The article outlines key developments in AI from 2025 and predictions for 2026. One significant highlight was the launch of DeepSeek-R1, which integrated Reinforcement Learning with Verifiable Rewards (RLVR) and the GRPO framework into mainstream use. This shift spurred advancements in model training, allowing for better reasoning capabilities with less supervision. The author notes that while RLVR/GRPO is powerful, the current implementations still face challenges like brittleness and sensitivity to infrastructure changes. In terms of hardware, the introduction of FP4 precision formats marked a major leap in efficiency. NVIDIA’s NVFP4 allowed for nearly double the inference speed without sacrificing accuracy. OpenAI's release of models quantized in this format emphasized a trend towards open weights and efficient deployment. However, the disappointment surrounding Meta’s Llama 4 is highlighted, as it struggled with performance and clarity in direction, leaving its future uncertain. Looking ahead to 2026, the author predicts a shift towards significantly faster token generation due to improved hardware rather than just software optimizations. The expectation is that models will handle reasoning more swiftly, alleviating current bottlenecks in latency and compute power. There's a call for better methods to manage context and cache information, suggesting that advancements will come from both refined architectures and specialized hardware. The author also mentions plans for their newsletter, emphasizing a focus on quality content and targeted model releases in the coming months.