Xuedong D. Huang

[𝗦𝗛𝗢𝗖𝗞𝗜𝗡𝗚!] 𝗧𝗵𝗼𝘂𝗴𝗵𝘁 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗱𝗮𝘆: 𝙂𝙤𝙤𝙙𝙗𝙮𝙚 𝘿𝙄𝙏𝘼 𝙍𝙚𝙪𝙨𝙚. 𝙃𝙚𝙡𝙡𝙤 𝘿𝙄𝙏𝘼! We recently asked a critical question: With agentic AI and a semantic graph knowledge layer, should we continue maintaining the DITA reuse model that has served us well for years? Or could we start fresh – resolving (normalizing) all reuse references while creating a more powerful, simpler, and content-source-agnostic model for reuse? This paper challenges one of the foundational strengths of DITA XML – its superior reuse model, and in the process, democratizes reuse for any format (with DOM formats such as DITA/XML remaining the superior choice). But DITA isn’t going away. Quite the opposite. DITA remains essential because its semantic structure and precision are what make this new model of knowledge-level reuse governable and precise. https://lnkd.in/evK-zMhZ

35

11 Comments

New work on Rethinking Thinking Tokens: LLMs as Improvement Operators: https://lnkd.in/errhNuuz Reasoning training encourages LLMs to produce long chains of thought (CoT), improving accuracy via self-checking but increasing context length, compute cost, and latency. This work studies whether frontier models can achieve better trade-offs, higher accuracy with lower cost. The paper develops a simple yet effective Parallel-Distill-Refine (PDR) procedure: Generate diverse drafts in parallel, Distill them into a compact textual workspace, and Refine conditioned on this workspace. This decouples context length from total token count, allowing control over compute via parallelism. PDR yields higher accuracy than long CoT at lower latency. Training an 8B model with RL to align with PDR further shifts the Pareto frontier. On math benchmarks, PDR achieves +11% (AIME 2024) and +9% (AIME 2025) over single-pass baselines. With Lovish Madaan, Aniket Didolkar, Suchin Gururangan, John Quan, Ruan Silva, Manzil Zaheer, Sanjeev Arora, and Anirudh Goyal.

79

1 Comment

The model was never the product. The runtime is. That is the thesis GTC 2026 quietly confirmed this week — and it has direct implications for every enterprise AI architecture decision being made right now. Three moves. One structural signal: → NVIDIA launched the Agent Toolkit at GTC on March 16 — an open-source stack comprising OpenShell (policy-based agent security runtime), AI-Q (hybrid agentic search blueprint cutting query costs by over 50%), and Nemotron open models. Seventeen enterprise software partners — Adobe, SAP, Salesforce, ServiceNow, Siemens, CrowdStrike, Red Hat, and more — committed at launch. NVIDIA is not selling chips here. It is claiming the infrastructure substrate beneath every enterprise AI agent. This is CUDA for the agentic era. → Mistral launched Forge on March 17 — a platform that lets enterprises train custom AI models from scratch on their own data, on their own infrastructure. On-premises. Air-gapped if required. No US cloud dependency. Early adopters include Ericsson, ASML, the European Space Agency, and Singapore's defence agencies. For sovereign AI strategies and Vision 2030 programmes, this is the first production-grade training platform that does not require a hyperscaler trust decision. → Microsoft restructured Copilot on March 17 — consolidating consumer and commercial teams and redirecting Mustafa Suleyman exclusively to frontier model development. His stated position: "the model is the product." Microsoft Copilot had 6 million daily active users in February. ChatGPT had 440 million. No organisational restructuring closes that gap. Only model parity does. The parallel signal: Cerebras CS-3 systems deployed on AWS Bedrock deliver 5× token throughput via disaggregated inference. GPT-5.4 matches DeepSeek V4 with a 1-million-token context window. Context depth is now table stakes across the frontier tier. The deeper pattern this week is not any single product. It is that enterprise AI is exiting the pilot phase and entering the deployment infrastructure phase — and the architecture choices made in the next 12 months will determine vendor dependencies for the next decade. The full strategic brief — with technical architecture breakdown, leadership pivots, and Harvard-referenced sourcing — is linked in the comments. One question for the enterprise AI architects and security leads reading this: As autonomous agent pipelines reach production inside regulated organisations, who in your organisation owns the translation of compliance obligations into runtime-enforceable agent policies — the CISO, the data governance team, or the AI product team? I am curious to hear how that accountability question is being resolved in practice. #ArtificialIntelligence #AgenticAI #AIStrategy #DigitalSovereignty #LLMOps #Vision2030 #GTC2026 #NVIDIA #MistralForge #MicrosoftCopilot #ICTLeadership #EnterpriseAI

13

2 Comments