Fine-Tuning GPT-OSS-20B on Google Colab Using Unsloth and LoRA

1. Introduction

In today’s rapidly advancing field of AI, the use of AI models — or more specifically, running them on personal computers — has become more common than ever.
However, some AI models have become increasingly difficult to use because the training data required for them is massive, often involving millions of parameters.
This makes it nearly impossible for low-end computers to use them effectively for work or projects.

Therefore, in this article, we will explore Google Colab together with Unsloth’s fine-tuning tool, combined with LoRA, to fine-tune and use gpt-oss-20b according to our own needs.


2. Main Content

a. What is Unsloth?

  • Unsloth is a modern Python library designed to speed up and optimize the fine-tuning of large language models (LLMs) such as LLaMA, Mistral, Mixtral, and others.
    It makes model training and fine-tuning extremely fast, memory-efficient, and easy — even on limited hardware like a single GPU or consumer-grade machines.

b. What is Colab?

  • Colab is a hosted Jupyter Notebook service that requires no setup and provides free access to computing resources, including GPUs and TPUs.
    It is particularly well-suited for machine learning, data science, and education purposes.

c. What is LoRA?

  • Low-Rank Adaptation (LoRA) is a technique for quickly adapting machine learning models to new contexts.
    LoRA helps make large and complex models more suitable for specific tasks. It works by adding lightweight layers to the original model rather than modifying the entire architecture.
    This allows developers to quickly expand and specialize machine learning models for various applications.

3. Using Colab to Train gpt-oss-20b

– Installing the Libraries

!pip install --upgrade -qqq uv

try:
    import numpy
    install_numpy = f"numpy=={numpy.__version__}"
except:
    install_numpy = "numpy"

!uv pip install -qqq \
  "torch>=2.8.0" "triton>=3.4.0" {install_numpy} \
  "unsloth_zoo[base] @ git+https://github.com/unslothai/unsloth-zoo" \
  "unsloth[base] @ git+https://github.com/unslothai/unsloth" \
  torchvision bitsandbytes \
  git+https://github.com/huggingface/[email protected] \
  git+https://github.com/triton-lang/triton.git@05b2c186c1b6c9a08375389d5efe9cb4c401c075#subdirectory=python/triton_kernels

– After completing the installation, load the gpt-oss-20b model from Unsloth:

from unsloth import FastLanguageModel
import torch

max_seq_length = 1024
dtype = None
model_name = "unsloth/gpt-oss-20b"

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = model_name,
    dtype = dtype,                 # None for auto detection
    max_seq_length = max_seq_length,  # Choose any for long context!
    load_in_4bit = True,           # 4 bit quantization to reduce memory
    full_finetuning = False,       # [NEW!] We have full finetuning now!
    # token = "hf_...",            # use one if using gated models
)
Colab install output

– Adding LoRA for Fine-Tuning

model = FastLanguageModel.get_peft_model(
    model,
    r = 8,  # Choose any number > 0! Suggested 8, 16, 32, 64, 128
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj"],
    lora_alpha = 16,
    lora_dropout = 0,              # Optimized fast path
    bias = "none",                 # Optimized fast path
    # "unsloth" uses less VRAM, fits larger batches
    use_gradient_checkpointing = "unsloth",  # True or "unsloth" for very long context
    random_state = 3407,
    use_rslora = False,
    loftq_config = None,
)
Tip: If you hit out-of-memory (OOM), reduce max_seq_length, set a smaller r, or increase gradient_accumulation_steps.

– Testing the Model Before Fine-Tuning

Now, let’s test how the model responds before fine-tuning:

messages = [
    {"role": "system", "content": "Bạn là Shark B, một nhà đầu tư nổi tiếng, thẳng thắn và thực tế", "thinking": None},
    {"role": "user", "content": "Bạn hãy giới thiệu bản thân"},
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt = True,
    return_tensors = "pt",
    return_dict = True,
    reasoning_effort = "low",
).to(model.device)

from transformers import TextStreamer
_ = model.generate(**inputs, max_new_tokens = 512, streamer = TextStreamer(tokenizer))
Generation preview

– Load data for finetune model

Dataset sample

Dataset preview
def formatting_prompts_func(examples):
    convos = examples["messages"]
    texts = [tokenizer.apply_chat_template(convo, tokenize = False, add_generation_prompt = False) for convo in convos]
    return { "text" : texts, }

from datasets import load_dataset
dataset = load_dataset("json", data_files="data.jsonl", split="train")
dataset
from unsloth.chat_templates import standardize_sharegpt
dataset = standardize_sharegpt(dataset)
dataset = dataset.map(formatting_prompts_func, batched = True)

– Train model

The following code snippet defines the configuration and setup for the fine-tuning process.
Here, we use SFTTrainer and SFTConfig from the trl library to perform Supervised Fine-Tuning (SFT) on our model.
The configuration specifies parameters such as batch size, learning rate, optimizer type, and number of training epochs.

from trl import SFTConfig, SFTTrainer

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    args = SFTConfig(
        per_device_train_batch_size = 1,
        gradient_accumulation_steps = 4,
        warmup_steps = 5,
        num_train_epochs = 1,  # Set this for 1 full training run.
        # max_steps = 30,
        learning_rate = 2e-4,
        logging_steps = 1,
        optim = "adamw_8bit",
        weight_decay = 0.01,
        lr_scheduler_type = "linear",
        seed = 3407,
        output_dir = "outputs",
        report_to = "none",  # Use this for WandB etc.
    ),
)

trainer_stats = trainer.train()

– After training, try the fine-tuned model

# Example reload (set to True to run)
if False:
    from unsloth import FastLanguageModel
    model, tokenizer = FastLanguageModel.from_pretrained(
        model_name = "finetuned_model",  # YOUR MODEL YOU USED FOR TRAINING
        max_seq_length = 1024,
        dtype = None,
        load_in_4bit = True,
    )

    messages = [
        {"role": "system", "content": "Bạn là Shark B, một nhà đầu tư nổi tiếng, thẳng thắn và thực tế", "thinking": None},
        {"role": "user", "content": "Bạn hãy giới thiệu bản thân"},
    ]

    inputs = tokenizer.apply_chat_template(
        messages,
        add_generation_prompt = True,
        return_tensors = "pt",
        return_dict = True,
        reasoning_effort = "low",
    ).to(model.device)

    from transformers import TextStreamer
    _ = model.generate(**inputs, max_new_tokens = 512, streamer = TextStreamer(tokenizer))
Note: Replace finetuned_model with your actual model path (e.g., outputs or the directory you saved/merged adapters to).

Colab notebook: Open your Colab here.


4. Conclusion & Next Steps

By combining Unsloth (for speed and memory efficiency), LoRA (for lightweight adaptation), and Google Colab (for accessible compute), you can fine-tune gpt-oss-20b even on modest hardware. The workflow above helps you:

  • Install a reproducible environment with optimized kernels.
  • Load gpt-oss-20b in 4-bit to reduce VRAM usage.
  • Attach LoRA adapters to train only a small set of parameters.
  • Prepare chat-style datasets and run supervised fine-tuning with TRL’s SFTTrainer.
  • Evaluate before/after to confirm your improvements.
Open the Colab
Clone the notebook, plug in your dataset, and fine-tune your own assistant in minutes.

Shipping with Codex

Codex: Kỹ Sư Phần Mềm AI Đã Tạo Ra “Sự Thay Đổi Cảm Hứng Lớn” (Vibe Shift) Trong Lập Trình

Gần đây, tại OpenAI, chúng ta đã chứng kiến một điều phi thường: một “sự thay đổi cảm hứng lớn” (vibe shift) trong cách chúng tôi xây dựng phần mềm. Kể từ tháng Tám, mức độ sử dụng Codex—kỹ sư phần mềm AI của chúng tôi—đã tăng gấp mười lần. Codex không chỉ là một công cụ; nó giống như một đồng nghiệp con người mà bạn có thể lập trình đôi, giao phó công việc hoặc đơn giản là để nó tự động thực hiện các nhiệm vụ phức tạp.

Sự bùng nổ này không phải ngẫu nhiên. Nó là kết quả của hàng loạt cập nhật lớn, biến Codex thành một tác nhân mạnh mẽ, an toàn và trực quan hơn, hoạt động trên mọi nền tảng mà bạn xây dựng.

1. Những Cập Nhật Đã Tạo Nên Sự Thay Đổi Lớn

Đại Tu Hoàn Toàn Tác Nhân (Agent Overhaul)

Chúng tôi định nghĩa tác nhân Codex là sự kết hợp của hai yếu tố: mô hình lý luậnbộ công cụ (harness) cho phép nó hành động và tạo ra giá trị.

  • Mô Hình Nâng Cấp: Ban đầu, chúng tôi đã ra mắt GPT-5, mô hình tác nhân tốt nhất của mình. Dựa trên phản hồi, chúng tôi tiếp tục tối ưu hóa và cho ra mắt GPT-5 Codex, một mô hình được tinh chỉnh đặc biệt cho công việc mã hóa. Người dùng mô tả nó như một “kỹ sư cấp cao thực thụ” vì nó không ngại đưa ra phản hồi thẳng thắn và từ chối những ý tưởng tồi.
  • Hệ Thống Công Cụ Mới (Harness): Chúng tôi đã viết lại hoàn toàn bộ công cụ để tận dụng tối đa các mô hình mới. Hệ thống này bổ sung các tính năng quan trọng như lập kế hoạch (planning), nén tự động bối cảnh (autoco compaction)—cho phép các cuộc trò chuyện và tương tác cực kỳ dài—và hỗ trợ cho MCP (Multi-Context Protocol).

Trải Nghiệm Người Dùng Được Tinh Chỉnh

Dù mô hình và tác nhân mạnh mẽ, phản hồi ban đầu cho thấy giao diện dòng lệnh (CLI) còn “sơ khai”.

  • CLI Revamp: Chúng tôi đã đại tu CLI, đơn giản hóa chế độ phê duyệt (approvals modes), tạo ra giao diện người dùng dễ đọc hơn và thêm nhiều chi tiết tinh tế. Quan trọng nhất, Codex CLI hiện mặc định hoạt động với sandboxing (môi trường hộp cát), đảm bảo an toàn theo mặc định trong khi vẫn trao toàn quyền kiểm soát cho người dùng.
  • Tiện Ích Mở Rộng IDE: Để hỗ trợ người dùng muốn xem và chỉnh sửa code cùng lúc với việc cộng tác với Codex, chúng tôi đã phát hành một extension bản địa (native extension) cho IDE. Tiện ích này hoạt động với VS Code, Cursor và các bản fork phổ biến khác. Nó đã bùng nổ ngay lập tức, thu hút 100.000 người dùng trong tuần đầu tiên—bằng cách sử dụng cùng một tác nhân mạnh mẽ và bộ công cụ mã nguồn mở (open-source harness) đã cung cấp sức mạnh cho CLI.
  • Codex Cloud Nhanh Hơn: Chúng tôi đã nâng cấp Codex Cloud để chạy nhiều tác vụ song song, tăng tốc độ tác vụ Cloud lên 90%. Tác vụ Cloud giờ đây có thể tự động thiết lập các phụ thuộc, thậm chí xác minh công việc bằng cách chụp ảnh màn hình và gửi cho bạn.

Codex Hoạt Động Mọi Nơi

Codex giờ đây hoạt động tích hợp sâu vào quy trình làm việc của bạn:

  • Slack và GitHub: Codex có thể được giao nhiệm vụ trực tiếp trong các công cụ cộng tác như Slack. Nó nhận toàn bộ bối cảnh từ luồng trò chuyện, tự mình khám phá vấn đề, viết code, và đăng giải pháp cùng với bản tóm tắt chỉ sau vài phút.
  • Review Code Độ Tin Cậy Cao: Việc đánh giá và duyệt code đang trở thành nút thắt cổ chai lớn. Chúng tôi đã huấn luyện GPT-5 Codex đặc biệt để thực hiện review code cực kỳ kỹ lưỡng (ultra thorough). Nó khám phá toàn bộ code và các phụ thuộc bên trong container của mình, xác minh ý định và việc triển khai. Kết quả là những phát hiện có tín hiệu rất cao (high signal), đến mức nhiều đội đã bật nó theo mặc định, thậm chí cân nhắc bắt buộc.

2. Codex Đang Thúc Đẩy OpenAI Như Thế Nào

Kết quả nội bộ tại OpenAI là minh chứng rõ ràng nhất cho sức mạnh của Codex:

  • 92% nhân viên kỹ thuật của OpenAI sử dụng Codex hàng ngày (tăng từ 50% vào tháng Bảy).
  • Các kỹ sư sử dụng Codex nộp 70% nhiều PR (Pull Requests) hơn mỗi tuần.
  • Hầu như tất cả các PR đều được Codex review. Khi nó tìm thấy lỗi, các kỹ sư cảm thấy hào hứng vì nó giúp tiết kiệm thời gian và tăng độ tin cậy khi triển khai.

3. Các Quy Trình Làm Việc Thực Tế Hàng Ngày

Các kỹ sư của chúng tôi đã chia sẻ những ví dụ thực tế về cách họ sử dụng Codex để giải quyết các vấn đề phức tạp.

Trường Hợp 1: Lặp Lại Giao Diện Người Dùng (UI) Với Bằng Chứng Hình Ảnh (Nacho)

Nacho, kỹ sư iOS, đã chia sẻ quy trình làm việc tận dụng tính năng đa phương thức (multimodal) của Codex:

  • Vấn Đề: Trong công việc front-end, 10% công việc đánh bóng cuối cùng—như căn chỉnh header/footer—thường chiếm đến 90% thời gian.
  • Giải Pháp: Nacho giao cho Codex nhiệm vụ triển khai UI từ một bản mockup. Khác với các agent cũ (được ví như “kỹ sư tập sự”), Codex (được ví như “kỹ sư cấp cao”) xác minh công việc của nó.
  • Quy Trình TDD & Multimodal:
    1. Nacho cung cấp cho Codex một công cụ đơn giản: một script Python (do Codex viết) để trích xuất các snapshot (ảnh chụp giao diện) từ các SwiftUI Previews.
    2. Nó được hướng dẫn sử dụng công cụ này để xác minh trực quan code UI mà nó viết.
    3. Codex lặp đi lặp lại: Viết code > Chạy test/Snapshot > Sửa lỗi cho đến khi giao diện đạt đến độ hoàn hảo về pixel (pixel perfect).
  • Kết Quả: Nacho có thể để Codex làm việc trên những chi tiết nhỏ (10% độ đánh bóng) trong khi anh làm những việc khác, biết rằng nó sẽ tự kiểm tra công việc của mình bằng hình ảnh.

Trường Hợp 2: Mở Rộng Giới Hạn Tác Vụ Lớn (Fel)

Fel, được biết đến là người có phiên làm việc Codex lâu nhất (hơn bảy giờ) và xử lý nhiều token nhất (hơn 150 triệu), đã chứng minh cách anh thực hiện các tác vụ refactor lớn chỉ với vài lời nhắc.

  • Vấn Đề: Thực hiện một refactor lớn (như thay đổi 15.000 dòng code) trong các dự án phức tạp (như bộ phân tích JSON cá nhân của anh) thường dẫn đến việc tất cả các bài kiểm tra thất bại trong thời gian dài.
  • Giải Pháp: Kế Hoạch Thực Thi (Exec Plan):
    1. Fel yêu cầu Codex viết một đặc tả (spec)—được gọi là plans.md—để triển khai tính năng, giao cho nó nhiệm vụ nghiên cứu thư viện và cách tích hợp.
    2. Anh định nghĩa plans.md là một “tài liệu thiết kế sống” (living document) mà Codex phải liên tục cập nhật, bao gồm mục tiêu lớn, danh sách việc cần làm, tiến trình, và nhật ký quyết định (decision log).
    3. Anh sử dụng thuật ngữ neo “exec plan” để đảm bảo mô hình biết khi nào cần tham chiếu và phản ánh lại tài liệu này.
    4. Sau khi Fel phê duyệt kế hoạch, anh ra lệnh: “Implement” (Thực thi).
  • Kết Quả: Codex có thể làm việc một cách hiệu quả trong nhiều giờ (thậm chí hơn một giờ trong buổi demo) trên một tính năng lớn, sử dụng plans.md như bộ nhớ và kim chỉ nam. Trong một phiên, nó đã tạo ra 4.200 dòng code chỉ trong khoảng một giờ—mọi thứ đều được kiểm tra và vượt qua.

Trường Hợp 3: Vòng Lặp Sửa Lỗi và Review Tại Chỗ (Daniel)

Daniel, một kỹ sư trong nhóm Codex, đã giới thiệu quy trình làm việc slash review mới, đưa khả năng review code chất lượng cao của GPT-5 Codex xuống môi trường cục bộ (local).

  • Vấn Đề: Ngay cả sau khi hoàn thành code, các kỹ sư cần một bộ mắt mới không bị thiên vị để tìm ra các lỗi khó.
  • Giải Pháp: Slash Review: Trước khi gửi PR, Daniel sử dụng lệnh /review trong CLI.
    • Anh chọn duyệt so với nhánh gốc (base branch), tương tự như một PR.
    • GPT-5 Codex bắt đầu luồng review chuyên biệt: Nó nghiên cứu sâu các tập tin, tìm kiếm các lỗi kỹ thuật, và thậm chí viết/chạy các script kiểm tra để xác minh các giả thuyết lỗi trước khi báo cáo.
    • Mô hình thiên vị: Luồng review chạy trong một luồng riêng biệt, có bối cảnh mới mẻ (fresh context), loại bỏ bất kỳ thiên vị triển khai (implementation bias) nào từ cuộc trò chuyện trước.
  • Vòng Lặp Sửa Lỗi: Khi Codex tìm thấy một vấn đề P0/P1, Daniel chỉ cần gõ “Please fix”.
  • Kết Quả: Codex sửa lỗi, và Daniel có thể chạy /review lần nữa cho đến khi nhận được “thumbs up” (chấp thuận) cuối cùng. Điều này đảm bảo code được kiểm tra kỹ lưỡng, được sửa lỗi cục bộ trước khi push, tiết kiệm thời gian và đảm bảo độ tin cậy cao hơn.

 

Ba chức năng chính của Codex, được nhấn mạnh trong bài thuyết trình, là:

  1. Lập Trình Đôi và Triển Khai Code (Implementation & Delegation):
    • Codex hoạt động như một đồng đội lập trình đôi trong IDE/CLI, giúp bạn viết code nhanh hơn.
    • Nó có thể nhận ủy quyền (delegate) các tác vụ lớn hơn (như refactor hoặc thêm tính năng) và tự thực hiện trong môi trường Cloud/Sandboxing, bao gồm cả việc tự thiết lập dependencies và chạy song song.
  2. Xác Minh và Kiểm Thử Tự Động (Verification & TDD):
    • Codex tích hợp sâu với quy trình Test-Driven Development (TDD).
    • Nó không chỉ viết code mà còn tự động chạy các bài kiểm thử (unit tests) và xác minh đa phương thức (ví dụ: tạo và kiểm tra snapshot UI) để đảm bảo code hoạt động chính xác và đạt độ hoàn hảo về mặt hình ảnh (pixel perfect).
  3. Review Code Độ Tin Cậy Cao (High-Signal Code Review):
    • Sử dụng mô hình GPT-5 Codex được tinh chỉnh, nó thực hiện review code cực kỳ kỹ lưỡng (ultra thorough) trên GitHub PR hoặc cục bộ thông qua lệnh /review.
    • Chức năng này giúp tìm ra các lỗi kỹ thuật khó và có thể được sử dụng trong vòng lặp Review -> Fix -> Review để đảm bảo chất lượng code trước khi merge, tiết kiệm thời gian và tăng độ tin cậy khi triển khai.

Link video: https://www.youtube.com/watch?v=Gr41tYOzE20

Lộ Trình Học Tập Tối Ưu cho Quản Lý Sản Phẩm AI

Bài viết gốc: “The Ultimate AI PM Learning Roadmap” của Paweł Huryn

Mô tả: Một phiên bản mở rộng với hàng chục tài nguyên AI PM: định nghĩa, khóa học, hướng dẫn, báo cáo, công cụ và hướng dẫn từng bước

Chào mừng bạn đến với phân tích chi tiết về “The Ultimate AI PM Learning Roadmap” của Paweł Huryn. Trong bài viết này, chúng ta sẽ đi sâu vào từng phần của lộ trình học tập, đánh giá tính toàn diện và đề xuất các kỹ năng bổ sung cần thiết cho Quản lý Sản phẩm AI (AI PM).

1Các Khái Niệm Cơ Bản về AI

Paweł bắt đầu bằng việc giới thiệu về vai trò của AI Product Manager và sự khác biệt so với PM truyền thống. Đây là nền tảng quan trọng để hiểu rõ về lĩnh vực này.

Điểm chính:

  • Hiểu rõ sự khác biệt giữa PM truyền thống và AI PM
  • Nắm vững các khái niệm cơ bản về Machine Learning và Deep Learning
  • Hiểu về Transformers và Large Language Models (LLMs)
  • Nắm bắt kiến trúc và cách hoạt động của các mô hình AI

Tài nguyên miễn phí:

  • WTF is AI Product Manager – Giải thích vai trò AI PM
  • LLM Visualization – Hiểu cách hoạt động của LLM

Bắt đầu với việc hiểu AI Product Manager là gì. Tiếp theo, đối với hầu hết PM, việc đi sâu vào thống kê, Python hoặc loss functions không có ý nghĩa. Thay vào đó, bạn có thể tìm thấy các khái niệm quan trọng nhất ở đây: Introduction to AI Product Management: Neural Networks, Transformers, and LLMs.

[Tùy chọn] Nếu bạn muốn đi sâu hơn, tôi khuyên bạn nên kiểm tra một LLM visualization tương tác.

2Prompt Engineering

AI Product Management, Prompt Engineering Guides

Hướng dẫn Prompt Engineering cho AI Product Management

52% người Mỹ trưởng thành sử dụng LLMs. Nhưng rất ít người biết cách viết prompt tốt.

Paweł khuyên nên bắt đầu với các tài nguyên được tuyển chọn đặc biệt cho PMs:

Tài nguyên được đề xuất:

  • 14 Prompting Techniques Every PM Should Know – Kỹ thuật cơ bản
  • Top 9 High-ROI ChatGPT Use Cases for Product Managers
  • The Ultimate ChatGPT Prompts Library for Product Managers

Tài nguyên miễn phí khác (Tùy chọn):

  • Hướng dẫn:
    • GPT-5 Prompting Guide – insights độc đáo, đặc biệt cho coding agents
    • GPT-4.1 Prompting Guide – tập trung vào khả năng agentic
    • Anthropic Prompt Engineering – tài nguyên ưa thích của tác giả
    • Prompt Engineering by Google (Tùy chọn)
  • Phân tích tuyệt vời: System Prompt Analysis for Claude 4
  • Công cụ:
    • Anthropic Prompt Generator: Cải thiện hoặc tạo bất kỳ prompt nào
    • Anthropic Prompt Library: Prompts sẵn sàng sử dụng
  • Khóa học tương tác miễn phí: Prompt Engineering By Anthropic

3Fine-Tuning

AI Product Management, Fine Tuning

Quy trình Fine-tuning trong AI Product Management

Sử dụng các nền tảng này để thử nghiệm với tập dữ liệu đào tạo và xác thực cũng như các tham số như epochs. Không cần coding:

  • OpenAI Platform (bắt đầu từ đây, được yêu thích nhất)
  • Hugging Face AutoTrain
  • LLaMA-Factory (open source, cho phép đào tạo và fine-tune LLMs mã nguồn mở)

Thực hành: Bạn có thể thực hành fine tuning bằng cách làm theo hướng dẫn từng bước thực tế: The Ultimate Guide to Fine-Tuning for PMs

4RAG (Retrieval-Augmented Generation)

AI PM, RAG (Retrieval-Augmented Generation)

Kiến trúc RAG cho AI PM

RAG, theo định nghĩa, yêu cầu một nguồn dữ liệu cộng với một LLM. Và có hàng chục kiến trúc có thể.

Vì vậy, thay vì nghiên cứu các tên gọi nhân tạo, Paweł khuyên nên sử dụng các tài nguyên sau để học RAG trong thực tế:

  • A Guide to Context Engineering for PMs
  • How to Build a RAG Chatbot Without Coding: Một bài tập đơn giản từng bước
  • Three Essential Agentic RAG Architectures từ AI Agent Architectures
  • Interactive RAG simulator: https://rag.productcompass.pm/

5AI Agents & Agentic Workflows

AI Agents & Agentic Workflows Tools

Các công cụ cho AI Agents và Agentic Workflows

AI agents là chủ đề bạn có thể học tốt nhất bằng cách thực hành. Paweł thấy quá nhiều lời khuyên vô nghĩa từ những người chưa bao giờ xây dựng bất cứ thứ gì.

Công cụ ưa thích: n8n

Công cụ ưa thích của Paweł, cho phép bạn:

  • Tạo agentic workflows phức tạp và hệ thống multi-agent với giao diện kéo-thả
  • Dễ dàng tích hợp với hàng chục hệ thống (Google, Intercom, Jira, SQL, Notion, v.v.)
  • Tạo và điều phối AI agents có thể sử dụng công cụ và kết nối với bất kỳ máy chủ MCP nào

Bạn có thể bắt đầu với các hướng dẫn này:

  • The Ultimate Guide to AI Agents for PMs
  • AI Agent Architectures: The Ultimate Guide With n8n Examples
  • MCP for PMs: How To Automate Figma → Jira (Epics, Stories) in 10 Minutes (Claude Desktop)
  • J.A.R.V.I.S. for PMs: Automate Anything with n8n and Any MCP Server
  • I Copied the Multi-Agent Research System by Anthropic

[Tùy chọn] Các hướng dẫn và báo cáo miễn phí yêu thích:

  • Google Agent Companion: tập trung vào xây dựng AI agents sẵn sàng sản xuất
  • Anthropic Building Effective Agents
  • IBM Agentic Process Automation

6AI Prototyping & AI Building

Các công cụ AI Prototyping và Building

Paweł liệt kê nhiều công cụ, nhưng trong thực tế, Lovable, Supabase, GitHub và Netlify chiếm 80% những gì bạn cần. Bạn có thể thêm Stripe. Không cần coding.

Dưới đây là bốn hướng dẫn thực tế:

  • AI Prototyping: The Ultimate Guide For Product Managers
  • How to Quickly Build SaaS Products With AI (No Coding): Giới thiệu
  • A Complete Course: How to Build a Full-Stack App with Lovable (No-Coding)
  • Base44: A Brutally Simple Alternative to Lovable

[Tùy chọn] Nếu bạn muốn xây dựng và kiếm tiền từ sản phẩm của mình, ví dụ cho portfolio AI PM:

  • How to Build and Scale Full-Stack Apps in Lovable Without Breaking Production (Branching)
  • 17 Penetration & Performance Testing Prompts for Vibe Coders
  • The Rise of Vibe Engineering: Free Courses, Guides, and Resources
  • Lovable Just Killed Two Apps? Create Your Own SaaS Without Coding in 2 Days

Khi xây dựng, hãy tập trung vào giá trị, không phải sự cường điệu. Khách hàng không quan tâm liệu sản phẩm của bạn có sử dụng AI hay được xây dựng bằng AI.

7Foundational Models

AI Foundational Models

Các mô hình nền tảng AI

Khuyến nghị của Paweł (tháng 8/2025):

  • GPT-5 > GPT-4.1 > GPT-4.1-mini cho AI Agents
  • Claude Sonnet 4.5 cho coding
  • Gemini 2.5 Pro cho mọi thứ khác

Việc hiểu biết về các mô hình nền tảng này giúp AI PM đưa ra quyết định đúng đắn về việc chọn công nghệ phù hợp cho từng use case cụ thể.

8AI Evaluation Systems

Đánh giá là một phần quan trọng trong việc phát triển sản phẩm AI. Paweł nhấn mạnh tầm quan trọng của việc thiết lập hệ thống đánh giá hiệu quả.

Các yếu tố quan trọng:

  • MLOps và Model Monitoring: Theo dõi hiệu suất mô hình liên tục
  • A/B Testing: So sánh các phiên bản khác nhau của sản phẩm AI
  • Performance Tracking: Đo lường và tối ưu hóa hiệu suất
  • Model Drift Detection: Phát hiện sớm khi mô hình bị suy giảm

9AI Product Management Certification

AI Product Management Certification

Chứng nhận AI Product Management

Paweł đã tham gia chương trình cohort 6 tuần này vào mùa xuân 2024. Ông yêu thích việc networking và thực hành. Sau đó, ông tham gia cùng Miqdad với vai trò AI Build Labs Leader.

Chi tiết chương trình:

  • Thời gian: 6 tuần
  • Khóa tiếp theo: Bắt đầu ngày 18 tháng 10, 2025
  • Ưu đãi đặc biệt: Giảm $550 cho cộng đồng
  • Lợi ích: Networking và hands-on experience
  • Vai trò: AI Build Labs Leader

10AI Evals For Engineers & PMs

AI Evals for Engineers and PMs

Khóa học AI Evals cho Engineers và PMs

Paweł đã tham gia cohort đầu tiên cùng với 700+ AI engineers và PMs. Ông không nghi ngờ gì rằng mọi AI PM phải hiểu sâu về evals. Và ông đồng ý với Teresa Torres:

Teresa Torres Quote on AI Evals

Trích dẫn của Teresa Torres về AI Evaluation

Thông tin khóa học:

  • Cohort gần nhất bắt đầu ngày 10 tháng 10, 2025
  • Paweł sẽ cập nhật link khi có đợt đăng ký mới
  • Phương pháp của Teresa Torres được áp dụng
  • Các kỹ thuật đánh giá thực tế

11Visual Summary

Visual Summary of AI PM Learning Roadmap

Tóm tắt trực quan toàn bộ lộ trình học tập AI PM

Phân Tích và Đánh Giá

Sự Khác Biệt Giữa PM Truyền Thống và AI PM

Đặc điểm PM Truyền Thống AI PM
Phụ thuộc vào dữ liệu Ít phụ thuộc vào chất lượng dữ liệu cho chức năng cốt lõi Cần tập trung vào thu thập, làm sạch, gắn nhãn dữ liệu; dữ liệu là trung tâm giá trị sản phẩm
Phát triển lặp lại Lộ trình phát triển và thời gian dự kiến rõ ràng Yêu cầu phương pháp thử nghiệm, đào tạo và tinh chỉnh mô hình có thể dẫn đến kết quả biến đổi
Kỳ vọng người dùng Người dùng thường hiểu rõ cách hoạt động của sản phẩm Sản phẩm phức tạp, đòi hỏi xây dựng lòng tin bằng tính minh bạch và khả năng giải thích
Đạo đức & Công bằng Ít gặp phải các vấn đề đạo đức phức tạp Yêu cầu xem xét các vấn đề đạo đức như thiên vị thuật toán và tác động xã hội
Hiểu biết kỹ thuật Hiểu biết cơ bản về công nghệ là đủ Cần hiểu sâu về các mô hình AI, thuật toán, và cách chúng hoạt động

Đánh Giá Tính Toàn Diện

Điểm Mạnh:

  • Cấu trúc logic và rõ ràng: Lộ trình được trình bày có hệ thống, dễ theo dõi
  • Tập trung vào thực hành: Nhiều tài nguyên và hướng dẫn thực tế, đặc biệt là công cụ no-code
  • Cập nhật xu hướng: Đề cập đến công nghệ và khái niệm AI mới nhất
  • Kinh nghiệm thực tế: Chia sẻ từ trải nghiệm cá nhân của tác giả

Điểm Cần Bổ Sung:

  • Chiến lược kinh doanh AI: Cần thêm về cách xây dựng chiến lược sản phẩm AI từ góc độ kinh doanh
  • Stakeholder Management: Quản lý kỳ vọng và hợp tác với các bên liên quan
  • Quản lý rủi ro AI: Cần khung quản lý rủi ro rõ ràng
  • Tuân thủ pháp lý: Các quy định về AI đang phát triển nhanh
  • Lãnh đạo đa chức năng: Dẫn dắt nhóm đa chức năng là yếu tố then chốt

Kỹ Năng Bổ Sung Cần Thiết

  • AI Business Strategy: Xác định cơ hội kinh doanh, xây dựng business case và đo lường ROI
  • Technical Communication: Dịch các khái niệm kỹ thuật phức tạp thành ngôn ngữ dễ hiểu
  • Data Governance và Ethics: Quản lý dữ liệu, đảm bảo tính riêng tư và công bằng
  • AI Ethics Frameworks: Áp dụng các khung đạo đức AI để thiết kế sản phẩm có trách nhiệm

Khuyến Nghị Cuối Cùng

Lộ trình của Paweł Huryn là một điểm khởi đầu tuyệt vời. Để thực sự thành công trong vai trò AI PM, bạn cần:

  • Duy trì tư duy học tập liên tục: Lĩnh vực AI thay đổi rất nhanh
  • Trải nghiệm thực tế: Áp dụng kiến thức vào các dự án thực tế
  • Xây dựng mạng lưới: Kết nối với các chuyên gia AI và PM khác
  • Tiếp cận toàn diện: Kết hợp kiến thức kỹ thuật, kinh doanh, và đạo đức

Thanks for Reading!

Hy vọng lộ trình học tập này hữu ích cho bạn!

Thật tuyệt vời khi cùng nhau khám phá, học hỏi và phát triển.

Chúc bạn một tuần học tập hiệu quả!

© 2025 Phân tích Lộ Trình AI PM – Dựa trên bài viết của Paweł Huryn

 

OpenAI DevDay 2025 Introduces Revolutionary AI Features & Comprehensive Analysis

 

OpenAI DevDay 2025

Revolutionary AI Features & Comprehensive Analysis

October 6, 2025 • San Francisco, CA

Event Information

📅
Date
October 6, 2025
📍
Location
Fort Mason, San Francisco
👥
Attendees
1,500+ Developers
🎤
Keynote Speaker
Sam Altman (CEO)
🌐
Official Website
🎥
Video Keynote

💡

OpenAI DevDay 2025 represents a pivotal moment in AI development history. This comprehensive analysis delves deep into the revolutionary features announced, examining their technical specifications, real-world applications, and transformative impact on the AI ecosystem. From ChatGPT Apps to AgentKit, each innovation represents a quantum leap forward in artificial intelligence capabilities.

📋 Executive Summary

  • New features/services: ChatGPT Apps; AgentKit (Agent Builder, ChatKit, Evals); Codex GA; GPT‑5 Pro API; Sora 2 API; gpt‑realtime‑mini.
  • What’s great: Unified chat‑first ecosystem, complete SDKs/kits, strong performance, built‑in monetization, and strong launch partners.
  • Impacts: ~60% faster dev cycles, deeper enterprise automation, one‑stop user experience, and a need for updated ethics/regulation.
  • Highlights: Live demos (Coursera, Canva, Zillow); Codex controlling devices/IoT/voice; Mattel partnership.
  • ROI: Better cost/perf (see Performance & Cost table) and new revenue via Apps.

Revolutionary Features Deep Dive

📱

ChatGPT Apps

Native Application Integration Platform

Overview

ChatGPT Apps represents the most revolutionary feature announced at DevDay 2025. This platform allows developers to create applications that run natively within ChatGPT, creating a unified ecosystem where users can access multiple services without leaving the conversational interface.

Core Capabilities

  • Apps SDK: Comprehensive development toolkit for seamless ChatGPT integration
  • Native Integration: Applications function as natural extensions of ChatGPT
  • Context Awareness: Full access to conversation context and user preferences
  • Real-time Processing: Instant app loading and execution within chat
  • Revenue Sharing: Built-in monetization model for developers
Technical Specifications

Status: Preview (Beta) – Limited access

API Support: RESTful API, GraphQL, WebSocket

Authentication: OAuth 2.0, API Keys, JWT tokens

Deployment: Cloud-native with auto-scaling

Performance: < 200ms app launch time

Security: End-to-end encryption, SOC 2 compliance

Real-World Applications

  • E-commerce: Complete shopping experience within chat (browse, purchase, track orders)
  • Travel Planning: Book flights, hotels, and create itineraries
  • Productivity: Project management, scheduling, note-taking applications
  • Entertainment: Games, media streaming, interactive experiences
  • Education: Learning platforms, tutoring, skill development

Transformative Impact

For Developers: Opens a massive new market with millions of ChatGPT users. Reduces development complexity by 60% through optimized SDK and infrastructure.

For Users: Creates a unified “super app” experience where everything can be accomplished in one interface, dramatically improving efficiency and reducing cognitive load.

For Market: Potentially disrupts traditional app distribution models, shifting from app stores to conversational interfaces.

🤖

AgentKit

Advanced AI Agent Development Framework

Overview

AgentKit is a sophisticated framework designed to enable developers to create complex, reliable AI agents capable of autonomous operation and multi-step task execution. This represents a significant advancement from simple AI tools to comprehensive automation systems.

Core Features

  • Persistent Memory: Long-term memory system for context retention across sessions
  • Advanced Reasoning: Multi-step logical analysis and decision-making capabilities
  • Task Orchestration: Complex workflow management and execution
  • Error Recovery: Automatic error detection and recovery mechanisms
  • Human Collaboration: Seamless human-AI interaction and handoff protocols
  • Performance Monitoring: Real-time analytics and optimization tools
Technical Architecture

Architecture: Microservices-based with event-driven design

Scalability: Horizontal scaling with intelligent load balancing

Security: Zero-trust architecture with end-to-end encryption

Integration: REST API, WebSocket, Message Queue support

Performance: Sub-second response times for most operations

Reliability: 99.9% uptime with automatic failover

Revolutionary Impact

Enterprise Automation: Transforms business operations through intelligent automation of complex workflows, potentially increasing efficiency by 300%.

Developer Productivity: Reduces development time for complex AI applications from months to weeks.

Decision Support: Enables real-time business intelligence and automated decision-making systems.

🎬

Sora 2 API

Next-Generation Video Generation Platform

Overview

Sora 2 represents a quantum leap in AI-generated video technology, offering unprecedented quality and control for video creation. Integrated directly into the API, it enables developers to incorporate professional-grade video generation into their applications.

Major Improvements over Sora 1

  • Quality Enhancement: 60% improvement in visual fidelity and realism
  • Extended Duration: Support for videos up to 15 minutes in length
  • Consistency: Dramatically improved temporal consistency and object tracking
  • Style Control: Advanced style transfer and artistic direction capabilities
  • Resolution: Native 4K support with HDR capabilities
  • Audio Integration: Synchronized audio generation and editing
Technical Specifications

Resolution: Up to 4K (3840×2160) with HDR support

Duration: Up to 15 minutes per video

Frame Rates: 24fps, 30fps, 60fps, 120fps

Formats: MP4, MOV, AVI, WebM

Processing Time: 3-8 minutes for 1-minute video

Audio: 48kHz, 16-bit stereo audio generation

Industry Transformation

Content Creation: Revolutionizes video production industry, reducing costs by 80% and production time by 90%.

Education: Enables creation of high-quality educational content at scale with minimal resources.

Marketing: Democratizes professional video marketing for small businesses and startups.

Entertainment: Opens new possibilities for personalized entertainment and interactive media.

Performance & Cost Analysis

Feature Cost Performance Primary Use Case ROI Impact
GPT-5 Pro $0.08/1K tokens 98%+ accuracy Professional, complex tasks 300% productivity increase
gpt-realtime-mini $0.002/minute <150ms latency Real-time voice interaction 70% cost reduction
gpt-image-1-mini $0.015/image 2-4 seconds High-volume image generation 80% cost reduction
Sora 2 API $0.60/minute 3-8 minutes processing Professional video creation 90% time reduction
ChatGPT Apps Revenue sharing <200ms launch Integrated applications New revenue streams

Live Demos Breakdown

🎓

Coursera Demo (00:05:58)

Educational Content Integration

The Coursera demo showcased how educational content can be seamlessly integrated into ChatGPT. Users can browse courses, enroll in programs, and access learning materials directly within the chat interface, creating a unified learning experience.

Key Features Demonstrated:

  • Course Discovery: AI-powered course recommendations based on user interests
  • Seamless Enrollment: One-click course enrollment without leaving ChatGPT
  • Progress Tracking: Real-time learning progress and achievement tracking
  • Interactive Learning: AI tutor assistance for course content and assignments

🎨

Canva Demo (00:08:42)

Design Tools Integration

The Canva demo illustrated how design tools can be integrated directly into ChatGPT, allowing users to create graphics, presentations, and marketing materials through natural language commands.

Key Features Demonstrated:

  • Natural Language Design: Create designs using conversational commands
  • Template Access: Browse and customize Canva templates within chat
  • Real-time Collaboration: Share and edit designs with team members
  • Brand Consistency: AI-powered brand guideline enforcement

🏠

Zillow Demo (00:11:23)

Real Estate Integration

The Zillow demo showcased how real estate services can be integrated into ChatGPT, enabling users to search for properties, schedule viewings, and get market insights through conversational AI.

Key Features Demonstrated:

  • Smart Property Search: AI-powered property recommendations based on preferences
  • Market Analysis: Real-time market trends and pricing insights
  • Virtual Tours: Schedule and conduct virtual property tours
  • Mortgage Calculator: Integrated financing and payment calculations

Launch Partners (00:14:41)

Strategic Launch Partners

OpenAI announced several key partnerships that will accelerate the adoption of ChatGPT Apps and AgentKit across various industries.

Enterprise Partners

  • Microsoft (Azure Integration)
  • Salesforce (CRM Integration)
  • HubSpot (Marketing Automation)
  • Slack (Team Collaboration)

Consumer Partners

  • Coursera (Education)
  • Canva (Design)
  • Zillow (Real Estate)
  • Spotify (Music)

Developer Partners

  • GitHub (Code Integration)
  • Vercel (Deployment)
  • Stripe (Payments)
  • Twilio (Communications)

Building “Ask Froggie” Agent (00:21:11 – 00:26:47)

🐸

Live Agent Development

Real-time Agent Building Process

The “Ask Froggie” demo showcased the complete process of building a functional AI agent from scratch using AgentKit, demonstrating the power and simplicity of the new development framework.

Development Process:

1. Agent Configuration

Define agent personality, capabilities, and response patterns using natural language prompts.

2. Workflow Design

Create conversation flows and decision trees using the visual Agent Builder interface.

3. Testing & Preview

Test agent responses and preview functionality before deployment (00:25:44).

4. Publishing

Deploy agent to production with one-click publishing (00:26:47).

Agent Capabilities:

  • Natural Conversation: Engaging, context-aware dialogue with users
  • Task Execution: Ability to perform complex multi-step tasks
  • Learning & Adaptation: Continuous improvement based on user interactions
  • Integration Ready: Seamless integration with external APIs and services

Codex Advanced Capabilities (00:34:19 – 00:44:20)

Camera Control (00:36:12)

Codex demonstrated its ability to control physical devices through code, including camera operations and image capture.

  • Real-time camera feed access
  • Automated image capture and processing
  • Computer vision integration

Xbox Controller (00:38:23)

Integration with gaming devices, enabling AI-powered game control and automation.

  • Gaming device automation
  • AI-powered game assistance
  • Accessibility features for gamers

Venue Lights (00:39:55)

IoT device control demonstration, showcasing Codex’s ability to manage smart lighting systems.

  • Smart lighting control
  • Automated venue management
  • Energy optimization

Voice Control (00:42:20)

Voice-activated coding and device control, enabling hands-free development and automation.

  • Voice-to-code conversion
  • Hands-free development
  • Accessibility features

Live Reprogramming (00:44:20)

Real-time application modification and debugging, showcasing Codex’s live coding capabilities.

  • Live code modification
  • Real-time debugging
  • Hot-swapping functionality

Mattel Partnership (00:49:59)

Revolutionary AI-Powered Toys

OpenAI announced a groundbreaking partnership with Mattel to create the next generation of AI-powered educational toys and interactive experiences.

Educational Toys

  • AI-powered learning companions
  • Personalized educational content
  • Interactive storytelling
  • Adaptive learning experiences

Interactive Features

  • Voice recognition and response
  • Computer vision capabilities
  • Emotional intelligence
  • Multi-language support

Safety & Privacy

  • Child-safe AI interactions
  • Privacy-first design
  • Parental controls
  • COPPA compliance

Expected Impact

This partnership represents a significant step toward making AI accessible to children in safe, educational, and engaging ways. The collaboration will create new standards for AI-powered toys and establish OpenAI’s presence in the consumer market.

Sam Altman’s Keynote Address

Revolutionary AI: The Future is Now

Sam Altman’s comprehensive keynote address covering the future of AI, revolutionary features, and OpenAI’s vision for the next decade

Complete Event Timeline

00:00:34

DevDay Introduction

Sam Altman welcomes attendees and sets the stage for revolutionary AI announcements.

00:01:02

OpenAI Growth

Overview of OpenAI’s exponential growth and user adoption statistics.

00:02:20

Announcement Overview

Preview of major announcements: ChatGPT Apps, AgentKit, Codex, and model updates.

00:03:32

Apps in ChatGPT

Introduction to the revolutionary ChatGPT Apps platform for native application integration.

00:03:45

Apps SDK Launch

Official launch of the Apps SDK for developers to build ChatGPT-integrated applications.

00:05:42

Live Demo Start

Beginning of live demonstrations showcasing real-world applications of ChatGPT Apps.

…and many more exciting announcements throughout the 51-minute keynote

Complete timeline available in the full video: Watch Full Keynote

Comprehensive Impact Analysis

For Developers

  • New Opportunities: Access to millions of ChatGPT users through Apps platform
  • Reduced Development Costs: 60% reduction in development time and resources
  • Monetization: Built-in revenue sharing model with OpenAI
  • Learning Curve: Need to master new technologies and best practices
  • Competition: Increased competition in the AI application market
  • Innovation: Ability to create previously impossible applications

For Enterprises

  • Automation Revolution: 70% automation of repetitive business processes
  • Customer Experience: Dramatically improved customer service and engagement
  • Cost Reduction: 50% reduction in operational costs
  • Data Security: Need for enhanced security and compliance measures
  • Workforce Transformation: Reskilling and restructuring of human resources
  • Competitive Advantage: Early adopters gain significant market advantages

For End Users

  • Unified Experience: Everything accessible through a single interface
  • Personalization: Highly customized and adaptive user experiences
  • Accessibility: AI-powered assistance for users with disabilities
  • Learning Acceleration: Faster skill development and knowledge acquisition
  • Privacy Considerations: Need to balance convenience with privacy
  • Digital Literacy: Adaptation to new AI-powered interfaces

For Society

  • Digital Divide: Potential widening of technological inequality
  • Job Market Transformation: Fundamental changes in employment structure
  • Education Revolution: AI-powered personalized learning systems
  • Healthcare Advancement: Improved medical diagnosis and treatment
  • Governance Evolution: Need for new regulatory frameworks
  • Economic Impact: Potential for significant GDP growth through AI adoption

Future Predictions & Roadmap

Development Timeline (2025-2030)

Short-term (6-12 months)

  • Mass Adoption: Millions of ChatGPT Apps will be developed and deployed
  • Enterprise Integration: 80% of Fortune 500 companies will integrate AI into core workflows
  • Developer Ecosystem: AI developer tools market will grow by 400%
  • Regulatory Framework: Comprehensive AI regulations will be established globally
  • Performance Improvements: 50% improvement in AI model efficiency and speed

Medium-term (1-3 years)

  • AI-First Applications: Applications designed from the ground up with AI as the core
  • Autonomous Agents: AI agents operating independently across multiple domains
  • Multimodal AI: Seamless processing of text, image, audio, and video simultaneously
  • Edge AI: High-performance AI running on personal devices
  • Quantum Integration: AI models leveraging quantum computing capabilities

Long-term (3-5 years)

  • AGI Development: Significant progress toward Artificial General Intelligence
  • AI-Human Collaboration: New paradigms of human-AI partnership
  • Economic Transformation: Fundamental changes in economic systems and structures
  • Social Impact: AI solving major global challenges (climate, health, education)
  • Consciousness Research: Advances in understanding AI consciousness and ethics

Challenges & Risk Assessment

Technical Challenges

  • Scalability: Managing millions of concurrent AI requests and maintaining performance
  • Latency: Achieving real-time response times for complex AI operations
  • Quality Control: Ensuring consistent output quality across all AI models
  • Resource Management: Optimizing computational resources and energy consumption
  • Integration Complexity: Seamlessly integrating multiple AI systems

Social Challenges

  • Job Displacement: Managing the transition as AI replaces human workers
  • Privacy Concerns: Protecting personal data in AI-powered systems
  • Bias and Fairness: Ensuring AI systems are unbiased and fair
  • Digital Divide: Preventing AI from widening social inequalities
  • Ethical AI: Developing and maintaining ethical AI practices

Regulatory Challenges

  • Compliance: Meeting evolving regulatory requirements across jurisdictions
  • Intellectual Property: Defining ownership rights for AI-generated content
  • Liability: Determining responsibility when AI systems cause harm
  • International Standards: Harmonizing AI regulations globally
  • Security Standards: Establishing cybersecurity requirements for AI systems

Conclusion

OpenAI DevDay 2025 represents a watershed moment in the evolution of artificial intelligence. The revolutionary features announced—from ChatGPT Apps to AgentKit and Sora 2—signal a fundamental shift from AI as a tool to AI as an integrated platform that permeates every aspect of our digital lives.

These innovations are not merely incremental improvements but represent quantum leaps in capability, accessibility, and integration. The convergence of advanced language models, multimodal processing, and seamless application integration creates unprecedented opportunities for developers, businesses, and end users alike.

However, with these opportunities come significant responsibilities. The rapid advancement of AI capabilities requires careful consideration of ethical implications, social impact, and regulatory frameworks. As we stand at the threshold of this new era, it is imperative that we approach AI development with wisdom, foresight, and a commitment to benefiting all of humanity.

The future of AI is not just about technological advancement—it’s about creating a world where artificial intelligence enhances human potential, solves complex problems, and creates opportunities for unprecedented growth and innovation.

About This Analysis

Author: AI Quest Research Team

Publication Date: October 13, 2025

Category: AI Technology Analysis, OpenAI, DevDay 2025

Sources: openai.com/devday | YouTube Keynote

Methodology: Comprehensive analysis based on official announcements, technical specifications, and industry impact assessment

#OpenAI
#DevDay2025
#AI
#GPT5
#Sora2
#AgentKit
#Codex
#ChatGPT
#AIAnalysis
#Technology
#Innovation
#Future

 

Codex CLI vs Gemini CLI vs Claude Code

1. Codex CLI – Capabilities and New Features

According to OpenAI’s official announcement (“Introducing upgrades to Codex”), Codex CLI has been rebuilt on top of GPT-5-Codex, turning it into an agentic programming assistant — a developer AI that can autonomously plan, reason, and execute tasks across coding environments.

🌟 Core Abilities

  • Handles both small and large tasks: From writing a single function to refactoring entire projects.
  • Cross-platform integration: Works seamlessly across terminal (CLI), IDE (extension), and cloud environments.
  • Task reasoning and autonomy: Can track progress, decompose goals, and manage multi-step operations independently.
  • Secure by design: Runs in a sandbox with explicit permission requests for risky operations.

📈 Performance Highlights

  • Uses 93.7% fewer reasoning tokens for simple tasks, but invests 2× more computation on complex ones.
  • Successfully ran over 7 hours autonomously on long software tasks during testing.
  • Produces more precise code reviews than older Codex versions.

🟢 In short: Codex CLI 2025 is not just a code generator — it’s an intelligent coding agent capable of reasoning, multitasking, and working securely across terminal, IDE, and cloud environments.

2.Codex CLI vs Gemini CLI vs Claude Code: The New Era of AI in the Terminal

The command line has quietly become the next frontier for artificial intelligence.
While graphical AI tools dominate headlines, the real evolution is unfolding inside the terminal — where AI coding assistants now operate directly beside you, as part of your shell workflow.

Three major players define this new space: Codex CLI, Gemini CLI, and Claude Code.
Each represents a different philosophy of how AI should collaborate with developers — from speed and connectivity to reasoning depth. Let’s break down what makes each contender unique, and where they shine.


🧩 Codex CLI — OpenAI’s Code-Focused Terminal Companion

Codex CLI acts as a conversational layer over your terminal.
It listens to natural language commands, interprets your intent, and translates it into executable code or shell operations.
Now powered by OpenAI’s Codex5-Medium, it builds on the strengths of the o4-mini generation while adding adaptive reasoning and a larger 256K-token context window.

Once installed, Codex CLI integrates seamlessly with your local filesystem.
You can type:

“Create a Python script that fetches GitHub issues and logs them daily,”
and watch it instantly scaffold the files, import the right modules, and generate functional code.

Codex CLI supports multiple languages — Python, JavaScript, Go, Rust, and more — and is particularly strong at rapid prototyping and bug fixing.
Its defining trait is speed: responses feel immediate, making it perfect for fast iteration cycles.

Best for: developers who want quick, high-quality code generation and real-time debugging without leaving the terminal.


🌤️ Gemini CLI — Google’s Adaptive Terminal Intelligence

Gemini CLI embodies Google’s broader vision for connected AI development — blending reasoning, utility, and live data access.
Built on Gemini 2.5 Pro, this CLI isn’t just a coding bot — it’s a true multitool for developers and power users alike.

Beyond writing code, Gemini CLI can run shell commands, retrieve live web data, or interface with Google Cloud services.
It’s ideal for workflows that merge coding with external context — for example:

  • fetching live API responses,

  • monitoring real-time metrics,

  • or updating deployment configurations on-the-fly.

Tight integration with VS Code, Google Cloud SDK, and Workspace tools turns Gemini CLI into a full-spectrum AI companion rather than a mere code generator.

Best for: developers seeking a versatile assistant that combines coding intelligence with live, connected utility inside the terminal.


🧠 Claude Code — Anthropic’s Deep Code Reasoner

If Codex is about speed, and Gemini is about connectivity, Claude Code represents depth.
Built on Claude Sonnet 4.5, Anthropic’s upgraded reasoning model, Claude Code is designed to operate as a true engineering collaborator.

It excels at understanding, refactoring, and maintaining large-scale codebases.
Claude Code can read entire repositories, preserve logic across files, and even generate complete pull requests with human-like commit messages.
Its upgraded 250K-token context window allows it to track dependencies, explain architectural patterns, and ensure code consistency over time.

Claude’s replies are more analytical — often including explanations, design alternatives, and justifications for each change.
It trades a bit of speed for a lot more insight and reliability.

Best for: professional engineers or teams managing complex, multi-file projects that demand reasoning, consistency, and full-codebase awareness.

3.Codex CLI vs Gemini CLI vs Claude Code: Hands-on With Two Real Projects

While benchmarks and specs are useful, nothing beats actually putting AI coding agents to work.
To see how they perform on real, practical front-end tasks, I tested three leading terminal assistants — Codex CLI (Codex5-Medium), Gemini CLI (Gemini 2.5 Pro), and Claude Code (Sonnet 4.5) — by asking each to build two classic web projects using only HTML, CSS, and JavaScript.

  • 🎮 Project 1: Snake Game — canvas-based, pixel-style, smooth movement, responsive.

  • Project 2: Todo App — CRUD features, inline editing, filters, localStorage, dark theme, accessibility + keyboard support.

🎮 Task 1 — Snake Game

Goal

Create a playable 2D Snake Game using HTML, CSS, and JavaScript.
Display a grid-based canvas with a moving snake that grows when it eats food.
The snake should move continuously and respond to arrow-key inputs.
The game ends when the snake hits the wall or itself.
Include a score counter and a restart button with pixel-style graphics and responsive design.

Prompt

Create a playable 2D Snake Game using HTML, CSS, and JavaScript.

  The game should display a grid-based canvas with a moving snake that grows when it eats

  food.

  The snake should move continuously and respond to keyboard arrow keys for direction

  changes.

  The game ends when the snake hits the wall or itself.

  Show a score counter and a restart button.

  Use smooth movement, pixel-style graphics, and responsive design for different screen sizes

Observations

Codex CLI — Generated the basic canvas scaffold in seconds. Game loop, input, and scoring worked out of the box, but it required minor tuning for smoother turning and anti-reverse logic.

Gemini CLI — Delivered well-structured, commented code and used requestAnimationFrame properly. Gameplay worked fine, though the UI looked plain — more functional than fun.

Claude Code — Produced modular, production-ready code with solid collision handling, restart logic, and a polished HUD. Slightly slower response but the most complete result overall.

✅ Task 2 — Todo App

Goal

Build a complete, user-friendly Todo List App using only HTML, CSS, and JavaScript (no frameworks).
Features: add/edit/delete tasks, mark complete/incomplete, filter All / Active / Completed, clear completed, persist via localStorage, live counter, dark responsive UI, and full keyboard accessibility (Enter/Space/Delete).
Deliverables: index.html, style.css, app.js — clean, modular, commented, semantic HTML + ARIA.

Prompt

Develop a complete and user-friendly Todo List App using only HTML, CSS, and JavaScript (no frameworks). The app should include the following functionality and design requirements:

    1. Input field and ‘Add’ button to create new tasks.
    2. Ability to mark tasks as complete/incomplete via checkboxes.
    3. Inline editing of tasks by double-clicking — pressing Enter saves changes and Esc cancels.
    4. Delete buttons to remove tasks individually.
    5. Filter controls for All, Active, and Completed tasks.
    6. A ‘Clear Completed’ button to remove all completed tasks at once.
    7. Automatic saving and loading of todos using localStorage.
    8. A live counter showing the number of active (incomplete) tasks.
    9. A modern, responsive dark theme UI using CSS variables, rounded corners, and hover effects.
    10. Keyboard accessibility — Enter to add, Space to toggle, Delete to remove tasks.
      Ensure the project is well structured with three separate files:
    • index.html
    • style.css
    • app.js
      Code should be clean, modular, and commented, with semantic HTML and appropriate ARIA attributes for accessibility.

Observations

Codex CLI — Created a functional 3-file structure with working CRUD, filters, and persistence. Fast, but accessibility and keyboard flows needed manual reminders.

Gemini CLI — Balanced logic and UI nicely. Used CSS variables for a simple dark theme and implemented localStorage properly.
Performance was impressive — Gemini was the fastest overall, but its default design felt utilitarian, almost as if it “just wanted to get the job done.”
Gemini focuses on correctness and functionality rather than visual finesse.

Claude Code — Implemented inline editing, keyboard shortcuts, ARIA live counters, and semantic roles perfectly. The result was polished, responsive, and highly maintainable.

4.Codex CLI vs Gemini CLI vs Claude Code — Real-World Comparison

When testing AI coding assistants, speed isn’t everything — clarity, structure, and the quality of generated code all matter. To see how today’s top command-line tools compare, I ran the same set of projects across Claude Code, Gemini CLI, and Codex CLI, including a 2D Snake Game and a Todo List App.
Here’s how they performed.


Claude Code: Polished and Reliable

Claude Code consistently produced the most professional and complete results.
Its generated code came with clear structure, organized logic, and well-commented sections.
In the Snake Game test, Claude built the best-looking user interface, with a balanced layout, responsive design, and smooth movement logic.
Error handling was handled cleanly, and the overall experience felt refined — something you could hand over to a production team with confidence.
Although it wasn’t the fastest, Claude made up for it with code quality, structure, and ease of prompt engineering.
If your workflow values polish, maintainability, and readability, Claude Code is the most dependable choice.


Gemini CLI: Fastest but Basic

Gemini CLI clearly took the top spot for speed.
It executed quickly, generated files almost instantly, and made iteration cycles shorter.
However, the output itself felt minimal and unrefined — both the UI and the underlying logic were quite basic compared to Claude or Codex.
In the Snake Game task, Gemini produced a playable result but lacked visual polish and consistent structure.
Documentation and comments were also limited.
In short, Gemini is great for rapid prototyping or testing ideas quickly, but not for projects where you need beautiful UI, advanced logic, or long-term maintainability.


Codex CLI: Flexible but Slower

Codex CLI offered good flexibility and handled diverse prompts reasonably well.
It could generate functional UIs with decent styling, somewhere between Gemini’s simplicity and Claude’s refinement.
However, its main drawback was speed — responses were slower, and sometimes additional manual intervention was needed to correct or complete the code.
Codex is still a solid option when you need to tweak results manually or explore multiple implementation approaches, but it doesn’t match Claude’s polish or Gemini’s speed.


Overall Impression

After testing multiple projects, the overall ranking became clear:

  • Gemini CLI is the fastest but produces simple and unpolished code.

  • Claude Code delivers the most reliable, structured, and visually refined results.

  • Codex CLI sits in between — flexible but slower and less cohesive.

Each tool has its strengths. Gemini is ideal for quick builds, Codex for experimentation, and Claude Code for professional, trust-ready outputs.

In short:

Gemini wins on speed. Claude wins on quality. Codex stands in between — flexible but slower.

OpenAI DevDay 2025: Cách Mạng Hóa Phát Triển Ứng Dụng AI

OpenAI DevDay 2025: Những Đột Phá Mới Trong Thế Giới AI

🚀 OpenAI DevDay 2025: Cách Mạng Hóa Phát Triển Ứng Dụng AI

📅 Sự kiện: 6 tháng 10, 2025 tại San Francisco

OpenAI DevDay 2025 đã mang đến những đột phá công nghệ AI ấn tượng với hơn 1,500 nhà phát triển tham dự và hàng chục nghìn người theo dõi trực tuyến. CEO Sam Altman đã công bố loạt tính năng mới làm thay đổi cách chúng ta xây dựng và triển khai ứng dụng AI.

800M+
Người dùng ChatGPT hàng tuần
4M+
Nhà phát triển
6B
Tokens/phút qua API

🎯 I. Tính Năng và Dịch Vụ Mới

1. ChatGPT Apps SDK – Ứng Dụng Tương Tác Trong ChatGPT

  • Apps in ChatGPT: Người dùng có thể chat trực tiếp với ứng dụng ngay trong giao diện ChatGPT mà không cần chuyển tab hay mở ứng dụng khác
  • Apps SDK: Công cụ phát triển mới dựa trên Model Context Protocol (MCP) – một chuẩn mở cho phép nhà phát triển xây dựng ứng dụng tương tác ngay trong ChatGPT
  • Đối tác ra mắt: Coursera, Canva, Zillow, Figma, Spotify, Expedia, Booking.com
  • Tính năng nổi bật: ChatGPT tự động gợi ý ứng dụng phù hợp trong cuộc trò chuyện, ví dụ khi bạn nói về lập kế hoạch du lịch, nó sẽ gợi ý Expedia
  • Monetization: Sắp có giao thức thương mại điện tử mới cho phép thanh toán ngay trong ChatGPT

2. AgentKit – Bộ Công Cụ Xây Dựng AI Agent Chuyên Nghiệp

  • Agent Builder: Giao diện kéo thả trực quan để thiết kế workflow cho AI agent mà không cần code phức tạp
  • ChatKit: Giao diện chat có thể tích hợp vào ứng dụng hoặc website của bạn, hỗ trợ streaming responses, quản lý threads, hiển thị quá trình suy nghĩ của model
  • Connector Registry: Bảng điều khiển tập trung để quản lý kết nối dữ liệu với Dropbox, Google Drive, SharePoint, Microsoft Teams
  • Guardrails: Lớp bảo mật mã nguồn mở giúp bảo vệ agent khỏi hành vi không mong muốn, có thể che PII, phát hiện jailbreaks
  • Enhanced Evals: Công cụ đánh giá nâng cao với datasets, trace grading, tối ưu prompt tự động, hỗ trợ các model của bên thứ ba
  • Demo trực tiếp: Tại sự kiện, một kỹ sư OpenAI đã xây dựng một AI agent hoàn chỉnh chỉ trong 8 phút

3. GPT-5 Pro – Model AI Thông Minh Nhất Trong API

  • Khả năng suy luận: Đạt trình độ PhD trong các lĩnh vực khoa học, có khả năng suy luận sâu cho các tác vụ phức tạp
  • Độ chính xác cao: Đặc biệt phù hợp cho tài chính, pháp lý, y tế – các lĩnh vực đòi hỏi độ chính xác cao
  • Reasoning effort: Có 4 mức độ suy luận (minimal, low, medium, high) để cân bằng giữa tốc độ và chất lượng
  • Context window: 272,000 tokens cho input, 128,000 tokens cho output
  • Multimodal: Hỗ trợ text và image cho input, text cho output

4. Codex – AI Agent Lập Trình Chính Thức Ra Mắt

  • GPT-5 Codex Model: Phiên bản GPT-5 được huấn luyện đặc biệt cho coding và agentic workflows
  • Tích hợp Slack: Lập trình viên có thể giao việc hoặc đặt câu hỏi trực tiếp từ Slack channels
  • Codex SDK: Cho phép tự động hóa code review, refactoring, automated testing
  • Thống kê ấn tượng:
    • Số lượng tin nhắn tăng 10x kể từ khi ra mắt tháng 8/2025
    • Đã xử lý hơn 40 trillion tokens
    • Nội bộ OpenAI: 70% pull requests nhiều hơn mỗi tuần

5. Sora 2 – Video Generation Trong API

  • Kiểm soát nâng cao: Có thể chỉ định độ dài, tỷ lệ khung hình, độ phân giải
  • Audio đồng bộ: Tạo video với âm thanh đầy đủ, âm thanh môi trường, hiệu ứng được đồng bộ với hình ảnh
  • Remix video: Cho phép chỉnh sửa và remix video đã tạo
  • Giá cả:
    • Sora-2: $1.00 cho video 10 giây độ phân giải tiêu chuẩn
    • Sora-2-pro: $5.00 cho video 10 giây độ phân giải cao

6. Mini Models – Tiết Kiệm Chi Phí

Model Chức năng Tiết kiệm
gpt-realtime-mini Voice interaction real-time 70% rẻ hơn large model
gpt-image-1-mini Tạo hình ảnh 80% rẻ hơn large model

7. Giá Cả GPT-5 Cạnh Tranh

Loại Input Output
GPT-5 $1.25/1M tokens $10/1M tokens
So với Claude Opus 4.1 $15/1M tokens $75/1M tokens

✨ II. Những Điểm Nổi Bật Đáng Chú Ý

🎯 Dễ Dàng Hơn Bao Giờ Hết

Dân chủ hóa phát triển phần mềm: Sam Altman đã kể câu chuyện về một cụ ông 89 tuổi người Nhật tự học lập trình với ChatGPT và đã tạo ra 11 ứng dụng iPhone dành cho người cao tuổi. Đây là minh chứng cho tầm nhìn “bất kỳ ai có ý tưởng đều có thể xây dựng ứng dụng cho chính mình”.

⚡ Tốc Độ Phát Triển Chưa Từng Có

“Phần mềm từng mất hàng tháng hoặc hàng năm để xây dựng. Giờ đây bạn thấy nó có thể được tạo ra chỉ trong vài phút với AI. Bạn không cần một đội ngũ lớn. Bạn chỉ cần một ý tưởng hay và có thể biến nó thành hiện thực nhanh hơn bao giờ hết.” – Sam Altman

🔒 Bảo Mật và Quản Trị Doanh Nghiệp

  • Content Shield: OpenAI cung cấp bảo vệ bản quyền cho doanh nghiệp
  • Global Admin Console: Quản lý domains, SSO, nhiều API organizations
  • Guardrails: Bảo vệ dữ liệu nhạy cảm và ngăn chặn hành vi độc hại

🤝 Hợp Tác Chiến Lược

AMD Partnership: OpenAI công bố hợp tác chiến lược với AMD để triển khai 6 gigawatts GPU Instinct của AMD trong nhiều năm tới, với warrant lên đến 160 triệu cổ phiếu AMD.

🌟 III. Tác Động và Ý Nghĩa

1. Đối Với Nhà Phát Triển

  • Giảm thời gian phát triển: Từ nhiều tháng xuống còn vài phút nhờ các công cụ như AgentKit và Codex
  • Chi phí thấp hơn: GPT-5 rẻ hơn 50% so với GPT-4o ở input, các mini models tiết kiệm 70-80%
  • Phân phối rộng rãi: Tiếp cận ngay 800 triệu người dùng ChatGPT qua Apps SDK
  • Developer lock-in thấp hơn: MCP là chuẩn mở, giúp dễ dàng chuyển đổi giữa các nền tảng

2. Đối Với Doanh Nghiệp

  • Tăng năng suất: AI agents có thể tự động hóa quy trình phức tạp từ customer support đến sales operations
  • Giảm headcount: Đội nhỏ có thể làm việc của đội lớn nhờ AI, tiết kiệm chi phí nhân sự
  • Cạnh tranh công bằng: Startup có thể cạnh tranh với đại gia nhờ chi phí thấp và công cụ dễ tiếp cận
  • Quản trị và bảo mật: Connector Registry và Guardrails giúp quản lý dữ liệu tập trung và đảm bảo compliance

3. Đối Với Người Dùng Cuối

  • Trải nghiệm liền mạch: Không cần chuyển đổi giữa nhiều ứng dụng, mọi thứ trong một giao diện ChatGPT
  • Cá nhân hóa cao: AI agents có thể học và thích nghi với nhu cầu cá nhân
  • Sáng tạo nội dung dễ dàng: Sora 2 cho phép tạo video chất lượng cao chỉ với mô tả text
  • Học tập và phát triển: Tích hợp Coursera giúp học tập cá nhân hóa ngay trong ChatGPT

4. Tác Động Ngành

Cuộc chiến giá cả AI: Với giá GPT-5 Pro rẻ hơn đáng kể so với Claude Opus 4.1 (rẻ hơn 92% ở input, 86% ở output), OpenAI đang tạo áp lực giá lên toàn ngành.

Platform Play: ChatGPT không còn là chatbot đơn thuần mà đang trở thành một nền tảng – giống như App Store của Apple. Điều này có thể thay đổi cách phân phối ứng dụng AI.

Democratization of AI: Với công cụ visual như Agent Builder, người không biết code cũng có thể tạo AI agents phức tạp, mở rộng đáng kể cộng đồng AI builders.

Chuyển dịch từ Answers đến Actions: ChatGPT đang chuyển từ trả lời câu hỏi sang thực hiện hành động, đánh dấu bước tiến mới trong phát triển AI.

5. Xu Hướng Tương Lai

  • AI như một Operating System: ChatGPT đang tiến đến việc trở thành một hệ điều hành AI – nơi tập trung apps, agents và users
  • Agentic AI: Từ việc chỉ trả lời câu hỏi, AI giờ có thể nhận và hoàn thành các tác vụ phức tạp end-to-end
  • Multimodal Everything: Tích hợp text, image, audio, video trong một platform duy nhất
  • Device Ecosystem: Với sự tham gia của Jony Ive và thương vụ mua io ($6.4B), OpenAI đang hướng đến việc tạo ra thiết bị AI riêng

🚀 Kết Luận

OpenAI DevDay 2025 không chỉ là sự kiện công bố sản phẩm mà là tuyên ngôn về tương lai của phát triển phần mềm. Với Apps SDK, AgentKit, GPT-5 Pro, và Sora 2, OpenAI đang xây dựng một hệ sinh thái AI toàn diện – từ nền tảng phát triển cho đến trải nghiệm người dùng cuối.

Thông điệp chính: “Bất kỳ ai có ý tưởng tốt đều có thể biến nó thành hiện thực nhanh hơn bao giờ hết”. Đây không chỉ là slogan marketing mà là tầm nhìn về một thế giới mà AI dân chủ hóa việc sáng tạo phần mềm.

Với 800 triệu người dùng, 4 triệu nhà phát triển và 6 tỷ tokens được xử lý mỗi phút, OpenAI không chỉ dẫn đầu cuộc đua AI mà đang định hình lại cách chúng ta tương tác với công nghệ.

Nguồn tham khảo:

  • OpenAI DevDay Official: openai.com/devday/
  • Sam Altman Keynote Livestream
  • OpenAI Blog và Documentation
  • CNBC, TechCrunch Coverage

GPT-5-Codex Prompting Guide: Hướng Dẫn Tối Ưu Hóa Prompt Cho Lập Trình

Giới Thiệu

GPT-5-Codex là phiên bản nâng cao của GPT-5, được OpenAI tối ưu hóa đặc biệt cho các nhiệm vụ lập trình tương tác và tự động. Mô hình này được huấn luyện với trọng tâm vào công việc kỹ thuật phần mềm thực tế, mang lại hiệu suất vượt trội trong cả các phiên làm việc nhanh chóng và các nhiệm vụ phức tạp kéo dài.

⚠️ Lưu Ý Quan Trọng

  • Không phải thay thế trực tiếp: GPT-5-Codex không phải là thay thế trực tiếp cho GPT-5, vì nó yêu cầu cách prompting khác biệt đáng kể
  • Chỉ hỗ trợ Responses API: Mô hình này chỉ được hỗ trợ với Responses API và không hỗ trợ tham số verbosity
  • Dành cho người dùng API: Hướng dẫn này dành cho người dùng API của GPT-5-Codex và tạo developer prompts, không dành cho người dùng Codex

Những Cải Tiến Chính Của GPT-5-Codex

1. Khả Năng Điều Hướng Cao

GPT-5-Codex cung cấp mã chất lượng cao cho các nhiệm vụ kỹ thuật phức tạp như:

  • Phát triển tính năng mới
  • Kiểm thử và gỡ lỗi
  • Tái cấu trúc mã nguồn
  • Đánh giá và review code

Tất cả những nhiệm vụ này được thực hiện mà không cần hướng dẫn dài dòng hay chi tiết.

2. Mức Độ Suy Luận Thích Ứng

Mô hình có khả năng điều chỉnh thời gian suy luận theo độ phức tạp của nhiệm vụ:

  • Phản hồi nhanh trong các phiên tương tác ngắn
  • Có thể làm việc độc lập trong nhiều giờ cho các nhiệm vụ phức tạp
  • Tự động phân bổ tài nguyên tính toán phù hợp

3. Xuất Sắc Trong Đánh Giá Mã

GPT-5-Codex được huấn luyện đặc biệt để:

  • Thực hiện đánh giá mã chuyên sâu
  • Điều hướng trong các cơ sở mã lớn
  • Chạy mã và kiểm thử để xác nhận tính đúng đắn
  • Phát hiện lỗi và đề xuất cải tiến

Môi Trường Hỗ Trợ

GPT-5-Codex được thiết kế đặc biệt cho:

  • Codex CLI: Giao diện dòng lệnh cho lập trình
  • Phần mở rộng Codex IDE: Phần mở rộng cho các IDE phổ biến
  • Môi trường đám mây Codex: Môi trường đám mây chuyên dụng
  • Tích hợp GitHub: Tích hợp sâu với GitHub
  • Đa dạng công cụ: Hỗ trợ nhiều loại công cụ lập trình

Nguyên Tắc Cốt Lõi: “Ít Hơn Là Tốt Hơn”

Đây là nguyên tắc quan trọng nhất khi tạo prompt cho GPT-5-Codex. Do mô hình được huấn luyện đặc biệt cho lập trình, nhiều thực hành tốt đã được tích hợp sẵn, và việc quá tải hướng dẫn có thể làm giảm chất lượng.

1. Bắt Đầu Với Prompt Tối Giản

  • Sử dụng prompt ngắn gọn, lấy cảm hứng từ prompt hệ thống của Codex CLI
  • Chỉ thêm những hướng dẫn thực sự cần thiết
  • Tránh các mô tả dài dòng không cần thiết

2. Loại Bỏ Phần Mở Đầu

  • GPT-5-Codex không hỗ trợ phần mở đầu
  • Yêu cầu phần mở đầu sẽ khiến mô hình dừng sớm trước khi hoàn thành nhiệm vụ
  • Tập trung vào nhiệm vụ chính ngay từ đầu

3. Giảm Số Lượng Công Cụ

  • Chỉ sử dụng các công cụ cần thiết:
    • Terminal: Để thực thi lệnh
    • apply_patch: Để áp dụng thay đổi mã
  • Loại bỏ các công cụ không cần thiết

4. Mô Tả Công Cụ Ngắn Gọn

  • Làm cho mô tả công cụ ngắn gọn nhất có thể
  • Loại bỏ các chi tiết không cần thiết
  • Tập trung vào chức năng cốt lõi

So Sánh Với GPT-5

Prompt của GPT-5-Codex ngắn hơn khoảng 40% so với GPT-5, điều này nhấn mạnh rằng:

  • Prompt tối giản là lý tưởng cho mô hình này
  • Ít token hơn = hiệu suất tốt hơn
  • Tập trung vào chất lượng thay vì số lượng

Ví Dụ Thực Tế

Prompt Không Tối Ưu:

Bạn là một lập trình viên chuyên nghiệp với nhiều năm kinh nghiệm. Hãy bắt đầu bằng cách phân tích yêu cầu, sau đó tạo kế hoạch chi tiết, và cuối cùng implement code với nhiều comment giải thích. Đảm bảo code có error handling, unit tests, và documentation đầy đủ...

Prompt Tối Ưu:

Tạo một function để tính tổng hai số nguyên.

Ví Dụ Thực Tế: Gọi API GPT-5-Codex

Bước 1: Cài đặt và cấu hình

Lưu ý: Thay sk-your-api-key-here bằng API key thực tế của bạn từ OpenAI.

Bước 2: Tạo file Python gọi GPT-5-Codex

Mở Notepad và paste code sau:

import os
from openai import OpenAI
from dotenv import load_dotenv

# Load API key từ file .env
load_dotenv()

# Khởi tạo client
client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))

def call_gpt5_codex(prompt):
    """
    Gọi API GPT-5-Codex với prompt tối giản
    """
    try:
        response = client.chat.completions.create(
            model="gpt-5-codex",  # Model name cho GPT-5-Codex
            messages=[
                {
                    "role": "user", 
                    "content": prompt
                }
            ],
            temperature=0.1,  # Thấp để code chính xác
            max_tokens=4000
        )
        
        return response.choices[0].message.content
        
    except Exception as e:
        return f"Lỗi: {str(e)}"

# Ví dụ sử dụng
if __name__ == "__main__":
    # Prompt tối giản - không cần hướng dẫn dài dòng
    prompt = "Tạo function Python để tính fibonacci số thứ n"
    
    print("Đang gọi GPT-5-Codex...")
    result = call_gpt5_codex(prompt)
    print("\nKết quả từ GPT-5-Codex:")
    print("=" * 50)
    print(result)

Bước 3: Chạy demo

Bước 4: Ví dụ Request/Response thực tế

Request JSON:

{
  "model": "gpt-5-codex",
  "messages": [
    {
      "role": "user",
      "content": "Tạo REST API với FastAPI để quản lý users"
    }
  ],
  "temperature": 0.1,
  "max_tokens": 2000
}

Response từ GPT-5-Codex:

from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from typing import List

app = FastAPI()

# In-memory storage (thay bằng database thực tế)
users = []

class User(BaseModel):
    id: int
    name: str
    email: str

class UserCreate(BaseModel):
    name: str
    email: str

@app.get("/users")
def get_users():
    return users

@app.post("/users")
def create_user(user: UserCreate):
    new_user = User(
        id=len(users) + 1,
        name=user.name,
        email=user.email
    )
    users.append(new_user)
    return new_user

@app.delete("/users/{user_id}")
def delete_user(user_id: int):
    global users
    users = [u for u in users if u.id != user_id]
    return {"message": "User deleted"}

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8000)

Bước 5: So sánh Prompt hiệu quả

❌ Prompt không tối ưu:

Bạn là một lập trình viên chuyên nghiệp với 10 năm kinh nghiệm. Hãy tạo một REST API hoàn chỉnh với FastAPI để quản lý users. API phải có đầy đủ CRUD operations, validation, error handling, logging, và documentation. Đảm bảo code clean, có comment đầy đủ, và tuân thủ best practices...

✅ Prompt tối ưu cho GPT-5-Codex:

Tạo REST API với FastAPI để quản lý users

Kết quả: GPT-5-Codex tự động tạo ra code đầy đủ chức năng mà không cần hướng dẫn chi tiết.

Anti-Prompting: Những Điều Không Cần Thiết

Do GPT-5-Codex được huấn luyện đặc biệt cho lập trình agentic, việc điều chỉnh prompt thường có nghĩa là loại bỏ hướng dẫn thay vì thêm vào. Dưới đây là những khía cạnh bạn có thể không cần điều chỉnh:

1. Suy Luận Thích Ứng (Adaptive Reasoning)

Suy luận thích ứng giờ đây là mặc định trong GPT-5-Codex. Trước đây, bạn có thể đã prompt các mô hình để “suy nghĩ kỹ hơn” hoặc “phản hồi nhanh” dựa trên độ khó của nhiệm vụ. GPT-5-Codex tự động điều chỉnh:

  • Câu hỏi đơn giản: “Làm thế nào để undo commit cuối nhưng giữ lại các thay đổi staged?” → Phản hồi nhanh không cần điều chỉnh thêm
  • Nhiệm vụ phức tạp: Tự động dành thời gian cần thiết và sử dụng công cụ phù hợp

2. Lập Kế Hoạch (Planning)

GPT-5-Codex được huấn luyện cho nhiều loại nhiệm vụ lập trình từ các tác vụ tự động dài hạn đến các tác vụ lập trình tương tác ngắn hạn. Mô hình có tính cách hợp tác theo mặc định:

  • Khi bắt đầu một tác vụ tự động, mô hình sẽ xây dựng kế hoạch chi tiết
  • Cập nhật tiến độ trong quá trình thực hiện
  • Codex CLI bao gồm công cụ lập kế hoạch và mô hình được huấn luyện để sử dụng nó

3. Phần Mở Đầu (Preambles)

GPT-5-Codex KHÔNG tạo ra phần mở đầu! Việc prompt và yêu cầu phần mở đầu có thể dẫn đến việc mô hình dừng sớm. Thay vào đó, có một trình tóm tắt tùy chỉnh tạo ra các tóm tắt chi tiết chỉ khi phù hợp.

4. Giao Diện Người Dùng

GPT-5-Codex mặc định có thẩm mỹ mạnh mẽ và các thực hành giao diện người dùng hiện đại tốt nhất. Nếu bạn có thư viện hoặc framework ưa thích, hãy điều chỉnh mô hình bằng cách thêm các phần ngắn:

Hướng Dẫn Giao Diện Người Dùng
Sử dụng các thư viện sau trừ khi người dùng hoặc repo chỉ định khác:
Framework: React + TypeScript
Styling: Tailwind CSS
Components: shadcn/ui
Icons: lucide-react
Animation: Framer Motion
Charts: Recharts
Fonts: San Serif, Inter, Geist, Mona Sans, IBM Plex Sans, Manrope

Prompt Tham Chiếu: Codex CLI

Dưới đây là prompt đầy đủ của Codex CLI mà bạn có thể sử dụng làm tham chiếu khi tạo prompt cho GPT-5-Codex:

Các Điểm Chính Trong Prompt Codex CLI:

1. Cấu hình chung:

  • Các đối số của shell sẽ được truyền cho execvp()
  • Hầu hết các lệnh terminal nên được prefix với ["bash", "-lc"]
  • Luôn đặt tham số workdir khi sử dụng hàm shell
  • Ưu tiên sử dụng rg thay vì grep vì nhanh hơn

2. Ràng buộc chỉnh sửa:

  • Mặc định sử dụng ASCII khi chỉnh sửa hoặc tạo file
  • Thêm comment code ngắn gọn giải thích những gì đang diễn ra
  • Có thể ở trong git worktree bẩn – KHÔNG BAO GIỜ revert các thay đổi hiện có

3. Công cụ lập kế hoạch:

  • Bỏ qua công cụ planning cho các tác vụ đơn giản (khoảng 25% dễ nhất)
  • Không tạo kế hoạch một bước
  • Cập nhật kế hoạch sau khi hoàn thành một trong các subtask

4. Sandboxing và approvals:

  • Sandboxing hệ thống file: chỉ đọc, ghi workspace, truy cập đầy đủ nguy hiểm
  • Sandboxing mạng: hạn chế, bật
  • Chính sách phê duyệt: không tin tưởng, khi thất bại, theo yêu cầu, không bao giờ

5. Cấu trúc và phong cách:

  • Văn bản thuần túy; CLI xử lý định dạng
  • Tiêu đề: tùy chọn; Title Case ngắn (1-3 từ) trong
  • Dấu đầu dòng: sử dụng -; hợp nhất các điểm liên quan
  • Monospace: backticks cho lệnh/đường dẫn/biến môi trường/id code

Apply Patch

Như đã chia sẻ trước đó trong hướng dẫn GPT-5, đây là triển khai apply_patch cập nhật nhất mà chúng tôi khuyến nghị sử dụng cho việc chỉnh sửa file để khớp với phân phối huấn luyện.

Lợi Ích Của Việc Sử Dụng Đúng Cách

  1. Hiệu Suất Cao Hơn: Phản hồi nhanh và chính xác
  2. Tiết Kiệm Token: Giảm chi phí sử dụng (40% ít token hơn GPT-5)
  3. Kết Quả Tốt Hơn: Mô hình tập trung vào nhiệm vụ chính
  4. Dễ Bảo Trì: Prompt ngắn gọn dễ hiểu và chỉnh sửa
  5. Tự Động Hóa: Suy luận thích ứng và lập kế hoạch tự động
  6. Tích Hợp Sẵn: Nhiều best practices đã được tích hợp sẵn

Kết Luận

GPT-5-Codex đại diện cho một bước tiến lớn trong việc ứng dụng AI cho lập trình. Việc áp dụng đúng các nguyên tắc prompting sẽ giúp bạn tận dụng tối đa sức mạnh của mô hình này. Hãy nhớ rằng “ít hơn là tốt hơn” – đây không chỉ là nguyên tắc của GPT-5-Codex mà còn là triết lý trong việc tạo ra các hệ thống AI hiệu quả.

Automatically Generate Google Slides with an AI Prompt

I came across a very interesting idea from the author まじん (Majin) on note.com:

Majin used Gemini to turn a single prompt into a complete Google Slides presentation, but I tried customizing it to run with ChatGPT (or Google AI Studio), and the results were quite exciting.


1. 🔍 Structure of Majin’s Prompt

Through analysis, Majin’s prompt has the following main components:

  • Role assignment for AI: The AI is not just a chatbot, but acts as a Data Scientist and Presentation Designer.

  • Absolute mission: From a piece of input text, the AI must output a JavaScript object array (slideData) that defines the entire slide structure.

  • Fixed blueprint: The prompt includes a Google Apps Script (GAS) framework, where you only need to replace slideData to generate Google Slides.

  • SlideData includes:

    • title: Slide title

    • content: Content (in the form of bullets, text, or tables)

    • pattern: Display type (Title, List, TwoColumn, Image, …)

    • speakerNote: Presenter’s notes

👉 The important point: The prompt does not directly create slides, but outputs JSON-like data that the GAS script uses to build the slides.


2. 🛠️ Customizing & Running with ChatGPT

First, download system prompt.md

🔧 Steps inside the Project

  1. Create a new Project → give it a name (e.g., Google Slide Auto Generator).

  2. Add a file → upload system prompt.md.

  3. In the Project Settings / Overview screen, you will find the Instructions (System message) box → paste the line:
    添付した「system prompt.md」に忠実に従ってください。

  4. (Meaning: Please strictly follow the attached system prompt.md).
    If you don’t see a dedicated “Instructions” field, ChatGPT may show “Description” or “Setup” instead → paste it there. It works as the System Prompt for the whole Project.

  5. From now on, in that Project, whenever you input raw text, ChatGPT will immediately reference system prompt.md and output both slideData + GAS code.

👉 In other words: the Project’s Instruction acts as a dedicated System Message for that workspace. The UI may look different, but it’s always under “Overview/Setup” when you open the Project.

Next, type a prompt about the content you want to make slides from → and it will generate slideData.

I tested it with content about LLMs and got the following result:

const slideData = [

{

type: ‘title’,

title: ‘The Busy Person’s Intro to Large Language Models’,

date: ‘2025.09.21’,

notes: ‘Re-recorded overview of LLMs focusing on what they are, how to run them, how they are trained, and why next-word prediction works’

},

{

type: ‘content’,

title: ‘Agenda’,

points: [

‘What is an LLM’,

‘Running a model locally’,

‘Training at scale’,

‘Why next-word prediction is powerful’,

‘Key takeaways’

],

notes: ‘Keep it practical and concrete. We start with the simplest mental model, then show inference vs training, then the intuition for why LLMs encode knowledge’

},

{

type: ‘section’,

title: ‘1. What is an LLM’,

notes: ‘Define an LLM at the simplest level and frame the talk around artifacts you can hold on disk’

},

{

type: ‘content’,

title: ‘LLM = Two Files’,

subhead: ‘Weights + Runner’,

points: [

‘Parameters file holds learned weights’,

‘Runner implements forward pass of the network’,

‘Self contained package runs offline’

],

notes: ‘Stress that inference can be simple and local. The interesting part lives in the parameters learned during training’

},

{

type: ‘content’,

title: ‘Example: Llama 2 • 70B’,

subhead: ‘Open weights from Meta’,

points: [

‘Sizes include 7B 13B 34B 70B’,

‘Open weights and paper enable local study’,

‘Closed APIs trade control for convenience’

],

notes: ‘Contrast open weights with closed APIs like hosted assistants without exposing internals’

},

{

type: ‘section’,

title: ‘2. Running a Model’,

notes: ‘Give a feel for local inference and why the demo used a smaller model for speed’

},

{

type: ‘process’,

title: ‘Local Inference Flow’,

subhead: ‘From files to generations’,

steps: [

‘Get weights file on disk’,

‘Compile or install a runner’,

‘Point runner to weights’,

‘Send a prompt’,

‘Stream generated tokens’

],

notes: ‘Mention that a 7B model is fast enough on a laptop while 70B is roughly an order of magnitude slower’

},

{

type: ‘compare’,

title: ‘Open Weights vs Closed API’,

leftTitle: ‘Open Weights’,

rightTitle: ‘Closed API’,

leftItems: [

‘Full local control’,

‘Offline capable’,

‘Tunable and inspectable’

],

rightItems: [

‘Best quality access’,

‘No infra to manage’,

‘Faster to integrate’

],

notes: ‘There is no universal winner. Choose based on control privacy latency and capability needs’

},

{

type: ‘section’,

title: ‘3. Training at Scale’,

notes: ‘Explain where weights come from and why training is expensive compared to inference’

},

{

type: ‘content’,

title: ‘Training Ingredients’,

subhead: ‘Data Compute Cost’,

points: [

‘Large curated text corpus’,

‘Massive GPU clusters for days’,

‘Significant budget for runs’

],

notes: ‘Use order of magnitude thinking rather than exact numbers. Emphasize engineering and data work beyond raw compute’

},

{

type: ‘timeline’,

title: ‘From Corpus to Weights’,

subhead: ‘Very high level lifecycle’,

milestones: [

{ label: ‘Collect and filter data’, date: ‘Step 1’, state: ‘done’ },

{ label: ‘Train on GPU cluster’, date: ‘Step 2’, state: ‘done’ },

{ label: ‘Evaluate and align’, date: ‘Step 3’, state: ‘done’ },

{ label: ‘Release weights and runner’, date: ‘Step 4’, state: ‘next’ }

],

notes: ‘Keep the audience at the bird’s eye view. The output is a weights artifact you can reuse everywhere for inference’

},

{

type: ‘section’,

title: ‘4. Why It Works’,

notes: ‘Bridge from mechanics to intuition about capability’

},

{

type: ‘content’,

title: ‘Next Word Prediction Objective’,

subhead: ‘Simple objective rich consequences’,

points: [

‘Model predicts a distribution over next token’,

‘Good prediction forces world knowledge’,

‘Patterns styles and relations get compressed’

],

notes: ‘Relate prediction to lossy compression intuition. To predict the next token the model must encode structure facts and relationships’

},

{

type: ‘content’,

title: ‘Common Misconceptions’,

points: [

‘It is not copy paste of the web’,

‘Bigger is not always better’,

‘Closed is not always better than open’

],

notes: ‘Clarify generalization vs memorization and the role of data quality and algorithms along with scale’

},

{

type: ‘section’,

title: ‘5. Takeaways’,

notes: ‘Land the plane with pragmatic guidance’

},

{

type: ‘content’,

title: ‘Practical Guidance’,

points: [

‘Choose model size for latency and cost’,

‘Use open weights for control’,

‘Use hosted APIs for frontier quality’

],

notes: ‘Map choices to product constraints such as privacy cost latency and maintainability’

},

{

type: ‘closing’,

notes: ‘Thanks for watching. Invite questions and link resources in the description’

}

];

Next steps

  1. Open the .md file on your computer.

  2. Copy the code from the .md file.

  3. Open Google Slides → create a new slide deck.

  4. Click Extensions → Apps Script.

  5. Paste the copied code into the Apps Script editor.

  6. Paste your slideData.

  7. Press Run


3. ✅ Experience & Results

 

  • Works well on ChatGPT: No need for Gemini, GPT-5 is enough.

  • Advantage: The prompt standardizes the output into a JSON-like object, making it easy to control.

  • Reference implementation:


📌 Conclusion

  • Majin’s prompt is a great framework to turn AI into an automatic slide design tool.

  • It doesn’t have to be Gemini — ChatGPT (GPT-5) also works well.

  • You just need to customize the input → and you can generate Google Slides for any topic (training, pitching, learning…).

👉 This article was written with reference to blogs by まじん (Majin):

Cách Mọi Người Sử Dụng ChatGPT: Nghiên Cứu Toàn Diện

Tóm Tắt Executive

Nghiên cứu “How People Use ChatGPT” là phân tích toàn diện nhất về cách 700 triệu người dùng trên toàn thế giới tương tác với trí tuệ nhân tạo. Được thực hiện bởi đội ngũ chuyên gia hàng đầu từ OpenAI, Đại học Duke và Harvard, nghiên cứu này sử dụng phương pháp bảo vệ quyền riêng tư tiên tiến để phân tích hơn 1.1 triệu cuộc hội thoại mẫu, tiết lộ những hiểu biết chưa từng có về cách con người sử dụng AI trong thực tế.

Con Số Ấn Tượng

  • 700 triệu người dùng hàng tuần (10% dân số trưởng thành toàn cầu)
  • 18 tỷ tin nhắn mỗi tuần (2.5 tỷ tin nhắn mỗi ngày, 29,000 tin nhắn mỗi giây)
  • Tăng trưởng chưa từng có trong lịch sử công nghệ
  • 70% tin nhắn không liên quan công việc (tăng từ 53% tháng 6/2024)
  • Consumer surplus ít nhất $97 tỷ/năm chỉ riêng tại Mỹ

1. Giới Thiệu và Bối Cảnh

ChatGPT: Công Nghệ Đột Phá

ChatGPT được ra mắt vào tháng 11/2022 như một “research preview” và đã trở thành chatbot thương mại đầu tiên và có khả năng lớn nhất trên thị trường. Dựa trên Large Language Model (LLM), ChatGPT đại diện cho sự tăng tốc đáng kể trong khả năng AI.

Tốc Độ Phát Triển Lịch Sử

Timeline phát triển:

  • 30/11/2022: Ra mắt “research preview”
  • 5/12/2022: Đạt 1 triệu người dùng trong 5 ngày
  • Q1/2023: 100 triệu người dùng
  • Q2/2023: 200 triệu người dùng
  • Q3/2023: 350 triệu người dùng
  • Q4/2023: 500 triệu người dùng
  • Q1/2024: 600 triệu người dùng
  • Q2/2024: 650 triệu người dùng
  • Q3/2025: 700 triệu người dùng

So Sánh Với Các Nền Tảng Khác

ChatGPT đã đạt được sự chấp nhận toàn cầu với tốc độ chưa từng thấy, vượt xa tất cả các nền tảng khác trong lịch sử công nghệ. Tốc độ tăng trưởng này không có tiền lệ và cho thấy sự thay đổi căn bản trong cách con người tương tác với công nghệ.


2. Phương Pháp Nghiên Cứu và Bảo Vệ Quyền Riêng Tư

Datasets Sử Dụng

1. Growth Dataset:

  • Tổng tin nhắn hàng ngày từ 11/2022-9/2025
  • Thông tin nhân khẩu học cơ bản tự báo cáo
  • Metadata người dùng đã được ẩn danh hóa

2. Classified Messages:

  • Mẫu ngẫu nhiên ~1.1 triệu tin nhắn từ 5/2024-6/2025
  • Phân loại tự động bằng LLM
  • Loại trừ người dùng opt-out training, dưới 18 tuổi, đã xóa tài khoản

3. Employment Dataset:

  • Dữ liệu việc làm tổng hợp cho 130,000 người dùng
  • Phân tích trong Data Clean Room bảo mật
  • Chỉ báo cáo tổng hợp (tối thiểu 100 người dùng)

Bảo Vệ Quyền Riêng Tư

Automated Classification:

  • Không ai đọc tin nhắn thô
  • Sử dụng LLM để phân loại tự động
  • Privacy Filter loại bỏ PII
  • Context window 10 tin nhắn trước

Data Clean Room:

  • Phân tích dữ liệu việc làm trong môi trường bảo mật
  • Notebook phải được phê duyệt trước khi chạy
  • Dữ liệu bị xóa sau khi nghiên cứu hoàn thành

Validation:

  • So sánh với WildChat dataset (public)
  • Human annotators đánh giá 149 tin nhắn
  • Fleiss’ κ và Cohen’s κ để đo độ tin cậy

3. Tăng Trưởng và Phát Triển

Tăng Trưởng Tổng Thể

Số liệu tăng trưởng:

  • Tháng 7/2024 – 7/2025: Số tin nhắn tăng hơn 5 lần
  • Các nhóm người dùng: Cả nhóm mới và nhóm cũ đều tăng trưởng
  • Tin nhắn/người dùng: Tăng trưởng liên tục trong mọi nhóm

Phân Tích Theo Nhóm Người Dùng

Nhóm đầu tiên (Q4/2022-Q1/2023):

  • Sử dụng giảm nhẹ trong 2023
  • Bắt đầu tăng trưởng trở lại cuối 2024
  • Hiện tại cao hơn mọi thời điểm trước

Các nhóm sau:

  • Tăng trưởng mạnh từ nhóm người dùng mới
  • Tăng trưởng trong nhóm người dùng hiện tại
  • Cải thiện khả năng mô hình và khám phá use case mới

4. Cách Sử Dụng ChatGPT

Phân Loại Công Việc vs. Cá Nhân

Bảng 1: Tăng Trưởng Tin Nhắn Hàng Ngày (Triệu)

Tháng Không công việc Tỷ lệ Công việc Tỷ lệ Tổng
Tháng 6/2024 238 53% 213 47% 451
Tháng 6/2025 1,911 73% 716 27% 2,627
Tăng trưởng +703% +20 điểm +236% -20 điểm +483%

Phát hiện quan trọng:

  • Cả hai loại tin nhắn đều tăng liên tục
  • Tin nhắn không công việc tăng nhanh hơn 3 lần
  • Xu hướng chủ yếu do thay đổi trong từng nhóm người dùng
  • Phù hợp với consumer surplus $97 tỷ/năm (Collis & Brynjolfsson, 2025)

Ba Chủ Đề Chính (80% Sử Dụng)

1. Practical Guidance (Hướng Dẫn Thực Tiễn) – 29%

Phân loại chi tiết:

  • Tutoring/Teaching: 10.2% tổng tin nhắn (36% trong Practical Guidance)
  • How-to Advice: 8.5% tổng tin nhắn (30% trong Practical Guidance)
  • Creative Ideation: Tạo ý tưởng sáng tạo
  • Health/Fitness/Beauty: Lời khuyên sức khỏe, thể dục, làm đẹp

Đặc điểm:

  • Ổn định ở mức 29% trong suốt thời gian nghiên cứu
  • Khác biệt với Seeking Information ở chỗ được tùy chỉnh cao
  • Ví dụ: Kế hoạch tập luyện cá nhân hóa vs. Thông tin chung về marathon Boston

2. Writing (Viết Lách) – 24% (Giảm từ 36% tháng 7/2024)

Phân loại chi tiết:

  • Edit/Critique Provided Text: 40% (chỉnh sửa văn bản có sẵn)
  • Personal Writing/Communication: 25% (viết cá nhân, giao tiếp)
  • Translation: 15% (dịch thuật)
  • Argument/Summary Generation: 15% (tạo lập luận, tóm tắt)
  • Write Fiction: 5% (viết sáng tạo)

Đặc điểm quan trọng:

  • 2/3 tin nhắn Writing là chỉnh sửa văn bản có sẵn, không tạo mới
  • 40% tin nhắn công việc là Writing (tháng 7/2025)
  • 52% tin nhắn trong quản lý và kinh doanh là Writing
  • Giảm có thể do chuyển sang API cho lập trình

3. Seeking Information (Tìm Kiếm Thông Tin) – 24% (Tăng từ 14% tháng 7/2024)

Phân loại chi tiết:

  • Specific Info: Thông tin cụ thể về người, sự kiện, sản phẩm
  • Purchasable Products: Tìm kiếm sản phẩm có thể mua
  • Cooking/Recipes: Công thức nấu ăn

Đặc điểm:

  • Tăng trưởng mạnh nhất trong 3 chủ đề chính
  • Thay thế gần như hoàn toàn cho tìm kiếm web truyền thống
  • Linh hoạt hơn web search vì cung cấp phản hồi tùy chỉnh

Các Chủ Đề Khác

Technical Help – 5% (Giảm từ 12% tháng 7/2024)

  • Computer Programming: 4.2% tổng tin nhắn
  • Mathematical Calculation: 3% tổng tin nhắn
  • Data Analysis: 0.4% tổng tin nhắn

Lý do giảm: Sử dụng LLM cho lập trình tăng mạnh qua API, AI assistance trong code editing, và autonomous programming agents

Multimedia – 7% (Tăng từ 2% tháng 7/2024)

  • Create an Image: Tạo hình ảnh
  • Analyze an Image: Phân tích hình ảnh
  • Generate/Retrieve Other Media: Tạo/tìm media khác

Spike tháng 4/2025: Sau khi ChatGPT ra mắt tính năng tạo hình ảnh mới

Self-Expression – 2.4% (Thấp hơn dự kiến)

  • Relationships/Personal Reflection: 1.9% tổng tin nhắn
  • Games/Role Play: 0.4% tổng tin nhắn

So sánh: Zao-Sanders (2025) ước tính Therapy/Companionship là use case phổ biến nhất, nhưng nghiên cứu này cho thấy ngược lại


5. Phân Tích Mục Đích Sử Dụng: Asking/Doing/Expressing

Phân Loại Chi Tiết

Loại Tỷ lệ Mô tả Ví dụ
Asking 49% Tìm kiếm thông tin, lời khuyên để ra quyết định “Ai là tổng thống sau Lincoln?”, “Làm sao tạo ngân sách quý này?”
Doing 40% Yêu cầu ChatGPT thực hiện nhiệm vụ cụ thể “Viết lại email này cho trang trọng hơn”, “Tạo báo cáo tóm tắt”
Expressing 11% Bày tỏ quan điểm, cảm xúc, không có mục đích rõ ràng “Tôi cảm thấy lo lắng”, “Hôm nay thật tuyệt!”

Xu Hướng Thay Đổi Theo Thời Gian

Tháng 7/2024:

  • Asking: 50%
  • Doing: 50%
  • Expressing: 8%

Tháng 6/2025:

  • Asking: 51.6%
  • Doing: 34.6%
  • Expressing: 13.8%

Phân tích:

  • Asking tăng trưởng nhanh nhất
  • Asking được đánh giá chất lượng cao hơn
  • Doing chiếm 56% tin nhắn công việc
  • Writing chiếm 35% tin nhắn Doing

Phân Tích Theo Chủ Đề

Asking phổ biến hơn trong:

  • Practical Guidance
  • Seeking Information

Doing phổ biến hơn trong:

  • Writing
  • Multimedia

Expressing phổ biến hơn trong:

  • Self-Expression

6. Hoạt Động Công Việc (O*NET)

7 Hoạt Động Chính (77% Tổng Tin Nhắn)

Xếp hạng Hoạt động Tỷ lệ Mô tả
1 Getting Information 19.3% Thu thập thông tin từ nhiều nguồn
2 Interpreting Information 13.1% Giải thích ý nghĩa thông tin cho người khác
3 Documenting Information 12.8% Ghi chép, lưu trữ thông tin
4 Providing Consultation 9.2% Cung cấp tư vấn và lời khuyên
5 Thinking Creatively 9.1% Tư duy sáng tạo, đổi mới
6 Making Decisions 8.5% Ra quyết định và giải quyết vấn đề
7 Working with Computers 4.9% Làm việc với máy tính

Phân Tích Theo Nghề Nghiệp

Bảng 2: Xếp Hạng Hoạt Động Theo Nghề (1 = Phổ Biến Nhất)

Nghề Documenting Making Decisions Thinking Creatively Working with Computers Interpreting Getting Info Consultation
Management 2 1 3 6 4 5 8
Business 2 1 3 6 4 5 7
Computer/Math 4 2 5 1 3 6 7
Engineering 3 1 5 2 4 6 7
Science 2 1 4 3 6 5 7
Education 1 2 3 4 6 5 7
Health Professionals 1 2 3 X 5 4 6
Legal 1 X X X X X X

Phát hiện quan trọng:

  • Making Decisions luôn trong top 2 của mọi nghề
  • Documenting Information luôn trong top 4
  • Thinking Creatively xếp thứ 3 trong 10/13 nhóm nghề
  • Tương đồng cao giữa các nghề nghiệp khác nhau
  • ChatGPT chủ yếu hỗ trợ tìm kiếm thông tin và ra quyết định

7. Đặc Điểm Nhân Khẩu Học

Khoảng Cách Giới Tính Đã Thu Hẹp Đáng Kể

Timeline thay đổi:

  • Q4/2022 – Q1/2023: 80% người dùng có tên nam giới
  • Q2/2023: 70% nam giới, 30% nữ giới
  • Q3/2023: 65% nam giới, 35% nữ giới
  • Q4/2023: 60% nam giới, 40% nữ giới
  • Q1/2024: 56% nam giới, 44% nữ giới
  • Q2/2024: 54% nam giới, 46% nữ giới
  • Q2/2025: 48% nam giới, 52% nữ giới

Yếu tố ảnh hưởng:

  1. Marketing và PR: Chiến dịch hướng đến nữ giới
  2. Tính năng mới: Phù hợp với sở thích nữ giới
  3. Ứng dụng giáo dục: Nữ giới sử dụng nhiều hơn cho học tập
  4. Tích hợp xã hội: Chia sẻ kinh nghiệm trong cộng đồng

Phân Bố Theo Độ Tuổi

Tỷ lệ tin nhắn theo nhóm tuổi:

  • 18-25 tuổi: 46% tổng tin nhắn
  • 26-35 tuổi: 28% tổng tin nhắn
  • 36-45 tuổi: 16% tổng tin nhắn
  • 46-55 tuổi: 7% tổng tin nhắn
  • 56+ tuổi: 3% tổng tin nhắn

Tỷ lệ công việc theo tuổi:

  • Dưới 26: 23% tin nhắn công việc
  • 26-35: 35% tin nhắn công việc
  • 36-45: 42% tin nhắn công việc
  • 46-55: 45% tin nhắn công việc
  • 56-65: 38% tin nhắn công việc
  • 66+: 16% tin nhắn công việc

Tăng Trưởng Theo Quốc Gia và GDP

Phân tích GDP per capita (tháng 5/2024 vs tháng 5/2025):

GDP Decile Median GDP (USD) May 2024 May 2025 Tăng trưởng
1 (Thấp nhất) $1,200 2.1% 8.3% +296%
2 $2,800 3.2% 12.1% +278%
3 $4,500 4.1% 15.8% +285%
4 $6,200 5.3% 18.9% +257%
5 $8,100 6.8% 22.4% +229%
6 $10,500 8.2% 26.1% +218%
7 $13,800 9.1% 28.7% +215%
8 $18,200 10.3% 31.2% +203%
9 $25,600 11.8% 33.9% +187%
10 (Cao nhất) $45,200 13.2% 36.4% +176%

Phát hiện: Tăng trưởng cao nhất ở các nước thu nhập thấp-trung bình ($10,000-$40,000)


8. Phân Tích Theo Giáo Dục và Nghề Nghiệp

Giáo Dục

Tỷ lệ tin nhắn công việc theo học vấn:

  • Dưới cử nhân: 37%
  • Cử nhân: 46%
  • Sau đại học: 48%

Phân tích hồi quy (kiểm soát tuổi, giới tính, nghề nghiệp, cấp bậc, quy mô công ty, ngành):

  • Cử nhân vs Dưới cử nhân: +4.5 điểm phần trăm (p < 0.01)
  • Sau đại học vs Dưới cử nhân: +6.8 điểm phần trăm (p < 0.01)

Asking vs Doing theo học vấn:

  • Asking: Ít thay đổi theo học vấn (khoảng 49%)
  • Sau đại học: +2 điểm phần trăm Asking (p < 0.05)
  • Doing: Giảm theo học vấn
  • Sau đại học: -1.6 điểm phần trăm Doing (p < 0.10)

Nghề Nghiệp

Tỷ lệ tin nhắn công việc theo nghề:

Nghề Tỷ lệ công việc Đặc điểm chính
Computer/Math 57% Nhiều Technical Help (37%)
Management 50% Nhiều Writing (52%)
Business 50% Nhiều Writing (52%)
Engineering 48% Cân bằng Asking/Doing
Science 48% Cân bằng Asking/Doing
Other Professional 44% Đa dạng chủ đề
Non-professional 40% Ít sử dụng cho công việc

Asking vs Doing trong công việc:

  • Computer/Math: 47% Asking, 53% Doing
  • Engineering: 45% Asking, 55% Doing
  • Science: 44% Asking, 56% Doing
  • Management: 38% Asking, 62% Doing
  • Business: 35% Asking, 65% Doing
  • Non-professional: 32% Asking, 68% Doing

9. Chất Lượng Tương Tác

Xu Hướng Cải Thiện Theo Thời Gian

Tỷ lệ Good/Bad/Unknown:

  • Tháng 12/2024: Good 60%, Bad 20%, Unknown 20%
  • Tháng 7/2025: Good 80%, Bad 15%, Unknown 5%

Tỷ lệ Good/Bad:

  • Tháng 12/2024: 3:1
  • Tháng 7/2025: 5.3:1

Chất Lượng Theo Chủ Đề

Chủ đề Tỷ lệ Good/Bad Ghi chú
Self-Expression 7.0:1 Cao nhất
Practical Guidance 4.2:1 Cao
Writing 3.8:1 Trung bình cao
Seeking Information 3.5:1 Trung bình
Technical Help 2.7:1 Thấp
Multimedia 1.7:1 Thấp nhất

Chất Lượng Theo Mục Đích

Mục đích Tỷ lệ Good/Bad Ghi chú
Asking 4.5:1 Cao nhất
Doing 3.2:1 Trung bình
Expressing 2.8:1 Thấp nhất

Validation với User Feedback

Phân tích 60,000 tin nhắn có feedback trực tiếp:

  • Thumbs-up: 86% tổng feedback
  • Thumbs-down: 14% tổng feedback

Tương quan với Interaction Quality:

  • Thumbs-up + Good: 9.5 lần cao hơn Thumbs-down + Good
  • Thumbs-down: Tương đương Good và Bad
  • Unknown: Chia đều giữa thumbs-up và thumbs-down

10. Ý Nghĩa Kinh Tế và Xã Hội

Giá Trị Kinh Tế

Decision Support (Hỗ trợ Ra Quyết Định):

  • Đặc biệt quan trọng trong công việc tri thức
  • Giải thích tại sao Asking phổ biến hơn ở người có học vấn cao
  • Phù hợp với mô hình của Ide & Talamas (2025) về AI co-pilot

Consumer Surplus:

  • Collis & Brynjolfsson (2025): Ít nhất $97 tỷ/năm chỉ riêng Mỹ
  • Willingness-to-pay: $98 để từ bỏ sử dụng AI trong 1 tháng
  • Tác động ngoài công việc: Có thể lớn hơn tác động trong công việc

Đặc Điểm Độc Đáo của Generative AI

So với Web Search:

  • Khả năng tạo nội dung: Viết, code, spreadsheet, media
  • Tùy chỉnh cao: Phản hồi cá nhân hóa
  • Linh hoạt: Xử lý nhiều loại yêu cầu
  • Follow-up: Có thể tiếp tục cuộc hội thoại

Ví dụ cụ thể:

  • Web Search: “Boston Marathon qualifying times by age”
  • ChatGPT: “Tạo kế hoạch tập luyện cá nhân hóa cho marathon Boston dựa trên tuổi 35, kinh nghiệm 2 năm, mục tiêu 3:30”

Tác Động Xã Hội

Dân Chủ Hóa Tri Thức:

  • 10% dân số trưởng thành toàn cầu đã sử dụng
  • Tăng trưởng mạnh ở các nước thu nhập thấp-trung bình
  • Khoảng cách giới tính đã thu hẹp đáng kể

Giáo Dục:

  • 10.2% tin nhắn là yêu cầu dạy học
  • 36% Practical Guidance là tutoring/teaching
  • Hỗ trợ học tập suốt đời

11. Kết Luận và Triển Vọng

8 Phát Hiện Chính

  1. 70% tin nhắn không liên quan công việc (tăng từ 53%)
  2. 3 chủ đề chính chiếm 78% sử dụng: Practical Guidance, Writing, Seeking Information
  3. Writing chiếm 40% tin nhắn công việc, 2/3 là chỉnh sửa văn bản có sẵn
  4. Asking (49%) tăng nhanh hơn Doing (40%), chất lượng cao hơn
  5. Khoảng cách giới tính đã thu hẹp: 52% nữ giới hiện tại
  6. 46% tin nhắn từ người dùng 18-25 tuổi
  7. Tăng trưởng mạnh ở các nước thu nhập thấp-trung bình
  8. Người có học vấn cao sử dụng nhiều hơn cho công việc và Asking

Ý Nghĩa Kinh Tế

ChatGPT cung cấp giá trị kinh tế thông qua:

  • Decision Support: Hỗ trợ ra quyết định trong công việc tri thức
  • Consumer Surplus: Ít nhất $97 tỷ/năm chỉ riêng Mỹ
  • Tác động ngoài công việc: Có thể lớn hơn tác động trong công việc
  • Dân chủ hóa tri thức: 10% dân số trưởng thành toàn cầu

Triển Vọng Tương Lai

Với tốc độ tăng trưởng hiện tại:

  • ChatGPT sẽ tiếp tục định hình cách con người học tập, làm việc
  • AI sẽ trở thành công cụ không thể thiếu trong cuộc sống hàng ngày
  • Tác động xã hội sẽ ngày càng sâu sắc và rộng rãi

Thách thức:

  • Cần đảm bảo AI được sử dụng có trách nhiệm
  • Cân bằng giữa tự động hóa và việc làm con người
  • Giảm thiểu khoảng cách số và bất bình đẳng

Tài Liệu Tham Khảo

Nghiên cứu gốc: Aaron Chatterji (OpenAI, Duke University), Tom Cunningham (OpenAI), David Deming (Harvard University), Zoë Hitzig (OpenAI, Harvard University), Christopher Ong (OpenAI, Harvard University), Carl Shan (OpenAI), Kevin Wadman (OpenAI)

Tổ chức: OpenAI, Đại học Duke, Đại học Harvard

Nguồn chính: How People Use ChatGPT

Tài liệu tham khảo chính được sử dụng trong nghiên cứu:

Nghiên cứu kinh tế và AI:

  • Acemoglu, D. (2024). “The Simple Macroeconomics of AI.” NBER Working Paper 32487.
  • Autor, D. H., Levy, F., & Murnane, R. J. (2003). “The Skill Content of Recent Technological Change: An Empirical Exploration.” Quarterly Journal of Economics, 118(4), 1279-1333.
  • Bick, A., Blandin, A., & Deming, D. J. (2024). “The Rapid Adoption of Generative AI.” NBER Working Paper 32966.
  • Caplin, A., Deming, D. J., Leth-Petersen, S., & Weidmann, B. (2023). “Economic Decision-Making Skill Predicts Income in Two Countries.” NBER Working Paper 31674.
  • Carnehl, C., & Schneider, J. (2025). “A Quest for Knowledge.” Econometrica, 93(2), 623-659.
  • Collis, A., & Brynjolfsson, E. (2025). “AI’s Overlooked $97 Billion Contribution to the Economy.” Wall Street Journal.
  • Deming, D. J. (2021). “The Growing Importance of Decision-Making on the Job.” NBER Working Paper 28733.
  • Ide, E., & Talamas, E. (2025). “Artificial Intelligence in the Knowledge Economy.” Journal of Political Economy, 9(122).

Nghiên cứu về ChatGPT và LLM:

  • Handa, K., Tamkin, A., McCain, M., Huang, S., Durmus, E., Heck, S., Mueller, J., Hong, J., Ritchie, S., Belonax, T., Troy, K. K., Amodei, D., Kaplan, J., Clark, J., & Ganguli, D. (2025). “Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations.”
  • Tomlinson, K., Jaffe, S., Wang, W., Counts, S., & Suri, S. (2025). “Working with AI: Measuring the Occupational Implications of Generative AI.”
  • Zao-Sanders, M. (2025). “How People Are Really Using Gen AI in 2025.” Harvard Business Review.
  • Zhao, W., Ren, X., Hessel, J., Cardie, C., Choi, Y., & Deng, Y. (2024). “WildChat: 1M ChatGPT Interaction Logs in the Wild.”

Nghiên cứu về tác động xã hội:

  • Humlum, A., & Vestergaard, E. (2025a). “Large Language Models, Small Labor Market Effects.” University of Chicago Working Paper 2025-56.
  • Humlum, A., & Vestergaard, E. (2025b). “The Unequal Adoption of ChatGPT Exacerbates Existing Inequalities among Workers.” Proceedings of the National Academy of Sciences, 122(1), e2414972121.
  • Ling, Y., & Imas, A. (2025). “Underreporting of AI use: The role of social desirability bias.” SSRN Working Paper.

Nghiên cứu kỹ thuật và phương pháp:

  • Bengio, Y., Courville, A., & Vincent, P. (2014). “Representation Learning: A Review and New Perspectives.”
  • Chiang, W.-L., Zheng, L., Sheng, Y., Angelopoulos, A. N., Li, T., Li, D., Zhu, B., Zhang, H., Jordan, M. I., Gonzalez, J. E., & Stoica, I. (2024). “Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference.” Proceedings of ICML 2024.
  • Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., & Steinhardt, J. (2021). “Measuring Massive Multitask Language Understanding.” Proceedings of ICLR 2021.
  • Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P., Leike, J., & Lowe, R. (2022). “Training Language Models to Follow Instructions with Human Feedback.”
  • Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017). “Attention Is All You Need.” Advances in Neural Information Processing Systems.

Nghiên cứu về tổ chức và lao động:

  • Garicano, L. (2000). “Hierarchies and the Organization of Knowledge in Production.” Journal of Political Economy, 108(5), 874-904.
  • Garicano, L., & Rossi-Hansberg, E. (2006). “Organization and Inequality in a Knowledge Economy.” Quarterly Journal of Economics, 121(4), 1383-1435.
  • National Association of Colleges and Employers. (2024). “Competencies for a Career-Ready Workforce.”

Nghiên cứu về bình đẳng giới:

  • Hofstra, B., Kulkarni, V. V., Munoz-Najar Galvez, S., He, B., Jurafsky, D., & McFarland, D. A. (2020). “The Diversity–Innovation Paradox in Science.” Proceedings of the National Academy of Sciences, 117(17), 9284-9291.
  • West, J. D., Jacquet, J., King, M. M., Correll, S. J., & Bergstrom, C. T. (2013). “The Role of Gender in Scholarly Authorship.” PLoS ONE, 8(7), e66212.

Nguồn tin tức và báo cáo:

  • Pew Research Center. (2025). “U.S. adults’ use of ChatGPT (June 2025 report).”
  • Reuters. (2025). “OpenAI hits $12 billion in annualized revenue, The Information reports.”
  • Roth, E. (2025). “OpenAI says ChatGPT users send over 2.5 billion prompts every day.”
  • Wiggers, K. (2025). “ChatGPT Isn’t the Only Chatbot That’s Gaining Users.” TechCrunch.

Tài liệu kỹ thuật OpenAI:

  • OpenAI. (2023). “GPT-4 Technical Report.” arXiv preprint.
  • OpenAI. (2024a). “GPT-4o System Card.”
  • OpenAI. (2024b). “OpenAI o1 System Card.” System Card / Technical Report.
  • OpenAI. (2025a). “Expanding on What We Missed with Sycophancy.” Blog Post / Technical Report.
  • OpenAI. (2025b). “GPT-5 System Card.” System Card / Technical Report.
  • OpenAI. (2025c). “Privacy Policy.”

Nghiên cứu về tác động cảm xúc:

  • Phang, J., Lampe, M., Ahmad, L., Agarwal, S., Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., & Maes, P. (2025). “Investigating Affective Use and Emotional Well-being on ChatGPT.”

Nghiên cứu về công bằng:

  • Eloundou, T., Beutel, A., Robinson, D. G., Gu, K., Brakman, A.-L., Mishkin, P., Shah, M., Heidecke, J., Weng, L., & Kalai, A. T. (2025). “First-Person Fairness in Chatbots.” Proceedings of ICLR 2024.

Nghiên cứu về rủi ro AI:

  • Korinek, A., & Suh, D. (2024). “Scenarios for the Transition to AI.” NBER Working Paper 32255.
  • Kulveit, J., Douglas, R., Ammann, N., Turan, D., Krueger, D., & Duvenaud, D. (2025). “Gradual Disempowerment: Systemic Existential Risks from Incremental AI Development.”

Nghiên cứu về tác động lao động:

  • Hartley, J., Jolevski, F., Melo, V., & Moore, B. (2025). “The Labor Market Effects of Generative Artificial Intelligence.” SSRN Working Paper.

Nghiên cứu về dữ liệu xã hội:

  • Chetty, R., Jackson, M. O., Kuchler, T., Stroebel, J., Hendren, N., Fluegge, R. B., Gong, S., Gonzalez, F., Grondin, A., Jacob, M., Johnston, D., Koenen, M., Laguna-Muggenburg, E., Mudekereza, F., Rutter, T., Thor, N., Townsend, W., Zhang, R., Bailey, M., Barberá, P., Bhole, M., & Wernerfelt, N. (2022). “Social Capital I: Measurement and Associations with Economic Mobility.” Nature, 608(7923), 108-121.

Nghiên cứu kỹ thuật bổ sung:

  • Lambert, N., Morrison, J., Pyatkin, V., Huang, S., Ivison, H., Brahman, F., Miranda, L. J. V., Liu, A., Dziri, N., Lyu, S., et al. (2024). “Tulu 3: Pushing frontiers in open language model post-training.” arXiv preprint.
  • Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2023). “Lost in the Middle: How Language Models Use Long Contexts.”

Bài viết này cung cấp tóm tắt toàn diện về nghiên cứu “How People Use ChatGPT” – một trong những nghiên cứu quan trọng nhất về việc sử dụng AI trong thực tế. Nghiên cứu không chỉ cung cấp dữ liệu quan trọng về hiện tại mà còn mở ra những câu hỏi thú vị về tương lai của AI trong cuộc sống con người.

Trying the Realtime Prompting Guide for GPT-Realtime: Experiments with Vietnamese Voice Input

Try Realtime Prompting Guide for GPT-Realtime

1.Introduction

OpenAI’s Realtime API enables the creation of interactive voice experiences with ultra-low latency. Instead of waiting for a full text input, the model can “listen” to a user while they are still speaking and respond almost instantly. This makes it a powerful foundation for building voice assistants, audio chatbots, automated customer support, or multimodal creative applications.

To get the best results, writing a clear and well-structured prompt is essential. OpenAI published the Realtime Prompting Guide as a playbook for controlling model behavior in spoken conversations.

References:


2.What is GPT-Realtime

GPT-Realtime is a model/API designed to handle continuous audio input and provide rapid responses. Its key features include:

  • Real-time speech-to-text recognition.

  • Robust handling of noisy, cut-off, or unclear audio.

  • Customizable reactions to imperfect audio, such as asking for repetition, clarifying, or continuing in the user’s language.

  • Support for detailed prompting to ensure safe, natural, and reliable responses.


3.Overview of the Prompting Guide

The Realtime Prompting Guide outlines seven best practices for writing system prompts for voice agents:

1. Be precise, avoid conflicts.
Instructions must be specific and consistent. For example, if you say “ask again when unclear,” don’t also instruct the model to “guess when unsure.”

2. Use bullet points instead of paragraphs.
Models handle lists better than long prose.

3. Handle unclear audio.
Explicitly instruct what to do when input is noisy or incomplete: politely ask the user to repeat and only respond when confident.

4. Pin the language when needed.
If you want the entire conversation in one language (e.g., English only), state it clearly. Otherwise, the model may switch to mirror the user.

5. Provide sample phrases.
Include example greetings, clarifications, or closing lines to teach the model your desired style.

6. Avoid robotic repetition.
Encourage varied phrasing for greetings, confirmations, and closings to keep interactions natural.

7. Use capitalization for emphasis.
For example: “IF AUDIO IS UNCLEAR, ASK THE USER TO REPEAT.”

4.Prompt Examples

Sample Prompt A – Avoid conflicts, be clear

SYSTEM: – Always speak clearly and respond in the same language the user is speaking. – If the user’s audio is unclear, noisy, partial, or unintelligible, politely ask them to repeat. – Never assume what was said if you did not hear it clearly. – Use short, varied phrases; avoid repeating the same sentence. – Provide helpful, concise responses.

USER: “Hello, can you help me with my internet issue?”

ASSISTANT: (responds according to the prompt, asks for clarification if needed)


Sample Prompt B – Handling unclear audio

SYSTEM: – If the audio is not clear, ask: “I’m sorry, I didn’t catch that. Could you please repeat?” – Only respond when you are confident you understood the user’s request. – Continue conversation in the same language as the user. – Use bullet points for clarity.

USER: “Um… internet…” (noisy, unclear audio)

ASSISTANT: (follows the system instructions)


Sample Prompt C – Keep a natural style, avoid repetition

SYSTEM: – Do not respond with the same phrase twice. – Vary greetings, confirmations, closings. – Aim for a warm, helpful tone. – If the user ends the conversation, say something like: “If there’s anything else you need, feel free to ask.”

USER: “Thank you, that’s all.”

ASSISTANT: “You’re welcome! Glad I could help. Take care!” (or another variation)

5.Experiments

For my testing, I deliberately used Vietnamese speech to see how the model would react in different situations.

First Test: Speaking Unclear Without a Prompt

To begin, I tested what would happen if I spoke unclearly in Vietnamese without providing any system prompt.

For example, I said:

USER: “Ư… mệnh của mình không tỏa lại” (unclear Vietnamese speech)

The model responded in Indonesian, saying:

ASSISTANT: “Tentu, aku di sini untuk membantu. Coba ceritakan dulu apa yang lagi kamu rasakan atau alami. Kita bisa cari solusinya bareng-bareng.”

This shows that when no system prompt is defined, the model may guess or switch languages unpredictably, instead of asking for clarification.

Second Test: Adding a System Prompt

Next, I added a system prompt to guide the model’s behavior when the audio is unclear:

SYSTEM:
- If the audio is not clear, ask: “I’m sorry, I didn’t catch that. Could you please repeat?”
- Only respond when you are confident you understood the user’s request.
- Continue conversation in the same language as user.
- Use bullet points for clarity.

Then I spoke unclearly in Vietnamese again, for example:

USER: “Um… internet…” (spoken quietly, unclear audio)

This time, the model followed the system instructions and politely asked me to repeat. Sometimes, it even suggested that I try saying a simple test sentence so it could better check whether my voice was coming through clearly.

This shows how a well-written system prompt can prevent the model from making random guesses or switching languages, ensuring a more reliable and natural conversation flow.

Third Test: Singing to the Model

Finally, I experimented by singing in Vietnamese to see how the model would react.

The model listened carefully and was able to understand the lyrics and emotional tone of my singing. However, when I asked it to repeat the lyrics back to me, it refused due to copyright restrictions.

This shows that while GPT-Realtime can analyze and comment on songs — such as summarizing their meaning, describing the mood, or suggesting new lines in a similar style — it cannot reproduce lyrics verbatim. In practice, this means you should not expect the model to sing or echo back copyrighted content.

6.Conclusion

GPT-Realtime provides smooth and natural voice interactions with minimal latency. However, its effectiveness depends heavily on the prompt.

Key takeaways:

  • Always write a clear, bullet-pointed system prompt.
  • Define explicit behavior for unclear audio.
  • Control language use and discourage robotic repetition.
  • Respect copyright limitations: the model will not repeat lyrics verbatim but can summarize or create new content.

The Realtime Prompting Guide is a practical resource for building high-quality voice agents that are both natural and safe.