11-14 Daily Briefing

AI News Daily · 14 Nov 2025

AI News | Daily Briefing | Tools | Research | Industry | Open Source | Human Impact

Highlights

  • Baidu releases ERNIE 5.0, calling it the world’s first native multimodal model; Google upgrades Gemini Live; Fei‑Fei Li’s World Labs Marble 3D enters public beta; SOLO ships its full version with a limited free tier.
  • ElevenLabs launches celebrity-licensed voice clones, Google commits USD 6.4B to a German AI datacentre, and Sam Altman tweets that GPT‑5.1 is live—yet early testers flag severe hallucinations.
  • New papers cover profile-poisoning attacks on recommenders, multi-agent GeoSQL translators, and an LLM-powered surgical copilot.

Product & Platform

  1. ERNIE 5.0 — At Baidu World, Robin Li declared ERNIE 5.0 the first “native multimodal” model, trained to simultaneously understand text, images, and audio instead of stitching modules. He noted, “intelligence itself is the biggest application,” signalling Baidu’s plan to infuse ERNIE across its ecosystem. [Source]
  2. Gemini Live voice upgrade — Google rolled out controllable pacing, tone, and even novelty accents (“tell it in a cowboy voice”), transforming Gemini Live from a utilitarian assistant into an emotional dialogue partner for interview practice or language learning. [Source]
  3. Marble 3D (World Labs) — Fei‑Fei Li’s World Labs opened public beta for Marble 3D, letting creators generate editable, interactive 3D worlds from text, images, or video in ~10 minutes. [Try it]
    Marble 3D world
  4. SOLO full release — SOLO is now generally available with a major UX overhaul and a limited-time free plan so everyone can test the upgraded workflow builder. [Overview]
    SOLO feature map SOLO UI

Research Notes

  1. Profile-poisoning attacks (CREAT) — A new paper reveals “profile pollution attacks” that subtly alter existing user interactions (without mass bot accounts) to mislead recommender systems. The reinforcement learning framework CREAT optimises for stealth and efficacy, warning that RS security must be upgraded. [Paper]
  2. Multi-agent GeoSQL — Researchers propose a multi-agent pipeline for translating natural-language queries into complex spatial SQL. Entity extraction, logic planning, code generation, and verification each get a dedicated agent, making geo-data analysis far more approachable. [Paper]
  3. Surgical AI Copilot — An “AI copilot” for image-guided surgery combines LLM reasoning with perception modules to segment tumours, track tools, and provide real-time suggestions. [Paper]

Industry / Capital

  1. ElevenLabs × Hollywood — The unicorn phonetics firm partnered with several A-list actors to create authorised celebrity voice clones, signalling that synthetic media is entering the mainstream with the blessing (and fees) of talent.
  2. Google’s EUR datacentre — Google will invest USD 6.4B in a German AI datacentre to bolster European capacity and comply with local regulations.
  3. GPT‑5.1 launch + hallucinations — Sam Altman announced GPT‑5.1, touting better instruction following. However, community testers quickly shared hallucination-heavy transcripts—reminding teams to validate outputs despite version bumps.

Community Signals

  • Users praised Marble 3D’s ability to conjure immersive spaces but warned that reliability and safety benchmarks must keep pace—especially when GPT‑5.1’s hallucinations still surface hours after launch.
Last updated on
11-13-Daily