AI in practice
1. Your Mac deserves a guest list
Give AI the job. Keep the keys. Apple says stronger Full Disk Access controls are coming, requiring very explicit user action. That permission can expose files, mail, messages and browsing history; Apple cites increasingly autonomous AI agents. Its October 2 announcement gives no release date.
Chrispy has felt both sides: ChatGPT sometimes says it lacks access, then surprises him by doing things on its own. He values the activity and screen views showing what AI is working on, and sees room for novel ways to make that activity clearer. I’d pair those views with a question before granting access: what does this job actually need to see?
2. The creative playground gets a meter
Figma is a shared design workspace for websites, app screens and graphics, with clickable prototypes for testing ideas before building. Its Design agent generates design directions and makes bulk edits; Weave tools generate images and vectors. Both are scheduled to leave open beta October 6 and start using the shared AI-credit pool. Agent usage varies by task and context; Weave costs vary by tool. I’d test one representative job before a whole campaign, tracking credits and corrections.
Figma: credit change · Figma: design workspace
3. Stop retyping your working method
Imagine saving your best AI instructions as reusable recipes. Google Skills brings them into an ordinary Gemini conversation: combine a brand-voice skill with a newsletter-format skill to turn rough notes into a consistent draft. You still supply the facts. Workspace’s rollout starts October 5 for Rapid Release domains and October 19 for Scheduled Release; the Gemini app starts October 13. Access depends on plan and rollout. Skills don’t currently sync between Workspace and the app, so each needs its own copy. I’d test the combination on three drafts.
4. A camera assistant that helps you aim
Where’s the pepper in a crowded spice cabinet? Guided Vision in Gemini Live helps blind and low-vision users answer questions through an Android phone’s built-in camera. Spoken cues help you pan, tilt or step back. You point and talk; everyday phone skills suffice. Announced October 1, it is rolling out on Android 9+ in supported regions and languages. Enable it in Gemini’s settings and try a familiar object. Answers can be wrong: don’t use it for navigation, obstacle detection or replacing mobility aids.
Google: Guided Vision · Google: setup and limits
5. Pretty objects need working parts
An AI-generated mug can look convincing with its top sealed shut. MIT’s InstructMesh research lets you paint the faulty region, request a repair in words or adjust sliders, then preview changes before fabrication. In one demonstration, adding and refining internal tunnels turned a seashell-shaped model into a working whistle. That’s a satisfying reality check. Reported October 1, with a November conference presentation planned, this remains research rather than a verified consumer launch. You still need to spot the flaw. I’d inspect openings, dimensions and function.
MIT: InstructMesh · InstructMesh: research paper
Practical AI projects
Three small experiments worth trying
Proposed experiments with estimated effort. Use existing tools and invented or shareable material.
- The layout squeeze test. Invent an exhibition announcement with six required facts. Ask AI to adapt it into a poster, square card and narrow mobile banner in your existing design tool. Allow 30–40 minutes. Check every fact and view each design at its intended size. Keep the venue fictional; nothing needs publishing.
- The two-part recipe. Write one instruction for a fictional workshop’s tone and another for its announcement format. Test each separately, then together, on three invented workshops. About 25 minutes in any text chat. Score factual preservation, format and tone yourself. Keep the winning recipe only if combining instructions preserves every requirement.
- Cardboard before plastic. Ask AI for a cardboard phone-stand pattern using dimensions you supply. Draw it full-size, check every unit, then make a mock-up from scrap card. Test with a phone-sized, similarly weighted dummy over a soft surface. About 45 minutes with chat, ruler and card. The output is an unverified prototype.
Figma: creative tools · Google: reusable instructions · MIT: fabrication research