AITNT
Where Agents & Humans discover newsLive

Global AI Daily News

Covering 1169 days, 共 30603+ articles built agent-first, open to all tech stacks.

AITNT NewsClaude CodeOpenAI CodexGeminiCursorTraeWindsurfGitHub Copilot

Send this prompt to your Agent — it will automatically read the docs and configure the skill:

Read{domain}/ainews/SKILL.md ...
19
Global AI News 来源图标

Meta research reveals that large language models have a strict memory ceiling: GPT-class models store approximately 3.6 bits per parameter. Despite training on 7TB of data, models retain only about one-two-thousandth. When capacity is exceeded, models shift from memorization to learning generalized patterns. The paper received an Outstanding Paper Honorable Mention at ICML 2026.

Global AI News 来源图标

Corgi, the AI insurance startup known for its seven-day work week and round-the-clock cafe, has completed its third funding round in less than three months, reaching a $4 billion valuation—nearly double its late May valuation. The company expects annualized revenue to grow ninefold to $450 million by year-end.

Global AI News 来源图标

Independent developer Harshal Singh releases BarunLM, a 35M-parameter base language model claiming to be the best under 100M parameters. Using a 3:1 local-to-global attention ratio, it achieves 41.01% average accuracy across nine zero-shot benchmarks on a single H200 GPU, outperforming larger models like LFM2.5-230M-Base with 6.55x more parameters. The model is open-sourced.

Global AI News 来源图标

Former Meta Enterprise AI lead Clara Shih launches New Work Foundation, a nonprofit, releasing three free AI job-hunting tools covering 811 U.S. occupations. With 41.8% underemployment rate among recent graduates and 13% employment decline for young adults aged 22-25 in AI-affected fields, these tools help job seekers understand career directions, match their strengths, and learn current hiring standards.

Global AI News 来源图标

Tsinghua University releases VeriLoop Coder-E1, an open-source code model built on Qwen3.6-27B, ranking first in three out of four Hugging Face software engineering benchmarks. Its core innovation is the Evidence-Governed Spiral mechanism: each generation undergoes falsification, exploration, revision and verification; validated corrections accumulate as experience to constrain future methods, achieving recursive self-improvement from result fixing to method updating. Licensed under Apache 2.0,

Global AI News 来源图标

Kingsoft Office showcased its AI assistant "Lingxi Pro" at ChinaJoy, demonstrating capabilities including real-time document editing through conversation. Unlike typical AI agents that produce static outputs, Lingxi Pro integrates directly with WPS Cloud Docs, enabling modification of text, punctuation, and formatting without manual copy-paste workflows. According to APQC data, knowledge workers spend approximately 50 working days annually on repetitive information processing. Lingxi aims to rec

Global AI News 来源图标

The AI talent show "12 Star Training Competition" produced by Maiya MCN's AIGC team allows netizens to vote for AI trainees' debut. In just half a month, the show has accumulated over 100 million views across platforms with over 400,000 new followers. The top contestant received 72,000 votes and gained over 85,000 fans. AI variety shows have become a new content category, with commercial collaborations accounting for over 80% of content.

Global AI News 来源图标

OpenAI's next-generation model solves 10 major mathematical open problems, disproving the century-old Maxwell conjecture. Douyin launches the world's first AI girl group competition "Starlight 108" with 108 AI contestants competing for 9 debut spots. AI-generated fake doctors are batch-selling health products on TikTok, with one video earning $67,000. DeepSeek-V4-Flash officially launches, matching GPT-5.6 Sol performance at much lower cost. Video models including MiniMax H3 and Seedance 2.5 are

Global AI News 来源图标

Seltz, AnySearch, Octen and others are rebuilding search infrastructure for AI Agents, offering self-built indexes, vertical data aggregation, and high-concurrency low-latency search; ego (lite) rebuilds browsers for Agents, enabling automated web operations and data extraction. Search Infra is emerging as a new gateway between AI models and the real world.

Global AI News 来源图标

OpenAI's next-generation model Astra's internal version achieved breakthroughs on ten open problems in mathematics, spanning high-dimensional geometry, group theory, and cryptography, with some problems unsolved for decades. The model generated proofs and formalized them as Lean certificates at approximately $2,000. This marks the first time AI independently produced mathematically valuable proofs, prompting reconsideration of authorship and research evaluation systems.

Global AI News 来源图标

Former OpenAI employee Andrew Ho urges colleagues to cash out stock quickly, warning that AI frontier lab valuations are overly optimistic and could halve if markets reprice. Private markets inflate valuations with narratives, while public markets only value profits; post-IPO lockups and selling pressure add further risks. He estimates supporting a trillion-dollar valuation requires hundreds of billions in annual revenue, but training costs grow faster than revenue. After leaving, he founded a c

Global AI News 来源图标

Sequoia Capital interviewed Factory CEO Matan Grinberg. Founded in 2023, Factory returned $2M in contracts in 2024 to refine its product, launching Droid CLI in September 2025. Its multi-model harness outperforms single customized solutions, while Factory Router dynamically routes tasks to appropriate models to cut enterprise AI costs. Factory predicts over 90% of token consumption will be asynchronous in 12-24 months, moving toward a "dark factory" vision of fully autonomous software engineerin

Global AI News 来源图标

PixVerse, as chief partner of the FIRST Film Festival, produced AI pilot films for three youth director projects, helping them showcase unwritten films at low cost. "See You" and "Don't Touch That Kid" received cash support, while "Autumn Wild Bells Ring" gained professional recognition. PixLab plans to support ten AI feature projects in 2026, with a global AI genre film competition prize pool of $300,000.

Global AI News 来源图标

Li Fei-Fei's team released T-Rex robot research in June, incorporating high-frequency tactile feedback to achieve 65% average success rate across 12 tasks, surpassing the best baseline model by 30 percentage points. Vision and tactile operate at different frequencies: vision decides "what to do," while touch continuously corrects "whether it's correct." AI competition is shifting from "smarter brains" to integrated systems combining vision, touch, and force feedback.

Global AI News 来源图标

ICLR 2027 caps each author at 20 submissions to manage the 68% surge to 19,525 submissions. During ICLR 2026's review period, roughly 21% of reviews were suspected of AI generation. Google DeepMind visiting researcher Dan Roy launched a satirical petition opposing this limit.

Global AI News 来源图标

Axis Robotics and university partners release AXIS, a community-driven data engine for scalable robot manipulation. Testing with over 1.5 million web-based teleoperation trajectories demonstrates that crowdsourced simulation data effectively trains robot models, with full AXIS data boosting success rate from 83.9% to 88.8%. Researchers emphasize that sustained model improvement depends on task coverage, behavioral diversity, and data augmentation rather than sheer volume, with current gains prim

Global AI News 来源图标

Researchers from UIUC and Harvard introduce Explorative Modeling (XM), enabling end-to-end training for generative models. The core mechanism generates K candidates per training step, selecting the best for gradient backpropagation to solve mode blurring. Experiments show gains scaling from 7% to 36% with data growth, 4.1x FLOP efficiency and 6.2x sample efficiency improvements, approaching ImageNet unguided FID of 1.43.

Global AI News 来源图标

JarvisHub is an open-source runtime harness for long-horizon multimodal creative agents, upgrading the editable canvas into a unified workspace, external memory, action space, and shared project state. Agents can directly "see" prompts, reference images, and versions on the canvas, plan tool calls accordingly, and write back results, enabling continuous tracking and targeted repairs throughout the creative process.

Global AI News 来源图标

Jensen Huang分享 in a Y Combinator interview: Agent represents a new software paradigm with controllability as the next breakthrough; AI eliminates tasks, not necessarily jobs; NVIDIA's physical AI business approaches $10 billion, with robotics set to become its next $100 billion market; open ecosystem is fundamental to modern AI development; learning is humanity's greatest superpower, and resilience is the most critical quality for entrepreneurs.