Frontier
Papers, benchmarks, and technical breakthroughs with real downstream impact.
Monitor
Google Research unveiled TurboQuant, a novel algorithm slashing LLM key-value cache memory by 6x and boosting inference speed up to 8x without quality degradation, pushing AI efficiency frontiers.
Marketeam.ai demos RL-KPI at GTC 2026, extending RL to non-deterministic business outcomes like revenue growth.
New data shows long-term Claude users iterate more carefully, attempt higher-value tasks, and achieve better success rates, but grant less full autonomy compared to new users.
Anthropic's engineering blog details a multi-agent system inspired by GANs, using planner, generator, and evaluator agents to enhance Claude's performance in complex, multi-hour software tasks. It improves coherence, quality, and functional correctness over single-agent approaches.
Allen Institute for AI (Ai2) launched MolmoWeb, an open-source agent using Molmo 2 vision models to navigate browsers via screenshots, enabling interpretable web tasks. It challenges proprietary agents amid Ai2's leadership transition.
Red Hat open-sources llm-d to the Cloud Native Computing Foundation, aiming to create a universal standard for running LLMs in cloud-native environments and accelerating AI adoption.
Security researchers jailbroke Anthropic's top Claude model rapidly, highlighting risks in agentic systems as they gain real-world tool access.
OpenAI's GPT-5.4 Pro has cracked a Tier 4 FrontierMath problem previously unsolved by humans or AI, as confirmed by Epoch AI. This milestone highlights advancing AI reasoning in pure mathematics research.
New challenging AGI test ARC-AGI-2 reveals limits of current models, with o1-pro and DeepSeek R1 scoring under 1.5%.
Sky's Tom Clarke explores if recent AI video advancements represent a true breakthrough, amid rapid progress in generative media.
Sky News explores if new AI video tools like Seedance 2.0 represent a pivotal moment in realistic video AI.