AI & LLMs
Sam Rose published another excellent interactive essay, this time explaining how LLM quantization works from first principles. A great resource for understanding how models are compressed for efficient inference.
An exercise in distinguishing AI-written from human-written design documents, offering insight into the telltale patterns and tropes of LLM-generated prose. Relevant for anyone trying to detect AI writing in professional contexts.
Symbolica achieved 36% on the new ARC-AGI-3 benchmark on its first day, using program synthesis approaches. This is a significant result on a benchmark designed to test general reasoning beyond pattern matching.
Reco used AI to port JSONata from JavaScript to Go in a single day, eliminating a costly dependency and saving $500K annually. A compelling case study of 'vibe porting' — using LLMs to translate entire codebases between languages.
Analysis shows the vast majority of Claude Code-linked commits go to tiny, low-star repos, suggesting AI coding tools are predominantly used for personal and small-scale projects rather than major open source work.
Privacy & Security
Research paper demonstrating how LLMs can be used to deanonymize online users at scale, posing serious implications for privacy. This intersects both AI capabilities and the erosion of online anonymity.
The EU's Chat Control proposal continues to push for client-side scanning of private messages and photos, raising major concerns about mass surveillance and encryption undermining. The campaign urges citizens to push back.
A clear explanation of prompt injection attacks where adversarial inputs instruct LLMs to ignore prior instructions. Highlights a fundamental and still-unsolved security problem in deploying LLMs in production systems.