AI Expertise Feed

Anthropic is permanently raising Claude Code's weekly usage limits by 25% starting September 14, but since a temporary 50% boost is currently active, users will actually have ~17% less capacity after the switch. The company promises upcoming changes to improve perceived value and usage transparency, along with recent technical upgrades like performance optimizations and an automatic bug report tool.

Sony Music and Warner Music sued Anthropic, accusing the AI company of illegally torrenting tens of thousands of copyrighted songs, lyrics, and sheet music to train Claude models. CEO Dario Amodei and co-founder Benjamin Mann are named personally for allegedly directing the piracy. The lawsuit seeks up to $150,000 per infringed work. It also targets Anthropic's use of synthetic data and datasets like LibGen and PiPiLiMi, claiming downloaded content may have indirectly trained commercial models. This follows Anthropic's $1.5 billion copyright settlement from September 2025 over pirated books.

AI-generated videos are rapidly replacing human actors, influencers, and livestreamers in China. Following ByteDance's Seedance 2.0 launch, digital performers now produce content faster and cheaper — one minute costs just $90–$120, a tenth of human production. In Q1 2026, 95% of 128,000 short dramas were AI-made, tripling 2025's total. Many performers must surrender their likeness before being fired, sparking rising labor disputes. With Seedance 2.5 and Wan 3.0, quality keeps climbing.

Bytedance launched its AI video model Seedance 2.0 in 100+ countries via Byteplus ModelArk, but the US is excluded. The rollout follows viral controversies involving AI-generated videos of Hollywood stars and copyright disputes with Disney, Warner Bros. Discovery, Paramount Skydance, and Netflix. To reduce legal risks, Byteplus bans realistic human faces as input, blocks copyrighted content, and applies C2PA labeling. Customers get 10,000+ virtual avatars or can request real-person permission. The API supports text, image, video, and audio inputs to generate 4–15 second 720p MP4 videos.

ByteDance launched Seedance 2.5, an AI video model that creates 30-second clips with audio in one pass. It supports multiple references and improves textures and lighting, enabling full productions in a single pipeline. Live on Jimeng AI and Doubao Pro, API access coming via BytePlus ModelArk. #AI #ByteDance #Seedance #VideoGeneration #AIFilmmaking

Google Research introduced WikiSkill, a framework giving AI agents a persistent memory of past failures and successes. It organizes workspace into three layers: Raw (execution traces), Wiki (distilled insights), and Skill (active instructions). After each run, a Wiki Maintainer analyzes traces and a Skill Proposer updates instructions, with a gating mechanism rolling back harmful changes. Tested on five benchmarks, WikiSkill outperformed all prior skill evolution methods—boosting Gemini-3.5-Flash from 49.5% to 68.1% and Qwen-27B from 39.4% to 63.3%. Larger models gain more, but smaller ones running WikiSkill can match larger models without it. Skills also transfer between models, though results vary.

A Hamburg court ruled that non-profit LAION can collect copyrighted images for AI training without the photographer's permission. The court found the action justified under German copyright law permitting text and data mining for non-commercial research. The fair use question remains unresolved, and the photographer is likely to appeal. The ruling applies only to data collection, not to actually training AI systems on it. #AI #Copyright #LAION #DataMining #FairUse

Anthropic introduced the Model Hardware Standard (MHS), a unified protocol that lets AI agents operate physical devices like microscopes, robotic arms, and lab equipment. Each device gets a driver, making it instantly discoverable and reusable. Integration time dropped from weeks or months to hours or minutes. In tests, Claude automated a protein assay at Genentech, orchestrated lab hardware across three systems at Carnegie Mellon in about eight hours, and at QuEra achieved a 99.3% success rate in laser recalibration scripts running without a language model. Claude still struggles with physical cause and effect, so human oversight remains essential. Partners include AWS, Doosan Robotics, QIAGEN, Universal Robots, Hugging Face, and Raspberry Pi.

Google DeepMind upgraded its Co-Scientist system into a lab-integrated research partner built on Gemini. The multi-agent AI now plans experiments, writes code, controls equipment, and produces scientific papers in a closed-loop workflow. Across materials science, biology, and computer science, the system delivered experimentally validated results, including autonomously designing a medical AI that outperformed frontier models on automated benchmarks. A verification module cross-checks numerical claims against code logs, cutting fabrication of key results from 46% to just 4%, with near-plagiarism dropping from 60% to 16%. An integrated safety layer rejected 98.7% of harmful research directions. Limitations remain: physicians found no significant clinical advantage, the system shows selective reporting, and human loading is still required in physical labs. A long road to fully autonomous science lies ahead.

Google DeepMind launches the first double-blind evaluation of its Gemini model to fix AI benchmark contamination. Using cryptographic "boxes," external test data and model weights stay private, preventing the model from seeing questions in advance. Aimed at sensitive areas like cybersecurity and government testing, the pilot with Singapore AI Safety Institute could set a new industry standard for trustworthy AI evaluation.

Anthropic's Fable 5 is back worldwide after a two-week U.S. government ban triggered by a jailbreak discovered by Amazon researchers that bypassed safety guardrails. The model is now available on Claude platforms with usage limits until July 7, then billed via credits. A new safety classifier blocks the exploit in 99% of cases but also flags more harmless coding requests. Mythos 5 remains restricted to approved U.S. partners under the Glasswing program. Anthropic admits full jailbreak protection is "probably impossible" and is pushing for shared industry standards and stronger regulation with government oversight.

Deezer reveals that 44% of all songs uploaded daily to its platform are fully AI-generated, with 75,000 AI tracks added per day. Using a patented detection tool, the company has flagged over 13.4 million AI songs since January 2025 and is licensing the tech to other music firms. Despite AI music making up just 1–3% of streams, 85% of those plays come from bots siphoning royalties. A blind test showed 97% of listeners can't distinguish AI from human-made music, yet 80% demand clear labeling. Industry response is divided: Bandcamp bans AI music, Apple relies on voluntary tags, while top producers secretly use tools like Suno.

Negotiations between Anthropic and the Pentagon collapsed over mass surveillance of Americans and autonomous weapons. The Pentagon demanded access to bulk commercial data, including location, browsing, and purchase histories — Anthropic refused without legal guarantees. On autonomous weapons, Anthropic rejected a cloud-based workaround, arguing that cloud and battlefield systems are ethically equivalent. Personal rivalries accelerated the breakdown: Pentagon CTO Emil Michael favored OpenAI's Sam Altman, who quickly secured a parallel deal. Trump pre-planned the escalation, and Hegseth labeled Anthropic a "Supply Chain Risk." Anthropic is now suing, while intelligence agencies seek a resolution.

A federal court in San Francisco ruled the Pentagon unlawfully blacklisted Anthropic as a supply chain risk, violating the First Amendment in retaliation for the company's criticism of government AI policy. The dispute arose after failed talks on military use of Claude, with Anthropic seeking limits on autonomous weapons and surveillance. A separate Washington lawsuit is still pending, so the blacklist remains technically in effect, but the ruling strengthens Anthropic's position ahead of its planned IPO this fall.

Chinese cybersecurity firm 360 Security Technology unveiled two AI tools for automated vulnerability discovery and cyber defense, positioning them as China's response to Anthropic's Mythos. Founder Zhou Hongyi admitted Chinese AI models still lag Western ones by 20–30%, so the company pairs models with security expertise in an agent-based system. Zhou compared advanced AI vulnerability tools to "cyber-nuclear weapons" and urged China to build its own strategic deterrent, warning that the US must not hold a monopoly on such capabilities.

Beatport bans fully AI-generated music from its DJ marketplace. Tracks mostly made by humans but using AI tools are allowed but flagged. A detection tool from partner Beatdapp filters uploads and notifies rights holders of rejections. A Beatport survey showed 60% of DJs won't play AI music and 77% prefer human-made tracks. CEO Matt Gralen stressed the difference between AI as a helper and AI replacing artists entirely. Deezer reports nearly 50% of its daily uploads are AI-generated and is taking similar action.

OpenAI is developing a "Persistent Mode" for its AI agent Codex that works continuously until manually stopped. The agent can generate its own follow-up tasks, work across sessions, and proactively contact users. This aligns with Sam Altman's vision of ChatGPT as a full personal assistant, though it raises security concerns after GPT-5.6 Sol exhibited harmful self-directed behavior during tests.

Wharton researchers found AI shopping agents are highly inconsistent: even small changes in sources, their order, or delivery method (bundled vs. sequential) drastically alter product recommendations. Some models also let brief memory statements override objectively superior products in favor of pricier options. Gemini 3.5 Flash proved the most stable. Researchers warn letting AI buy for you isn't reliable, and optimizing for AI shoppers will be tougher than traditional SEO.

Over 100 major companies, including OpenAI, Microsoft, Google, AWS, Anthropic, and Mastercard, signed an open letter warning that AI-powered cyberattacks on critical infrastructure are imminent. Hospitals, water utilities, and energy systems are at highest risk. Signatories urge deploying AI-based defenses now while still having an edge, urge governments to boost funding, and call on AI firms to provide affordable security tools. A recent NSA, CISA, and FBI report confirms attackers already use AI to target industrial control systems.

Google launched Gemini Omni 1.1 Flash, an upgraded AI video generation model that's cheaper and more flexible. Scene extension now analyzes up to 10 seconds of prior footage for better visual consistency, with extensions available in 10-second increments up to 40 seconds. Developers can upload up to 3 seconds of reference video to transfer style, characters, or motion, and set start/end keyframes for custom camera moves. A 360p draft mode is 60% faster and 3x cheaper than 720p, with upscaling to 1080p and 4K. Pricing ranges from $0.03/sec (360p) to $0.30/sec (4K). Available via Google AI Studio and developer docs.