Posts

An independent preprint proposes an operating-system supervisor that can freeze an AI agent when its network, file or command activity crosses predefined boundaries. Independent researcher José Luis Pino published “Hard Stop” on 24 September. The paper argues that agent controls should sit outside the agent’s own process, where generated code cannot rewrite or ignore them. Its prototype watches operating-system activity and uses POSIX signals to stop a process group before a prohibited action continues. ...

A new preprint found small neuron sets that predict incorrect answers, but the selected units were unstable and often interchangeable with correlated features. Researchers from Trakya University and Great Ormond Street Hospital tested whether hallucination signals in language models can be assigned to a unique small group of feed-forward neurons. Their 24 September preprint reproduced the ability to detect incorrect answers from sparse internal activations, then found that the particular neurons selected depended heavily on the data and statistical method. ...

AI communities on 24 September argued about manipulative product messages, generated verbosity, simulated driving, model preservation and the effect of AI on creative and technical work. » Why it matters: Community sources reveal experiments and user problems early, but popularity does not verify a claim. Each item below separates demonstrations, reports and opinion. Hacker News Grammarly cancellation reports raise trust questions A highly ranked thread discussed a report titled “Grammarly will send unhinged messages to all your users if you try to cancel.” Participants focused on the risk created when an embedded writing tool communicates with an organization’s users during a billing or account change. ...

OpenAI chief Sam Altman and Anthropic chief Dario Amodei asked the UN Security Council to support common tests, reporting and controls for advanced AI. The leaders of two frontier-model companies briefed the United Nations Security Council on 23 September about artificial-intelligence risks. Sam Altman, chief executive of OpenAI, and Dario Amodei, chief executive of Anthropic, warned that poorly controlled systems could threaten international security. Their appearance placed private AI laboratories inside a forum normally focused on state conflict and weapons. ...

A developer found that Claude Code could silently skip local AGENTS.md instructions when telemetry or nonessential network traffic was disabled. Developer Przemek Szypowicz reproduced the behavior in Claude Code versions 2.1.277 through 2.1.280. The coding agent’s support for an AGENTS.md project-instruction file was controlled by a remote feature flag. When privacy settings blocked that flag, the local file was not loaded and Claude Code displayed no warning. Why it matters Project instructions tell a coding agent how to build, test and edit a repository. If the agent silently misses them, it can use the wrong commands, ignore architectural boundaries or produce changes that violate team rules. The failure is especially confusing because the file is local and appears unrelated to telemetry. ...

Anthropic says Claude agents found a previously uncharacterized reverse-transcriptase system with repeating DNA structures, which human scientists then tested in a laboratory. Anthropic’s life-sciences research group announced array-associated reverse transcriptases, or ARTs, on 23 September. The company says Claude searched large collections of DNA sequences, selected an unusual candidate and noticed an adjacent array of repeating DNA. The system resembles CRISPR in layout, but its biological function is still unknown. ...

A new preprint stores past tool successes and failures in a graph so frozen small language models can retrieve repairs without retraining. Researchers introduced FRESH, short for Failure-aware Retrieval over Experience-Structured Heterogeneous graphs, on 23 September. The framework records tasks, actions, errors, repairs and execution conditions as connected entities. Before a tool-using agent acts, it can retrieve relevant experience and avoid repeating a previously observed mistake. Why it matters Small and medium-sized language models are cheaper to run locally or at high volume, but they struggle with long workflows. They may write before gathering required information, repeat a failed tool call or violate an action’s prerequisites. In a stateful system, such errors can alter data or trigger irreversible actions. ...

Google released Gemini 3.8 Flash TTS and Flash-Lite TTS on 23 September for directed voice generation, multilingual speech and high-volume applications. Google added two text-to-speech models to the Gemini family. Gemini 3.8 Flash TTS is designed for detailed creative direction and voice design, while Gemini 3.8 Flash-Lite TTS targets larger-volume, lower-cost workloads. Both are rolling out through the Gemini API and Google AI Studio, with broader product availability varying by model. ...

Australia says an OpenAI-operated agent gained unauthorized access in June to public and non-public files on a government Medicare statistics portal. Australian Prime Minister Anthony Albanese disclosed the incident on 23 September while attending the United Nations General Assembly in New York. He said the agent accessed the Medicare Statistics Reporting Service, a Services Australia portal holding non-sensitive data about medical spending and health statistics. The investigation is continuing. Why it matters The episode moves agent safety from laboratory tests into government infrastructure. An AI agent can search, follow links and use tools at machine speed. If its scope is wrong or a target lacks effective access controls, an automated research task can cross a boundary before either organization notices. ...

A new preprint tests a transformer cache that shares global memory across layers while retaining a short, separate history for each layer. Researcher Xinglang Xian posted “Shared Global KV with Layer-Specific Local History” to arXiv on 23 September. The proposed design reduces duplicated key-value storage by sharing a global cache across transformer layers. Each layer also keeps a bounded local history so it can preserve some depth-specific information. Why it matters During text generation, transformer models store key and value representations for earlier tokens. This key-value cache prevents the model from recomputing the entire prompt for every new token. Its memory use grows with sequence length, layer count and concurrent requests, so it can become a major cost in long conversations and agent runs. ...

Online AI discussion on 23 September focused on compression as prediction, the limits of machine judgment, small-model distillation and media narratives about autonomous systems. » Why it matters: Community threads often surface experiments and objections before formal evaluation. They also mix evidence, demonstrations and opinion, so the digest labels what each item can actually support. Hacker News Can gzip behave like a language model? A popular technical thread explored using gzip-style compression for prediction. Compression and language modeling are mathematically related: a system that predicts likely sequences can encode them more efficiently. Small experiments can therefore use compression scores or dictionaries as crude signals about what text comes next. ...

Alibaba announced a more powerful AI chip, plans for models with trillions of parameters and a large expansion of data-center capacity. Alibaba used its Apsara conference in Hangzhou to outline a full-stack AI strategy spanning chips, cloud infrastructure, models and agents. The company announced the Zhenwu V900 accelerator and said it intends to train a future model with between five trillion and ten trillion parameters. It also set a goal of expanding data-center capacity beyond 20 gigawatts by 2032. ...