OpenAI has snapped up Ona to plug secure cloud execution and orchestration into its Codex platform. The goal: let customer-owned agents run persistently across sessions without manual intervention. It’s a move toward long-running AI workflows you can control end to end. Meanwhile Anthropic quietly reversed a policy that muffled its Claude Fable 5 model. Researchers spotted that certain requests—training rival models, debugging AI code, even tweaking neural nets—were silently rerouted to a weaker model. After complaints about wasted tokens and veiled safeguards, Anthropic made its frontier-LLM guardrails transparent again. On the research front, there’s a 15-minute guide to finding optimal tokenizers—algorithms that map byte sequences to integer tokens for model training. And a 50-minute “vintage LLM” walkthrough shows how one developer built a full transformer from scratch for about $80, leveraging a decent home PC. A short essay from CoreWeave’s Brannin McBee argues compute isn’t a fungible commodity; he pinpoints where pricing spreads still hide and hints at why his company’s custom infrastructure matters. Xiaomi just open-sourced MiMo Code V0.1.0, a terminal-native coding assistant that outperforms Claude Code on tasks over 200 steps. It uses a memory subagent to track decisions and context across sessions. GitHub hosts the code under an MIT license. Silico’s Predictive Data Debugging tool analyzes preference datasets before training to flag safety failures, hallucinations, or sycophancy, letting you tweak data or training loops in advance. Finally, there’s a 29-minute deep dive on fusing multilayer perceptrons in PyTorch, showing how merging linear ops boosts throughput and cuts latency in large models.