AI Daily — April 24, 2026
2026-04-24
openai.com
Introducing GPT-5.5
OpenAI has released GPT-5.5, positioned as their most capable model to date with improvements in coding, research, and data analysis tasks. The model is designed for complex multi-step workflows and integrates tightly with tools across the OpenAI ecosystem. A system card was published simultaneously, providing safety evaluation details.
deepmind.google
Decoupled DiLoCo: A new frontier for resilient, distributed AI training
Google DeepMind introduces Decoupled DiLoCo, an extension of the DiLoCo distributed training framework that decouples inner and outer optimization steps to improve resilience and efficiency in large-scale training across heterogeneous or geographically distributed compute clusters. The approach reduces synchronization overhead and is designed to tolerate slow or intermittent inter-node communication without degrading model quality. This is a notable step toward practical federated and geo-distributed LLM training at scale.
rss.arxiv.org
FairyFuse: Multiplication-Free LLM Inference on CPUs via Fused Ternary Kernels
FairyFuse presents an inference system that exploits ternary weight quantization ({-1, 0, +1}) to eliminate floating-point multiplications entirely during LLM inference on commodity CPUs, replacing them with fused conditional additions. The system addresses the memory bandwidth bottleneck of autoregressive generation on CPU-only hardware by fusing multiple sub-GEMVs per layer rather than dequantizing on the fly. Benchmarks suggest meaningful throughput gains over existing ternary and 4-bit CPU inference approaches.