All Updates
Page 246 of 929
April 13, 2026
Buffett & Munger Open-Sourced as AI Agent
An AI agent embodying Warren Buffett and Charlie Munger's investment wisdom has been developed. It's now open-sourced, providing retail investors with a 'master model' for decision-making. Enables accessible AI-driven investing advice.
86% Non-Coders Deploy Apps via Claude Code
Goodpatch mandated Claude Code use for all employees, leading to 86% of those with zero coding experience successfully deploying apps. The apps mostly addressed small daily or work troubles. The company publicly shared these impressive results.
Korean Chip Suppliers Sunk in Price Cut Quagmire
Korean chip suppliers are deeply trapped in a price reduction swamp. They confront continuous price cuts for the second consecutive year. This exacerbates semiconductor market pressures.
China ADAS Chips Enter Cutthroat Competition
Domestic Chinese NEV brands push into high-end markets with rapid iterations. Fierce competition heats up, thrusting国产 automotive-grade chips into a survival battle.
Yin Qi Sees Himself in Young AI Peer
SenseTime founder Yin Qi recognizes his past self in Yang Zhilin. The two Tsinghua alumni, five years apart in age, are propelled into distinct AI eras by destiny.
Duolingo Stock Crashes 80% Post-6x Boom
Duolingo, a former 6x bull stock, plummeted 80% in one year. The analysis attributes both its meteoric rise and sharp decline to AI.
Kling Creator Returns to Alibaba with HappyHorse Hit
The creator of Kuaishou's Kling AI returns to Alibaba and launches a new dark horse product, HappyHorse. ByteDance faces heightened competition, beyond just HappyHorse topping the charts.
Windows 11 Lifts 30-Year FAT32 32GB Limit
Microsoft's latest Windows 11 Insider previews in Dev channel (26300.8170) and Beta channel (26220.8165) optimize storage settings page performance and user interactions. A key change removes the artificial 32GB partition size limit on FAT32 file system, in place for 30 years.
Nippon Columbia Launches ¥10M AI Anime Contest
Nippon Columbia Group announced the launch of 'COLOTEK' contest soliciting works utilizing AI, particularly AI-generated anime. The contest offers a total prize pool of 10 million yen. It aims to promote innovative AI applications in animation.
Claude Tops EPL Prediction; Grok Flops
General Reasoning's report ranks Claude Opus 4.6 best in 2023-24 Premier League betting simulations among eight AI models. Grok ranked last, wiping out funds in one test and failing others. AI models overall underperform humans in dynamic prediction tasks.
Vidu Q3 Revives Reference Image King
Vidu's Q3 update quietly reintroduces powerful reference image generation capabilities. It promises seamless video creation from user stories, ideal for animation, dramas, short videos, films, and ads. Videos are delivered ready-to-use out of the box.
Wait Out AI Super-Spending False Start
Merryn Somerset Webb interviews Fractal Brain CEO Janusz Marecki on AI hype versus reality. Discussion centers on LLMs hitting limits like data ceilings, diminishing compute scaling returns, hallucinations, and probabilistic errors.
ViSA-R2 Infers Physics from Visual Fields
ViSA-R2 recovers analytical SymPy expressions from 2D linear steady-state field visualizations using a self-verifying chain-of-thought pipeline mimicking physicist reasoning. It introduces ViSA-Bench, a synthetic benchmark with 30 verifiable scenarios for VLM evaluation. Built on 8B Qwen3-VL, it outperforms open-source baselines and frontier VLMs.
StaRPO: Stability RL for Reliable Reasoning
StaRPO introduces a stability-augmented RL framework for LLMs, incorporating reasoning stability via ACF for local coherence and PE for global goal-directedness. It combines these lightweight metrics with task rewards for process-aware feedback. Experiments on four benchmarks show superior accuracy and logical stability over baselines.
SPPO: Efficient PPO for Long Reasoning
SPPO is a scalable algorithm that improves Proximal Policy Optimization (PPO) for aligning LLMs on long-horizon reasoning tasks. It reformulates the process as a sequence-level contextual bandit, using a decoupled scalar value function for stable, low-variance advantages without multi-sampling. Experiments show it outperforms standard PPO and matches compute-heavy methods on math benchmarks.
Scaling Agentic AI to Enterprise Production
Article details building production-grade agentic AI enterprises by starting with KPIs as agent goals, decomposing tasks from human roles, and using robust multi-modal integrations. It emphasizes governance, observability, and reliability to move beyond pilots. Edgeverve shares production examples in finance and facilities.
RAMP: Hybrid DRL for Numeric Action Learning
RAMP strategy enables online learning of numeric action models via DRL policy training, interaction-based model refinement, and integrated planning in a feedback loop. It introduces Numeric PDDLGym, a framework converting numeric PDDL problems to Gym environments. Experiments show RAMP outperforms PPO on IPC numeric domains in solvability and plan quality.
PilotBench: Safe Aviation AI Benchmark
PilotBench is a benchmark for evaluating LLMs on safety-critical flight trajectory and attitude prediction using 708 real-world aviation trajectories. It introduces Pilot-Score, weighting 60% accuracy and 40% instruction/safety compliance. LLMs excel in controllability but lag in precision versus traditional forecasters, especially in complex flight phases.
PETITE: Tutor-Student Boosts LLM Coding
PETITE introduces tutor-student multi-agent interaction using the same LLM to enhance coding problem-solving. The student agent generates and refines code, while the tutor provides feedback without ground-truth access. It matches or exceeds SOTA methods like Self-Consistency on APPS benchmark with far fewer tokens.
OpenKedge: Safe Agent Mutation Protocol
OpenKedge redefines AI agent mutations as governed processes via declarative intent proposals evaluated against system state and policies before execution. It enforces safety through execution contracts with bounded actions, resources, and time using ephemeral identities. The protocol introduces Intent-to-Execution Evidence Chain (IEEC) for cryptographic auditability, proven effective in multi-agent and cloud scenarios.