Jan 4, 2026

Looped and recursive LLMs, plus a memory IPO and a TypeScript Torch

🧩 The Gist

Two new arXiv papers push different paths to stronger reasoning and long‑context handling. One builds iterative reasoning into pretraining with looped latent computation, the other treats long inputs as an external environment and lets the model recurse over snippets. A small educational project recreates core ML components in TypeScript, signaling continued grassroots tooling. On the hardware side, a major Chinese DRAM maker is pursuing an IPO to fund capacity and next‑gen memory, underscoring how compute and memory supply sit behind AI advances.

🚀 Key Highlights

• Ouro, a family of Looped Language Models, integrates reasoning during pretraining through iterative latent computation, an entropy‑regularized objective for learned depth, and scaling to 7.7T tokens. The 1.4B and 2.6B models match results reported for models up to 12B parameters across a wide range of benchmarks. The authors attribute gains to superior knowledge manipulation rather than capacity, and report reasoning traces that align more closely with final outputs than chain of thought.
• Recursive Language Models propose an inference strategy that programmatically examines, decomposes, and recursively calls the model over snippets of long prompts. The approach handles inputs up to two orders of magnitude beyond context windows and outperforms base models and common long‑context scaffolds on four tasks, with comparable or lower cost per query.
• Torch.ts reimplements core PyTorch‑style pieces in TypeScript as a learning project. It is currently a tensor manipulation library with an autograd engine planned, and documents hands‑on work with strides and manual matmul outside NumPy.
• Chinese DRAM maker CXMT filed for a STAR Market IPO to upgrade manufacturing and advance next‑generation memory technologies. The linked page presents inconsistent raise figures, the headline cites USD 4.2B while the page description cites CNY 295B, about USD 42.2B, both framed as funding to strengthen China’s memory supply chain.

🎯 Strategic Takeaways

• Model design and scaling
  • Baking iterative reasoning into pretraining suggests an alternative to relying on explicit chain of thought, with smaller models matching larger baselines on reported benchmarks.
  • Emphasis on knowledge manipulation over raw capacity points to new levers for performance without proportional parameter growth.
• Inference‑time orchestration
  • Treating long inputs as an environment and enabling recursive self‑calls can extend usable context far beyond native windows while keeping costs comparable, a practical path for long‑document and code tasks.
• Infrastructure and supply chain
  • Planned DRAM investment highlights memory as a gating factor for AI workloads, and continued build‑out of next‑gen memory aims to support training and inference capacity.
• Developer tooling and education
  • Lightweight reimplementations like Torch.ts help demystify tensors, strides, and autograd, and broaden ML experimentation in the TypeScript ecosystem.

🧠 Worth Reading

• Scaling Latent Reasoning via Looped Language Models: Introduces LoopLM, which trains models to perform iterative computation in latent space with an entropy‑regularized objective and learned depth. Reported results show smaller models matching larger baselines across benchmarks and producing reasoning traces that better align with final answers, a practical signal that pretraining for reasoning can reduce reliance on verbose inference‑time scaffolds.