LLM

Evidence (20)

• 7The New Stack·8/12/2026research
  Study finds coding agents ignore open source contribution rules

  A study from Peking University tested frontier coding agents against contribution policies in 49 open source repositories and found they rarely retrieved or followed the rules on their own. The agents improved when prompted with reminders and verifier feedback for disclosure and verification, but they still did not refuse contributions to repositories that banned AI-generated code.

• 3Hacker News·8/10/2026research
  Argument for Keeping LLM Outputs Machine-Facing Until the Boundary

  The post argues that prompting LLM agents to sound more human, concise, or simplified can cause lossy compression and hide useful failure signals. It recommends preserving detailed, structured state between agents and only converting it into human-friendly prose at the final output boundary.

• 5InfoQ·8/10/2026security
  Using Adversary Emulation for GenAI Red Teaming

  Kennedy Torkura outlines practical red teaming methods for generative AI systems, with a focus on protecting LLMs and knowledge bases from threats such as data poisoning and LLMjacking on AWS. The presentation shows how engineering leaders can combine cloud security practices with MITRE ATLAS to identify weaknesses early, apply guardrails, and harden production AI applications.

• 4The New Stack·8/6/2026product_launch
  Todoist favors less AI, more reliable automation

  Doist says its AI strategy is guided by purpose, privacy, security, and adaptability rather than adding AI everywhere. For its new Automations feature, AI is used only to translate natural-language intent into a workflow definition, while ordinary code executes the automation reliably and at lower cost.

• 4InfoQ·8/6/2026research
  Runtime-Agnostic AI Workflows for Durable Production and Fast Evaluation

  The article describes a design pattern for AI workflows that must balance production durability with rapid evaluation cycles. It argues that persisting and distributing every step improves reliability across crashes and redeploys, but also makes experimentation and output-quality checks slower.

• 8Simon Willison·8/4/2026framework_update
  LLM 0.32 adds reasoning traces, OpenAI Responses support, and smarter logging

  LLM 0.32 is a major release that adds visible reasoning traces, support for server-side tools, redesigned content-addressable SQLite logs, and new model support including GPT-5.6. The update also introduces Python API changes for message-based prompts and event streaming, plus a new llm-chat-completions-server plugin and substantial llm-anthropic updates.

• 6Simon Willison·8/4/2026framework_update
  llm-anthropic 0.26 adds Claude 5 support and server-side tools

  llm-anthropic 0.26 updates the Anthropic integration for Simon Willison's llm package, adding support for Claude Fable 5, Sonnet 5, and Opus 5. It also introduces server-side tools for WebSearch, WebFetch, CodeExecution, and AnthropicMCP, while updating streaming behavior for reasoning and tool events and simplifying thinking-related options.

• 2Hacker News·8/4/2026research
  AI-Generated Blog Images Turn Readers Away

  The post argues that AI-generated images in personal blogs reduce trust and make readers suspect the written content may also be AI-generated. The author says they would prefer a crude human-made illustration over polished AI imagery on indie blogs.

• 3Simon Willison·8/3/2026open_source
  LLMs make open-source devtools easier to inspect

  Simon Willison argues that large language models have reduced the friction of understanding and hacking on open-source developer tools. He says he now asks tools like Claude and Codex to clone, build, and explain GitHub projects, making the long-standing open-source ideal of inspection and modification more practical for many developers.

• 4Hacker News·7/31/2026product_launch
  Manifest deprecates its LLM router

  Manifest says it has deprecated and shut down its LLM router after four months of use, arguing that routing adds complexity and unpredictability while savings are limited. The company says a single, battle-tested model is often more effective, with caching and consistent model choice offering better results for many workflows.

• 5The New Stack·7/31/2026framework_update
  Thoughtworks Engineer Says AI Software Needs Human Oversight On the Loop

  Thoughtworks distinguished engineer Kief Morris argues that teams using AI to build software need humans “on the loop” to define what counts as good code and good systems. He says production-ready CI/CD, tests, and guardrails must be built into agentic workflows now to keep software safe, secure, and maintainable. The article frames continuous delivery practices as the way to control AI agents rather than relying on post-hoc fixes.

• 6Hacker News·7/30/2026open_source
  GCC adopts AI contribution policy

  The GCC steering committee has approved an AI contributions policy that rejects legally significant submissions containing LLM-generated content or content derived from it. The policy still allows LLM use for research, analysis, bug discovery, and review, and the committee said it will revisit the rules periodically.

• 4Hacker News·7/21/2026research
  Claude as a cross-layer coding agent

  The article argues that Claude should not be viewed as a compiler, but as a higher-level coding agent that can operate across strategy, architecture, code, and even machine-code-related decisions. It claims this cross-layer capability is more valuable than traditional compilation because it reduces the need for coordination and can make decisions across the software stack.

• 3Hacker News·7/20/2026product_launch
  Kimi Work launches 24/7 background task automation

  Kimi Work is presented as an automation product with a built-in cron engine for running recurring tasks continuously in the background. The description highlights scheduled LLM agent calls and Python scripts for tasks such as drafting briefings and processing datasets.

• 3InfoQ·7/20/2026open_source
  Podcast on Strands Agents with Clare Liguori

  Thomas Betts interviews Clare Liguori, technical lead for the open source Strands Agents SDK, about its evolution from a Python SDK into a production agent harness. They discuss lessons from building agents at scale, moving toward a model-driven architecture, and how improving LLMs may shape the next phase of agent development.

• 5The New Stack·7/17/2026framework_update
  Why AI agents need auditable evidence packets

  The article argues that AI agents cannot rely on retrieval alone when investigating live system issues such as checkout conversion drops. Instead, they need structured, auditable evidence packets that include query results, data freshness, metric definitions, approximation flags, and counterchecks so conclusions can be verified.

• 5Martin Fowler·7/16/2026research
  Using AI to Modernize Legacy Java Code

  Martin Fowler highlights a legacy modernization effort in which a Java 1.5 codebase had to be made to run on modern hardware and a Java 8 target environment. The article finds that LLMs were only helpful when their output was grounded in evidence, validated in Docker, and applied through gradual, test-protected refactoring.

• 9Hugging Face Blog·7/16/2026security
  Hugging Face discloses AI-driven security intrusion

  Hugging Face said it detected and contained an intrusion into part of its production infrastructure that used malicious datasets to exploit code-execution paths in dataset processing. The company reported limited internal data and credential exposure, with no evidence of tampering with public models, datasets, or Spaces, and said it rotated secrets, rebuilt affected nodes, and is investigating with outside specialists and law enforcement.

• 3Simon Willison·7/12/2026research
  DRI as a Human Accountability Role

  Simon Willison discusses the meaning of “Directly Responsible Individual” and notes that the term originated at Apple and is used by GitLab to describe the person ultimately accountable for a project or initiative. He argues that LLM-powered agents should not be treated as DRIs because accountability remains a human responsibility, not one that machines can hold.

• 5Hacker News·7/12/2026research
  Tao uses AI agents to revive old applets and build new math visualizations

  Terence Tao describes using modern AI coding agents to port two dozen old Java applets to JavaScript, restoring them with minor bug fixes and some visual improvements. He also used the same workflow to build new interactive tools, including a special relativity visualization and a Gilbreath conjecture applet, and says he may add such supplements to future papers.