Moonshot AI shipped a 2.8-trillion-parameter open-weight model with a 1M-token context, native vision, and a new attention stack. What's actually new in the architecture, what the benchmarks say, and why it matters for builders.
Amazon built two radically different approaches to predicting the future — a proprietary supply chain optimization pipeline (SCOT) and an open-source time series foundation model (Chronos). This post compares their architectures, trade-offs, and when each philosophy applies.
A decision framework for choosing between Amazon's Chronos-2 foundation model and custom XGBoost many-models pipelines for demand forecasting. Based on real patterns from SKU-level supply chain work.
A discovery call with a global specialty chemicals company revealed that the real AI bottleneck isn't models — it's data. Here's what enterprise chemistry teams actually need versus what the hype promises.
What TFLOPS actually measures, why FP16 matters for LLMs, and why the most important GPU bottleneck for inference isn't compute at all.
A practical walkthrough of how large language models are aligned with human values — from collecting feedback to PPO optimization and the reward hacking pitfalls.
Understand how LLM fine-tuning works, when to use it, and how to choose between full fine-tuning, LoRA, soft prompts, and other PEFT methods.
Building a production forecasting application without the complexity of traditional ML model training and feature engineering.