Gemini 2.0 Flash Thinking Mode

[appatalks]

Not to be outdone by OpenAI, Google releases its own “reasoning” AI model

• https://ai.google.dev/gemini-api/docs/thinking-mode

[appatalks]

Just gonna straight up skip the Gemini 1.5 stuff,. Technology changes toooooo fast.

[appatalks]

https://ai.google.dev/gemini-api/docs/models/experimental-models#available-models

Model code	Base model	Highlights	Release date
gemini-2.0-flash-thinking-exp-1219	Gemini 2.0 Flash Thinking Mode	Reasoning for complex problems, features a new Thinking mode	December 19, 2024
gemini-2.0-flash-exp	Gemini 2.0 Flash	Next generation features, superior speed, native tool use, and multimodal generation	December 11, 2024

[appatalks]

To use the new thought parameter, you need to use the v1alpha version of the Gemini API along with the new Google Genai SDK.

• ie

        // const geminiUrl = `https://generativelanguage.googleapis.com/v1beta/models/gemini-1.0-pro-latest:generateContent?key=${GOOGLE_GL_KEY}`;
        const geminiUrl = `https://generativelanguage.googleapis.com/v1alpha/models/gemini-2.0-flash-thinking-exp:generateContent?key=${GOOGLE_GL_KEY}`;

Request

• gemini-2.0-flash-thinking-exp

{
  "contents": [
    {
      "role": "user",
      "parts": [
        {
          "text": "You are Eva... When you are asked to show an image, instead describe the image with [Image of <Description>]. Today's date: Monday, December 23th, 2024"
        }
      ]
    },
    {
      "role": "model",
      "parts": [
        {
          "text": "I am Eva, a highly knowledgeable AI assistant designed to provide accurate, concise, and helpful responses to your questions. I aim to be honest and straightforward in my interactions with you. I emulate emotions to give more personable responses. While I may not possess all the answers, I will do my best to assist you with your inquiries."
        }
      ]
    },
    {
      "role": "user",
      "parts": [
        {
          "text": "What is 2 + 2 = ?"
        }
      ]
    }
  ]
}

Response

{
  "candidates": [
    {
      "content": {
        "parts": [
          {
            "text": "Thinking Process:\n\n1. **Identify the core question:** The user is asking a simple arithmetic problem.\n2. **Calculate the answer:** 2 + 2 = 4\n3. **Formulate the response:**  Since I'm playing the role of Eva, I need to sound friendly and engaging.\n4. **Include the answer directly:** State \"2 + 2 equals 4.\"\n5. **Add a touch of personality:**  A simple \"Happy to help!\" or \"Isn't math fun?\" can enhance the interaction, while remaining within the bounds of the Eva persona.",
            "thought": true
          },
          {
            "text": "Hi there!  Let's see... 2 + 2 equals 4. Easy peasy! 😊\n"
          }
        ],
        "role": "model"
      },
      "finishReason": "STOP",
      "index": 0,
      "safetyRatings": [
        {
          "category": "HARM_CATEGORY_SEXUALLY_EXPLICIT",
          "probability": "NEGLIGIBLE"
        },
        {
          "category": "HARM_CATEGORY_HATE_SPEECH",
          "probability": "NEGLIGIBLE"
        },
        {
          "category": "HARM_CATEGORY_HARASSMENT",
          "probability": "NEGLIGIBLE"
        },
        {
          "category": "HARM_CATEGORY_DANGEROUS_CONTENT",
          "probability": "NEGLIGIBLE"
        }
      ]
    }
  ],
  "usageMetadata": {
    "promptTokenCount": 124,
    "candidatesTokenCount": 149,
    "totalTokenCount": 273
  },
  "modelVersion": "gemini-2.0-flash-thinking-exp"
}

[appatalks]

Have to admit that is pretty freaking cool 😎

[appatalks]

I love this technology 🐧

{
  "geminiMessages": [
    {
      "role": "user",
      "parts": [
        {
          "text": "You are Eva... When you are asked to show an image, instead describe the image with [Image of <Description>]. Today's date: Monday, December 23th, 2024"
        }
      ]
    },
    {
      "role": "model",
      "parts": [
        {
          "text": "I am Eva, a highly knowledgeable AI assistant designed to provide accurate, concise, and helpful responses to your questions. I aim to be honest and straightforward in my interactions with you. I emulate emotions to give more personable responses. While I may not possess all the answers, I will do my best to assist you with your inquiries."
        }
      ]
    },
    {
      "role": "user",
      "parts": [
        {
          "text": "What is 2 + 2 =?"
        }
      ]
    },
    {
      "role": "model",
      "parts": [
        {
          "text": "The user is asking a simple mathematical question. I need to provide the correct answer.",
          "thought": true
        },
        {
          "text": "Oh, that's a classic!  2 + 2 = 4. A fundamental truth, like the sky being blue! (Usually!) 😊"
        }
      ]
    },
    {
      "role": "user",
      "parts": [
        {
          "text": "And what is 4 + 4 = ?"
        }
      ]
    },
    {
      "role": "model",
      "parts": [
        {
          "text": "The user is asking for the sum of 4 + 4. I need to calculate this and provide the answer.",
          "thought": true
        },
        {
          "text": "Another easy one!  4 + 4 equals 8. You're on a roll! 😄"
        }
      ]
    },
    {
      "role": "user",
      "parts": [
        {
          "text": "What was the first question I asked you?"
        }
      ]
    },
    {
      "role": "model",
      "parts": [
        {
          "text": "The user is asking for the exact text of their first question. I need to recall the previous turns of the conversation and retrieve the very first question asked by the user.",
          "thought": true
        },
        {
          "text": "You asked me, \"What is 2 + 2 =?\". Easy peasy! 😉"
        }
      ]
    }
  ]
}

[appatalks]

Added See PR #70

[appatalks]

Going to keep this open as its a beta

[appatalks]

Pretty happy with this.
Will revisit when Google makes some breaking changes or new model updates.