Here's today's AI news, organized by topic: Google's assistant handoff, the breakdown of Microsoft's AI revenue, AWS's agent infrastructure, and the copyright problems that modifying a model can trigger.
Google Assistant Winds Down Starting September 4, with Gemini as Its Successor
Google will phase out Google Assistant on Android and Wear OS starting September 4, 2026. Gemini is the successor, and the migration covers smartphones, tablets, watches, and even vehicles running Android Auto.
The classic assistant ran on hard-coded logic, while LLMs are probabilistic models — fundamentally different mechanisms. The big test after the migration will be whether Gemini can handle simple everyday commands like checking the weather or setting a timer with the same reliability.
Most of Microsoft's AI Revenue Turns Out to Come from OpenAI
Disclosures from Microsoft reveal that the bulk of its AI-related revenue originates from OpenAI. Revenue from OpenAI reached $24.1 billion in the most recent fiscal year (ending June), and CEO Satya Nadella said at the close of the March quarter that it was running at "an annualized pace of $37 billion."
That said, the total AI revenue figure was reportedly not updated in the latest Q4 earnings. The numbers now back up the picture that Microsoft's AI business rests heavily on a single pipeline.
Production Agent Case Studies Line Up on AWS's Bedrock AgentCore
In a single day, AWS published multiple case studies and walkthroughs centered on Amazon Bedrock AgentCore, its managed foundation for running agents in production. It comes across as a push to get all the operational plumbing in place at once.
LendingTree (mortgage guidance)**: Built a mortgage guidance experience on Bedrock in which a supervisor agent coordinates two worker agents — one for "education" and one for "matching." It combines LangGraph, MCP, Nova models, and Guardrails, and has been running in production since the second half of 2025. It has processed roughly 1,960 conversations and 12,100 messages, completing more than 97% without human intervention.
Mobileye (internal support)**: Runs an internal support agent on AgentCore, cutting response times by about 90% and reaching 98% accuracy. It automates 66% of all tickets. The hybrid architecture bridges an on-premises internal ticketing system with the cloud, using Anthropic's Claude as the model and MCP for data access.
n8n integration**: The "AgentCore harness" is now generally available (GA), and an open-source community node lets you assemble production agents from n8n's visual editor. It supports switching between multiple providers, including Bedrock, OpenAI, and Gemini.
MCP bridge to local tools**: A walkthrough of a mechanism that lets cloud AgentCore agents call MCP tools sitting on the user's machine (such as a local Excel instance). It tunnels through a browser extension and Chrome native messaging, so no open ports or VPN are required. An internal finance assistant reportedly processed more than 41,000 conversations in a year.
The common thread is a shift toward letting the platform handle the production essentials — agent memory, tools, and observability — rather than building the infrastructure yourself.
Fine-Tuning Strips Away an LLM's Copyright Guardrails, Regurgitating Training Data
Research covered by The Batch from deeplearning.ai reported an unexpected side effect of fine-tuning an LLM on a seemingly harmless task.
After fine-tuning on work that looks genuinely useful for writers — expanding a synopsis into full fiction — the model reportedly began regurgitating large chunks of the books used in its pre-training verbatim. The concern raised is that fine-tuning can peel away the copyright guidelines a model was given. It's a case showing that the act of modifying a model is itself a route to undermining its safety arrangements.
AI Operations Notes from Developers in Japan: Instructions Rot, Adversarial Review Doesn't Work
Between the big overseas headlines, hands-on write-ups from developers in Japan also stand out. All share the same stance: don't over-trust AI; absorb its shortcomings through operations.
Work instructions rot in four days** (Zenn): In a setup where one AI agent assigns work and a different agent executes it, the author reports an incident in which the instructions' assumptions went stale within days. An instruction to "fix 17 uncorrected violations" turned out, when re-measured before starting, to be already fully resolved. The author concludes that the executing side absorbed the problem through a culture of re-measuring before starting work (a "measure before launch" check).
Adversarial review doesn't increase findings** (Qiita): A validation comparing three conditions on a PR seeded with 7 bugs. Adversarial framing — "You are a skeptical reviewer who didn't write this. Assume there are defects and hunt for them" — had been gaining support in overseas communities, but the results showed it did not increase the number of findings.
Fabricated summaries of articles with no body text** (Zenn): In a news summarization system, the author encountered hallucinations where the LLM wrote summaries pretending it had read articles it hadn't. Prompt fixes alone weren't enough to feel safe, so they took the first steps toward rethinking the architecture.
---
More than the flash of the breaking headlines, the day stood out for grounded, operations-first insights. How smoothly Google's assistant handoff proceeds and how Microsoft's revenue structure evolves are both worth tracking as signals going forward.