Multimodal AI
Multimodal AI breakthroughs and products
Articles
The Download: Tricking LLMs and Reviving Geothermal Plants⭐8
A fundamental flaw makes large language models vulnerable to attacks that can't be fixed. Plus, reviving a failing geothermal plant with AI and novel sensing.
A fundamental flaw leaves LLMs strikingly vulnerable to attack⭐9
A fundamental flaw in LLMs enables simple attacks, tricking models into prohibited tasks like sabotage, with >95% success across major AI systems.
TraceCLIP: Recovering Local Semantics from Patch-to-CLS Contributions⭐8
TraceCLIP recovers local semantics from CLIP's patch-to-CLS contributions, achieving 1.3-4.5 mIoU gains on zero-shot segmentation without training or external m...
Weight and Height Estimation from a Single Human Image Captured⭐8
This paper introduces a new dataset and deep learning models for estimating BMI, weight, and height from a single unconstrained human image, showing
Knowledge-guided Disentanglement with Atomic Actions for Action⭐9
Proposes KDA, using LLMs to decompose actions into atomic actions with spatial-temporal semantics for enhanced, disentangled action recognition, achieving
How an Overlooked Geothermal Plant Got a Second Chance⭐9
An abandoned geothermal plant in New Mexico is revived by AI-driven startup Zanskar, now generating 25 MW of clean baseload power for thousands of homes.
The Download: A Chip Talent Battle, and Deflating AI Hype⭐8
Samsung engineers defect to SK Hynix for record bonuses amid an AI chip talent war, plus a reality check on deflating AI hype in 2026.
The AI Hype Index: Unsexy AI⭐8
MIT Technology Review’s AI Hype Index ranks the most overhyped and underrated AI trends, celebrating practical “unsexy” AI over flashy promises.
DisasterTD: Disaster Toponym Disambiguation Using Multimodal⭐7
DisasterTD integrates multimodal LLMs and cross-view geolocalization to disambiguate disaster toponyms, achieving 71.62% accuracy within 1 km on Hurricane Harve...
Gradient-Based Latent Decomposition Reveals Mechanisms of⭐7
A new method reveals why fine-grained malignancy cues degrade in weakly supervised mammography, with 95.6% of latent features lying orthogonal to supervisory si...