Original AI News | Winzheng

Terence Tao Criticizes OpenAI for Releasing 722 Mathematical Manuscripts at Once: Solving Problems Is Just a Tool, Mathematics Is Losing Fertile Ground

OpenAI has published 722 model-generated mathematical manuscripts on GitHub, prompting Fields Medalist Terence Tao to warn that AI's focus on solving problems is stripping mathematics of the byproducts that sustain new methods and questions. The controversy over OpenAI's claimed solution to Navier-Stokes and a joint statement by 25 Fields Medalists have exposed a deeper conflict between AI-driven benchmarks and the mathematical community's standards.

Pwn2Own Ireland Day One: AI Infrastructure Falls in Succession as LiteLLM Is Breached Twice in a Single Day

On day one of Pwn2Own in Cork, Ireland, security researchers launched a concentrated assault on AI infrastructure, with LiteLLM broken twice by two independent teams, OpenAI Codex felled by a single vulnerability, and Oracle's autonomous AI database also failing to hold.

The Common Method for Measuring AI Bias Is Logically Self-Contradictory—New Paper Proposes a Dual-Axis Minimal Pair Framework

A new arXiv paper from researchers at the University of Edinburgh, the University of Amsterdam, and other institutions argues that the most widely used methods for measuring stereotypes in LLMs have a systematic logical flaw, and proposes a dual minimal pair framework and a mutual-information-based metric to address it.

Anthropic Loosens AI Safety Access With Three-Tier System, Giving Security Experts Early Access to Opus 5.5

Anthropic has expanded its Cyber Verification Program into a three-tier access framework, opening early access to Claude Opus 5.5 and other unreleased models to qualified security professionals. The program has surfaced more than 129,000 vulnerabilities over the past six months, 33,000 of them rated critical or high severity.

Claude Haiku 5.5 Officially Released: Beyond an Average 75% Price Cut, a 100K-Token Price Cliff

Anthropic released Claude Haiku 5.5 with steep short-context price cuts, but requests over 100K tokens trigger a 5x price jump, and a more aggressive tokenizer further complicates real savings for RAG and long-context workloads.

GPT-6 Rolls Out Globally: The Interface Is the App, and OpenAI Changed More Than the Model

OpenAI has begun rolling out GPT-6 globally, introducing an Intelligent UI that lets ChatGPT generate interactive layouts, tools, and mini-apps inside the chat window. The launch positions OpenAI as a larger-scale chaser in the interactive AI interface race, while raising unresolved questions about compute costs and platform boundaries.

GPT-6 Sol First Test: YZ Index 18-Question Targeted Evaluation

OpenAI has begun rolling out GPT-6 with Intelligent UI, making GPT-6 Sol the default model for paid tiers. The YZ Index's 18-question targeted evaluation gives GPT-6 Sol a composite score of 84, though this is not yet a full weekly baseline.

Claude Sonnet 4.6 Leads with 87.41: 2026-10-08 Smoke Quick Test Data Brief

Claude Sonnet 4.6 tops the 2026-10-08 YZ Index Smoke quick test with 87.41 points. This daily 10-question test covers Code Execution and Material Constraints, offering short-term monitoring signals rather than long-term conclusions.

Claude Comes to Google Workspace's Three Core Apps in Public Beta, Challenging Microsoft Copilot on Gemini's Home Turf

Anthropic has launched Claude for Google Workspace in public beta for all paid plan users, embedding it in Docs, Sheets, and Slides sidebars and connectors. It competes directly with Google's native Gemini and Microsoft 365 Copilot, with review-before-edit controls and enterprise-focused limits.

Google Signs Largest AI Energy Contract in History: The Grid Sovereignty Battle Behind a 3,590 MW Nuclear Bet

Google and Constellation Energy have completed a 3,590 MW long-term power agreement covering PJM, the largest U.S. grid, triggering more than $4.3 billion in new investment. The deal combines 890 MW of nuclear power from uprates at 11 existing reactors with 2,700 MW of long-term supply from Constellation’s existing PJM fleet, reflecting an era in which AI data centers must increasingly bring their own power.

DeepSeek's Funding Target Doubles to $15 Billion: The Chip War and Pre-IPO Outpost Behind State Capital's Scramble for Stakes

DeepSeek is considering expanding its current funding round from an initial RMB 50 billion (about US$7.4 billion) to as much as RMB 100 billion (about US$14.9 billion), driven by heavy subscriptions from state funds, listed-company investment arms, and venture capital. The round is taking shape as a pre-IPO move, as DeepSeek builds a large Huawei Ascend cluster in Inner Mongolia and prepares for a STAR Market listing in 2027.

OpenAI Publishes 722 AI Mathematics Manuscripts: The Transparency Game Behind the Largest Public Benchmark

OpenAI has released 722 AI-generated mathematics manuscripts grouped into 372 problem families, with Lean formalizations for some results. The release tests academic transparency and governance as mathematicians assess the claims and the formal verification gaps.

Meta and Sierra Launch AI Agent Identity Protocol; Walmart and Stripe Join as Founding Partners

Meta and Sierra, together with six companies including Walmart, Stripe, Shopify, and Rocket, released the Personal Agent Protocol (PAP), an industry standard for AI agent identity and trust. The protocol uses OAuth-based tiered identity authentication and aims to reduce friction between businesses and AI agents.

OpenAI Decisions API Opens Public Beta: Powered by GPT-6 Luna, Up to 10x Faster Decisions, Input-Only Pricing

OpenAI has opened its Decisions API to all developers in public beta, powered by GPT-6 Luna and returning typed answers through a dedicated endpoint. It delivers decisions up to 10 times faster than comparable tasks through the Responses API and charges only for input tokens.

EmbeddingGemma 2 Released: 740M Parameters Unify Five Modalities, Open-Source On-Device Embedding Models Enter a New Stage

Google DeepMind has released EmbeddingGemma 2, its first native multimodal open-source embedding model for on-device use, mapping text, code, images, video frames, and audio into a single 768-dimensional vector space. The 740M-parameter model is available under Apache 2.0 on Hugging Face and Kaggle.

Research Lab

WDCD Run #365: Average Instruction Decay Hits -53.3%, GLM-4.6 Leads 15-Model Field

WDCD Run #365 (2026-10-07) measured multi-turn commitment across 15 AI models and recorded an average instruction decay of -53.3% from Round 1 to Round 3, with GLM-4.6 taking the top score at 95.2 points.

Qwen3 Max WDCD Plunges 21.9 Points; 6 Models Decline, Only Doubao Pro Rises 9.5

In WDCD v3.1 testing, Qwen3 Max fell 21.9 points from Run #360, with six models declining overall; only Doubao Pro rose 9.5 points. The drops were concentrated in middle-to-late pressure rounds, while GLM-4.6 and Grok 4 remained in the 95-point range.

WDCD Cross-Scenario Review: Safety Compliance Falls to 1.1, 15 Models Show Imbalance Gaps of Up to 2.9 Points

WDCD v3.1's five-scenario evaluation finds safety compliance to be the weakest area among 15 models, with qwen3-max scoring just 1.1/4, and ten models showing clear scenario-specific imbalances exceeding one point.

15 Models, Only One Full R3 Collapse — Yet Qwen3 Max Fails 10% of the Time: The Real Story Behind WDCD v2 Anchor Three-Round Decay

Across 150 worst-of-3 samples on 8 v2 anchor questions, 15 models show a clear three-round decay — 100% average R1 confirmation, 60% average R2 resistance, 76.7% average R3 integrity — with just one full R3 collapse, but Qwen3 Max accounts for a 10% collapse rate on the data-boundary question.

WDCD Compliance Leaderboard: GLM-4.6 Leads with 95.2 as Qwen3 Max Trails at 65.6

In WDCD v3.1 compliance testing, GLM-4.6 scored 95.20 to lead 15 evaluated models, while Qwen3 Max scored 65.60 to rank last, a gap of 29.6 points. The results show a tightly clustered top tier and a sharply separated lower tail.