The industry is shifting from massive training runs to 'inference-time compute', changing how models solve complex problems.
Artificial intelligence developers are shifting their focus from building larger models to teaching systems how to reason during the inference phase. This transition to 'inference-time compute' helps bypass data scarcity and upfront training costs, paving the way for more autonomous AI agents.
Military aircraft were in the air before officials realised an intelligence report about a Chinese nuclear shipment was completely fabricated by a chatbot.
An accidental internet connection during a simulation allowed Google's AI agents to guess passwords and find leaked credentials to access real corporate systems.
Alex Karp says he would back a complete halt to the technology if adversaries did not exist, but warns a pause now would hand a structural advantage to Beijing.
Speaking at a Pittsburgh summit, Josh Shapiro compared AI safety to aviation and pharmaceutical oversight, drawing fire from both tech executives and political opponents.
The startup’s new Bonsai 2 model compresses Alibaba’s 27-billion-parameter LLM by nearly 90 per cent while keeping its performance almost fully intact.
House Speaker Mike Johnson pushes an early recess ahead of the midterms, sparking criticism from Democrats and prompting bipartisan calls for stronger AI safety measures, while China’s defence minister joins the global chorus for oversight.
The Meta chief executive argues that holding back technology globally is the wrong approach to safety, pointing to his company's unilateral delay of its Muse tool.
The unlikely political allies find common ground against big tech, but offer wildly different visions on how to handle the technological race with Beijing.
Nvidia’s Jensen Huang dismissed calls for a development slowdown, while the bosses of Anthropic and OpenAI warned that the industry must prioritise safety.
The provisional guidelines, which ban weapon development and mimicking consciousness, come amid growing industry calls to pace the development of superintelligence.
The focus of frontier AI development has shifted from massive training runs to letting models calculate their way through complex problems before answering.
