# Feb 9, 2026

## TSMC to make advanced AI semiconductors in Japan
- **TSMC** will manufacture **3-nanometer semiconductors** in Japan, targeting the surging demand for **AI technologies** and enhancing Japan's chipmaking capabilities.

## Hard-braking events as indicators of road segment crash risk
- **Hard-braking events (HBEs) serve as a reliable leading indicator of crash risk**, with a significant correlation established between HBE frequency and road segment crash rates, suggesting their potential for proactive safety assessments.

## Eight more months of agents
- **Agents have evolved significantly**, with the latest models like Opus now capable of writing **90% of code**, a stark improvement from Claude Code's 25% last year, indicating a rapid advancement in coding capabilities.

## [R] Really nice interactive explanation of Speculative Decoding
- **Speculative decoding** enhances model performance by predicting future tokens, allowing for faster and more efficient text generation, as visualized in recent studies.

## [D] Are autoregressive video world models actually the right foundation for robot control, or are we overcomplicating things?
- **LingBot-VA's autoregressive video world model** claims to provide a unique foundation for robot learning, achieving **92.9% accuracy on RoboTwin 2.0** and outperforming π0.5 by over **20%** on long horizon tasks with minimal adaptation.

## [P] Built a real-time video translator that clones your voice while translating
- **Real-time video translation** allows users to speak in one language while their voice is cloned and translated into another, achieving a latency of approximately **545ms**, making it nearly imperceptible during calls.

## InftyThink+: Effective and Efficient Infinite-Horizon Reasoning via Reinforcement Learning
- **InftyThink+** introduces a novel **reinforcement learning framework** that optimizes iterative reasoning by strategically managing summarization and continuation, addressing the limitations of traditional chain-of-thought models.

## [D] Advice on journal for work between ML, data infrastructures, and robotics
- The work presents **dynamic and semantic "data-to-Knowledge pipelines"** that enhance robot control by training a robust base model for inverse kinematics, demonstrating practical feasibility across multiple research institutes.

## 🤗Transformers.js v4 Preview: Now Available on NPM!
- **Transformers.js v4** is now available on NPM, featuring a **new WebGPU Runtime** that enhances performance across ~200 model architectures, enabling local execution in various JavaScript environments like Node and Deno.

## [R] Identifying the "Complexity Kink": An Econometric Analysis of AI Marginal Productivity Collapse in Multi-Asset Tasks
- The study reveals a **significant Benchmark Curation Bias (p=0.03)**, indicating that current benchmarks favor low-coordination tasks, obscuring the true limits of AI productivity compared to human experts.

## Endogenous Resistance to Activation Steering in Language Models
- **Endogenous Steering Resistance (ESR)** allows large language models, particularly **Llama-3.3-70B**, to recover mid-generation from task-misaligned steering, enhancing response quality despite ongoing steering efforts.

## Halluverse-M^3: A multitask multilingual benchmark for hallucination in LLMs
- **Halluverse-M^3** is a novel dataset that enables systematic analysis of **hallucinations** in large language models across **four languages** (English, Arabic, Hindi, Turkish) and two tasks (question answering and dialogue summarization), addressing a critical gap in multilingual model evaluation.

## Rethinking Multi-Condition DiTs: Eliminating Redundant Attention via Position-Alignment and Keyword-Scoping
- The proposed **Position-aligned and Keyword-scoped Attention (PKA)** framework significantly enhances multi-condition control in text-to-image models by eliminating **redundant attention**, achieving a **10.0× inference speedup** and **5.1× VRAM saving**.

## POP: Online Structural Pruning Enables Efficient Inference of Large Foundation Models
- **POP (Partition-guided Online Pruning)** introduces a novel framework for **context-conditioned dynamic pruning**, enhancing the efficiency of large foundation models (LFMs) during inference without the need for extensive preprocessing.

## HyPER: Bridging Exploration and Exploitation for Scalable LLM Reasoning with Hypothesis Path Expansion and Reduction
- **HyPER** introduces a **dynamic expand-reduce control** mechanism that optimizes the exploration-exploitation trade-off in multi-path reasoning, enhancing accuracy without the need for extensive retraining.