AI Research, Models & Agents
Anthropic Research · www.anthropic.com
Anthropic's interpretability research investigates how emotion-like concepts arise and function inside large language models, examining whether models that say they're happy or sorry have internal representations corresponding to those expressions.
Andrej Karpathy published a public gist he calls an "LLM Wiki" — an example of an "idea file" collecting notes and observations about LLMs.
A research paper demonstrates that a straightforward self-distillation technique can meaningfully improve LLM code generation performance without requiring additional training data or complex pipelines.
Sebastian Raschka breaks down the architectural components that make up modern coding agents, covering tool use, planning, memory, and evaluation.
AI's Impact on Software Engineering & Security
The post argues that AI-accelerated code writing doesn't address the real bottlenecks in software engineering, which lie in understanding requirements, system design, and organizational coordination.
An essay arguing that as AI makes code generation increasingly commoditized, the value of software engineers shifts away from writing code toward problem-solving, architecture, and judgment.
A developer documents how Claude Code discovered a Linux vulnerability that had been present in the codebase for 23 years, illustrating AI's growing capability in security-relevant code analysis.
Simon Willison · simonwillison.net
Simon Willison highlights Thomas Ptacek's assessment of how frontier AI models are suddenly having an enormous impact on vulnerability research, alongside quotes from Linux kernel and curl maintainers about the surge in AI-generated security reports.
Privacy, Security & Supply Chain
Research reveals that LinkedIn is detecting and cataloging users' installed browser extensions, raising privacy concerns about the platform's data collection practices.
Simon Willison · simonwillison.net
The Axios team's postmortem reveals the NPM supply chain attack was carried out via individually targeted social engineering rather than automated methods, resulting in malware being shipped in an official release.
Germany's planned eIDAS digital identity wallet implementation will require citizens to have either an Apple or Google account, raising concerns about dependence on private US companies for government identity services.