Anthropic is aggressively expanding its infrastructure footprint—a $1.8B Akamai cloud deal and the xAI/Colossus data center arrangement signal massive capacity bets that will likely shape model availability and pricing for enterprise customers in regulated industries
Domain-specific AI continues to outperform frontier generalists in high-stakes verticals: a curated pharma platform beat GPT, Claude, and Gemini on drug-asset discovery, reinforcing that "build vs. buy" decisions for regulated use cases deserve rigorous benchmarking rather than defaulting to the biggest model
Academic AI evaluations are systematically lagging frontier capabilities by months or years, meaning capability assessments cited in compliance, procurement, or risk reviews may already be obsolete—PMs should scrutinize evaluation vintage before trusting published benchmarks
Anthropic has struck a seven-year, $1.8 billion computing agreement with Akamai Technologies to handle surging inference demand, on the heels of also securing capacity from xAI's Colossus data center
OpenAI continues its GPT-5 family rollout with new state-of-the-art real-time voice, translation, and transcription APIs, expanding multimodal coverage at the API layer
A benchmarking study pit Gosset, a platform backed by curated drug-asset annotations, against Claude Opus, GPT-5.5, Gemini 3.1 Pro, and Perplexity across ten niche oncology/immunology targets, finding the domain-specific system consistently superior
Anthropic published research turning Claude's internal reasoning representations into human-readable natural language, advancing interpretability methods that expose what the model is "thinking."
A widely discussed post (568 HN points) argues that reliable agentic systems require explicit programmatic control flow—branching, loops, error handling—rather than relying on ever-larger prompts to manage complexity
A new open-source project, re_gent, proposes Git-style version control semantics applied to agent state and execution history, enabling branching and rollback for agentic runs
Anthropic's Thariq Shihipar documents how prompting Claude Code to produce HTML artifacts—rather than Markdown—yields richer, more structured, and more reviewable outputs for tasks like PR reviews and data summaries
A new arXiv survey maps the emerging paradigm of agentic reinforcement learning, where LLMs act as autonomous agents trained beyond narrow reward functions across open-ended environments
Mozilla deployed Anthropic's Claude Mythos preview to identify and fix hundreds of Firefox security vulnerabilities, marking a shift from AI-generated security noise to actionable, high-quality bug reports
The AI drug-discovery company spun out of DeepMind is in advanced discussions for a Thrive Capital-led $2B+ round, underscoring continued mega-investment in AI-for-science verticals
Simon Willison's notes surface that the Colossus facility initially operated gas turbines without Clean Air Act permits or pollution controls, a compliance gap flagged as Anthropic expands its reliance on the site