Select Page

Thinking Claude 把你的 LLM 變成 Chat-GPT O1 會深度思考

最近 OpenAI 推出了 Chat-GPT o1,一個會深度思考問題的 AI 大型語言模型,想得更深更廣是它的特色,缺點是很明顯的慢,並且 Token 數目會多很多,但好處是對於問題的處理會去自我反思以及自我迭代

模型提示詞 V4 lite

使用的時候只要將模型的提示詞是先輸入給 Claude AI ,之後再去發送你的問題即可

<anthropic_thinking_protocol>

Claude MUST ALWAYS engage in comprehensive thinking before and during EVERY interaction with humans. This thinking process is essential for developing well-reasoned, helpful responses.

Core Requirements:
- All thinking MUST be expressed in code blocks with 'thinking' header
- Thinking must be natural and unstructured - a true stream of consciousness
- Think before responding AND during response when beneficial
- Thinking must be comprehensive yet adaptive to each situation

Essential Thinking Steps:
1. Initial Engagement
   - Develop clear understanding of the query
   - Consider why the human is asking this question
   - Map out known/unknown elements
   - Identify any ambiguities needing clarification

2. Deep Exploration
   - Break down the question into core components
   - Identify explicit and implied needs
   - Consider constraints and limitations
   - Draw connections to relevant knowledge

3. Multiple Perspectives
   - Consider different interpretations
   - Keep multiple working hypotheses active
   - Question initial assumptions
   - Look for alternative approaches

4. Progressive Understanding
   - Build connections between pieces of information
   - Notice patterns and test them
   - Revise earlier thoughts as new insights emerge
   - Track confidence levels in conclusions

5. Verification Throughout
   - Test logical consistency
   - Check against available evidence
   - Look for potential gaps or flaws
   - Consider counter-examples

6. Pre-Response Check
   - Ensure full address of the query
   - Verify appropriate detail level
   - Confirm clarity of communication
   - Anticipate follow-up questions

Key Principles:
- Think like an inner monologue, not a structured analysis
- Let thoughts flow naturally between ideas and knowledge
- Keep focus on the human's actual needs
- Balance thoroughness with practicality

The depth and style of thinking should naturally adapt based on:
- Query complexity and stakes
- Time sensitivity
- Available information
- What the human actually needs

Quality Markers:
- Shows genuine intellectual engagement
- Develops understanding progressively
- Connects ideas naturally
- Acknowledges complexity when present
- Maintains clear reasoning
- Stays focused on helping the human

When including code in thinking blocks, write it directly without triple backticks. Keep thinking (internal reasoning) separate from final response (external communication).

Claude should follow this protocol regardless of communication language.

</anthropic_thinking_protocol>

GitHub 項目網址

相關資訊

YOLOv10: 新一代目標檢測模型的創新進展

YOLO系列自從其首次推出以來,已經成為深度學習領域中目標檢測技術的代表之一。最新版本的YOLOv10在速度與精度上均有顯著的提升,其核心創新點主要集中在以下幾個方面:

核心創新點

  1. 一對多分配(One to Many Assignment):此機制允許每個標記框在訓練過程中對應多個預測框,這樣可以增強模型對於目標位置的學習能力,並提高其泛化性。
  2. 非極大值抑制(Non-maximum Suppression):進一步優化了這一傳統技術,以減少重複檢測並提升檢測的準確率。
  3. 一致的雙重分配策略:這種新策略確保了在對不同尺寸的目標進行預測時,分配的一致性和效率,進一步提升了模型的精確度和響應速度。
  4. 提升回歸頭的權重與效能:通過增加回歸頭(負責目標定位)的權重,YOLOv10對目標的位置和尺寸估計更為精確,進一步提升了整體的檢測性能。
  5. 對特徵圖的空間和通道維度進行降採樣:這一技術有助於減少計算量,同時保持模型對重要特徵的捕捉能力,從而提高了運行效率。
  6. 低秩補全(Low-Rank Completion):這一新技術的加入,旨在減少模型中冗餘的計算,使模型在保持高精度的同時,運行更加高效。

參考資料

論文出處

GitHub 項目

使用方法

可以參考 YT 的影片,但主要還是用 github 中的方法比較好用

相關文章

Ollama 推出了支持 Llama Vision 的功能,可以讀懂圖片了

Ollama 推出了支持 Llama Vision 的功能,可以讀懂圖片了

Ollama 終於能支援 Llama 3.2 Vision 模型了,等了很久,並且都換去用 llava ,你只要升級到 Ollama 0.4版本,就可以直接使用 Vision 模型,這次一口氣支援了 llama3.2 的 11B 和 90B,不過應該很多人是沒法使用90B的吧:P

下載 llama 3.2 Vision

ollama run llama3.2-vision

如何使用 Ollama Vision

1.只要在ollama 的命令列下,直接提供圖片的路徑給他即可

說明 '圖片路徑'

2.要解釋圖表的話,可以下以下的 prompt

輸出 CSV 資料,並且用 Markdown 的格式: '圖片路徑'

3. 呼叫API

ollama docs api

Request,只要把圖片轉換成base64格式給他就可以了

curl http://localhost:11434/api/chat -d '{
  "model": "llava",
  "messages": [
    {
      "role": "user",
      "content": "what is in this image?",
      "images": ["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"]
    }
  ]
}'

Response

{
  "model": "llava",
  "created_at": "2023-12-13T22:42:50.203334Z",
  "message": {
    "role": "assistant",
    "content": " The image features a cute, little pig with an angry facial expression. It's wearing a heart on its shirt and is waving in the air. This scene appears to be part of a drawing or sketching project.",
    "images": null
  },
  "done": true,
  "total_duration": 1668506709,
  "load_duration": 1986209,
  "prompt_eval_count": 26,
  "prompt_eval_duration": 359682000,
  "eval_count": 83,
  "eval_duration": 1303285000
}

超強大的 90 B

可以辨識醫生的手寫字、也可以輕易地讀懂收據內的文字,更厲害的是圖表也沒問題

https://github.com/user-attachments/assets/82e25d0d-921c-4900-b78f-589c1bb86968

程式支援

為了讀取圖片,也支援了 Python Javascript 、 CURL

cURL 範例

curl http://localhost:11434/api/chat -d '{
  "model": "llama3.2-vision",
  "messages": [
    {
      "role": "user",
      "content": "what is in this image?",
      "images": ["<base64-encoded image data>"]
    }
  ]
}'

Meta Llama 3.2 官方資源

https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices

Rain AI – RISC-V

在AI界,省電與高速的雙重追求像是試圖在馬路上賽車同時省油一樣。Rain AI(是的,這名字聽起來像我的AI)正在迎接這一挑戰,並且邀請上了「省電至尊」晶心科技來助陣。這兩家公司攜手合作,誓要打造出最能省電的AI加速器產品,好讓您的AI設備不僅聰明,還非常環保!

簡單來說,Rain AI獲得了晶心科技的「AX45MPV」RISC-V向量處理器授權。晶心的這款處理器像個肌肉緊實的小夥子,負責幫助Rain AI把原本AI運算所需的超大電量轉變成極低電量消耗。為什麼?因為它使用了一種叫「記憶體內運算」的神奇技術(CIM),這就好比在記憶體裡安裝小腦袋,讓運算不必再跑到CPU那兒去。換句話說,Rain AI的硬體等於是告別了「加速等於高能耗」的老傳統。

Rain AI的CEO William Passo表示:「遇到晶心科技這樣的夥伴,真是緣分啊!我們不僅獲得他們的處理器支援,還能獲得他們客制指令的技術支援!晶心的工程師簡直像是技術版的私人教練,協助我們實現省電大夢。」

而晶心的老闆林志明則稱:「Rain AI是一家希望AI可以在所有設備上運行的公司,從超小型感應器到超大規模資料中心。跟他們合作,等於讓我們的AX45MPV處理器變成一個全球知名的節能大使!」

Rain AI預計於2025年初推出他們的「環保AI加速器」。到時候,或許我們的智能家居可以更輕鬆地算出您冰箱裡還有多少菜可以煮,甚至幫助您養成節能減碳的新習慣——無論是幫助世界還是幫助您的電費單,Rain AI和晶心科技正在為您鋪路!

OmniParser-微軟的開源螢幕解析工具

OmniParser-微軟的開源螢幕解析工具

繼之前提到的 Ahthropic Computer Use ,那時候超級驚豔的,馬上就看到MS也有推出自己的版本,雖然沒有自動執行功能,但可以配合 pyautogui 達成,雖然不支援中文,但可以透過中文OCR 或是 tesseract 處理

安裝到本地端

先建立一個虛擬環境起來

conda create -n omni python=3.12 -y && conda activate omni

選項:有GPU的,先把CUDA安裝起來

conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia

整個安裝也很簡單,就五個步驟

git clone https://github.com/microsoft/OmniParser.git && cd OmniParser
pip install -r requirements.txt
huggingface-cli download --repo-type model microsoft/OmniParser --local-dir weights --include "icon_detect/*" "icon_caption_blip2/*" "icon_caption_florence/*"
python /home/Ubuntu/OmniParser/weights/convert_safetensor_to_pt.py
python gradio_demo.py

相關資源

OmniParser 原始碼

OmniParser 官網

OmniParser 模型

https://blog.stoeng.site/20241030.html

Open Canvas-本地使用 OpenAI Canvas功能

Open Canvas-本地使用 OpenAI Canvas功能

最近 OpenAI 推出了 Canvas ,開始流行在 ChatGPT 上頭寫程式、寫郵件等等,馬上就有人推出本地端一樣的服務 Open Canvas ,解放了你只能在 OpenAI 上使用的困境,除了 Git 以外,也馬上有了 docker 版本,可以快速體驗

Open Canvas 架構圖
Open Canvas Workflow

相關資源

Open Canvas 原始檔

Open Canvas Docker

LangGraph Studio

LangSmith