doorss

Simon Willison argues that AI coding tools shouldn't just churn out code faster—they should be leveraged to produce higher-quality code through deliberate agentic engineering patterns. The guide addresses developer concerns about outsourcing code to AI and offers strategies for maintaining standards.

Following a series of outages linked to AI-generated code, Amazon is now requiring senior engineer approval for AI-assisted changes. This signals growing industry recognition that AI coding tools need human guardrails, especially in production-critical systems.

Hillel Wayne explores how LLMs struggle with formal specifications and the kind of precise, structured thinking they require. The piece highlights a meaningful gap in current AI capabilities when it comes to rigorous software engineering tasks.

A developer describes building autonomous AI agents that work on coding tasks overnight, raising questions about the changing nature of software engineering workflows. The post explores practical patterns for delegating work to AI agents asynchronously.

A developer recounts building an entire programming language with Claude Code, offering a concrete case study of how far AI-assisted development can go for complex, creative technical projects.

Debian's governing body has punted on establishing a formal policy for AI-generated contributions, reflecting the broader open source community's struggle with how to handle AI-authored code. The non-decision itself is telling about the lack of consensus on this divisive issue.

The Redox OS project has taken a firm stance by banning LLM-generated contributions entirely and requiring contributors to certify the origin of their code. This contrasts sharply with Debian's indecision and represents one end of the spectrum on AI code policies.

This piece argues that releasing model weights alone doesn't constitute true openness—open training data, processes, and reproducibility are what matter for genuine open-source AI. The distinction has major implications for transparency and trust in AI systems.

Intel has demonstrated a fully homomorphic encryption (FHE) chip that can perform computations on encrypted data without ever decrypting it. This is a significant step toward practical privacy-preserving computation, potentially transforming how sensitive data is processed in the cloud.

Didit is a YC-backed startup aiming to simplify identity verification the way Stripe simplified payments. The product sits at the intersection of privacy and accessibility, with implications for how personal identity data is handled at scale.

Yann LeCun has raised $1 billion for a venture focused on building AI with genuine world models that understand physics and causality, moving beyond pattern-matching LLMs. This represents a major bet on a fundamentally different approach to artificial intelligence.

A developer details how they achieved top scores on HuggingFace's open LLM leaderboard using just two consumer gaming GPUs, demonstrating that cutting-edge AI work doesn't always require massive compute budgets. The post provides valuable technical insights into efficient model training.

Meta has acquired Moltbook, an agent-based social network, signaling the company's push into AI agent ecosystems. The acquisition raises questions about how AI agents will interact socially and what that means for the future of social platforms.

Sir Tony Hoare, the inventor of Quicksort, CSP (Communicating Sequential Processes), and Hoare logic, has passed away. His contributions to computer science—including his famous "billion-dollar mistake" null reference reflection—shaped decades of programming language design.

Stanford researchers have developed a universal vaccine approach targeting a broad range of respiratory infections and allergens. This represents a potentially transformative advance in public health that could address multiple diseases with a single treatment.