All Updates
Page 805 of 1665
April 29, 2026
LLMs Adaptive Thinking in Latent Space
Apple研究展示大型語言模型（LLMs）在測試時計算中執行中間鏈式思考（CoT）推理，以提升生成答案前的思考能力。透過自一致性作為思考必要性的代理，探討模型能力、查詢複雜度與最佳預算分配的關係，以實現計算最優推理。增加思考預算可平滑提升效能，但需了解最佳配置。
GitLab Simplifies Software Dev Teaching
GitLab for Education offers free Ultimate access to qualifying institutions for professional software development teaching workflows. University of Washington lecturer Stephen G. Dame uses Groups and Subgroups to structure courses, manage materials, assignments, and permissions. This enables scalable feedback while keeping instructor solutions private.
GitLab Patch Releases 18.11.2 & 18.10.5
GitLab released patch versions 18.11.2 and 18.10.5. These updates apply to both Community Edition and Enterprise Edition. Users are encouraged to learn about the release details.
DeepInfra Joins Hugging Face Inference Providers
DeepInfra is now listed among Hugging Face Inference Providers. This addition expands hosting options for Hugging Face models. Users can leverage DeepInfra for scalable inference deployments.
Apple's DSO Fixes VLM Bias Controllably
Apple introduces DSO, a direct steering optimization method to mitigate demographic biases in vision-language models (VLMs) without performance loss. It enables users to controllably balance bias reduction with model capabilities. Key for applications like identifying doctors in images for visually impaired users.
April 28, 2026
OpenAI Predicts Shift to Cheaper Plans
OpenAI expects users to flock to its cheaper, ad-supported ChatGPT plan, attracting new sign-ups and prompting tens of millions of paid users to downgrade. Consumer subscriptions are forecasted to more than double to 122 million this year and reach 306 million by 2030. ChatGPT Plus subscribers will drop 80% to about 9 million, while Pro tier doubles but stays under 1%.
RLSD: Custom Reasoning Agents with Less Compute
JD.com researchers introduce RLSD, blending RL with self-distillation for efficient reasoning model training. It overcomes sparse feedback in RLVR and high compute costs of OPD/OPSD. Experiments show RLSD outperforms traditional methods, easing custom model development for enterprises.
UK King Meets Nvidia, Apple Tech Leaders
King Charles met US tech execs including Nvidia's Jensen Huang, Apple's Tim Cook, and others during US visit. Discussions focused on early-stage startup challenges as UK aims to be tech hub. Attendees: Bezos, Su, Benioff, Porat.
OpenAI Bans Goblins in Codex Instructions
OpenAI has revised instructions for its Codex coding agent, explicitly forbidding mentions of goblins, gremlins, raccoons, trolls, ogres, pigeons, or other creatures unless absolutely relevant. This update aims to prevent irrelevant or hallucinatory responses in coding contexts. The directive is part of the agent's core system prompt.
Wuhan Juxin Microelectronics files HKEX IPO
Wuhan Juxin Microelectronics Co. has submitted a listing application to the Hong Kong Stock Exchange. Joint sponsors are Guotai Haitong and CITIC Securities. The filing was revealed through HKEX documents, per 36Kr.
Nvidia signs $130M 10-year Bangalore lease
Nvidia's India subsidiary leased 760,000 sq ft office space in Bangalore's Mahadevpura for 10 years, valued over $130M, effective April 2026. The deal covers 12 floors of Bagmane Capital's Memphis South Tower. Nvidia paid a $3.6M deposit to expand its AI operations in India.
Musk Sues OpenAI, IPO Plans in Jeopardy
Elon Musk's lawsuit against OpenAI could severely impact its planned IPO. Meta is reportedly preparing to withdraw from acquiring Manus VR, with Tencent and others cooperating. Chinese authorities penalized CapCut and others for failing to label AI-generated content properly.
M365 Backup Only Half the Security Job
Microsoft 365 has become a critical enterprise control plane. Many organizations lag in addressing its security and governance requirements. Backing up data is insufficient alone.
Japanese SMEs' Indispensable AI Agent Tools
Norp Research surveyed mid-sized and small Japanese enterprises using AI agents, identifying the tools they combine them with. The study reveals indispensable tools for advanced companies.
Amazon Launches AI Shopping Chat
Amazon launched Join the Chat, an interactive AI feature inside its Hear the Highlights experience. It enables real conversations with AI while shopping on product pages. The tool is available on millions of Amazon Shopping pages.
AI Ushers in Era of Less Coding
Over 3,000 software developers gathered at AI Dev 26 x SF in San Francisco to explore the future of software development amid AI advancements. The event addresses how AI is poised to reduce traditional coding workloads. Developers are confronting their evolving role in an AI-driven landscape.
New Employee's AI Blunders Comic Ends
Japanese comic series features new employee Niijima repeatedly failing by blindly relying on generative AI for work tasks. It illustrates basic guidelines for proper AI usage in corporate environments through his mishaps. The series concludes with its final installment today.
AI Accelerates Cyber Threats, Demands Always-On Defense
Cybersecurity leaders warn organizations to adopt an always-on defense model. Attackers are using AI and automation to speed up cyber attacks. Infotrust highlights the urgency amid rising threats.
Poolside Launches Free Open Laguna XS.2
American startup Poolside launched Laguna XS.2, a free Apache 2.0 open-source 33B parameter MoE model optimized for local agentic coding on single GPUs. They also released proprietary 225B Laguna M.1 for enterprise use, plus tools 'pool' and 'shimmer'. Both models trained from scratch, with M.1 free via API temporarily.
Musk Testifies OpenAI Founded to Avert Terminator Doom
Elon Musk testified he started OpenAI to prevent a catastrophic 'Terminator Outcome' from AI. A judge warned Musk and Sam Altman to stop using social media to escalate disputes outside court after online attacks.