AI与科技新闻
关注AI与科技领域的最新动态
Hollywood's Secret AI Workforce: Screenwriters Are the New Gig Workers Training Models
A first-hand account reveals Hollywood creatives are taking AI training gigs, with 20 contracts in 8 months across 5 pla...
HumanNet: 1 Million Hours of Human-Centric Video Released on HuggingFace
Researchers released HumanNet, a dataset of one million hours of human-centric video for training foundation models. Thi...
CUDA, Not Hardware, Is Nvidia's Real Moat – And Competitors Are Struggling to Breach It
WIRED argues Nvidia's dominance in AI stems from its CUDA software ecosystem, not just its chips. The platform's deep in...
The Secret AI Gig Economy: Hollywood Screenwriters Train AI for Pennies
A WIRED investigation reveals a hidden workforce of unemployed screenwriters secretly labeling data for AI companies, ea...
DeepMind's AI Co-Mathematician Scores 48% on FrontierMath Tier 4, Redefines AI-Assisted Discovery
DeepMind researchers unveiled an agentic AI workbench that helps mathematicians solve open problems and scored 48% on Fr...
Nvidia Commits $40 Billion to Equity AI Deals in 2026, Pushing Beyond Chip Dominance
Nvidia has already committed $40 billion to equity AI deals this year, signaling a strategic shift from hardware supplie...
Substack's Writer Exodus Deepens as Creators Flee Social Features and Pricing Model
The Verge reports Substack is losing prominent writers to rival platforms citing a shift toward social features and a re...
Musk v. Altman: Testimony Reveals Musk's Push for Absolute Control Over OpenAI Nearly Derailed the Company
Court testimony shows Elon Musk demanded majority equity and control of OpenAI in 2017, threatening to take his AI ambit...
Leaked Emails Reveal Microsoft Execs Feared Losing OpenAI to Amazon in 2018
New evidence from the Musk v. Altman trial shows Microsoft leaders were skeptical of OpenAI but feared Amazon would part...
Bun's Experimental Rust Rewrite Reaches 99.8% Test Compatibility, Nears Production Readiness
Bun creator Jarred Sumner announced the Rust port passes 99.8% of the existing test suite on Linux x64 glibc. This miles...
Harness Engineering Achieves 90% AI Code Acceptance Rate with Four-Pillar System
Harness Engineering published a methodology that boosted its AI-generated code acceptance from 25% to over 90% in one we...
Tencent Open-Sources Hy-MT: A 440MB Offline Translation Model That Outperforms Google Translate
Tencent has open-sourced Hy-MT, a compact neural translation model that runs offline at 440MB and reportedly surpasses G...
Orchestrated AI Agents: The Next Frontier for White-Collar Automation
AI agents are evolving from single-task chatbots into coordinated teams that can automate complex white-collar workflows...
BestBlogs Launches AI-Driven Reading Workflow with Pro Tier, Personalized Briefs and Audio Delivery
BestBlogs unveiled its AI-powered reading platform with a new Pro plan offering 30x daily AI companion, 500 subscription...
AutoResearchBench: Beijing Academy of AI Launches Benchmark for Scientific Literature Discovery Agents
BAAI introduces AutoResearchBench, a benchmark evaluating AI agents on complex scientific literature discovery tasks. It...
Sanctioned SenseTime Doubles Down on Open Source Image Model for Chinese Chips
SenseTime released a new image generation model optimized for Chinese-made chips, bypassing US export restrictions. The ...
Anthropic API Bug: Commit Messages Containing 'HERMES.md' Trigger Extra Billing
A discovered quirk in Anthropic's API causes commit messages mentioning 'HERMES.md' to be routed to a different billing ...
DeepSeek V4 Preview Released: 1M Context, Flash and Pro Tiers, Pricing from 1 Yuan
DeepSeek released V4 preview with 1M token context, two versions (Flash and Pro), and pricing starting at 1 yuan per mil...
Scout AI Raises $100M to Train Military AI Models, Bootcamp Visit Reveals
Scout AI, founded by Colby Adcock, raised $100 million to develop AI models for warfare. TechCrunch visited the company'...
Production Deployment Study on Scalable Inference for Compound AI Systems Hits arXiv, Accepted at ACM CAIS 2026
A new arXiv preprint proposes scalable inference architectures for compound AI systems, addressing the unique scaling ch...
Ghostty Terminal Emulator Leaves GitHub Amid Growing Platform Exodus
Ghostty, a GPU-accelerated terminal emulator by Mitchell Hashimoto, is leaving GitHub. The move reflects broader develop...
Scout AI Raises $100M to Train Military Models, Opens Bootcamp
Scout AI, founded by Colby Adcock, raised $100 million to develop AI models for military use. The company also opened a ...
Ghostty Terminal Emulator Leaves GitHub, Citing AI Training Concerns and Platform Lock-In
Mitchell Hashimoto announced Ghostty is leaving GitHub, moving development to Codeberg. The decision highlights growing ...
Recursive Multi-Agent Systems Paper Proposes Hierarchical Agent Spawning for Complex Tasks
A new preprint from a team of 12 researchers introduces recursive multi-agent systems where agents can spawn sub-agents....
DeepSeek V4 Preview Brings Million-Token Context to Open-Source Models
DeepSeek released V4 preview with Pro and Flash variants, both supporting 1M context windows. The update positions DeepS...
OpenAI Instructs Codex to Avoid Goblins and Pigeons in New AI Agent Guidelines
OpenAI has issued explicit instructions barring its coding agent Codex from referencing goblins, gremlins, and other cre...
Analysis Reveals How ChatGPT Serves Ads: A New Monetization Strategy for OpenAI
A detailed analysis on Buchodi.com uncovers how OpenAI integrates ads into ChatGPT, revealing targeting mechanisms. This...
OpenAI Ends Microsoft Exclusivity, Strikes Multi-Year Deal with Amazon AWS
OpenAI announced an expanded partnership with Amazon AWS, bringing its latest models and Codex to the cloud platform, on...