Amazon Objects to "Human-in-the-Loop" AI Governance

According to reporting by The Register, Amazon has taken a strong stance against "human-in-the-loop" requirements in the governance of AI systems. The company reportedly questions regulatory approaches that mandate human involvement in AI decision-making processes, citing concerns about their effectiveness and cost.

Human-in-the-loop has long been regarded as one of the foundational principles of AI governance — a technique that improves safety by pairing AI autonomy with human oversight. Amazon, however, appears to argue that this requirement actually reduces system efficiency while failing to adequately mitigate substantive safety risks.

The move is drawing attention as an example of how major tech companies are working to shape the direction of AI regulation. The debate between regulators and industry over the "right balance" of AI oversight seems set to continue.

The Atlantic Investigation: Millions of Songs Were Used for AI Music Training

An investigation by The Atlantic has revealed that millions of songs — including tracks from major record labels — were used without permission as training data for AI music generation models. Engadget covered the story.

The investigation confirmed cases in which AI music startups used songs collected from Spotify, Apple Music, and other platforms as training data to generate new tracks mimicking the styles of existing artists. There are also reports of cases where neither licensing nor compensation was provided to rights holders.

The music industry is already pursuing legal action over AI-generated music, and these findings are likely to have a significant impact on future copyright lawsuits and regulatory debates. The question of how to balance protecting creators' rights with technological innovation is once again being put to the test.

FERNme: An Agent Memory Layer with Near-Zero LLM Calls

A project called "FERNme" was announced on Hacker News. It is a memory layer for AI agents: whereas conventional systems such as Mem0 call an LLM every turn to extract and summarize memories, FERNme is designed to use no LLM at all on the write path.

FERNme represents each user as a node in a "fuzzy preference graph" and updates edges using Hebbian co-occurrence rules. It incorporates an ACT-R-style decay mechanism and ultimately compiles everything into "memory cards" of roughly 40 tokens. Retrieval is performed via "spreading activation" rather than vector search.

According to the developer, FERNme ingested 86 free-form journal entries and built memories with zero LLM calls, achieving roughly four times the accuracy of frequency and recency baselines on a LoCoMo-style QA benchmark. That said, a full head-to-head comparison with Mem0 has not yet been conducted, so validating its practical utility is still at an early stage.

FunnyBench: A Benchmark for Whether AI Gets the Joke

A benchmark called "FunnyBench" has been released to measure whether AI models can understand funny jokes. How well AI can grasp humor — a subjective, culturally shaped element — is one of the important frontiers of natural language understanding.

It may look like a simple idea, but humor is packed with elements AI tends to struggle with, such as wordplay, sarcasm, and context-dependent nuances. How this benchmark gets used to improve models will be worth watching.

Debate Over Why Desktop AI Apps Are So Heavy

A question sparked discussion on Hacker News: "Why are desktop AI apps so heavy?" The poster asked why an app that is essentially just a text application runs slowly on a modern PC.

In the comments, commenters pointed to Electron-based architectures, unnecessary dependency libraries, and the overhead of local inference. With inference itself often running remotely, the observation that the UI layer has become the bottleneck is an interesting one that may prompt a rethink of how desktop AI tools are designed.

---