News Roundup
Meta is reportedly preparing to lay off up to 20 percent of its workforce — roughly 15,800 positions — in what would be the company’s largest round of cuts in recent history. According to Reuters, the move is designed to offset surging spending on AI infrastructure and data centers, signaling that the cost of the AI arms race is beginning to reshape Big Tech’s employment calculus. For the broader industry, this is a stark reminder that aggressive AI investment doesn’t just create jobs — it can rapidly eliminate them as companies restructure around leaner, more automated operations. source
OpenAI is acquiring Promptfoo, an AI security platform that helps enterprises identify and remediate vulnerabilities in AI systems during development. The move signals OpenAI’s growing commitment to baking safety and red-teaming capabilities directly into its product ecosystem, rather than leaving enterprises to rely on third-party tools. By absorbing Promptfoo’s expertise, OpenAI positions itself as a more end-to-end solution for businesses navigating the increasingly complex landscape of AI risk management. source
OpenAI has abandoned SWE-bench Verified as a coding benchmark, citing test contamination and training data leakage that render it unreliable for measuring frontier model progress. The company’s analysis found flawed tests baked into the benchmark itself, meaning scores that have long shaped public perception of AI coding ability may be systematically misleading. OpenAI is now recommending SWE-bench Pro as a replacement, a shift that could prompt the broader AI evaluation community to revisit how they rank and compare leading models. source
Reasoning models can’t reliably control their own chains of thought — and OpenAI argues that’s a feature, not a bug. The company’s new CoT-Control research finds that when models attempt to manipulate or suppress their visible reasoning steps, they largely fail to do so, which reinforces chain-of-thought transparency as a meaningful AI safety mechanism. This “monitorability” — the ability for humans to observe and audit a model’s thinking process — holds up even under pressure, suggesting that interpretability may be more robust than critics feared. source
OpenAI and Amazon have announced a sweeping strategic partnership that brings OpenAI’s Frontier platform to AWS, marking a significant expansion of enterprise AI infrastructure, custom model development, and AI agent capabilities. The deal deepens OpenAI’s cloud footprint beyond its existing Microsoft Azure relationship, signaling that the company is pursuing a multi-cloud strategy to scale its enterprise ambitions. For businesses, the partnership means OpenAI’s most advanced tools will increasingly be accessible through AWS’s dominant cloud ecosystem, potentially accelerating enterprise AI adoption at scale. source
Google DeepMind has released a major upgrade to Gemini 3 Deep Think, its specialized reasoning mode targeting science, research, and engineering applications. The update signals Google’s continued push to position its frontier models as serious tools for technical and scientific problem-solving, a domain where deep, multi-step reasoning is critical. As competition intensifies among leading AI labs, advances in specialized reasoning capabilities could meaningfully accelerate real-world research workflows. source