Here are the show notes for episode 0047. You can save them to data/episodes/0047/show_notes.md:

---

Episode 0047: Trillion-Parameter Agentic Intelligence

Why it matters. Ant Group's inclusionAI team has released "Ling and Ring 2.6 Technical Report: Efficient and Instant Agentic Intelligence at Trillion-Parameter Scale", describing two open-weight model families — Ling 2.6 for low-latency generation and Ring 2.6 for deep agentic reasoning — culminating in Ring 2.6-1T, a one-trillion-parameter model purpose-built for multi-step tool use, code execution, search, and workflow orchestration. The paper's central contribution is a co-design approach spanning architecture, training, serving, and agent environments: rather than training from scratch, they migrate weights from a prior model generation into a new hybrid linear attention architecture, then apply a suite of token-efficiency techniques and a trillion-scale RL training framework called KPop to produce an agentic model that is both capable and practically servable. The weights are open and available on HuggingFace.

Ant Group / inclusionAI. This work comes from Ant Group, Alibaba's financial technology subsidiary (the company behind Alipay), through their inclusionAI research initiative. The paper is available on arXiv (2606.15079). Open weights for both model families are available on HuggingFace (inclusionAI/ling-26 collection). The paper lists over 200 contributing authors, reflecting the institutional scale required to build, train, and serve a trillion-parameter agentic model.

The Researchers. The author list includes Ang Li, Ben Liu, Bin Han, and over 200 additional contributors from Ant Group's inclusionAI team. The scale of the collaboration — spanning architecture design, training infrastructure, serving systems, RL environments, and post-training pipelines — reflects the organizational breadth required for trillion-parameter model development. The full author list is available in the paper.

Key Technical Concepts. The core architectural innovation is hybrid linear attention, combining Lightning Attention — a hardware-efficient linear attention mechanism that uses IO-aware tiling for GPU memory hierarchies — with Multi-head Latent Attention (MLA), the KV cache compression technique introduced by DeepSeek-V2 that projects keys and values into a shared low-dimensional latent space. This hybrid replaces standard quadratic transformer attention, enabling tractable long-context operation at trillion-parameter scale. Rather than training from scratch, the models are built via architectural migration pre-training: weights from the existing Ling 2.0 base model are migrated into the new architecture and training continues from the migrated checkpoint, dramatically reducing total compute cost — a technique related to prior work on progressive training and model recycling. For token efficiency, the paper introduces three techniques: Evolutionary Chain-of-Thought, which iteratively selects for reasoning traces that are both correct and concise (building on chain-of-thought prompting); Linguistic Unit Policy Optimization (LUPO), which applies policy gradient updates at the level of meaningful linguistic units rather than individual tokens, improving credit assignment; and shortest-correct-response distillation, which trains on only the most concise correct outputs from a teacher model, drawing on knowledge distillation principles. The KPop RL training framework enables stable reinforcement learning at trillion-parameter scale by using asynchronous scheduling across four heterogeneous environment types (coding, search, tool use, workflow execution), keeping training infrastructure continuously utilized despite unpredictable environment latencies. The work sits at the intersection of linear attention research, mixture-of-experts scaling, and the emerging paradigm of agentic AI systems that integrate tool use with deep reasoning.

Daily Tech Feed: From the Labs is available on Apple Podcasts, Spotify, and wherever fine podcasts are distributed. Visit us at pod.c457.org for all our shows. New episodes daily.

---

Link inventory (~20 links): arXiv ID from paper.json (reliable). HuggingFace collection URL from paper.json. Ant Group corporate site. inclusionAI GitHub org. arXiv links for: DeepSeek-V2/MLA (2405.04434), Attention Is All You Need (1706.03762), chain-of-thought prompting (2201.11903), InstructGPT/RLHF (2203.02155), Hinton knowledge distillation (1503.02531), linear transformers (2006.16236). Lightning Attention GitHub repo. I omitted Google Scholar links for individual researchers since with 200+ authors and only 3 named in paper.json, I couldn't verify Scholar IDs. The progressive training and model recycling arXiv IDs (2404.19737, 2401.12863) and agentic AI systems (2309.07864) and MoE scaling (2401.04088) are less certain — I included them as plausible related work but flag them for verification.