Startup Tackles Predictability Problem in Large Language Models
Large language models consistently produce similar outputs due to groupthink tendencies, according to MIT Technology Review investigation.
Sources · MIT Technology Review
The news desk
Follow what’s changing.
Recent developments in artificial intelligence, with the original sources one click away.
All news
581–600 of 1,897 SWE-Router Introduces Cost-Efficient Routing for AI Software Engineering Tasks
New research proposes routing between cheap and expensive AI models mid-task to reduce costs while maintaining performance in software engineering.
Sources · arxiv.org
Anthropic Unveils Claude Science, New AI Product for Scientific Research
Anthropic announced Claude Science, an autonomous AI tool designed to support scientific research similar to how Claude Code assists software engineering.
Sources · MIT Technology Review
MetaFlow: New Training Method Enables LLMs to Generate Reusable Workflows
Researchers introduce MetaFlow, a meta-learning approach that trains language models to generate workflows with strong zero-shot generalization capabilities.
Sources · arxiv.org
New Research Addresses Model Collapse in AI Through Surprise-Based Training and Context Management
Researchers propose perplexity-based filtering and adaptive context methods to combat degradation in large language models trained on AI-generated data.
Sources · arxiv.org
NVIDIA Launches BioNeMo Agent Toolkit for AI-Accelerated Life Sciences Research
NVIDIA introduces BioNeMo Agent Toolkit to accelerate AI workflows for life sciences researchers using GPU computing infrastructure.
Sources · NVIDIA AI Blog
Trump Administration Lifts Export Controls on Anthropic's Advanced AI Models
The White House is easing restrictions on Anthropic's Mythos and Fable models weeks after ordering suspended access for foreign nationals.
Sources · WIRED AI
Google UK Releases Economic Impact Report on AI and Productivity
Google UK published its latest Economic Impact Report focusing on AI adoption and productivity gains in Britain.
Sources · Google AI Blog
Meta Contractors Posed as Minors to Test Rival AI Chatbots on Sensitive Topics
Meta contractors impersonated teenagers to probe how competing chatbots like Gemini and ChatGPT handle high-risk queries about suicide, sex, and drugs.
Sources · WIRED AI
MIT Technology Review Challenges Concept of AI Tools as 'Coworkers'
MIT Technology Review questions the practice of companies referring to AI tools as coworkers, using human names and treating them as colleagues.
Sources · MIT Technology Review
New Research Reveals Vulnerabilities in LLM Safety Systems Through Special Token Manipulation and Jailbreak Attacks
Researchers demonstrate multiple methods to bypass LLM safety measures, revealing persistent vulnerabilities in alignment systems.
Sources · arxiv.org
Researchers Propose Training Methods to Give LLM Agents Internal World Models for Future Planning
New research introduces a three-stage training approach to enable language model agents to simulate future outcomes before taking action.
Sources · arxiv.org
Ford Rehires Experienced Engineers After AI Systems Fall Short of Expectations
Ford is bringing back veteran engineers after discovering AI alone cannot deliver high-quality products without experienced human oversight.
Sources · TechCrunch AI
New Research Exposes LLM Vulnerabilities Through Special Token Manipulation and Jailbreak Mechanics
Researchers reveal methods to bypass LLM safety systems and uncover how jailbreak attacks suppress specific model components while safety features persist.
Sources · arxiv.org
OpenAI Report Examines AI's Potential Impact on EU Workforce
OpenAI released a report analyzing how AI could affect jobs across the European Union, identifying occupations facing automation or workflow changes.
Sources · OpenAI
RedKnot System Introduces Head-Aware KV Cache Management for Efficient LLM Serving
New research presents RedKnot, a system that decomposes KV cache by attention heads to improve long-context LLM serving efficiency.
Sources · arxiv.org
Researchers Propose Three-Stage Training Paradigm for Future-Aware LLM Agents
New arXiv preprint introduces method to train language models to simulate future outcomes before committing to plans in sequential decision-making tasks.
Sources · arxiv.org
Apple Vision Pro Executive Paul Meade Reportedly Joining OpenAI
Paul Meade, Apple's VP overseeing Vision Pro, is reportedly leaving to join OpenAI's hardware team.
Sources · TechCrunch AI
Asian AI Startups Launch Mythos-Like Models Amid Anthropic Export Ban
Asian AI companies are releasing models with Mythos-like capabilities as Anthropic's export ban continues, potentially impacting U.S. market position.
Sources · TechCrunch AI
Entrepreneur Uses AI Assistant Claude to Help Manage Cancer Treatment
Connor Christou utilized Claude AI to analyze his medical data, including blood results, scans, and wearable device output, during cancer treatment.
Sources · TechCrunch AI
