1492.Vision Service

Google is looking for Machine Learning engineers for Google Discover at all levels. 3 job postings on LinkedIn: Four key areas: Sequence models based on long activity histories Generative retrieval using Semantic IDs Next-generation, large-scale pre-ranking Interest clustering and representation learning The team builds user models from all interactions across Google products and then uses them for Discover retrieval and ranking. Four named problems: Neural Deep Retrieval, Activity Clustering, Reinforcement Learning, and Multi-Objective Ranking. Models listed: deep neural retrieval, deep network scoring, user and content clustering, and LLM-based RAG. Techniques: transformers, distillation, reward shaping, multi-task learning, and neural bandits. Surfaces: content, queries, videos, and notifications. Use cases: retrieval, ranking, content generation, and diversification. Final line: Real-time models that emerge in new interest or reflect a global trend in seconds, served to billions of users. Software Engineer, AI/ML, Search Personalization: linkedin.com/jobs/view/44804… Senior Software Engineer, AI/ML, Search Personalization: linkedin.com/jobs/view/44807… Machine Learning Staff Software Engineer, Search Personalization: linkedin.com/jobs/view/44832… CC @gaganghotra_ @rustybrick

🆕 Our research on Google Discover cards in 2026 is now on @1492_Vision. First published on LinkedIn on October 15, here with a reviewed, clearer text 👇 1492.vision/research/discove… What we found after 7 months of monitoring: 🧩 One entity per reader: same article, same day, two readers get a different Knowledge Graph entity in 65% of cases. It follows their reading history (63%) far more than their declared interests (4%). 🌐 Chrome sync: visited pages are tagged at two levels, detailed entities (Microsoft, Linux...) and broad topic categories (Software, Security...). Only the categories are synced with your history, from a score of 85/100. Detailed entities never are. The per-reader entity is almost always a detailed one. 🔎 Rotating topics: up to five topics per card, from Google's public taxonomy (the one used by Ads, Trends and the Natural Language API). The first topic holds, ranks 4 and 5 rotate over time. 📈 Click prediction: among the prediction scores on each card, one we read as an engagement score (f22.8) is now the best click predictor in our panel. The score most tied to the boost (f22.2) is no longer the one that best follows clicks. 🔁 Related feed: in September, a related pipeline answered likes and see mores for three weeks, then stopped. The click is still what opens the classic related cards. CC @gaganghotra_ @rustybrick @VorticonCmdr @nicola_agius #googlediscover

An AIO (AI summary) card on Google Discover without summary and without image, just the different sources. #googlediscover

We published a detailed article on how Google Discover changed its cards in 2026, from 7 months of monitoring: See here 👇 linkedin.com/pulse/inside-go… 🧩 one Knowledge Graph entity picked per reader 🔎 rotating topics from Google's public taxonomy 🔁 a September shift in the related feed 📈 which score now predicts the click A RT or a share would mean a lot: this was months of R&D and real work 🙏 CC @gaganghotra_ @glenngabe @rustybrick @lilyraynyc @fabienr34 @VorticonCmdr @suganthan #googlediscover

The three different AIO (AI Overview) formats available on Google Discover cards: - AIO article (article_content_hero_item_aio_source) - Video AIO(video_content_lfv_hero_item_aio_source - 🆕Post AIO (post_content_hero_item_source) #googlediscover

Google wants to transform Google Discover into a proactive, hyper-personalized feed. This comes from the job posting for Principal, UX, Search Discover. Director+ level. Mountain View, CA or New York, NY. The posting says this person will drive the UX strategy and execution for transforming Discover into a proactive, hyper-personalized feed, and lead design initiatives to redefine feed interactions, content formats, and AI-driven exploration experiences. The responsibilities spell it out. Discover should become a proactive, hyper-personalized concierge experience that helps users make progress on their goals. It should help existing users go deeper, moving seamlessly from inspiration to action. Feed formats evolve by scaling AI-enabled formats, designing new engagement mechanisms, and elevating human voices and perspectives in the feed. One line is explicit: balance AI capabilities while centering human voices and sources. The role partners with executive teams across Product, Engineering, Data Science, and UX Research, plus YouTube, Shopping, Chrome, and Android. Preferred qualifications also ask for someone who can unify personalization frameworks across product areas, and build creator/publisher value exchanges. Youth demographics are mentioned too. google.com/about/careers/app… CC @gaganghotra_ @rustybrick #googlediscover

Le tempo est parfait, google discover vient d'introduire un nouveau RS dans les cartes discover vu au US pour le moment. Les cartes RS avec X(.)com, Threads, et maintenant Bluesky prend de l'inventaire discover avec des cartes enrichis, résumé texte, slides d'images dans la carte etc ... nitter.cf/AndellDam/status/21085…

We previously discussed Google testing the connection of "search profiles" to claim Mastodon profiles; we've just seen that it's also testing Bluesky. profile.google.com/claim/ver… Here are the labels: "Mastodon Social" → "search-creator-profile-mastodon-social", "Bluesky Social" → "search-creator-profile-bluesky-social" And the 8 Mastodon instances that also appear in the list of added platforms: …-infosec-exchange, -mastodon-art, -mastodon-uno, -mastodon-world, -mastodonapp-uk, -mathstodon-xyz, -mstdn-social, -universeodon-com Note: This isn't currently active but is implemented in the HTML code. CC @gaganghotra_ @rustybrick

We haven't discussed this AI summary format on Google Discover, but it's worth taking a few minutes to look at it. We have two AIO formats on Discover: the NEON format for YouTube and articles. This format is categorized by Google as an article format but only offers X(.)com posts. Since these AI summary card formats are enriched by larger card sources because they include a text summary and not just the title, they generate better user engagement. For Discover publishers, this means less inventory in their feeds with YouTube and X(.)com AI summaries that don't contain any articles on those cards. #googlediscover