All Updates
Page 36 of 658
April 7, 2026
LGBM Beats LLMs in Discharge Prediction
This arXiv study evaluates lightweight LLMs and traditional models for next-day discharge prediction using postoperative clinical notes in spine surgery. TF-IDF with LGBM delivered the best F1-score (0.47), recall (0.51), and AUC-ROC (0.80). Results favor interpretable, resource-efficient models over compact LLMs for imbalanced clinical tasks.
IC3-Evolve: LLM Evolves IC3 Heuristics Safely
IC3-Evolve is an offline LLM-driven framework that proposes small, auditable patches to improve IC3 hardware model checking performance. Patches are validated strictly via proofs for SAFE results or replayable counterexamples for UNSAFE, ensuring soundness. The evolved checker has zero runtime ML overhead and shows gains on HWMCC and industrial benchmarks.
Contextual Control Sans Memory Growth
Researchers introduce an intervention-based recurrent architecture for contextual control in sequential decision-making without enlarging recurrent memory. A shared latent state is modified by additive, context-indexed operators. It matches baselines on context-switching tasks under partial observability.
Christian Benchmark Reveals AI Bias
Introduces Flourishing AI Benchmark: Christian Single-Turn (FAI-C-ST) to evaluate LLMs on Christian human flourishing across seven dimensions. Testing 20 frontier models shows 17-point average decline vs. Christian criteria, with 31 points in Faith and Spirituality. AI defaults to Procedural Secularism, lacking theological coherence.
AI Evaluation Needs Item-Level Data
This position paper argues that item-level benchmark data is essential for a rigorous science of AI evaluation, addressing systemic validity failures in current paradigms. It draws from psychometrics and computer science to advocate for granular diagnostics. OpenEval is introduced as a repository to support evidence-centered AI evaluation.
NetEase Launches Localized AI Knowledge Base
NetEase Youdao launched Youdao Baoku, China's first localized AI knowledge base on April 7. It acts as an AI research assistant for deep knowledge understanding and restructuring. Offers knowledge management, smart Q&A, and multimodal generation, filling a domestic market gap.
Qwen3.6-Plus Tops OpenRouter Weekly Leaderboard
Alibaba's Qwen3.6-Plus secured the top spot on OpenRouter's global big model weekly call volume leaderboard. It led the daily rankings for four consecutive days and became the first model on the platform to exceed 1 trillion tokens in a single day.
US Ends Tesla Smart Summon Safety Probe
US auto safety agency closed its probe into Tesla's remote parking feature. Collisions were found extremely rare, all low-speed with minor consequences. This clears regulatory hurdles for the AI-driven function.
Lenovo Bets Big on AI Native Shift
Lenovo CEO Yang Yuanqing launches 'Smarter AI for all' drive, aiming for $100B revenue in two years via hybrid AI and Nvidia ties. AI revenue surges 72% with $155B server backlog, but lacks proprietary LLMs. Past hype flags raise skepticism on true transformation.
Samsung Orders 70+ Litho Machines for P5 HBM Fab
Samsung ordered over 70 lithography machines for Pyeongtaek P5 fab's PH1 phase, targeting 2027 operation. Includes ~20 ASML EUV systems from ASML and Canon. Will produce 1c nm DRAM, including HBM and general memory.
AI Fraud Debuts Big in FBI Stats
AI-assisted fraud has made its major debut in the FBI's cybercrime statistics. Losses from these schemes surpassed $1.4 billion in 2025. This highlights AI's growing role in enabling sophisticated scams.
Beginner Earns 1M RMB in 23 Days with AI Shorts
A non-film-school background creator earned 1 million RMB in 23 days using AI-generated short films, breaking into the competitive video content space. Director Guo Fan praised the work highly. This story signals a golden era for AI creators.
Ilya Leaks 70 Pages of OpenAI Secrets
Ilya Sutskever has exposed 70 pages of confidential OpenAI documents. He publicly accused Sam Altman of being a habitual liar. The leak comes from Chinese source Quantum Position.
Kazik Open-Sources AI Content Skill
Author 'Digital Life Kazik' open-sources 'khazix-writer' skill, distilling 3 years of writing methods, style rules, pitfalls, and a four-layer self-inspection system. GitHub repo installable via Agent; best with Claude Opus 4.6 Chat mode. Shares iteration process to build personal skills beyond imitation.
New Yorker Probes Sam Altman's Trust
The New Yorker published a long-form profile on OpenAI CEO Sam Altman based on interviews with nearly 100 associates. It contrasts his rare persuasion skills with accusations of deceptive patterns. The article details his ouster drama, power concentration, and questions accountability in superintelligence leadership.
Microsoft Launches 4 AI Certifications
Microsoft is rolling out new and revised certifications from March, including 4 AI-focused ones in April-June. These cover Copilot Studio, Dynamics 365, and Power Platform. The move addresses the booming demand for AI skills in job markets.
Hongxiong AI Raises 210M RMB Series A
Hongxiong AI, an enterprise AI provider, completed a 210M RMB Series A funding round at over 1.5B RMB valuation, led by Huayu Venture. It focuses on 'memory science + multimodal LLMs' for physical AI, enabling human-like episodic memory. Plans HKEX IPO in 2027-2028.
Doubao Hits 120T Daily Tokens in 3 Months
ByteDance's Doubao AI model exceeded 120 trillion daily tokens by March, doubling usage in three months and surging 1,000-fold since its May 2024 launch. Volcano Engine launched Seedance 2.0 API public beta for enterprise users.
Saifuros Raises 300M RMB for AR Micro-LED
AR display firm Saifuros completed 300M RMB C-round funding to scale full-color Micro-LED production for AR glasses. Their silicon-based tech achieves 5x efficiency over Micro-OLED, targeting outdoor AR needs. Future plans include AI data center optical comms with <5% power use.
LG Chairman Meets Valley AI Execs
LG Group chairman Koo Kwang-mo met Silicon Valley AI company leaders last week. Aim is to speed up LG's AI transformation across the group.