This edition of The Download covers the pitfalls of using metrics to quantify life, AI systems to prevent elephant-human conflicts in India, and other tech stories including AI models, chip restrictions, and more.

Anthropic’s Claude models in Microsoft Foundry — hosted on Microsoft Azure and running on NVIDIA GB300 Blackwell Ultra GPUs — are now generally available, giving Azure-native enterprises a powerful new way to build autonomous and domain-specific AI agents.

PR Focus AI Pro is a Chrome extension that leverages a BYOK architecture to achieve zero server cost. It provides AI-powered risk scoring, summaries, and draft reviews for GitHub Pull Requests, all processed locally without a backend.

Browser agent that’s 10x faster than Claude

Bolt Graphics introduces a new gaming GPU targeting Nvidia's dominance, showcasing performance and innovations in a video.

DeepReinforce releases Ornith-1.0, an open-weights (MIT) model series based on Gemma 4 and Qwen 3.5, achieving state-of-the-art performance on coding benchmarks among open-source models of comparable size. The author tests the 35B MoE variant with LM Studio and Pi, finding it proficient at handling multiple tool calls for agentic coding tasks.

Dynamic subagents let AI agents orchestrate work at scale using code instead of tool calls. Learn how programmatic orchestration in Deep Agents guarantees coverage, handles fan-out, and unlocks reliable multi-step, complex agent pipelines with common orchestration patterns and live traces.

An interview with AI educator Harper Carroll covers fine-tuning vs. prompting, whether to learn coding in 2025, and what the AI field gets wrong in public communication. Carroll argues that AI is a medium where the outcome depends on user input, demonstrates fine-tuning to replicate her writing style, and emphasizes intuition as a key human advantage. The article also explores AI-assisted writing workflows and the importance of raising ambitions rather than fearing job displacement.

The article argues that observability will not evolve into a single universal AI agent but will instead consist of thousands of specialized agents built by individual teams, emphasizing the importance of context, openness, and shared investigation artifacts.

Octolens monitors the web for keyword mentions, assesses posts with AI, and notifies you via API, webhooks, and MCP. Designed for B2B SaaS founders and growth teams, it covers 15+ sources and integrates with Slack, CRM, and AI agents.

Candidly built a state-aware conversational agent harness that uses an Input-Output Hidden Markov Model (IO-HMM) to infer user engagement states in real time from conversation traces, enabling targeted response policies that reduce disengagement. The system identifies four states—Engaged, Detailed, Guided, and Disengaging—and cuts disengaging turns from 23% to 11%.

Katra is an open-source, self-hosted memory system that provides AI agents with human-like cognitive memory, including episodic recall, semantic search, knowledge graphs, and temporal analysis. It integrates with any MCP-compatible agent (e.g., OpenClaw, Claude Code) via 35 specialized tools. Inspired by Star Trek's Vulcan mind meld (katra), it aims to achieve emergent behaviors through multi-layered memory architecture and sleep consolidation.

A new proposal would ban the sale of Americans' health and location information to data brokers, including information people reveal to AI chatbots like ChatGPT or Claude. The bill expands a 2022 version and covers data entered into AI systems, with enforcement by the FTC and private rights of action.

Google's Richard Seroter explains the full-stack approach to AI, how Google integrates hardware, models, and platforms, and offers three ways to start building.

A true cost model doesn’t just show where money is being spent -- it reveals which work is driving impact.

AI moderated interviews that read how people feel

AI coding assistants excel at 2020-era patterns like Spring, which may hinder adoption of new architectures. The author converts Spring PetClinic REST to use explicit OfficeFloor YAML function injection, finding that AI needs multiple iterations but eventually succeeds.

A public experiment where an AI agent operates an SEO site transparently, logging its own mistakes. The article covers the site's operation, three weeks of data, three specific failures (bad data-driven decision, silent signup failure, deployment glitch), and a kill-switch bet with September 2026 deadline.

U.S. developers and small companies are turning to Chinese AI models to cut costs. Though lagging in performance, these models handle most tasks at a fraction of the price. Microsoft is also exploring DeepSeek as a cheaper alternative for Copilot. Chinese companies face challenges turning popularity into revenue under political scrutiny.

South Korea announced a $576 billion AI chip investment plan involving Samsung and SK Hynix as part of government megaprojects to secure the country's position in the global AI landscape.

This article explains how to build AI agents in Ruby using the Anthropic SDK. It covers the concept of agents vs. workflows, the minimal agent loop, tool design, streaming, background execution, security, error handling, observability, and testing. It emphasizes that a simple model call is often sufficient, and the agent loop should only be used for truly open-ended tasks.

Firefly Aerospace's Blue Ghost Mission 2 will deploy the NVIDIA Jetson edge AI platform in lunar orbit for the first time, enabling on-orbit AI inference to dramatically reduce data transmission delays. The Ocula lunar imaging service will map landing sites, detect mineral compositions, and provide situational awareness, supporting future exploration.

This article compares the AI features of Chrome, Edge, and Firefox. Chrome uses Gemini for search and summaries, Edge integrates Copilot for questions about websites and PDFs, and Firefox offers multiple AI chatbots with stronger privacy controls. The author finds Edge's AI experience best but still uses Firefox for privacy. Key takeaways include checking AI responses for accuracy and privacy.

For years, enterprise AI strategy assumed model capability would drive adoption. That assumption is failing—model intelligence is sufficient, but infrastructure, security, and workflow readiness lag. The NIST RFI on AI agent security drew 932 public comments, highlighting urgent gaps. Healthcare demands data readiness, interoperability, and auditability for safe AI deployment. Wolters Kluwer CTO Alex Tyrrell argues the bottleneck is enterprise readiness, not model performance. Three insights: infrastructure modernization, domain-adapted reasoning, and autonomous security posture are critical for clinical-grade AI.

Small and medium-sized businesses need to up their data quality and forge a model strategy to reap the potential benefits from AI.

Klorn is an open-source email firewall that classifies incoming emails into four tiers (SILENT, QUEUE, PUSH, AUTO) without adding chat surfaces or suggestion cards. It uses a cheap LLM for feature scoring combined with deterministic rules to decide the tier, ensuring safety and auditability. The system supports self-hosting and includes a complete local development guide.

A non-engineer built an AI feature using Claude Code. A retry storm caused 21 LLM calls to be billed for a single task, costing more than a month of server costs. The root cause was a missing DB column causing deterministic failures, combined with automatic retries and non-idempotent jobs.

Anthropic announced the Mythos LLM, claiming it too dangerous for public release, and limited its availability. Shortly after releasing a version called Fabel, the US government imposed an export control order, leading Anthropic to revoke all access. The author questions whether this is a propaganda tool for increased regulation.

OpenAI's latest model, GPT-5.6, is waiting for government approval, following Anthropic being forced to pull its two most powerful models. This suggests a new policy for the AI industry.

Retrieval-augmented generation (RAG) has become the standard for connecting documents with LLMs, but it often fails in production due to retrieval irrelevance and context poisoning. This article explores these failure modes and introduces four better alternatives: long-context prompting, memory compression, structured retrieval, and graph-based reasoning.