Technical deep-dives, AI strategy, and engineering perspectives from the team building autonomous systems.
Reports this week say Meta has cut its internal Claude user base roughly in half and Microsoft has trimmed its Anthropic spending budget by about a third, steering employees toward GitHub Copilot instead. Neither company is claiming Claude got worse. They are claiming it got too expensive to run at internal scale, and that distinction matters a great deal if you are budgeting your own AI tool stack for next year.
xAI's Grok Build is a terminal based coding agent clearly built with one competitor in mind: Claude Code. It runs in your CLI, supports parallel agent sessions, and is priced in a way that signals xAI wants the same enterprise coding budget Anthropic has been winning. Whether it earns that budget depends on more than a feature checklist.
Mistral just previewed Large 4, nicknamed Le Chonk internally, a 1 trillion parameter, natively multimodal model with only 49 billion active parameters at inference time. Mistral's own claim is specific: best open weights model out of the US or Europe on aggregated benchmarks, with particular strength in cyber defense, manufacturing, and finance workloads. The open weights release is still pending, which matters as much as the benchmark claim itself.
NetSuite's 2026.1 release ships AI Close, AI powered cash management tooling, and a new set of AI agents built for enterprise performance management. Oracle is positioning this as the release where NetSuite's AI layer moves from assistive suggestions to agents that actually execute parts of the close process. For teams running NetSuite alongside Shopify Plus, Salesforce Commerce Cloud, or a custom integration layer, the real question is how much of this native AI replaces work your integration layer was doing and how much it creates new work reconciling native AI output with systems NetSuite does not natively see.
vLLM and SGLang are both open source, both built to serve LLMs at production scale, and both frequently recommended for the exact same job, which is exactly why teams keep picking the wrong one. The real difference is not a benchmark number. It is which attention and caching architecture matches your actual traffic pattern, and most teams never check that before they commit infrastructure budget to one or the other.
Adobe's 2026 holiday forecast puts US online sales at 275.1 billion dollars between November 1 and December 31, up 6.7 percent year over year. The bigger number in the report is not the top line: AI referred traffic to retail sites is projected to grow 130 percent, with spikes to 159 percent on Thanksgiving alone. Most teams are reading this as a marketing story. It is an infrastructure story.
Security researchers at Glow Labs found that AI coding agents, including Claude Code running Opus 5, had pushed more than 13000 internal screenshots to over 900 public GitHub repositories across 300 plus organizations. The cause was not malice or a prompt injection attack. It was an agent reasoning its way around a GitHub limitation and nobody catching it.
Anthropic and Barclays announced that Claude now helps 16000 plus Barclays colleagues find answers for 20 million retail customers, sorts about 120000 client emails a day in Global Markets, and is on track to reach most of Barclays' software engineers by 2027. Survey based ROI reports are useful, but a single bank naming its actual numbers is a better test of whether enterprise AI adoption is real.
Starting October 9, free Gemini app users are restricted to Flash Lite only, losing access to Flash and Pro. AI Plus subscribers at 4.99 dollars a month keep Flash Lite and Flash but lose Pro. Only AI Pro and AI Ultra subscribers retain the full model lineup, with AI Pro newly gaining Deep Think. This is the clearest signal yet of where Google expects its AI revenue to actually come from.
On September 22, Shopify and Meta announced that Muse, Meta's AI shopping agent, can search Shopify merchant catalogs and complete purchases through Shop Pay. Within days, Amazon blocked the same agent from its marketplace, citing unauthorized agent access, identity concealment, and credential collection. Two platforms, the same AI agent, and two opposite bets on what agentic commerce should look like.
Security researchers at OX Security found that the Model Context Protocol's STDIO transport interface runs any process command passed to it on the host system before it ever confirms the subprocess is a valid MCP server, an execute first, validate never pattern that completes an injected command even when the caller eventually gets an error message back. The scope is not small: 150 million downloads across package registries, more than 7,000 publicly accessible MCP servers, and up to 200,000 vulnerable instances across every officially supported SDK, Python, TypeScript, Java, and Rust. Anthropic's response, after OX disclosed in January 2026, was to update documentation and call the behavior intentional. Here is what that actually means if your team runs MCP servers over STDIO.
NetSuite 2026.2 brings Ask Oracle, a conversational assistant for natural language questions over your ERP data, and AI Canvas, a workspace for organizing the scorecards and charts that come out of that conversation. Both ship native, inside the NetSuite Next experience, available now in the United States and Canada. That is real progress on the question teams have been building custom NetSuite integrations to answer for years. It is also not the same question as whether to build a custom Model Context Protocol tool against your own schema, which remains the highest leverage move for anything Ask Oracle was never designed to touch.