--- title: "AI engineering, November 15, 2026" description: "The Forward Pass daily issue, November 15, 2026." canonical: https://theforwardpass.net/archive/daily/2026-11-15 updated: 2026-11-15 --- # AI engineering, November 15, 2026 The Forward Pass daily issue, November 15, 2026. Source: https://theforwardpass.net/archive/daily/2026-11-15 > This issue is researched and written by AI models, and every fact is checked against its cited source. No human edits it before it is sent. _Fable 5.2 Max outscores Gemini 4 Hydrogen, Hugging Face halves Qwen3 step time, ASE plans NT$100 billion expansion, CodeRabbit exposes 150+ operations through MCP_ ASE is buying artificial noses by the thousand. Its latest US$9 million order adds 5,000 chemical sensing systems. Meanwhile, Artificial Analysis scores Claude Fable 5.2 Max at 65 versus Gemini 4 Hydrogen at 54 on its Intelligence Index. Hugging Face’s TRL fix halved a Qwen3 training step on an H100. If you read one thing today: the Fable 5.2 versus Gemini 4 Hydrogen comparison. The per-task costs and response times give you concrete numbers for choosing an API against your budget and latency limits. ## Artificial Analysis scores Fable 5.2 Max 65 versus Gemini 4 Hydrogen’s 54 **Top News** · 378 HN points Choosing between Claude Fable 5.2 and Gemini 4 Hydrogen means trading benchmark performance against cost and latency. Artificial Analysis puts Fable 5.2 Max at 65 on its Intelligence Index, ahead of Gemini 4 Hydrogen at 54 with xhigh effort. Here’s how they compare: - Fable 5.2 API pricing is $8 per million input tokens and $40 per million output tokens. Gemini 4 Hydrogen costs $1.50 and $7.50, respectively. - Artificial Analysis measured $26.54 per Intelligence Index task for Fable 5.2, compared with $3.16 for Gemini 4 Hydrogen. - The index combines nine evaluations spanning knowledge work, agentic workflows, coding, science, reasoning, and professional documents. One catch: Fable 5.2’s maximum-effort score comes with a wait. Artificial Analysis measured 44.5 minutes per task and 262.07 seconds to first answer token, versus Gemini 4 Hydrogen’s 14.8 minutes and 46.64 seconds at xhigh. Sources: [artificialanalysis.ai](https://artificialanalysis.ai/models/claude-fable-5-2) · [artificialanalysis.ai](https://artificialanalysis.ai/models/gemini-4-hydrogen) ## Hugging Face halves Qwen3 training-step time on H100 with TRL fix **Top News** · Reported · runtimewire.com For standard-attention models using best-fit-decreasing packing, the TRL fix changes the default batch layout. Hugging Face halved a Qwen3 training step on an H100 by changing how packed batches are arranged for standard-attention models. Here’s what changed: - Best-fit-decreasing packed examples now stay in separate batch rows, rather than being flattened into one long row per packed batch. - The new layout produced step timings on AWS Trainium2, where the previous layout crashed. - Hugging Face merged the TRL change on November 13, 2026. One catch: The halved step time comes from a reported Qwen3 H100 test, and the layout change applies to standard-attention models. Why care? The layout change both halved the reported Qwen3 H100 step time and produced timings on Trainium2, where the old layout crashed. Sources: [runtimewire.com](https://runtimewire.com/article/hugging-face-trl-packed-batches-rows-h100-trainium) ## ASE plans NT$100 billion expansion across its Luzhu plant phases **Top News** · Reported · ctee.com.tw · translated from Chinese ASE is expanding its Luzhu plant in two phases, with total investment expected to reach about NT$100 billion when fully loaded. The expansion is expected to contribute up to NT$80 billion in annual revenue. Here’s what is happening: - Phase 1 has opened after ASE spent NT$6.5 billion acquiring an existing factory building and associated facilities, then integrated the site and built production lines. - Phase 2 has broken ground, with vibration control, separately configured plant facilities, and space reserved for later expansion. - ASE also signed a US$9 million agreement with Ainos for 5,000 additional AI Nose systems, which sense chemicals. - The Ainos order brings the combined total to 6,400 systems, including 1,400 already delivered. One catch: Phase 2 is still at the groundbreaking stage. The projected annual revenue contribution is for the fully loaded expansion, not the newly opened Phase 1 alone. Sources: [ctee.com.tw](https://www.ctee.com.tw/news/20261114700124-430502) ## Signals 1. [AWS adds OpenFold3-preview to SageMaker JumpStart for 3D biomolecular structure prediction](https://aws.amazon.com/blogs/machine-learning/openfold3-on-amazon-sagemaker-jumpstart-predict-biomolecular-structures/) 2. [CopilotKit adds server-side agent runs through Intelligence in version 1.88.0](https://github.com/CopilotKit/CopilotKit/releases/tag/v1.88.0) · 38,692 Stars 3. [NVIDIA proves error bounds for double-precision matrix multiplication emulated on low-precision hardware](https://research.nvidia.com/publication/2026-11_providing-accuracy-guarantees-fixed-point-emulation-matrix-multiplication) 4. [CodeRabbit adds a hosted MCP server exposing 150+ product operations to AI agents](https://docs.coderabbit.ai/changelog#code-rabbit-api-and-mcp) 5. [DeepSeek V4 Pro 0813 hits 87.9 on Terminal Bench 2.1, beating its preview’s 72.1](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813) · 25 HN points 6. [SlopCodeBench tests 15 coding agents, the best passes 14.8% of iterative checkpoints](https://arxiv.org/abs/2603.24755)