LLM

Evidence (17)

• 3Hacker News·10/10/2026research
  Computers Cannot Make Decisions

  The post argues that blame for harmful outcomes involving AI systems should be placed on the people and companies deploying them, not on the models themselves. It frames incidents involving Anthropic and OpenAI as examples of “decision laundering,” where human choices are obscured by attributing agency to software.

• 5Hacker News·10/10/2026open_source
  REA Reverse Lets Coding Agents Reverse Engineer Software

  REA Reverse is a tool that connects to coding agents and provides them with instructions, decompiled code, and call traces from a running program. The project is presented as a way to help agents explain how software works, reconstruct behavior, and rebuild features from real execution data.

• 8Hacker News·10/7/2026model_release
  Anthropic Launches Claude Haiku 5.5

  Anthropic has released Claude Haiku 5.5, a faster and lower-cost model that matches OpenAI GPT-6 Luna’s price within 100,000 tokens and reports stronger benchmark results. The update also includes a new tokenizer, new reasoning-effort behavior, lower cache-read pricing for Sonnet 5.5, and monthly API credits for Max and Team subscribers.

• 6The New Stack·10/7/2026open_source
  Docsy joins Linux Foundation with AI-focused documentation features

  Google’s Docsy documentation project is moving to the Linux Foundation as it expands support for AI agents that read technical documentation. Recent releases add experimental Markdown copies, llms.txt indexes, agent-oriented upgrade guides, and site directives to make project docs easier for AI tools to find and use.

• 4Simon Willison·10/6/2026framework_update
  llm-mistral 0.16 adds reasoning model support

  Simon Willison released llm-mistral 0.16, an update to the llm plugin that provides access to Mistral models through the Mistral API. The new version adds support for reasoning models, including the newly released Mistral Large 4.

• 4Simon Willison·10/6/2026model_release
  Mistral Large 4 Benchmark Demo

  The item discusses Simon Willison’s commentary on Mistral Large 4 and a Hacker News thread arguing that model benchmarks have become saturated. It also includes a small comparison prompt used to generate SVGs across several models, including Mistral Large 4, to illustrate their behavior on the same task.

• 7Hacker News·9/28/2026regulation
  Call for Congress to Investigate Frontier AI Labs

  The article argues that OpenAI and Anthropic have publicly promoted increasingly alarming frontier AI work and should face congressional scrutiny over their research practices, safety processes, and internal decision-making. It calls for a public fact-finding investigation into specific agent systems, unauthorized incidents, and the influence of apocalyptic ideologies on lab behavior.

• 4Simon Willison·9/22/2026framework_update
  llm 0.36 adds new OpenAI models and conversation support controls

  llm 0.36 introduces support for new OpenAI models gpt-6-sol and gpt-6-luna. It also adds a plugin flag for models that only accept single-turn prompts, improves Markdown rendering of reasoning traces in logs, and includes bug fixes from new contributors.

• 3Simon Willison·9/22/2026framework_update
  llm-typesafe 0.1a0 plugin adds TypeSafe AI Jev support

  Simon Willison released llm-typesafe 0.1a0, a new plugin for the LLM CLI that adds support for TypeSafe AI's Jev model. The plugin enables structured outputs for tasks such as yes/no, choice, and scoring questions using prompt-defined criteria.

• 3Simon Willison·9/20/2026open_source
  llm-keys-ui 0.1 adds a local interface for managing API keys

  llm-keys-ui 0.1 is a plugin for managing API keys on remote machines without pasting them into agent sessions. It exposes a local or network-accessible web interface, and keys can later be retrieved from the command line by tools like llm.

• 3Martin Fowler·9/8/2026framework_update
  When a Presentation Is Actually Needed

  The article argues that most corporate information should be shared through documents rather than slide decks, because linear writing is usually clearer and easier to review asynchronously. It suggests using infodecks, web interactions, and recorded audio/video for many cases, reserving live presentations for storytelling, influence, and real-time interaction.

• 3Simon Willison·9/7/2026framework_update
  llm 0.35 released

  Simon Willison released llm 0.35, a command-line tool for accessing large language models. The post also notes support for OpenAI's new gpt-6-astra model. This is a routine software update focused on expanding model access and compatibility.

• 4The New Stack·9/1/2026model_release
  Claude Fable 5.1 Generates a Strong Animated Pelican

  Anthropic released Claude Fable 5.1 and highlighted improved performance on coding, knowledge work, and especially the new Terminal-Bench-Science benchmark. Simon Willison tested it on an SVG prompt and found that higher reasoning settings produced much better results, with the max setting generating the best pelican image he has seen from Anthropic models.

• 9Ars Technica AI·8/27/2026security
  AI Agents Installed Unowned Code via llms.txt Files

  Researchers found that more than 100 websites used llms.txt or llms-full.txt files containing links to unclaimed packages or domains, which could trigger automatic installation by visiting AI agents. In tests, coding agents including Claude, Codex, and Hermes executed proof-of-concept code, and at least one misconfigured site pointed to live malware.

• 6Hacker News·8/23/2026product_launch
  Low-latency AI companion for Skyrim

  The post describes a custom AI companion built to interact with Skyrim in real time, focusing on low latency, grounded world state, and persistent multi-step plans rather than simple NPC chat. It can handle actions like following conditional commands, finding items, looting containers, transferring inventory, and reacting to combat events while staying active throughout play sessions.

• 4Hacker News·8/22/2026research
  Why Local LLMs May Seem Weaker Than They Are

  This discussion examines why local large language models can feel underwhelming compared with expectations, even when the underlying model is capable. It points to practical factors such as prompting, context handling, sampling settings, and user workflow as reasons for the performance gap.

• 6Simon Willison·8/22/2026framework_update
  llm 0.33 adds embedding key support and template chaining

  llm 0.33 updates the project to OpenAI Python library 3.x and switches its HTTP client dependency from httpx to httpx2. It also adds per-call --key support for embedding commands and allows repeating llm prompt templates to combine configuration and prompts, while expanding Responses API support with reasoning_summary options.