SK Hynix and Sandisk to Standardize New AI Memory

- The new HBF (High-Bandwidth Flash) memory is designed to complement HBM by providing up to 16 times more capacity at a similar cost, addressing the memory capacity limitations of GPUs in large-scale AI inference. HBF is built for read-intensive workloads due to its limited write endurance of about 100,000 cycles. - AI inference is frequently constrained by memory bandwidth and capacity, not just GPU processing power. Inefficient memory systems cause expensive GPUs to sit idle, waiting for data, which creates a significant bottleneck in AI workloads. - Reinforcement Learning from Human Feedback (RLHF) faces a significant bottleneck in the speed, cost, and volume of high-quality human data annotation required. This has led to a shift from large-scale, crowd-sourced data labeling to a focus on smaller, higher-quality datasets annotated by domain experts. - To overcome the limitations of RLHF, AI labs are increasingly adopting Constitutional AI (CAI), which uses an AI model to critique and revise its own outputs based on a set of predefined principles. This method, also known as Reinforcement Learning from AI Feedback (RLAIF), is more scalable and cost-effective than relying solely on human feedback. - While synthetic data can be generated quickly and cost-effectively, it often lacks the nuance and real-world complexity that human-annotated data provides. Research indicates that hybrid approaches, which use a large amount of synthetic data supplemented by a smaller set of high-quality human-labeled data, can achieve the best results. - The evaluation of agentic AI systems requires new benchmarks that go beyond traditional language model metrics. Frameworks like AgentBench, WebArena, and GAIA are being developed to assess agents on their ability to perform complex, multi-step tasks, use tools effectively, and recover from errors. - Go-to-market strategies for AI infrastructure startups are shifting to focus on demonstrating clear value and outcomes rather than just the underlying technology. Successful strategies often involve deep customer discovery to identify specific, painful problems and then articulating how the AI solution can lead to measurable improvements, such as a 40% reduction in debugging time. - The fundraising climate for AI companies has seen a massive influx of capital, with AI-related startups capturing nearly half of all global venture funding in 2025. However, this funding is heavily concentrated in a few large foundation model and infrastructure companies, creating a more challenging environment for application-layer startups.