Interconnects
·
3 months ago
Open and closed AI models have different performance gaps depending on which benchmarks and real-world tasks are being measured, rather than a single measurable distance between them. The industry shifts its focus every 12 to 18 months—moving from chat and math capabilities to coding tasks to specialized domain work in accounting and law—making benchmark relevance constantly change. Frontier labs like OpenAI and Anthropic must continuously develop new valuable use cases to justify their infrastructure investments, while open-source models struggle most in specialized domains requiring private training data and complex evaluation environments.
Import AI
·
3 months ago
Huawei developed HiFloat4, a 4-bit precision format for AI training that outperformed the Western-developed MXFP4 format on its Ascend chips, suggesting Chinese companies are building specialized hardware-software combinations to work around export restrictions. In direct comparison testing on models up to 30 billion parameters, HiFloat4 achieved approximately 1.0% relative loss error versus MXFP4's 1.5% when measured against a full-precision baseline. This reflects broader Chinese efforts to maximize efficiency of homegrown semiconductors as access to advanced Western chips remains limited.
BAIR
·
3 months ago
Researchers propose GRASP, a gradient-based planning method for world models that addresses long-horizon planning challenges by optimizing over both actions and states in parallel, adding stochasticity for exploration, and reshaping gradients to avoid brittle state-input optimization through vision models. The method lifts the dynamics constraint to a soft penalty, enabling parallel computation across time steps and avoiding exponential Jacobian conditioning from sequential rollouts that causes exploding/vanishing gradients. By treating states as optimization variables rather than computed through repeated model applications, GRASP mitigates adversarial robustness vulnerabilities in deep learning-based dynamics models that make naive state optimization unstable.
Allen Institute (AI2)
·
3 months ago
Researchers from AI2 introduced BAR (Branch-Adapt-Route), a method for modular post-training that trains separate domain experts through independent pipelines and merges them using mixture-of-experts architecture. The approach achieved an average score of 49.1 across 19 benchmarks, outperforming monolithic post-training-only baselines (47.8) while enabling individual experts to be upgraded independently without retraining the entire model. This enables linear cost scaling for domain updates compared to quadratic costs in traditional monolithic retraining, allowing teams to upgrade specific capabilities like code or math without affecting others.
OpenAI Blog
·
3 months ago
Hyatt has deployed ChatGPT Enterprise across its global workforce to improve internal productivity and guest experiences. The hotel chain is using GPT-5.4 and Codex models across its operations. This deployment enables Hyatt employees to automate tasks and streamline workflows, though specific productivity gains are not quantified in the announcement.
