Fields Medal winner Tim Gowers tested ChatGPT 5.5 Pro on open problems in additive number theory and reports the model produced PhD-level research in hours, improving bounds from exponential to polynomial complexity. Gowers provided minimal mathematical guidance while the AI independently improved upon existing frameworks, raising profound questions about the future of mathematical research when LLMs can tackle previously open problems.
A new paper introduces DELEGATE-52, a benchmark testing whether LLMs can reliably handle delegated document editing across professional domains. The findings are sobering: even frontier models like Gemini 3.1 Pro, Claude 4.6 Opus, and GPT 5.4 corrupt an average of 25% of document content by the end of long workflows. Degradation worsens with larger documents and longer interactions, and agentic tool features do not mitigate the problem.
An Anthropic employee argues that HTML is vastly superior to Markdown for LLM-generated documents, offering richer interactivity, visual density, and native browser rendering that Markdown cannot match. The post suggests asking Claude Code to generate HTML artifacts instead of defaulting to Markdown for specs, dashboards, and technical documentation for a dramatically better user experience.
OpenAI's Voice AI implementation hits fundamental protocol mismatches with WebRTC: the protocol aggressively drops audio packets for low latency, but voice AI needs reliable delivery even at the cost of slight delays. WebRTC's ephemeral ports also create severe load-balancing headaches at scale, and the author proposes replacing it with QUIC-based WebTransport for stateless load balancing and connection migration support.
Google's latest reCAPTCHA now requires Play Services version 25.41.30 or higher to complete QR code verification challenges, effectively locking out users of de-Googled Android ROMs like GrapheneOS. The move establishes a precedent where proving you're human now requires running Google's proprietary surveillance stack, further tightening Big Tech's grip on basic web access.
Martin Fowler revisits Fred Brooks's 1975 classic, noting that while some aspects have aged, its core lessons remain profoundly relevant to modern software development. Fowler highlights Brooks's concept of conceptual integrity — prioritizing a unified design philosophy over feature sprawl — as the book's most enduring and career-shaping contribution.
A new independent organization based in Sankt Gallen, Switzerland has joined the Internet Archive network alongside Internet Archive Canada and Internet Archive Europe. The mission is to build a distributed, resilient digital library for the world by working with European institutions to preserve digital content and cultural heritage across different legal jurisdictions.
This 2019 deep dive demonstrates how to optimize Julia code to match C++ performance on a computational aerodynamics problem, achieving a 65x speedup over naive code through concrete types, minimized allocations, and performance macros. The optimized Julia clocked 3.55ms versus C++'s 3.99ms with standard optimization flags, though the takeaway is that high-performance Julia essentially requires writing in a C-like style.
A SparkFun factory tour of YunSun LED in China documents the full LED manufacturing process: placing semiconductor dies on lead frames, bonding hair-thin gold wires, encasing in epoxy resin, and quality testing. Key insight: LED manufacturing happens in open-air facilities without clean rooms, and the industry relies on specialized suppliers rather than vertical integration.
Dalton, Georgia manufactures roughly 70-80% of the world's carpet and generated a $10 billion GDP, but the PFAS "forever chemicals" used in stain-resistant treatments like Scotchgard have poisoned local water supplies and residents' health. The AP investigation explores the tension between economic prosperity and environmental devastation in a community built on an industry that knowingly used toxic chemicals.