v1beta（公式フォーマット - ストリーミング）

data: {"candidates":[{"content":{"parts":[{"text":"AI, or Artificial Intelligence, fundamentally works by creating systems that can simulate human intelligence. Instead of explicit, step-by-step instructions for every possible scenario (as in traditional programming), AI systems are designed to **learn, adapt, and make"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":49,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1143}}

data: {"candidates":[{"content":{"parts":[{"text":" decisions** or predictions based on patterns in data.\n\nHere's a breakdown of the core principles and common approaches:\n\n### The Core Idea: Learning from Data\n\nAt its heart, most modern AI (especially Machine Learning) operates by learning from"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":100,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1194}}

data: {"candidates":[{"content":{"parts":[{"text":" large amounts of data. Think of it like a student:\n\n1.  **Input Data (the \"Textbooks\"):** The AI is fed vast quantities of relevant data – images, text, numbers, sounds, sensor readings, etc."}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":149,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1243}}

data: {"candidates":[{"content":{"parts":[{"text":" This data usually contains examples of the problem it's trying to solve.\n2.  **Algorithms (the \"Learning Rules\"):** These are the mathematical models and procedures that define *how* the AI learns from the data. Different"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":197,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1291}}

data: {"candidates":[{"content":{"parts":[{"text":" algorithms are suited for different types of problems (e.g., predicting a number vs. classifying an image).\n3.  **Training (the \"Studying\"):** The algorithm processes the input data to identify patterns, relationships, and features"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":246,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1340}}

data: {"candidates":[{"content":{"parts":[{"text":" within it. During training, the model's internal parameters are adjusted repeatedly until it can accurately perform its task. If it's a \"supervised learning\" task, the data comes with \"correct answers\" (labels), and the AI tries"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":295,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1389}}

data: {"candidates":[{"content":{"parts":[{"text":" to predict those answers, adjusting itself when it's wrong.\n4.  **Model (the \"Learned Knowledge\"):** The output of the training process is a \"model.\" This model encapsulates the knowledge the AI has gained about"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":343,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1437}}

data: {"candidates":[{"content":{"parts":[{"text":" the data's patterns.\n5.  **Inference/Prediction (the \"Exam\"):** Once trained, the model can be presented with *new, unseen* data. Using the knowledge it gained during training, it can then"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":391,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1485}}

data: {"candidates":[{"content":{"parts":[{"text":" make predictions, classify items, generate new content, or make decisions without being explicitly programmed for that specific new input.\n\n### Key Components of How AI Works:\n\n1.  **Data:**\n    *   **Volume:** Modern AI often requires"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":441,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1535}}

data: {"candidates":[{"content":{"parts":[{"text":" massive amounts of data.\n    *   **Quality:** The data must be accurate, relevant, and representative to avoid biases and ensure the AI learns correctly.\n    *   **Variety:** Different types of data (text, images, audio"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":490,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1584}}

data: {"candidates":[{"content":{"parts":[{"text":", video) lead to different AI specializations.\n\n2.  **Algorithms:** These are the \"recipes\" for learning and problem-solving. Some popular categories include:\n    *   **Machine Learning (ML):** The most"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":538,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1632}}

data: {"candidates":[{"content":{"parts":[{"text":" common approach to AI today.\n        *   **Supervised Learning:** The model learns from labeled data (input-output pairs). E.g., given images of cats and dogs, labeled as such, it learns to distinguish them.\n        *"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":589,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1683}}

data: {"candidates":[{"content":{"parts":[{"text":"   **Unsupervised Learning:** The model finds patterns or structures in unlabeled data. E.g., grouping customers into segments without prior definitions.\n        *   **Reinforcement Learning:** An \"agent\" learns by trial and error in an"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":636,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1730}}

data: {"candidates":[{"content":{"parts":[{"text":" environment, receiving rewards for good actions and penalties for bad ones. E.g., training an AI to play chess or control a robot.\n    *   **Deep Learning (a subset of ML):** Uses artificial neural networks with many layers (hence"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":687,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1781}}

data: {"candidates":[{"content":{"parts":[{"text":" \"deep\"). Inspired by the structure of the human brain, these networks are excellent at recognizing complex patterns in raw data (like images, speech, and text) without extensive manual feature engineering.\n\n3.  **Computational Power:**\n    "}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":733,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1827}}

data: {"candidates":[{"content":{"parts":[{"text":"*   Training complex AI models (especially deep learning models) requires enormous computational resources, often utilizing specialized hardware like GPUs (Graphics Processing Units) and TPUs (Tensor Processing Units).\n\n### A Simplified Example: Image Recognition\n\nLet's say you want"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":784,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1878}}

data: {"candidates":[{"content":{"parts":[{"text":" to teach an AI to recognize a \"cat\":\n\n1.  **Data Collection:** Gather millions of images, some containing cats, some not. Each cat image is **labeled** as \"cat.\"\n2.  **Algorithm (e"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":833,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1927}}

data: {"candidates":[{"content":{"parts":[{"text":".g., Deep Neural Network):** You choose an appropriate deep learning architecture.\n3.  **Training:**\n    *   The network is fed one image at a time.\n    *   Initially, it makes random guesses about"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":881,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":1975}}

data: {"candidates":[{"content":{"parts":[{"text":" whether there's a cat.\n    *   If it guesses \"cat\" for a \"dog\" image (or vice-versa), it compares its guess to the **label** and realizes its mistake.\n    *   The network'"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":931,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":2025}}

data: {"candidates":[{"content":{"parts":[{"text":"s internal connections (weights) are slightly adjusted to reduce the chance of making that same mistake in the future.\n    *   This process repeats for millions of images and many cycles (epochs).\n4.  **Model Formation:** Eventually"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":979,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":2073}}

data: {"candidates":[{"content":{"parts":[{"text":", the network \"learns\" to identify the features and patterns associated with cats (like whiskers, pointy ears, specific fur textures, etc.) and dogs (different features).\n5.  **Inference:** Now, you show the"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":1027,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":2121}}

data: {"candidates":[{"content":{"parts":[{"text":" trained model a *new image it has never seen before*. It processes the image through its learned network and predicts with a certain probability whether it contains a cat or not.\n\n### How AI Is NOT Working (Yet):\n\nIt's important to"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":1077,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":2171}}

data: {"candidates":[{"content":{"parts":[{"text":" note that most AI systems today are **Narrow AI** or **Weak AI**. They are exceptionally good at *one specific task* (e.g., playing chess, identifying objects, generating text). They don't possess general human-like intelligence,"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":1128,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":2222}}

data: {"candidates":[{"content":{"parts":[{"text":" consciousness, self-awareness, or common sense. That's the realm of **Artificial General Intelligence (AGI)**, which is still a major research goal, not a current reality.\n\nIn essence, AI works by using algorithms to"}],"role":"model"},"index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":1176,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":2270}}

data: {"candidates":[{"content":{"parts":[{"text":" find patterns and make predictions from vast datasets, enabling systems to perform tasks that typically require human intelligence, but usually within a defined domain."}],"role":"model"},"finishReason":"STOP","index":0}],"modelVersion":"gemini-2.5-flash","responseId":"LMgeaeujMp2dvdIPm8_G6Qc","usageMetadata":{"candidatesTokenCount":1203,"promptTokenCount":6,"promptTokensDetails":[{"modality":"TEXT","tokenCount":6}],"thoughtsTokenCount":1088,"totalTokenCount":2297}}