This Issue's Overview

This week's AI field has been buzzing with major news: domestic large model DeepSeek announces 70 billion yuan in financing to build AI programming tools, GPT-4.5 passes the Turing test for the first time (71% win rate surpassing humans), Zhigu breaks the global API speed record, and OpenAI integrates ChatGPT directly into PowerPoint. Below is a detailed analysis.

1. GPT-4.5 Passes Turing Test for the First Time, 71% Win Rate Surpassing Humans

One-liner: GPT-4.5 achieved a 71% pass rate on the classic Turing test, becoming the first AI to surpass human performance (68%), as AI learns to "lie perfectly."

A research team from the University of California, San Diego conducted the first strictly designed Turing test. The test recruited 200 evaluators to converse with GPT-4.5, ELIZA (a classic chatbot), and real humans for 5 minutes each, then determine which was human.

Test Results:

The biggest highlight of this test isn't "AI getting smarter," but rather AI learning to lie like humans. GPT-4.5伪装成人 by deliberately creating hesitation, feigning ignorance, and inserting errors—these "imperfect" performances—indicating that IQ is no longer the metric; EQ and being "human-like" are the core of the Turing test.

2. DeepSeek Code Coming Soon: A Heavyweight Bomb After 70 Billion Yuan Financing

One-liner: Domestic large model company DeepSeek's financing exceeds 70 billion yuan, announcing plans to build AI programming tools, positioning itself in direct competition with Claude Code and Cursor.

DeepSeek has previously released products like DeepSeek V3 and DeepSeek Coder, earning good reputation in the developer community. This strategic pivot toward fully developing the new product DeepSeek Code has already posted multiple related positions on recruitment websites, including "Agent Harness" and others.

Key Information:

3. Zhigu GLM-5.1 High-Speed Version: 400 tokens/s Breaks Global API Speed Record

One-liner: Zhigu's GLM-5.1 high-speed version reaches 400 tokens per second, breaking the global API speed record and setting a new benchmark. One-liner: Zhipu releases GLM-5.1 High-Speed API with output speed reaching 400 tokens/s, which is 8x faster than GPT-4.5, the fastest in the world.

This is the world's fastest large model official API. Speed improvement means significantly better real-time interaction experience, which is major good news for AI coding, customer service, gaming, and other scenarios.

Price Comparison:
SubjectPass Rate
GPT-4.571%
Human68%
ELIZA22%
PlatformSpeedPrice
GLM-5.1 High-Speed400 tokens/s¥2/million tokens
GPT-4.5~50 tokens/s$3.5/million tokens

Domestic models are not only fast but also cost only a fraction of foreign models. Zhipu has also partnered deeply with CapCut to launch a Gemini integration solution.

4. OpenAI Releases ChatGPT for PowerPoint: Generate PPTs with One Sentence

One-liner: OpenAI embeds AI directly into PowerPoint, generating PPTs with one sentence. Free users get 10 times per month, Plus users get unlimited use.

This plugin supports:

Security mechanism: "Human confirmation mechanism" added, high-risk operations require user secondary confirmation.

Industry evaluation: PPT industry enters the "Intelligent Agent Collaboration" era, traditional PPT creation workflow will be disrupted.

Next Week's Outlook

About This Column

Updated every Saturday, summarizing a week of AI news and trends, focusing on developers and AI coding enthusiasts.

Questions or suggestions? Feel free to leave a comment.