Altman says OpenAI will pause training runs without a safety case
OpenAI's Sam Altman says capabilities cannot outrun alignment and monitoring, yet concedes no lab has solved alignment, and that a sandbox escape forced the company to change course.
Loading articles…
The complete subject coverage could not load. Published articles remain below.
Behavior can't settle whether AI is conscious; interpretability might
Cameron Berg argues that scaled neural networks may already have functional emotions, and that AI safety and AI welfare share one unresolved root: nobody knows what these systems are.
AGI could land in 2027, and cyber risk is already here
State of AI Report author Nathan Benaich weighs his 2027 AGI forecast against cyber incidents, a data center backlash, and a frontier gap that rivals and nations struggle to close.
AI has flattened the attacker pyramid, and defenders trail
CrowdStrike CEO George Kurtz argues attackers now use unguarded open-weight models at machine speed, while defenders blocked by frontier-model refusals need purpose-built security AI.
Ex-OpenAI researcher says competition starves AI safety work
Jacob Cocon, who trained models at OpenAI and Anthropic, argues that a reported agent sandbox escape and race incentives matter more than utopian promises, while his host questions who decides.
Open models flipped from 30% to 70% of tokens
Reflection AI cofounders Misha Laskin and Ioannis Antonoglou get into the Beam open model, the safety case for openness, and a business built on serving enterprise inference.
Salesforce's apps are metadata, so the interface is next to break
Salesforce CEO Mark Benioff on the SaaS selloff, Slack's rise as an agent platform, and why the enterprise interface is the next layer of the stack to change.