Beyond Fixed Budgets: Characterizing the Inelasticity and Limitations of Tree-of-Thought Reasoning Strategies
The study shows that when you squeeze compute tokens, Tree-of-Thought-style reasoning strategies don’t degrade gracefully – DPTS can stall from exploration issues while SSDP burns its frontier through heavy semantic merging. Across scales and benchmarks, the results argue that “set it once” search parameters won’t generalize across budgets, pushing toward adaptive controllers that change strategy as time and progress shift.
Trump’s attempt to rename AI is looking awfully artificial
The Verge digs into how President Donald Trump is reframing “artificial intelligence” language into “super” while using “artificial” as a pejorative – a rhetorical move with real influence on how the public and policymakers talk about AI. The piece also highlights the political alignment risk when major AI players mirror the same framing.
Sophos cuts threat investigation time by 96% with OpenAI Daybreak
OpenAI and Sophos describe how Daybreak helps automate parts of managed threat detection and response, claiming a 96% reduction in investigation time while still keeping humans in the loop. The reporting focuses on operational workflow integration rather than raw model demos, emphasizing speed, triage assistance, and measurable throughput gains.
Plan-and-Patch: Diffusion Language Models for Agentic Planning
Plan-and-Patch tackles the reality that long-horizon agents need to revise plans when environments and tool outputs don’t match assumptions. Instead of regenerating everything, a diffusion language model repairs only the affected plan region while preserving the surrounding prefix and suffix, improving plan repair success and reducing latency compared with autoregressive approaches.
California is trying to shut down robot vs. human cage matches
The Verge reports that California’s state athletic commission has sent a cease-and-desist letter to a company running robot-versus-human fights, arguing the events violate rules around who can compete and how. The coverage spotlights how quickly “robot sports” are colliding with existing governance frameworks.
Impactful scheduling for GPU clusters
This Hugging Face/AllenAI technical blog argues that GPU cluster scheduling is often the bottleneck that determines whether AI systems feel responsive – not just model efficiency. It outlines scheduling tactics that better match workload characteristics to hardware availability, aiming to increase utilization and reduce latency under real-world training and inference mixes.
Keyframe Mnemonics: Self-Supervised Keyframe Discovery for Horizon-Invariant Behavior Cloning
Keyframe Mnemonics reframes behavior cloning in memory-heavy, non-Markovian settings by learning a small set of “mnemonics” (keyframes) that preserve long-horizon context without relying on long attention windows. Experiments show strong horizon generalization and practical gains on robot manipulation tasks, including results that retain much higher success rates when evaluation horizons stretch far beyond training.
OpenAI doubles down on decision to fire three AI safety researchers
The Verge reports that OpenAI is standing by its decision to dismiss three safety researchers after concluding they committed a “significant breach of trust.” The article frames the dispute around internal handling of sensitive information, responding to a public letter from the researchers and pushing the broader question of transparency versus enforcement in safety governance.