Every story we have kept, newest first.
Looking for the daily editions? → Past editions
Page 260 of 1375
OpenAI explains how it built GPT-Live, a realtime system for continuous voice interaction, in six months. The system combines a turnless speech model with a low-latency architecture to make AI conversations more responsive and natural.
OpenAI News · 55d ago
Elon Musk suggests that money may become obsolete within a decade due to AI advancements and predicts China will emerge as a leader in the AI sector despite geopolitical tensions.
虎嗅 · 55d ago
HanZhi has officially launched its GEO product, marking its entry into the competitive AI solution market. The move highlights the growing trend of enterprise software brands pivoting toward AI-native answers.
量子位 · 55d ago
Microbial fermentation toilets are failing to maintain odor-free environments during summer due to poor maintenance, design flaws, and system overload. The article highlights the gap between theoretical environmental benefits and the reality of poor operational management.
虎嗅 · 55d ago
Detroit automakers are seeing increased profits from high demand for fuel-powered trucks and SUVs. In contrast, European manufacturers are struggling with declining sales and restructuring pressures.
36氪 · 55d ago
The AI industry is moving away from a 'who is strongest' pricing model to one based on 'replaceability' and total cost of ownership. Kimi K3's open-weight strategy and Anthropic's aggressive pricing cuts signal a new era of market competition.
虎嗅 · 55d ago
Chery Automobile has agreed to a $75 million investment in South Korea's KG Mobility via convertible bonds. This strategic move aims to expand Chery's international footprint and deepen automotive partnerships.
TechNode · 55d ago
Despite a decline in shipment volume, global smartphone revenue grew 7% to a record $109 billion in Q2 2026. This growth is driven by a 17% increase in average selling price (ASP) as consumers shift toward premium segments.
36氪 · 55d ago
The article analyzes the shift from model-centric development to agent-based workflows, highlighting the economic challenges of input token costs and the importance of simulation in achieving AGI.
虎嗅 · 55d ago
E-Vite Aviation has completed a Series A+ funding round to advance its 2-ton class eVTOL, the ET9. The company focuses on a self-developed 800V electric propulsion system and follows a rigorous civil aviation system engineering approach.
36氪 · 55d ago
GCL Integration has established a new energy management company in Suzhou with a registered capital of 10 million RMB. The firm focuses on solar, wind, and energy storage technology services.
36氪 · 55d ago
Shanghai has officially registered 11 additional generative AI services, bringing the city's total to 211. This update reflects the ongoing regulatory oversight and expansion of the local AI ecosystem.
TechNode · 55d ago
Shanghai-based BCI startup Active Technology has secured RMB330 million in a record-setting angel round. The company specializes in invasive brain-computer interface technology.
TechNode · 55d ago
Couwa Robotics' Chief Scientist Yang Xue discusses the transition of embodied AI from lab demos to real-world deployment. The focus is on reducing task completion time variance and improving error recovery in complex urban environments.
虎嗅 · 55d ago
Smartphone manufacturers are moving toward a tiered pricing model for AI features, where basic capabilities are included but advanced, cloud-dependent AI agents require subscriptions. This reflects the high operational costs of running large-scale cloud models.
虎嗅 · 55d ago
This Perspective argues that LLMs lack sufficient evidence of safety for autonomous triage of undifferentiated patients. The central problem is not medical knowledge, but failure to gather missing information, detect red flags, and prioritize rare yet catastrophic diagnoses under uncertainty.
ArXiv AI · 55d ago
ViSAGE is a multimodal agentic memory framework for long-form video understanding that creates entity-centric, self-correcting memories. It improves identity consistency and evidence verification, achieving 5.9% higher accuracy than the strongest baseline.
ArXiv AI · 55d ago
ThinkReset introduces a reusable intermediate interface that allows language models to continue solving long-horizon tasks after discarded reasoning history. It optimizes post-reset continuation success to reduce redundancy, context overflow, error anchoring, and premature guessing.
ArXiv AI · 55d ago
TAPR is a Task-Aware Prompt Rewriter trained to transform user prompts into task-optimized instructions for downstream LLMs. Experiments across question answering, summarization, and arithmetic reasoning show consistent performance gains, including on Natural Questions and GSM8K.
ArXiv AI · 55d ago
Researchers introduce two sound encodings for planning multi-agent paths under spatio-temporal and topological constraints expressed in STL-GO. The mixed-integer programming and satisfiability modulo theory approaches are evaluated on multi-UAV search-and-rescue scenarios with dynamic graph interactions.
ArXiv AI · 55d ago
This research presents a production extraction layer that uses ontology-guided prompting, deterministic deduplication, and embedding-based entity resolution to build validated knowledge graphs from heterogeneous documents. Tested on intelligence corpora, it raised search recall from about 70% to 95% without false merges.
ArXiv AI · 55d ago
SARE is a framework for measuring computational effort at individual chain-of-thought steps rather than only across an entire reasoning trajectory. Evaluations across six benchmarks and three open-weight LLMs found that reasoning energy varies by step type, drops at critical points in incorrect solutions, and can outperform output-based confidence signals.
ArXiv AI · 55d ago
This study evaluates whether LSR-Synth can distinguish genuine symbolic discovery from equation memorization and conventional operator search. It finds that a fixed, semantics-free vocabulary solves most current tasks, while language-model-generated candidates add substantial value mainly when parts of that vocabulary are deliberately removed.
ArXiv AI · 55d ago
GenCDSR is a generative framework for cross-domain sequential recommendation that combines hybrid tokenization with serial-parallel decoding. On three public datasets, it reports a 1.5% average accuracy improvement and an 85.1% reduction in inference latency versus existing methods.
ArXiv AI · 55d ago
Factory runs its cloud backend on Vercel and Next.js, serving tens of millions of daily API requests with a p95 response time of 350ms. API routes, middleware, webhooks, analytics pipelines, observability, and WAF protection operate without a dedicated infrastructure team.
Vercel News · 55d ago
This study introduces a rigorous benchmarking protocol for autonomous AI research systems using multi-model LLM peer review. It evaluates four frameworks against commercial benchmarks, revealing significant performance gaps and validating the reliability of automated evaluation.
ArXiv AI · 55d ago
Researchers have developed a three-stage pipeline that uses LLMs to systematically generate, validate, and formally verify new mathematical conjectures. The framework ensures novelty and significance by integrating Lean 4 and Mathlib for rigorous formal proof checking.
ArXiv AI · 55d ago
Ulanzi founder William discusses the strategy behind scaling a hardware company to global leadership in 11 years, focusing on rapid product iteration and AI-driven management.
虎嗅 · 55d ago
The article explores the tension between industrial automation, such as Boston Dynamics' Atlas, and worker anxiety, contrasting it with Elon Musk's vision of a post-labor society.
虎嗅 · 55d ago
The trend of high school students entering corporate internships is accelerating due to competitive job markets and AI-lowered skill barriers. Companies are increasingly formalizing these programs to capture talent earlier.
虎嗅 · 55d ago