ChatGPT search expands, Google Images shifts: July 15

ChatGPT can now find work buried across chats and files, Google is turning image search into a personalized discovery feed, and a 27-billion-parameter model fits on a phone. In 30 minutes, you can clean up how staff retrieve AI work, prepare images for discovery, and identify one private task worth testing locally.

ChatGPT searches chats, projects, images, and files

OpenAI made one search box cover four kinds of saved work on July 14. The official ChatGPT release notes say search now spans chats, projects, images, and documents on web, iOS, and Android. Filters let you narrow results by content type, and the change is available on every ChatGPT plan worldwide.

The business gain is retrieval time. If your team uses ChatGPT for proposals, customer research, meeting notes, and campaign drafts, useful work no longer has to disappear into an old conversation. But broader search also makes careless naming more expensive. A result called “new chat” tells the next person nothing.

This is a filing change, not a reason to move your whole business into ChatGPT. Keep the approved asset in your normal shared storage and treat ChatGPT as the workbench. If customer information is involved, use the minimum needed for the task. The broader AI tools setup lane is the next place to audit what belongs in each system.

Google Images becomes a personalized discovery feed

Google introduced a browsable Google Images home on July 14. Its 25th-anniversary announcement says the desktop page will show a real-time gallery shaped by a signed-in user’s interests. Saved collections become tabs above that gallery. The rollout starts in US English over the coming weeks.

Google also said image generation is coming to AI Overviews, the AI-written answer that can appear above normal search results. A user will be able to describe an image and have Google create it with the Nano Banana model. That feature will roll out in English across regions where AI Mode already supports image creation.

For an owner, the first change matters more. Images can now reach people while they browse, before they type a precise query. Product sellers, venues, home-service companies, and design-led businesses should act. A business with generic stock photos can ignore the generation feature and fix its source images first.

Open your website’s five highest-value pages. Give each page one original, sharp image that shows the actual product, result, location, or process. Rename each file in plain words, then write a one-sentence image description in your website editor. That description is often called alt text, and it tells search systems what the picture contains. Pair this with the Google AI Overviews content build and the AI recommendation visibility setup.

Bonsai 27B moves private AI work onto local devices

PrismML released Bonsai 27B on July 14 in two compact versions. The company announcement lists a 3.9 GB 1-bit version and a 5.9 GB ternary version. “Bit” describes how tightly the model’s stored numbers are compressed. Both accept text and images, and the downloadable model files use the Apache 2.0 license.

This is not a simple replacement for ChatGPT. Running a model locally means the AI works on your own device instead of sending every request to a cloud provider, but setup and quality checks become your responsibility. The practical opportunity is a narrow private job: sorting internal documents, classifying feedback, or drafting from sensitive source material without uploading that material elsewhere.

Do not install it across the company yet. Ask whoever manages your systems to test the 5.9 GB version on copied, non-sensitive documents and compare 20 outputs with your current cloud model. Record accuracy, processing time, and where the files travel. If you cannot explain that path, keep the model away from live customer records. Use the AI automation tools comparison to decide where a local model would sit beside the rest of the workflow.

On the bench

Claude Code versus n8n is worth mapping next: use code for flexible one-off work and a visible automation tool for repeated jobs an owner needs to inspect.

Zapier’s new AI pilot survey deserves a process check. Pick one stalled experiment and write down its owner, required data, approval step, and weekly operating cost.

Google’s image-generation rollout needs a brand review once it reaches your account. Test whether generated concepts save briefing time, while keeping real product and customer imagery as the source of truth.