AWS Went Down, Sora Under Fire (Again!?), Claude 4.5, Music Without Musicians

What we covered this week: check out YouTube, listen wherever you get your podcast or read below 👇

Musicians miss out as Belgian retailers turn to AI-generated music in stores | https://www.brusselstimes.com/1798519/musicians-miss-out-as-belgian-retailers-turn-to-ai-generated-music-in-stores

Major Belgian chains like Brico and Carrefour are swapping licensed playlists for AI‑generated, royalty‑free background music to cut costs and tailor in‑store sound. Rights agency Sabam warns this could erase roughly a quarter of songwriters’ public‑performance income and flags legal risks around models trained on copyrighted works.

Japan wants OpenAI to stop ripping off manga and anime | https://www.theverge.com/news/799938/japan-government-openai-sora 

Japan’s Cabinet Office, led by IP minister Minoru Kiuchi, formally asked OpenAI to curb Sora videos that imitate manga/anime, calling those cultural forms “irreplaceable treasures.” The move escalates backlash to OpenAI’s opt‑out approach and waves of look‑alike content, sharpening global debates over IP in generative media.

Opera’s Neon shows just how confusing AI browsers still are | https://www.theverge.com/tech/801899/opera-neon-ai-browser-trial-run 

Opera’s paid AI browser Neon bundles three agents—Chat (chatbot), Do (controls the browser), and Make (content builder)—for $19.90/month. Early testing found the multi‑agent idea promising but confusing, with verbose answers, misfires on page interactions, and uncertainty over which agent to use; Opera says it taps OpenAI and Google models without detailing which.

[R] Plain English outperforms JSON for LLM tool calling: +18pp accuracy, -70% variance | https://www.reddit.com/r/MachineLearning/comments/1o8szk0/r_plain_english_outperforms_json_for_llm_tool/ 

A community paper (NLT) claims LLMs pick tools more accurately with plain‑English instructions than JSON schemas: across 6,400 trials on 10 models, accuracy rose ~18pp (to ~87.5%) and variance fell ~70%. Tests were single‑turn and parameter‑free, suggesting natural language can cut format burden and token overhead but may not generalize to complex tool calls.

Introducing Claude Haiku 4.5 | https://www.anthropic.com/news/claude-haiku-4-5 

Anthropic’s Claude Haiku 4.5 targets fast, low‑cost coding and agentic tasks, aiming for near‑Sonnet‑4 code quality at one‑third the price and more than double the speed. Available via the Claude API, Amazon Bedrock, and Google Vertex AI, it’s priced at $1/$5 per million input/output tokens and ships with safety evaluations and an ASL‑2 classification.

Amazon outage breaks much of the internet | https://techcrunch.com/2025/10/20/amazon-dns-outage-breaks-much-of-the-internet/ An AWS DNS‑related outage in us‑east‑1 disrupted DynamoDB API endpoints and cascaded across major apps—Amazon, Coinbase, Fortnite, Signal, Venmo, Zoom and more—before Amazon said the issue was fully mitigated. The incident highlights the internet’s dependence on AWS and directs users to the AWS Health Dashboard for remediation details.