All Updates
Page 787 of 1666
April 30, 2026
Financial Engineering Aids Cancer Cures
Explores how financial engineering mitigates biotech risks like 95% drug failure rates via Andrew Lo's strategies, royalties, PRV, and CVR. Highlights AI-driven drug research and prediction markets to fund uncertain life sciences. Spotlights longevity biotech like Impetus Grants and General Control.
Beijing Drone Sales Ban Sparks Rush
Beijing's May 1 drone sales ban leads to store clearances, stockpiling up to 3-unit limits, and resales. DJI stores pivot to products like Osmo Pocket 4; users turn to AI-generated aerial views as alternative.
DingTalk Launches 180-Hour Power Bank Recorder
DingTalk has launched a power bank version recording card in response to user feedback. It supports continuous recording for 180 hours. The device can standby for half a year without charging.
vivo X300 Adds Disney Watermarks & AI Telephoto
vivo rolled out imaging upgrades for X300 Ultra and X300s, featuring Shanghai Disney watermarks and travel styles. Cloud-based large model enhances long-focal clarity and details. New editing, video stabilization, and external storage support added.
Horizon Capital Surges 67% to $2.5B
Beijing Horizon Information Technology increased registered capital from $1.5B to $2.5B, up 67%. Founded in 2015 by Yu Kai, it develops AI chips, software/hardware, and related tech for autos and more. Fully owned by Horizon Robotics Holdings Limited.
Tesla Semi Enters Mass Production
Tesla announced its first Semi truck has rolled off the mass production line. The pure electric long-haul truck enters a new mass production phase. Long-range version offers about 500 miles per charge, with deliveries starting this year.
Claude Pro Faces Feature Cuts and Price Hikes
Anthropic tests removing Claude Code from $20 Pro tier, shifting it to $100 Max plan amid compute cost pressures. Recent changes include blocking third-party agents and switching enterprise to usage-based billing. Rapid feature rollouts like Cowork and Connectors build user lock-in before tiered pricing extracts more value.
Qwen-Code v0.15.6-preview.0 Bug Fix Release
Qwen-Code releases preview version v0.15.6 with multiple bug fixes enhancing CLI stability, memory handling, and core functionality. Key improvements include preventing UI flicker, preserving reasoning content across operations, and correcting model precedence. New contributor @cyphercodes adds directory persistence features.
Xiaohongshu Ramps Up AI with New Dept
Xiaohongshu announced an organizational upgrade integrating community, e-commerce, commercialization, and tech systems. It established AI Level 1 department Dots and Enterprise Intelligence Dept to boost AI from product and org sides. Conan appointed President to oversee core businesses, reporting to CEO Shiya.
Onchain LLM Agents Trade $20M Real ETH
In a 21-day deployment, 3,505 user-funded LLM agents traded real ETH in DX Terminal Pro, generating $20M volume and 99.9% settlement success. Reliability stemmed from operating-layer controls like prompt compilation, policy validation, and execution guards, not just the base model. Pre-launch testing fixed key failures, boosting capital deployment from 42.9% to 78%.
OMEGA Generates ML Algorithms Beating Scikit-Learn
OMEGA is a full end-to-end framework automating AI research from idea generation to executable code, producing novel ML classifiers. It outperforms scikit-learn baselines across 20 benchmark datasets in infinity-bench. Models are available via pip install omega-models.
LLM Safety Flunks in Health Robots
Researchers created a dataset of 270 harmful instructions based on medical ethics to test 72 LLMs for robotic health attendant control. Mean violation rate hit 54.4%, with proprietary models far safer than open-weight ones. Safety defenses like fine-tuning and prompts offered limited gains, blocking clinical use.
LLM Persuadability in Legal Decisions
Researchers examine how frontier LLMs respond to legal arguments from contending parties, testing if advocate quality influences decisions. Experiments reveal factors affecting persuadability in open- and closed-weight models. Results question LLMs' readiness for judicial and administrative roles.
iLTN Proves Grounding Insufficient for Reasoning
New research challenges the assumption that symbol grounding alone enables compositional reasoning in neuro-symbolic AI. Introduces iLTN, a differentiable architecture for multi-step deduction, showing grounding-trained models fail to generalize while joint grounding-reasoning training succeeds in zero-shot tasks.
Google Identifies 10 Cognitive Abilities for AGI
Google DeepMind published a paper proposing a cognitive framework to measure AGI progress. It addresses the shortage of empirical tools for evaluating AI intelligence. The new method is grounded in cognitive science.
Evidence-Grounded Multi-Persona from User Logs
A hierarchical framework aggregates user actions into intent memories and induces multiple evidence-grounded personas via clustering and labeling. It optimizes persona quality—cohesion, alignment, truthfulness—using groupwise Direct Preference Optimization (DPO). Experiments on large logs show more coherent, trustworthy personas improving interaction prediction.
DreamProver Evolves Lemma Libraries via Wake-Sleep Agent
DreamProver is an agentic framework using a wake-sleep paradigm to discover reusable lemmas for formal theorem proving. In the wake stage, it proves theorems and proposes candidate lemmas; in the sleep stage, it abstracts and optimizes them into a compact library. It significantly boosts proof success on math benchmarks with concise proofs and lower compute costs.
Distill-Belief Tackles Reward Hacking in Source Localization
Distill-Belief introduces a teacher-student framework for closed-loop inverse source localization and characterization (ISLC) in physical fields, addressing belief-space challenges under time constraints. A Bayes-correct particle-filter teacher provides accurate posteriors and information-gain signals, distilled into a compact student for efficient control and uncertainty certification. Deployment uses only the student for constant per-step cost, outperforming baselines in sensing efficiency and accuracy across seven field modalities.
BTF-2 Benchmark Tests AI Forecasting Reasoning
Introduces Bench to the Future 2 (BTF-2), a benchmark with 1,417 pastcasting questions and a frozen 15M-document corpus for reproducible AI forecasting evaluation. It detects tiny accuracy differences (0.004 Brier score) and separates research from judgment skills. Builds a superior forecaster and pinpoints frontier agents' failures in strategic reasoning like leader incentives and institutional processes.
Apriori Uncovers Learned Helplessness in Math Tutoring
This ArXiv paper uses the Apriori algorithm to mine behavioral patterns in math tutoring logs linked to learned helplessness (LH). Low-LH students show persistence and hint use leading to success, while high-LH exhibit skipping tied to failure. Interventions influence pattern strengths, with skipping without hints predicting unsolved problems.