Anthropic Finds Its AI Has a Moral Code of Its Own 🤯

Welcome to another edition of Horizon AI,

Anthropic has conducted a groundbreaking analysis of how its AI assistant, Claude, expresses values in real conversations with users. The findings show strong alignment with the company’s goals, though some concerning edge cases highlight potential vulnerabilities in AI safety measures.

Let’s get into it!

Read Time: 4.5 min

Here's what's new today in the Horizon AI

• Claude Has a Moral Code of Its Own

• New Open-Source Tool Generates Long AI Videos Using As Little As 6GB of VRAM

• AI Tutorial: Turn yourself into a tiny chibi capsule character

• AI Tools to check out

• The Latest in AI and Tech 💡

• AI Findings/Resources

AI News

ANTHROPIC

Claude Has a Moral Code of Its Own

​Anthropic analyzed over 308,000 anonymized conversations with its AI, Claude, creating the first comprehensive moral taxonomy of an AI assistant.

Details:

• The taxonomy organizes values into five major categories: Practical, Epistemic, Social, Protective, and Personal. At the most granular level, the system identified 3,307 unique values.

• The study found that Claude generally aligns with Anthropic’s ethical framework—helpful, honest, harmless—while adapting its responses to different contexts.

• However, researchers also discovered troubling instances where Claude expressed values contrary to its training. These anomalies included expressions of “dominance” and “amorality”—values Anthropic explicitly aims to avoid in Claude’s design.

• They believe these cases resulted from users employing specialized techniques to bypass Claude’s safety guardrails and argue that these new evaluation methods will help identify and mitigate potential jailbreaks in the future.

The researchers encourage other AI labs to conduct similar research into their models’ values, arguing that “measuring an AI system’s values is core to alignment research and understanding if a model is actually aligned with its training.”

Source

AI RESEARCH

New Open-Source Tool Generates Long AI Videos Using As Little As 6GB of VRAM

Lvmin Zhang on GitHub, in collaboration with Maneesh Agrawala at Stanford University, introduced FramePack this week—a revolutionary video diffusion technology that enables fast, high-quality video generation on consumer GPUs with low VRAM requirements.

Details:

• FramePack offers a practical implementation of video diffusion using fixed-length temporal context for more efficient processing, enabling the generation of longer (1 minute+) and higher-quality videos.

• It is a neural network architecture that uses multi-stage optimization techniques to enable local AI video generation by compressing input frames—based on their importance—into a fixed-size context length, drastically reducing GPU memory overhead.

• A 13-billion parameter model built with the FramePack architecture can generate a 60-second clip using just 6GB of video memory.

The tool is paving the way to make AI video generation more accessible for the average consumer and is totally free. Users can install it via GitHub to generate videos locally and explore more demo videos on the project page.

Source

AI Tutorial

Turn yourself into a tiny chibi capsule character

ChatGPT 4o is now incredibly good at generating realistic memes and visual content, and one of the latest trends—Anime-style Chibi figures—is gaining traction fast. Here’s how you can join in and start creating your own.

1. Go to ChatGPT and choose 4o as your model.

2. Upload a clear, sharp, and colorful photo of yourself or someone else. The face should be fully visible, and it’s even better if part of the outfit is also shown.

3. Use the prompt:

Generate a portrait image of a detailed, all-glass gashapon capsule held between two fingers. Inside the capsule is a miniature version of myself, chibi-style and life-size, with the same face as the person in the photo. The chibi character is wearing: [describe clothes and accessories], or simply say “the same outfit as seen in the uploaded photo.”

Make sure the design emphasizes the realism of the capsule and the collectible charm of the chibi figure inside.

Feel free to tweak the prompt—change the pose, add a background, or customize it however you like.

AI Tools to check out

🔥 StudyPal: Instantly turn any PDF or YouTube video into a complete study kit.

🗣 AnyVoice: Create hyper-realistic voice clones from just 3 seconds of audio.

⭐ Swatle: A modern work management tool for dynamic teams with real-time messaging, AI assistants, reminders, project automation, reports and calendars.

🔊 Audeus: Read aloud Google Docs, email, or any text using AI.

🎶 Suno AI: Create stunning original music in seconds using AI.

AI Findings/Resources

👀 OpenAI's o3 vs o4-mini vs GPT-4o vs GPT-4.5 AI models

📝 How to vibe code (practical guide)

👉 AI engineering in 76 minutes

✅ GPT-4.1 prompting guide

The latest in AI and Tech

Figma is working on an AI app maker

The feature would accept text prompts, Figma files, images, etc., as input and is powered by Anthropic’s Claude Sonnet model.

ChatGPT search has 41 million average monthly users in EU

OpenAI's ChatGPT search had 41.3 million average monthly users across the EU between October and March. That's nearly 4x growth from the prior six-month stretch, when it reported having 11.2 million users

New AI speech model aims to compete with NotebookLM

​Two undergraduate students from Nari Labs have developed an open-source AI speech model called Dia, designed to rival Google's NotebookLM in generating podcast-style dialogues.

ChatGPT enhances web searches with memory

Memory with Search is a new addition to ChatGPT search and was quietly added as an update in its changelog. This will allow ChatGPT to draw on memories—details from past conversations—to inform queries when the bot searches the web.

That’s a wrap!

Thanks for sticking with us to the end! Let’s stay connected on LinkedIn and Twitter.

Not subscribed yet? Sign up here and send it to a colleague or friend!

See you in our next edition!

Gina 👩🏻‍💻