AI Hallucinations Are a Test-Taking Problem, Says OpenAI
OpenAI researchers argue that language models hallucinate because current training and evaluation methods statistically reward guessing over expressing uncertainty.
Research
A collection of our articles covering the latest in AI research. Get comprehensive analyses and insights into the cutting-edge AI models and breakthroughs from leading tech companies and research labs.
Microsoft's MAI-DxO Crushes Doctors at Medical Diagnosis while Cutting Costs
Microsoft researchers developed an AI system that achieved 85.5% accuracy on challenging medical cases versus 20% for human physicians, while reducing estimated diagnostic costs in a controlled study.
DeepMind Launches AlphaGenome to Better Understand Gene Regulation
DeepMind’s AlphaGenome AI model decodes how mutations affect non-coding DNA, potentially transforming our understanding of disease.
Databricks Co-founder Launches Laude Institute, Stakes His Own $100 Million
Databricks and Perplexity co-founder Andy Konwinski is putting $100 million of personal funds behind Laude Institute, a grant-style venture that’s kicking off with a five-year, $15 million commitment to an AI Systems Lab at UC Berkeley.
Google's DeepMind Unveils AlphaEvolve, an AI System that Designs and Optimizes Algorithms
Google DeepMind’s AlphaEvolve is an AI agent that evolves algorithms with LLMs and automated evaluation, already improving infrastructure, math problems, and its own AI training systems.
AI in Higher Ed: 7 Major Takeaways from OpenAI’s ChatGPT Student Usage Report
More than one-third of college-aged students in the US use ChatGPT, with a significant portion relying on it for learning. OpenAI’s new report reveals how this impacts workforce readiness, economic competitiveness, and the role of AI in higher education.
Microsoft’s Muse AI Model Can Generate Video Game Environments in Real-Time
Microsoft has unveiled Muse, a generative AI model designed to create video game environments that respond to player actions, potentially transforming game development and preservation.
Google Unveils AI Co-Scientist to Accelerate Scientific Discovery
Google has introduced an AI co-scientist built on Gemini 2.0, designed to assist researchers in generating novel hypotheses, synthesizing literature, and formulating experimental plans, with early access available through the Trusted Tester Program.
NVIDIA and Arc Institute Introduce Evo 2, A State of the Art Foundation Model for Biology
The Arc Institute and NVIDIA unveil Evo 2, the largest AI model for biology ever, trained on 9.3 trillion DNA base pairs from over 128,000 species to advance synthetic biology, medicine, and genome design.
OpenAI's New Benchmark Tests AI Models Against Real-World Software Engineering Tasks
OpenAI has launched SWE-Lancer, a benchmark evaluating AI models on over 1,400 real-world software engineering tasks sourced from Upwork, collectively worth $1 million in payouts.
Adobe's New AI, TransPixar, Adds Transparency to Generated Videos
TransPixar generates videos with RGBA channels, integrating transparency for effects like smoke, fire, and glass reflections.
Apple's Ferret-UI is an AI that Can Understand and Navigate Mobile Interfaces
Ferret-UI can discuss specific parts of the interface, infer the overall function of a screen, and even reason about how a user might interact with it.
Living Skin Makes Robots More Human-Like
To show how versatile their technique is, University of Tokyo researchers made two prototypes: a 3D facial mold fully covered in living skin, and a simple robotic face that can smile.
PhysDreamer Brings Realistic, Interactive Dynamics to AI-Generated Video
The key idea behind PhysDreamer is to generate a plausible video of the object in motion and then optimize the material properties to match this synthesized motion.
Microsoft's VASA can Create a Realistic Talking Head Video from a Single Photo
VASA demonstrates superior performance compared to existing methods, delivering high-quality video frames with precise lip-audio synchronization and a diverse array of lifelike facial dynamics.
Apple Introduces MobileCLIP, a State-of-the-Art Image-Text Model for Mobile Devices
Apple's multi-modal reinforced training demonstrates 10-1000x improved learning efficiency compared to standard CLIP training.
Anthropic Shares Research on Technique to Exploit Long Context Windows to Jailbreak Large Language Models
Many-shot jailbreaking works by prompting the model with a large number of fictitious question-answer pairs that depict the AI assistant providing harmful or dangerous responses.
Google DeepMind's New Research Shows LLMs Can Outperform Humans in Fact-Checking
The research introduces a novel method for evaluating long-form factuality in LLMs, demonstrating that bigger models are more factual and that LLMs are 20 times cheaper than human annotators.
NVIDIA's LATTE3D Can Generate 3D from Text Prompts in Seconds
Like a virtual 3D printer, LATTE3D turns text prompts into 3D representations of objects and animals within a second.
Meet SIMA: Google's New AI that Can Play Video Games with You
The research is building towards more general AI systems and agents that can understand and safely carry out a wide range of tasks in a way that is helpful to people online and in the real world.
Tencent Proposes Technique to use LLMs to Make Diffusion Models More Accurate
By enabling diffusion models to more faithfully visualize the content of natural language prompts, this approach opens up exciting possibilities for creative applications and beyond.
Orca-Math Shows the Potential of Specialized Small Language Models
The research shows the value of smaller, specialized models in specific domains, where they can match or even surpass the performance of much larger models.
Meet EMO, Alibaba's New AI That Can Make Any Photo Talk or Sing
By enabling the generation of highly expressive and lifelike videos from a single reference image and audio input, EMO opens new avenues in entertainment, telepresence, and beyond.
Smart Gloves From MIT Can Teach You to Play the Piano
The adaptive smart glove is capable of sending tactile feedback to teach users new skills, guide robots with more precise manipulation, and assist in the training of surgeons and pilots.