AI Daily: Claude's New Constitution, Microsoft VibeVoice Challenges Long Audio, and Gemini's SAT Prep Tool

This AI Daily covers three key developments: How Anthropic is reshaping Claude’s core values via a ‘New Constitution’, Microsoft’s VibeVoice model solving the 60-minute transcription challenge, and Google Gemini partnering with Princeton Review to help students prepare for the SAT smarter.

Teaching AI “Why”: Claude’s New Constitution and Value Reshaping

In the development of artificial intelligence, ensuring that models are both smart and kind has always been a major question. Anthropic recently took a quite interesting move: they released a brand new “Constitution” for their AI model, Claude. This is not just a list of rules, but more like a detailed declaration of values, explaining what kind of existence Anthropic wants Claude to be.

From Rigid Rules to Flexible Principles

Training AI in the past often relied on specific rules, like telling a child “don’t do this, don’t do that.” But the real world is too complex, and rigid rules often appear clumsy or even counterproductive when facing unforeseen situations. Anthropic realized this.

The new approach is somewhat different. They no longer just tell Claude what to do, but try to let the model understand why it should do so. This Claude’s New Constitution contains detailed explanations of values, aiming to help the model use judgment to weigh options when facing dilemmas. For example, how to balance between “honesty” and “compassion”? Or how to provide as much help as possible while protecting sensitive information? This document is mainly used to give Claude the knowledge and understanding needed to act in a complex world.

Balancing Safety, Ethics, and Utility

This new constitution revolves around four core priorities, with a clear order of precedence:

1. Broadly safe: Most importantly, do not undermine mechanisms for human oversight of AI.
2. Broadly ethical: Be honest and trustworthy, avoiding harm or danger.
3. Compliant: Follow specific developer guidelines in specific contexts.
4. Genuinely helpful: Benefit users from interactions.

Interestingly, Anthropic admits that this document is not perfect. They view it as a “living document” that will be continuously revised over time. Moreover, to achieve true transparency, this constitution is released under the Creative Commons CC0 1.0 license, which means anyone is free to use it without permission. For those worried about unpredictable AI behavior, this provides a window to examine the AI’s internal logic.

Understanding One-Hour Conversations: Microsoft VibeVoice-ASR’s Long Recording Breakthrough

Transcribing long meeting recordings has always been a pain point for users. Traditional Automatic Speech Recognition (ASR) models usually chop long audio files into small pieces for processing. While simple, this often leads to lost context, disjointed semantics, and confusion about who is speaking.

Breaking the 60-Minute Coherence Limit

Microsoft’s VibeVoice-ASR was born to break this limitation. This is a unified speech-to-text model, and its strength lies in being able to process audio up to 60 minutes long in a “single pass” without chopping it up. This ensures the model maintains a coherent understanding of semantics throughout the entire hour of recording and accurately tracks the speaker’s identity.

This model can generate structured transcription content, containing three key elements:

• Who: Accurately distinguishes different speakers.
• When: Provides precise timestamps.
• What: Complete content record.

Customized Hotwords and Open Source Resources

In addition to handling long recordings, VibeVoice also supports “Customized Hotwords”. Imagine a meeting full of obscure technical terms or specific names; ordinary AI often mishears them. But VibeVoice allows users to provide a specific list of words to guide the recognition process, which greatly increases accuracy in professional fields.

For developers and researchers, the good news is that relevant resources are already public. You can find the VibeVoice-ASR model on Hugging Face, or check the codebase directly on GitHub. If you want to experience its capabilities directly, there is also an online Demo to try. This ability to combine speech recognition, speaker diarization, and timestamping really takes the utility of automated note-taking to the next level.

A Boon for Examinees: Google Gemini Launches Free SAT Practice Tests

For many high school students, standardized tests are like a mountain that must be climbed. At this year’s BETT UK (British Educational Training and Technology Show), Google announced a practical update for students: Gemini can now act as your personal SAT prep coach.

Professional Support from The Princeton Review

The quality of practice questions on the market varies. To ensure students practice with “authentic material,” Google chose to partner with the education authority The Princeton Review. This means that the practice questions in Gemini have been strictly reviewed, and their difficulty and format highly replicate real exam scenarios.

This feature is currently completely free. Students can take full, on-demand practice tests on Gemini. Although it currently mainly supports the SAT, Google says more types of exams will be added in the future.

Personalized Guidance Learning from Mistakes

The true value of Gemini is revealed after finishing the questions. It doesn’t just give you a score and end there, but provides immediate feedback, pointing out where you performed well and which concepts need strengthening.

If there is a doubt about an answer, students can ask Gemini directly to explain the logic behind the correct answer. This is like having a tutor on standby, helping students identify knowledge blind spots and turning these insights into concrete action plans. Whether preparing for the SAT for the first time or planning to retake it for a higher score, this tool can make the preparation process more directional and reduce the anxiety of blindly doing questions.

FAQ

Q1: Why does Anthropic think the new “Constitution” is better than the old list of rules?

Anthropic believes that for AI to behave like a “good person” when facing various novel and unforeseen situations, it needs to understand the “why” behind it, not just rote memorize “what to do”. Broad principles allow the model to learn to use judgment for generalization and trade-offs, which is better adaptable to the complex real world than rigidly following specific rules.

Q2: What is the biggest advantage of Microsoft VibeVoice-ASR compared to traditional speech recognition models?

The biggest advantage is that it can process audio up to 60 minutes in a single pass without cutting it into small fragments. Traditional model slicing processing easily loses global context, leading to incoherent speaker tracking or broken semantics. VibeVoice maintains semantic coherence for the entire hour of recording while outputting structured information of “who, when, and what”.

Q3: Are the SAT practice questions on Google Gemini reliable?

Quite reliable. Google partners with the renowned educational institution The Princeton Review, using strictly reviewed materials. This ensures the quality and difficulty of the practice questions are close to the real exam, avoiding candidates practicing with poor quality or outdated questions.

Q4: What is the practical use of VibeVoice’s “Customized Hotwords”?

This feature is very useful for specific fields. For example, in medical, legal, or engineering meetings, many proper nouns or names that general models cannot understand appear. Users can provide these vocabularies (such as drug names, technical terms) to VibeVoice in advance to guide the model to pay special attention to these words, thereby significantly improving recognition accuracy on specific domain content.