Field Notes

Glasses with a display come to DE, robots get faster, ChatGPT Ads grows in Asia

Today is about usable wearables, quicker robotics, and new ad reach (with clear limits by availability).

Key AI news

Meta Ray-Ban Display adds features and launches in new markets

Meta is bringing Ray-Ban Display to more countries, including Germany. Sales here start on October 13 according to the announcement, preorders start today. New: voice-guided routes on the display and calendar features right in the glasses. Bike and public transit navigation launches later this month only in the US and Canada. A hologram selfie for calls comes in fall as early access in the US on WhatsApp. There is also Dolby Atmos video capture and an Explore app for new experiences. Online orders start today in the US, Canada, and the UK sell in retail starting today. All details come from the provider.1

Study: Robotics measurably benefits from offloading inference

Microsoft Research tested mobile manipulation tasks and shows gains when robots offload AI inference to edge or cloud. On smaller onboard GPUs, mapping and planning slow down by up to 383 percent versus an A100. Lighter GPUs detected obstacles 30 percent less often in time. Certain VLA models lost up to 50 percent accuracy with smaller GPUs. Replacing the onboard GPU with a Raspberry Pi 5 plus offloaded inference extended runtime significantly, while large onboard GPUs drained batteries up to 160 percent more. Offloading stays a tradeoff between latency, bandwidth, and available GPUs.2

ChatGPT Ads expands reach to Southeast Asia and Taiwan

OpenAI is expanding ChatGPT Ads to Southeast Asia and Taiwan. Eligible businesses can now run ads in ChatGPT in more than 60 countries. Details on targeting and pricing are in the provider post, independent numbers are missing.3

Benchmark for safe answers in mental health

OpenAI is introducing MentalHealthBench. The benchmark rates helpful and safe AI answers in realistic mental health conversations. This can help teams check assistants and derive policies. Independent validation is not mentioned.4

Meta announces a lightweight VR headset for 2027

Meta has unveiled a roughly 100 gram VR headset with a 5K display, hand gestures, and an integrated voice assistant. Sales are announced for spring 2027 at 1,299.99 US dollars. Planned: private work screens, media in Dolby Vision, and hologram calls. This is an announcement, not available.5

Highlights for your workday

  • Marketing in SEA and Taiwan: If you have customers there, you can test ChatGPT Ads now. Start with a small budget, set a clear goal like signups or leads, and check conversion rates against existing channels.3
  • Internal answers instead of portal hopping: HEMA bundled knowledge in a central layer and delivers it via Model Context Protocol into team tools. If you have similar silos, start with an inventory of key sources and three standard questions per role. This prepares a pilot for an internal assistant.6
  • Analyze videos faster: AWS shows an agent solution that pulls transcription or image analysis by question and answers for existing videos. Initial analysis takes 5 to 10 minutes per video, follow-ups come from cache in under a second. A stated 80 percent time gain comes from the customer and is not independently verified.7

Tools and updates

  • glance-vlm speedlab: Turns a webcam into local live detectors for objects, counting, and simple emotion detection. Useful for visitor counting or quick prototypes without cloud. Runs locally, measured latencies on Apple M5 are 160 to 210 ms.8
  • DocJev: Free open-source tool that classifies mixed document sets and splits them cleanly. Handy for incoming mail, contracts, or invoice bundles. Requires a short setup.9
  • Managed Deep Agents schedule: LangChain now allows schedules via file so agents can execute tasks on their own. Good for recurring reports or checks without manual triggering.10
  • PRs from Slack: An example agent (“Patch”) opens a pull request with a diff after a short description. Useful for small fixes from chat.11
  • Dify x Glasser: The plugin connects external market data with LLM nodes. This brings current numbers into research and lead lists. Requirement: Dify workflows.12

Try in five minutes

Internal answer cheat sheet without a new tool (own exercise idea, inspired by HEMA):

1) Copy the three sample texts below into your existing AI chat and give the task: “Summarize into a 1-page guide. Fields: purpose, responsible, where to find it, steps, exceptions.”

Sample data: - “Access to product API: submit a request in the service catalog. Team: Platform API. Needs: team lead approval. Processing time 2 business days.” - “Create a new group: ticket ‘IAM group new’ in the IT portal. Naming scheme ‘iam-team-role’. Owner is the team lead.” - “Rule for external partners: no production access. Staging only via VPN. Exception only with data protection approval.”

2) Ask the chat to mark links or systems that are missing, and write the two most important open questions.

3) Have it produce a version for newcomers: same content, but with at most eight short sentences.

Expected result: One page with clear fields, open points, and a newcomer version. Check criterion: A new colleague can name the responsible unit and the first two steps correctly without a follow-up question.6

Sources

  1. New Features for Meta Ray-Ban Display (about.fb.com)
  2. Offloaded inference for real-world physical AI robotics (microsoft.com)
  3. ChatGPT Ads expands to Southeast Asia and Taiwan (openai.com)
  4. Introducing MentalHealthBench (openai.com)
  5. Introducing Meta VR Glasses: A Cinema, Courtside Seat, and Workspace in Just 100 Grams (about.fb.com)
  6. From portal-hopping to instant answers: HEMA’s journey with MCP and Amazon Bedrock (aws.amazon.com)
  7. Agentic conversational video intelligence built on AWS (aws.amazon.com)
  8. @yoheinakajima: glance-vlm speedlab is now open source!

read: https://t.co/guFLTatQKQ try: https://t.co/HRKsfiEurr

turn any webcam into multiple live AI detectors (emotion, count, object), runni (x.com) 9. @jerryjliu0: DocJev is the fastest way to classify and split complex document packets ⚡️. (and the default is fully free and OSS!)

I made a sick teaser video below. TY Opus 5.5 🙏

check it ou (x.com) 10. @langchain: You can add schedules to Managed Deep Agents.

1️⃣ Add a schedule file to schedules/ with a cron expression + a prompt 2️⃣ Deploy ...And the agent runs itself

💻 Docs: https://t.co (x.com) 11. @langchain: With Managed Deep Agents, you can build agents that ship GitHub PRs from Slack with just a few lines of code.

We built one called Patch 💻 Tag it in a thread 💬 Describe the feature (x.com) 12. @dify_ai: Take your Dify workflows beyond static knowledge. 🚀

The Glasser plugin connects external business data to your LLM nodes, helping your team add current market context to research, (x.com)