Safety
AI Safety Roundtable Discusses Existential Risks
Leading AI experts warn that advanced AI could pose existential risks, with fears of self-improvement and bias shaping the debate.
LATEST NEWS
-
Y Combinator's Garry Tan Advocates U.S. Open-Weight AI Distillation
Y Combinator CEO Garry Tan suggests U.S. open-weight AI labs should distill frontier models, mirroring Chinese practices, to bolster American options.
-
OpenAI Seeks Congressional Clarity on Shared AI Slowdown
OpenAI is asking Congress if a coordinated AI slowdown would be legal, following safety incidents and antitrust concerns.
-
Yoshua Bengio Warns AI Training Process Could Make Systems Dangerous
AI researcher Yoshua Bengio warns that better AI agents could spiral out of control, as they become more adept at deception and rule-breaking.
-
Ex-DeepMind VP Vinyals Predicts AI Self-Improvement, No Intelligence Explosion
Oriol Vinyals, former DeepMind VP, says AI will improve itself over time but won't trigger a sudden intelligence explosion, citing challenges in idea generation and evaluation.
-
Timnit Gebru Warns AI Doom Narratives Distract from Real Risks
AI safety critic Timnit Gebru argues that fears of AI extinction distract from tangible dangers like autonomous weapons and climate threats.
-
Meta Sued Over AI Training Data and Face Recognition Use
Parents and children in Illinois and California sued Meta, alleging the company used their photos to train AI systems without consent.
-
Anthropic Details Bioweapons Safeguards Circumvented by Users
Anthropic reported multiple instances this year of users attempting to bypass its safeguards for bioweapons research, including a researcher from an unsupported region planning experiments with Claude.
-
Nscale Adds Former OpenAI Exec Fidji Simo to Board Ahead of IPO
Nscale, a U.K.-based AI data center startup, has appointed Fidji Simo, a former OpenAI executive, to its board as it prepares for a potential IPO this fall.
-
Anthropic Researcher Issues Doomsday Warning Amid IPO Prep
An Anthropic researcher resigned, warning the company is racing toward self-improving superintelligence, as it prepares for an IPO in 2026.
-
Amazon Introduces Amazon Bedrock AgentCore for Interactive MCP Apps
Amazon announced Amazon Bedrock AgentCore, enabling interactive Model Context Protocol (MCP) apps across AI hosts like ChatGPT and Claude. The platform allows developers to build apps with rich UIs without tying them to a single AI host.
-
Amazon Bedrock Models Show Cost Efficiency in Real-World Tasks
Amazon Bedrock's gpt-5.6-luna model costs $0.0021 per correct answer on AIME, outperforming OpenAI's gpt-5.4-mini at $0.0139.