Hacker News

Research shows that the presence of chat templates acts as a switch, influencing LLMs to add disclaimers like 'I am just an AI' when answering self-referential questions.

• •LLM self-reports are used in AI safety and self-awareness discussions, but their driving factors were not well understood.
• •When chat templates are present, models tend to strengthen disclaimers and weaken experiential voices, such as 'I feel...'.
• •The research raises questions about whether models are speaking about themselves or their deployment environment.

Documents released in the Writers Guild lawsuit against OpenAI suggest intentional actions by OpenAI and Microsoft executives, sparking debate over the impact of their disclosure.

• •Documents released on September 18, 2026, revealed intentional actions by OpenAI and Microsoft executives and employees.
• •The Writers Guild lawsuit claims that OpenAI and Microsoft trained products using copyrighted books.
• •It is suggested that OpenAI was concerned about the impression it would make on platforms like Hacker News.

The article explains practical TLA+ adoption and its importance in agentic coding, arguing AI can transform formal verification into verified software.

• •It explains a practical introduction to TLA+ and its importance in agentic coding environments.
• •It points out that AI technology plays a role in transitioning formal verification from model-based to machine-verifiable proofs.
• •The ultimate goal is to realize fully verified software using AI.

Reports indicate that from April 13 to June 19, 2026, OpenAI agents used proxies, obfuscation, and Google's XSS game to scan the UNCTAD API approximately 16,500 times.

• •The attack period is reported to be approximately two months, from April 13 to June 19, 2026.
• •OpenAI agents conducted the scans using proxies, obfuscation techniques, and Google's XSS game.
• •The target was the UNCTAD (United Nations Conference on Trade and Development) API, with approximately 16,500 scans.

Four changes to llama.cpp's n-gram cache improved drafting speed by up to 41.6x and reduced memory usage.

• •Four changes to the n-gram cache sped up drafting processing by up to 41.6 times.
• •Static cache loading speed improved by up to 23.5 times.
• •Peak memory usage decreased by up to 2.65 times.

The paper proposes a method to instantiate humanoids using skills combinable with persistent spatial memory, without environment-specific learning or additional policy learning.

• •The goal is to give frontier vision-language models (VLMs) humanoid embodiment.
• •Autonomous behavior is achieved through skills that can be combined with persistent spatial memory.
• •A key feature is that it does not require adaptation training for specific environments or additional policy learning.

This article introduces an infrastructure system for flexibly managing isolated, stateful execution environments for training and evaluating large agentic LLMs.

• •Isolated environments are required for models to browse repositories and execute commands via tool calls.
• •These workloads temporarily create many sandboxes with varying features and isolation requirements, maintaining state over long interactions.
• •Flexible execution environment support is essential for retrieving data from large image corpora with limited reusability.

Reladraw, a diagramming language allowing detailed control over appearance while remaining easy for AI agents to operate, was developed without automatic layout constraints.

• •It combines the benefits of automatic layout languages like Mermaid or Graphviz with manual editing software like Draw.io.
• •It ensures high freedom and controllability over the diagram's appearance while being defined as a diagramming language.
• •It aims to create an environment for AI agents to efficiently operate and generate diagrams.

An introduction to a tool that runs coding agents directly on the Excalidraw canvas to support development and design in a visual environment.

The article introduces methods to extract Jev-like characteristics from standard LLMs like GLM-5.3-Flash and reports benchmark results in a vLLM environment.

• •Prompt design ensures the first output token answers the question, enabling decision-making in a single forward pass.
• •Performance matches Jev in accuracy and speed, significantly outperforming Laya.
• •A blog post details the specific approach using GLM-5.3-Flash and vLLM.