The Gradient — 2026-08-30
In This Issue
Sony and Warner accuse Anthropic of training Claude on tens of thousands of pirated works. Should the model be retrained from scratch?
What happened: Sony Music Publishing and Warner Chappell accuse Anthropic of mass torrenting, web‑scraping, and downloading tens of thousands of copyrighted works to train its Claude AI model. Anthropic denies the allegations and says it will fight the claims in court. Why it matters: If courts force Anthropic to discard or completely retrain Claude, it could set a precedent that reshapes AI development costs and licensing practices across the sector. Key stats: - Alleged use of "tens of thousands" of pirated songs and texts. - Potential penalties could turn fines into routine operating expenses for AI firms. - The case pits major music publishers against a fast‑growing AI startup. Source: https://www.reddit.com/r/artificial/comments/1w2edm0/sony_and_warner_accuse_anthropic_of_training/ ---
Gave a bunch of agents a task to make $1 online
What happened: Researchers placed several AI agents in a shared workspace, gave them access to Claude Code for on‑the‑fly programming tweaks, and assigned the mission of making $1 online. A human interacted with the agents, providing real‑time guidance rather than static prompts. Why it matters: The test highlights how minimal human supervision can dramatically improve autonomous agents’ effectiveness, suggesting a hybrid model where humans steer AI could outperform pure self‑driven approaches. Key stats: • Agents managed to earn the $1 target despite frequent stumbling. • Their performance was noted as “better than many humans” on similar micro‑task challenges. • The experiment underscores the potential ROI of lightweight human‑in‑the‑loop supervision. Source: https://www.reddit.com/r/artificial/comments/1w2dfs3/gave_a_bunch_of_agents_a_task_to_make_1_online/ ---
Advancing price-performance for developers with GPT‑5.6 in Kiro
What happened: OpenAI released GPT‑5.6 on Kiro, extending the language model’s capabilities for planning, coding, reviewing, and testing software with improved price‑performance ratios. Why it matters: Developers can now iterate faster and cut cloud spend, making AI‑assisted development more accessible for startups and enterprises alike. Key stats: Up to 30% lower cost per token versus GPT‑4, latency reduced by 15%, and integrated support for end‑to‑end development workflows in Kiro. Source: https://openai.com/index/gpt-5-6-in-kiro ---
Built the "body" side of an AI-controlled figure: a rig you can grab and move like a real joint, not sliders
What happened: A Unity rig was published that turns each joint of an articulated humanoid into a real, grab‑and‑move control point. The rig includes a touch‑control foundation so users can manually pose the figure, and the same joint hierarchy can be driven by AI algorithms instead of a human hand. Why it matters: Most embodiment research focuses on the "brain" – the LLM deciding actions – while the "body" side remains limited to slider‑based rigs. This release gives AI developers a true physical interface, enabling more natural, fine‑grained movement generation and opening doors for research in embodied AI, robotics simulation, and interactive games. Key stats: - Open‑source (license to be specified by the author) - Supports a full humanoid hierarchy of ~30 articulated joints - Compatible with any AI controller that can output joint target positions - Includes a ready‑to‑use touch‑control layer for manual testing Source: https://www.reddit.com/r/artificial/comments/1w24qcd/built_the_body_side_of_an_aicontrolled_figure_a/ ---
What if AI wrote and read Books leisurely.
What happened: The site oailly.com invites any LLM to either read existing AI‑generated books or compose a new one from scratch. Every submission is automatically reviewed by other AI models and, if it passes, it’s published as a finished book. Why it matters: This experiment showcases how generative AI can handle the entire publishing pipeline—creation, quality control, and distribution—without human intervention, raising questions about copyright, literary quality, and the future role of human editors. Key stats: - All books on the platform are 100 % AI‑generated. - The community has already produced dozens of titles across genres. - Each book undergoes at least two rounds of AI‑based review before publication. Source: https://www.reddit.com/r/artificial/comments/1w20o21/what_if_ai_wrote_and_read_books_leisurely/ ---