What stood out this week was not a new model or product launch. It was how AI is already embedded in routine systems, including healthcare use, memory supply, and online purchasing.

These are no longer pilot deployments. They are part of normal operations. As AI becomes infrastructure, attention shifts to cost, supply, and regulation. That process is already in progress.

OpenAI reports that more than 40 million people use ChatGPT for health-related questions each day. Over 5% of all messages fall into healthcare, including symptom checks, medical terminology, billing questions, insurance disputes, and preparation for doctor visits.

This week, OpenAI launched ChatGPT Health. It is a separate environment for medical queries that allows users to connect health records and apps such as Apple Health. Health data is handled in an isolated environment, stored separately, can be deleted on demand, and is not used for model training.

Access is limited. The same report also references the need for clearer FDA approval pathways for AI medical devices.

Takeaway: Health-related use already accounts for a measurable share of ChatGPT traffic. OpenAI is separating this usage into a distinct product with defined data handling and regulatory positioning.

OpenAI has reached preliminary agreements with Samsung and SK Hynix for approximately 900,000 DRAM wafers per month, about 40% of global supply. The memory is intended for data centre expansion, including the Stargate project.

DRAM supply is shared across consumer and enterprise hardware. Allocating capacity at this level reduces availability elsewhere.

Pricing has shifted. A 64GB DDR5-5600 kit priced at around $180 last May now sells for about $710. Analysts expect pressure on pricing to continue into 2026.

The same supply chain supports phones, laptops, servers, and embedded systems.

Takeaway: Data centre demand is drawing on the same memory supply used by consumer hardware, with visible pricing effects.

Google is rolling out Gemini-powered features in Gmail. These include thread summaries, natural-language search, and inbox prioritisation.

Previously paid tools such as Help Me Write and Suggested Replies are now free for personal users. Google states that personal email content is not used for model training and that all AI features are optional.

The rollout is currently limited to the US.

Takeaway: Gemini features are being added directly to an existing email workflow rather than released as a separate product.

Amazon has launched Alexa.com, allowing access to the Alexa+ assistant through a browser. The web version supports research, planning, and task execution.

Alexa+ integrates with services including Expedia, Yelp, Angi, Square, Uber, and OpenTable. Amazon reports 3–5x increases in shopping and cooking-related engagement since rollout.

The Alexa mobile app is being redesigned around a chatbot interface.

Takeaway: Amazon is extending Alexa beyond dedicated devices, relying on existing distribution rather than new model features.

For buyers: Browse, compare, and buy products inside a Copilot chat or on a brand’s site without leaving the conversation.

For sellers: Keep control of checkout, payments, and customer data while using AI agents in Copilot and on your own site to guide buyers.

Copilot Checkout handles the full flow inside chat, including payments, integrates with PayPal, Shopify, and Stripe, and is currently available in the US only.

Takeaway: Checkout is being integrated directly into conversational interfaces.

Step 1: Generate breadth, not quality

I start with full context and a loose objective:

This works for LinkedIn posts, presentation titles, UI concepts, product angles, or writing hooks.

The goal is not to use these ideas directly. It’s to surface patterns, directions, and unexpected combinations.

Step 2: Converge deliberately

After reviewing the list, I pick the ideas or parts of ideas that feel interesting.

Then I prompt again:

This second pass usually produces stronger results than asking for “good ideas” upfront, because the model is refining from contrast rather than inventing from scratch.

The key is separating divergence from convergence instead of trying to do both in one prompt.