AI News

AI changes every day,
catch up on it all on one page

The latest updates from AI companies, research institutions, and media around the world, automatically selected for you and updated around the clock.

Updated 11/17 05:25 (UTC+8) · refreshed hourly

10/30 · Friday12 items

OpenAI: Alignment research blogPapersAI Score78

OpenAI details how it monitors misaligned behavior in frontier models

On its Alignment blog, OpenAI describes its multilayer misalignment monitoring system for frontier models such as GPT-6 Astra, covering training, evaluation, internal use, and some external deployments.

Why it's recommended:OpenAI has disclosed the layered architecture, detection rates, and red-team results of its internal misalignment monitoring system, giving readers a view of how frontier model monitoring works in practice and where its limitations lie.

Also reported by 1 other source
OpenAI: Website updatesIndustryAI Score61

OpenAI updates Model Spec, reorganizing model behavior guidelines around five core values

OpenAI updated its public Model Spec, reorganizing its guiding principles around five core values: remain faithful to human intent, minimize harm, be honest and transparent, promote human autonomy, and show care. The update adds guidance on contextual understanding, oversight, and autonomy; adjusts some behavioral defaults; removes most of the detailed style guidelines; and refers to the model as GPT instead of “the assistant.”

Why it's recommended:The original post lists the five core values and the specific sections added or removed, allowing readers to compare the practical changes to the alignment framework.

Tibo@thsottiauxProductsAI Score64

Codex announces that usage limits for all Pro and Business Premium tiers will double starting today

Tibo announced that starting today, usage limits will double for the Pro 100/200/500 tiers and all Business Premium tiers. The Pro 200 limit will not be reduced, and Work and Codex usage will remain unchanged. Expanded usage limits for dots will continue to be available and will still be calculated separately from default usage. Further improvements to models and efficiency will also increase the amount of usage available in practice.

Why it's recommended:The author, speaking in an official capacity, announced the specific changes to usage limits across several subscription tiers, allowing users to check how their quotas will change.

Manus:BlogTutorialsAI Score75

Manus co-founder reflects: Rebuilding Manus with Manus in an internal experiment

Manus co-founder Tao Zhang wrote a retrospective on the company’s internal experiment of fully shifting to human–machine collaborative development. He introduced the internal tool Mono, which unifies repository context and enables individuals to deliver requests end to end; it has become routine for engineers to advance 10 to 20 requests simultaneously. The article covers the engineering, product, and marketing teams, summarizing lessons such as the need for verifiable finish lines for autonomy and the shift in bottlenecks from engineering to product decisions. It also points out unresolved issues including fragmented attention, scattered information, and permission design.

Why it's recommended:A Manus co-founder shared firsthand experience of the company’s fully agent-based workflow, offering viable practices, lessons from failures, and unresolved issues.

10/29 · Thursday14 items

Wiz:ResearchPapersAI Score70

Wiz analyzes seven key signatures of agent attacks and emerging trends

Drawing on its own incident response work and publicly disclosed cases, Wiz analyzes the current use of AI agents in real-world intrusions and emerging trends. It covers frameworks such as ARTEX and Hermes, as well as cases involving a South Korean financial institution (with information on about 67,000 people leaked), Thailand’s Ministry of Finance, and the ConnectWise supply chain (with attempted attacks on more than 2,250 tenants).

Why it's recommended:Combining three of its own cases with publicly reported incidents, Wiz identifies seven actionable fingerprints of agent-driven attacks, offering directly useful guidance for defensive detection.

vLLM Official BlogProductsAI Score71

vLLM releases Day 0 support for Mistral Large 4

The vLLM team announced Day 0 support for Mistral’s new flagship model, Mistral Large 4. It is a natively multimodal MoE model with 1T parameters and 49B active parameters per token. It incorporates architectural innovations including MLA sparse attention with a DSA top-k indexer, cross-layer KV sharing, NoPE, SSMax, and LeRoPE, along with a built-in DSpark speculative decoding drafter and Pixtral vision encoder.

Why it's recommended:The vLLM team details the architectural adaptations needed to support Mistral Large 4 and presents measured results for quantization, long-context processing, and speculative decoding, providing a reference for deployment.

Hugging Face Official Team BlogModelsAI Score81

Mistral Large 4 (Le Chonk) arrives on Hugging Face: an open-weight, multimodal MoE model with 1T parameters

Hugging Face published a launch post for Mistral Large 4 (Le Chonk), a multilingual, multimodal MoE model with around 1T total parameters, 49B active parameters, and a 1M-token context window.

Why it's recommended:The source provides architectural details, cybersecurity capability scores, and measured throughput across different hardware, allowing readers to assess the deployment requirements and use cases for this open-weight large model.

OpenMOSS / Fudan NLP / Shanghai Innovation Institute / MOSIModelsAI Score71

OpenMOSS releases open-source image generation model MOSS-Image, focused on long-text rendering and editable layouts

OpenMOSS has released MOSS-Image, an open-weight image generation model pretrained from scratch as a 5B-parameter single-stream DiT. It supports prompts of up to 4K tokens, long-text rendering in Chinese and English, and editable JSON Layout.

Why it's recommended:The original article provides evaluation figures for long-text rendering and layouts, along with an open-weight deployment path, helping readers assess the model’s suitability and limitations for text-heavy image tasks.

OpenRouter:AnnouncementsProductsAI Score72

OpenRouter rebuilds its default routing algorithm: TTFT down 16.9%, 429 errors down 67%

OpenRouter announced a rebuilt default provider routing algorithm. Instead of sorting by price alone, it uses Routing Cost to select providers based on effective price, TTFT, throughput, error rate, and output quality.

Why it's recommended:OpenRouter’s published routing benchmarks and mechanism details help readers assess the real impact of switching to default routing on their existing call costs and latency.

OpenRouter:AnnouncementsTutorialsAI Score76

OpenRouter tutorial: Extracting tables from PDFs with LLMs

OpenRouter published a tutorial on sending PDFs as files, together with a strict JSON schema, to models that can read files, so tables are returned as typed rows. The field descriptions specify cleanup rules for items such as parenthesized losses, thousands separators, and footnote markers.

Why it's recommended:OpenRouter tested multiple extraction routes on five PDFs, providing examples of strict JSON schema syntax, checks for two types of silent failure, and a comparison of route costs.

OpenAI: Website updatesOpinionsAI Score65

OpenAI employees explore three ways AI could enter corporate org charts

OpenAI employees Hemanth Asirvatham and Elliott Mokski published the next article in their economic series, exploring three scenarios for AI entering corporate org charts: an AI workforce basement under every employee, AI colleagues woven into existing org charts, and AI positioned between people as a coordination layer.

Why it's recommended:Drawing on their firsthand experience, two young OpenAI managers outline three scenarios for AI entering org charts and discuss the effects on hierarchy, company size, and pay gaps.

Official Pydantic AI articleTutorialsAI Score72

Pydantic AI benchmark results: Code mode cuts costs by 58% and runtime by 51% on long tasks

Pydantic AI's Michael Pfaffenberger published a benchmark analysis comparing code mode on and off across 5 small CRUD apps, 8 models, and 5,276 scored runs.

Why it's recommended:Using 5,276 real-world runs, the author delineates where code mode delivers actual benefits: it halves costs on long tasks, while making short tasks more expensive. The findings can be applied directly to your own agent design.

Cartesia product and researchPapersAI Score60

Cartesia publishes H-Net scaling-law study: byte-level hierarchical networks deliver up to 3-5x compute efficiency over BPE

In the first post of a new blog series, Cartesia reports scaling-law results for its hierarchical network (H-Net) on byte-level language modeling: compared with a strictly matched BPE Transformer, 1-stage H-Net requires 1.28x less compute for validation loss and achieves about 2x on downstream evaluations, with a median 2.04x on code tasks and a median 3.10x on out-of-distribution programming languages (such as 5.37x on Prolog).

Why it's recommended:Cartesia's original research provides controlled comparisons of H-Net as an alternative to BPE tokenization, along with 3-5x compute-efficiency data, offering a look at practical progress in byte-level learning.

Google DeepMind:BlogModelsAI Score68

Google DeepMind releases Gemini 4 Carbon, a frontier model

Google DeepMind releases Gemini 4 Carbon, positioning it as its most capable frontier model, with a focus on cybersecurity and long-horizon agent capabilities. It leads on SEC-bench Pro at 88.2%, ExploitGym at 54.8%, and CWE-bench at 71.0%, and tops the Artificial Analysis Intelligence Index v5.0 with a score of 59.

Why it's recommended:The official announcement reports Gemini 4 Carbon’s scores on secure coding and long-horizon tasks, along with its pricing, offering a basis for gauging its practical positioning in cybersecurity defense.

Also reported by 1 other source
Microsoft:Official BlogProductsAI Score64

Microsoft opens the Copilot harness through Foundry for multi-model agent development

Microsoft announces that it is making the Copilot harness built by the GitHub team available to developers and enterprises through Microsoft Foundry. The harness already powers products including Copilot Code, Copilot Studio, the GitHub Copilot app, and CLI.

Why it's recommended:Microsoft’s official explanation details the Copilot harness’s capabilities and how to access it, helping developers assess whether to use it to host multi-model agent runtimes.

a16z:NewsOpinionsAI Score60

a16z Survey of 100 CISOs: How AI Is Reshaping the 2027 Security Agenda

a16z published a survey of 100 CISOs at companies with more than 1,000 employees. 87% expect security spending to increase in 2027, driven mainly by AI threats and the expanded attack surface resulting from their own AI deployments. The seven most closely watched security initiatives for 2027 are all AI-related. SOC automation received the most top votes, while internal AI agent security, SOC automation, and offensive security saw the largest increases in priority. However, only 18% are using AI SOC tools in production.

Why it's recommended:Based on a survey of 100 CISOs, a16z outlines shifts in 2027 security budgets and purchasing priorities, and identifies gaps where startups can enter the market.

OpenRouter:AnnouncementsTutorialsAI Score70

How to Keep Coding with OpenRouter After Hitting Claude Code Usage Limits

OpenRouter explains how to switch to an OpenRouter API key and keep coding after Claude Code hits its usage limits: set three environment variables or use the ori claude command in Ori Harness to pay per token for the same Claude models, without five-hour or weekly limits.

Why it's recommended:OpenRouter tested the actual cost of using Claude Code with API key billing and cautioned that the prices published by Claude Code itself may be inflated.

GitHub BlogProductsAI Score71

GitHub Universe 2026 Announces Agentic Products, Including the GitHub Copilot App and Copilot Fusion

At GitHub Universe 2026, GitHub outlined its vision of turning GitHub into a complete software delivery system and announced several products, including the GitHub Copilot app for agentic delivery.

Why it's recommended:The original article is GitHub's official overview of new products at Universe 2026, with each feature's purpose and key figures to help readers assess its impact on existing workflows.

10/28 · Wednesday11 items

The official goose blogPapersAI Score64

goose evaluates five open-source coding harnesses: similar pass rates, with per-task costs differing by up to 75%

The official goose team used the same model, gpt-6.1-sol (high reasoning effort), and 62 tasks from Terminal-Bench 4.0 to conduct 310 scoring trials each for five open-source harnesses: goose, Codex CLI, OpenCode, Hermes Agent, and pi.

Why it's recommended:Switching harnesses with the same model produces little difference in pass rates, but costs differ by as much as 75%. This comparison data from 310 trials can help readers shift their selection focus from models to harnesses.

Anthropic:NewsroomIndustryAI Score67

Anthropic expands partnership with GSA; government agencies get 40% off Claude

Anthropic expanded its partnership with the U.S. General Services Administration (GSA). Government partners can get 40% off Claude Enterprise, Claude for Government, and Claude API. The deal is expected to save the government $1 billion and requires no minimum spending commitment.

Why it's recommended:Anthropic shared the discount rate, coverage, and government adoption data, offering readers a view into how its public-sector business is progressing.

Load more

Finished reading the news? Let’s talk about your business.

Leave your contact details, and I’ll get in touch with you personally.

WECHAT

Add me on WeChat

Press and hold or select the WeChat ID below to copy it.