧪 Step 3: Call the Model from C++ (Code Example)

3-1. Simple C++ Code (Inference)

🧱 Step 4: Compilation Method (Example)

✅ Example Structure of Output (Organized)

🧠 Additional Notes for Applying to Use Cases

📌 Summary

How to generate illustration images using AI (Stable Diffusion)

What is Stable Diffusion?

What you need

Generation steps

Generation results (some)

References

✅ TinyLLaMA × C++ Setup Steps (using `llama.cpp`)

2-2. Convert to GGUF format (for `llama.cpp`)