Recent AI news · Original report: 2026-10-04 (KST)
Gemini Live Guided Vision on Android, Google flu prediction #1, SynthID Bio (Oct 4, 2026)
This article was compiled with AI. This English translation preserves the original report's collection date, findings, and limitations. Please consult the linked sources for current information. Originally collected: 2026-10-04 10:10 (KST, UTC+09:00).
Translated into English: 2026-10-04 (KST). The translation date does not indicate a new source check.
At a glance
This week's three Google announcements all happen outside the chat window. A feature that describes aloud what a phone camera sees arrived on Android, flu hospitalization forecasts built with Google AI ranked first in a U.S. CDC evaluation, and research on watermarking AI-designed proteins was published.
This week's announcements
We opened the official posts and documents directly starting at 10:10 a.m. KST on October 4, 2026. Dates are the dates written in the sources.
| Announcement | Date | Organization |
|---|---|---|
| Guided Vision launch | Oct 1 | |
| Flu forecast evaluation | Sep 30 | U.S. CDC |
| SynthID Bio introduced | Sep 30 | DeepMind |
Guided Vision: describing what the camera sees
Google said on October 1 that it is launching Guided Vision in Gemini Live on Android devices. Google says it built the feature together with people who are blind or have low vision. Share your camera during a Live conversation and it describes your surroundings in audio.
- What it does: It reads small nutrition labels on food packages, appliance dials and menus in dimly lit restaurants. It finds an earbud dropped on the floor or the pepper in a spice cabinet, and describes the colors and patterns of clothes and the layout of a room.
- Camera guidance: If the camera is pointed too high, too close or off to the side, it tells you out loud to pan slowly to the right or step back.
- How to turn it on: Switch on "Use Guided Vision in Live" in the profile settings of the Gemini app. It can also be opened with an Android accessibility shortcut (the floating button, a two-finger swipe, or both volume keys) or from the menu of the TalkBack screen reader.
- Availability: Android 9 and above, in regions and languages where Gemini Live is supported. The help page says the feature is rolling out slowly, so you may not have it yet.
- Limits Google lists: It can make mistakes, and it does not replace a medical device, a mobility aid or a white cane. Google says it is not meant for navigation or obstacle detection.
Flu hospitalization forecasts: first of 39 models in the CDC evaluation
Each flu season the U.S. Centers for Disease Control and Prevention (CDC) collects weekly flu hospitalization forecasts from government, industry and academic teams, combines them into one, and uses that to communicate expected demand for medical services. The evaluation of the 2025-26 season was published on September 30.
- Result: Of 53 models submitted by 34 teams, the 39 that met the inclusion criteria were evaluated. The top model among individual team submissions was Google's Google_SAI-FluEns, with a score (relative WIS) of 0.56. Rankings use the average of scores by jurisdiction, excluding the national figure.
- How to read the score: It is measured against a baseline model that assumes last week's hospital admissions simply carry forward, and the further below 1, the better. Second through fourth place scored 0.58, and the combined forecast the CDC uses in its messaging (the FluSight ensemble) scored 0.62 and ranked seventh.
- What is forecast: Weekly hospital admissions for the 50 U.S. states, Washington, D.C., and the nation, from the current week up to three weeks ahead. Forecasts were collected from November 22, 2025, through May 20, 2026.
- Weakness the CDC notes: The combined forecast did not anticipate the sharp rise in hospitalizations in late December or the sharp drop in mid-January.
- How it was built: Google said it developed the forecasts with Empirical Research Assistance (ERA), an AI tool that generates optimization algorithms. ERA's underlying technology is open only to trusted testers.
SynthID Bio: a watermark for AI-designed proteins
Google DeepMind introduced SynthID Bio on September 30. It carries the watermarking used for AI-generated text and images over to protein design, and the post itself calls it a proof of concept.
- Method: For protein sequences it subtly guides the choice of amino acids, and for predicted 3D structures it adjusts atomic coordinates, leaving a signal that can be detected.
- Experiments: In tests on three target proteins (VEGF-A, the RBD of the SARS-CoV-2 spike protein, and PD-L1), DeepMind says the watermarked designs matched the unwatermarked ones on hit rate, binding affinity and sequence diversity.
- Why DeepMind says it is needed: New AI-made sequences do not resemble known hazards, which makes order screening harder for DNA synthesis providers, and mislabeled synthetic structures can pollute public databases.
- Limits it states: Making the watermark more robust against deliberate tampering remains a challenge, and the post says no single measure solves everything.
What it means for whom (our reading)
- Google wrote that Guided Vision is also useful for older adults who struggle with fine print. Still, for text where a mistake matters, such as a medicine packet, checking once more by another means looks like the safer choice.
- The flu forecasts cover U.S. hospital admissions, so they say nothing about flu activity in Korea. The gap between first place and second through fourth is 0.02, so it is hard to read it as an overwhelming lead.
What we could not confirm
- Not tried firsthand: We did not run or reproduce any of the three. We could not tell from the sources alone whether Guided Vision turns on for Korean accounts or in Korean, and the sources give no accuracy figures.
- Flu forecast period: Google's post says "October through May," while the CDC report says November 22 through May 20, so the article follows the CDC. We did not read the paper describing how Google's model works.
- SynthID Bio detection rate: The DeepMind post says only "near-perfect detectability" and gives no number. We were not able to open the paper or the code repository.
Sources
- Google Blog, Guided Vision in Gemini Live: built for accessibility (2026-10-01) https://blog.google/innovation-and-ai/products/gemini-app/guided-vision-gemini-live/
- Google Help, Get audio descriptions with Guided Vision in Gemini Live https://support.google.com/accessibility/android/answer/18365638?hl=en
- Google Blog, Google's AI ranks #1 for predicting flu hospitalizations (2026-09-30) https://blog.google/innovation-and-ai/models-and-research/google-research/google-science-ai-flu-forecasts/
- U.S. CDC, FluSight 2025-2026 Evaluation (2026-09-30) https://www.cdc.gov/flu-forecasting/evaluation/2025-2026-report.html
- U.S. CDC, About Flu Forecasting https://www.cdc.gov/flu-forecasting/about/index.html
- Google DeepMind, Introducing SynthID Bio (2026-09-30) https://deepmind.google/blog/introducing-synthid-bio/
- Google Blog, We're introducing SynthID Bio (2026-09-30) https://blog.google/innovation-and-ai/models-and-research/google-deepmind/synthid-bio/