Meta's AI Training Documents Expose Ethical Concerns

Good morning

It's Tuesday, April 22, 2025, and you're reading the Agentive Daily Report, where we cut through the noise of the AI sphere to bring you what actually matters. Let's dive into what's caught our eyes the most today.

TL;DR for busy people

• Meta's internal AI training documents leaked in a lawsuit, revealing the usage of copyrighted books to improve Llama model's performance
• OpenAI's new o3 and o4-mini reasoning models show better performance but hallucinate significantly more than previous models
• ByteDance claims its Seedream 3.0 text-to-image model outperforms GPT-4o and Midjourney in speed and quality
• Google DeepMind proposes "streams" architecture for AI agents to learn through continuous real-world interaction
• China held the world's first humanoid robot half-marathon, with only 4 out of 21 robots completing the race
• Google released Gemini 2.5 Flash, a cost-optimised version of Gemini that matches Claude on STEM benchmarks

Today's Top Stories

Meta's AI Training Secrets Revealed Through Lawsuit

Meta's internal AI training process has been exposed through court documents, showing how the company used "ablation" tests to determine which data types improve model performance. The company discovered that adding science, tech, and fiction books could improve benchmark scores by up to 6%, and some of this testing included books from unauthorised distributors like LibGen.

This revelation raises serious questions about fair compensation for content creators and the ethics of using copyrighted materials for commercial AI training. The leaked documents provide rare insight into major AI companies' usually secretive training methodologies, potentially setting precedents for how training data is sourced and credited in the future.

ByteDance's Seedream 3.0 Claims to Outperform Leading Image Generators

ByteDance has introduced Seedream 3.0, a text-to-image model that reportedly surpasses competitors like GPT-4o and Midjourney v6.1 in both speed and quality. The model generates high-resolution 2K images in roughly three seconds and achieves 94% accuracy in rendering text in both English and Chinese.

Beyond image generation, ByteDance has also launched SeedEdit for precise in-image editing capabilities. These advancements represent a significant competitive challenge to established players in the generative AI space, and the technology will be integrated into ByteDance's Doubao chatbot platform, potentially expanding its reach to hundreds of millions of users.

OpenAI's New Reasoning Models Trade Accuracy for Performance

OpenAI's newest reasoning models, o3 and o4-mini, demonstrate better performance on complex tasks but come with a concerning tradeoff: significantly increased hallucinations. Tests show o3 hallucinating on 33% of questions, double the rate of its predecessor, while o4-mini performed even worse at 48%.

This unexpected correlation between improved reasoning capabilities and increased factual inaccuracy poses a serious challenge for applications requiring reliability. OpenAI acknowledges that they don't fully understand why hallucination rates increase alongside reasoning improvements, highlighting the ongoing tension between advancing AI capabilities and ensuring trustworthy outputs. Recent independent testing also revealed discrepancies between OpenAI's reported benchmark scores and real-world performance.

Other Developments Worth Noting

• Google DeepMind's "Era of Experience": Researchers Richard Sutton and David Silver propose a new "streams" architecture where AI agents learn through continuous interaction with dynamic environments rather than from static data, potentially enabling more human-like learning and adaptation.
• AI Avatars Evolution: AI avatars are advancing beyond basic lip-syncing into fully expressive digital characters with complete body movement and emotional nuance, transforming content creation, marketing, and communication for businesses of all sizes.
• Anthropic's Agentic Coding Guidelines: Anthropic released detailed best practices for using Claude Code, their command-line tool for AI-assisted programming, emphasising context organisation, workflow patterns, and techniques like "ultrathink" to maximise performance.
• Chinese Robot Marathon: China held the world's first humanoid robot half-marathon in Beijing, where only 4 out of 21 robots completed the 13-mile course. The fastest robot finished in 2 hours and 40 minutes, while the human champion completed it in just over 1 hour.
• Google's Gemini 2.5 Flash: Google released an optimised version of its Gemini model that balances performance with efficiency, matching Claude on several important STEM benchmarks while costing less to operate.

New Tools Discovered

• Google Whisk 2.0: Transform still images into smooth eight-second animated clips with AI-powered motion and transitions.
• Seedream 3.0 by ByteDance: Generate high-resolution images with 94% text accuracy in English and Chinese, outperforming competitors in photorealistic rendering.
• Gemini 2.5 Flash: Google's new cost-optimised model for developers who need solid reasoning capabilities without the premium price of full-sized models.
• Ad Clone: An AI tool that scrapes LinkedIn's Ad Library and generates customised ad creatives with your brand elements in just 30 seconds.
• MCP Containers: Open-source scripts to build and manage Model Context Protocol server containers, making it easier to create cross-model AI applications.

Discover more tools at Agentive.Directory

That's a wrap for today! Thank you for reading this report. Have thoughts on today's developments? Hit reply and let me know what you're thinking. Or if you've discovered a cool AI tool we should feature, drop me a line.

Until tomorrow,
Hak from Agentive.Studio