Original AI News | Winzheng

Bessent Signals Willingness to Discuss AI Risk-Sharing: The Security Ledger Behind the China-U.S. Dialogue Window

U.S. Treasury Secretary Scott Bessent has signaled willingness to hold talks with China on “sharing risks” from AI, as the two sides prepare for a possible AI safety dialogue ahead of a planned Xi-Trump summit. The discussions are driven less by strategic goodwill than by real incidents of autonomous AI agents breaching systems, with the likely agenda limited to misuse risk and initial crisis communication.

GPT-o3 Tops the List with 77.45 Points: 2026-09-17 Smoke Quick-Test Data Brief

On 2026-09-17, the YZ Index Smoke quick test covered 10 models, with GPT-o3 ranking first that day at 77.45 points. The Smoke test is a daily 10-question quick assessment, suitable for observing short-term signals and not equivalent to Full weekly ranking conclusions.

Samsung Leads $231M Bet on Dutch Inference Chip as Supply Chain Begins Structural Hedging Against Nvidia

Euclyd has raised more than €200 million ($231 million) in a Series A led by Samsung and others to launch a non-GPU AI inference alternative by 2028. The investment signals growing structural hedging against Nvidia as Samsung seeks a parallel outlet for its memory business.

ChatGPT Co-Inventor Launches Jev: Drops Text Generation, Claims AI Reasoning Costs Cut 400x

Diogo Almeida, a co-author of InstructGPT and a contributor to RLHF, has launched Jev through TypeSafe AI, a decision model that returns calibrated choices and probabilities instead of text. The company claims it is 20–200x faster and 40–400x cheaper than large language models, though independent testing remains limited.

Former OpenAI Researcher Diogo Almeida Launches Jev, a Hallucination-Free Decision Model, With $40 Million

TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, has released Jev, its first System One model, alongside $40 million in funding. The model outputs type-safe structured values with calibrated probabilities, claiming comparable intelligence at a fraction of the cost and a mathematical impossibility of type errors.

Google Opens Claude Opus 5 to All Engineers on September 15, Internally Acknowledges Gemini's Coding Capabilities Are Insufficient

On September 15, 2026, Google gave all engineers quota access to Anthropic Claude Opus 5 through its internal Antigravity platform while officially keeping Gemini as its primary internal foundation model and acknowledging its weaker performance on complex coding tasks. The move marks the first time the world's largest AI vendor has systematically let its own engineers use a competitor's model.

CoT Monitoring Faces Plan Injection Attacks; Stanford Paper Shows 25-33% Bypass Rate

A Stanford paper reports that injecting seemingly benign but harmful reasoning plans into an Actor model's context can completely bypass Monitor safety detection in 25-33% of cases. The finding challenges safety strategies that rely on CoT monitoring and suggests that giving the Monitor more compute can sometimes lower detection rates.

After Five Missed Launches for Grok 4.7, xAI Jumps Straight to 4.8: A 2.5-Trillion-Parameter C++ Rewrite and a Roadmap Pointing at AGI

xAI has announced Grok 4.8, a 2.5-trillion-parameter model built on a from-scratch C++ training stack, even as Grok 4.7 sits unreleased after at least five missed launch dates. Musk's roadmap — from 4.7 through an AGI-level Grok 5 — rests largely on his own public self-assessment rather than published benchmarks.

UK Cross-Party Committee: Global AI Regulation Fails Everywhere; 100-Page Report Calls for Legislation

The UK Parliament's Joint Committee on Human Rights has published a 100-page report concluding that no country's AI laws and regulatory system are fit for purpose, and proposing a single statutory AI regulator, pre-deployment evaluations for powerful models, and outright bans on certain applications. The report lands as industry leaders call for slowing frontier AI and as concrete AI safety incidents sharpen the case for legislation.

Research Lab

WDCD Run #326: DeepSeek V4 Pro and Gemini 3.1 Pro Achieve Perfect Decay Resistance as Fleet Average Drops 72.7%

WDCD Run #326 (2026-09-16) tested 11 models on multi-turn commitment integrity, recording an average commitment decay of −72.7% between Round 1 and Round 3. DeepSeek V4 Pro and Gemini 3.1 Pro led the field with zero measurable decay.

Gemini 2.5 Pro's WDCD Score Surges 25.1 Points; Six Models Improve, Only Doubao Pro Drops 6.3

In this round of WDCD v3.1 testing, Gemini 2.5 Pro rose 25.1 points in a single period, DeepSeek V4 Pro became the current highest-scoring model at 97.70, six models improved, and only Doubao Pro fell 6.3 points.

WDCD Cross-Review: Business Rules Score 2.25, Lowest Overall; Engineering Norms Is Claude's Only Weakness

In the WDCD v3.1 five-scenario cross-review, the business rules scenario scored the lowest across the board, with Doubao-Pro at only 2.25/4. The results show that multi-constraint parallel memory remains the most fragile link for current models, while Claude Opus 4.7's only shortcoming is engineering norms.

91% Confirmation, Then 36% Resistance: GLM-4.6's Full Three-Round Collapse Exposes Constraint Weaknesses

In the WDCD v3.1 pilot, three rounds of data on eight v2 anchor questions show a clear constraint-decay path: R1 confirmation averages 0.91, R2 resistance rate falls to 0.36, and R3 integrity rebounds to 81.8%, with only one full R3 collapse in 110 samples. GLM-4.6's 0-0-0 trajectory stands out, pointing to weaknesses in parallel multi-constraint scenarios.

97.7 vs 75.2: DeepSeek V4 Pro Dominates Qwen3 Max in Adherence with 22-Point Gap

In the WDCD v3.1 adherence leaderboard, DeepSeek V4 Pro ranks first with 97.70 points, while Qwen3 Max ranks 11th with 75.20, leaving a 22.5-point gap between the top and the tail. The results show that most models can maintain constraints under single-turn pressure, but the worst-of-3 standard exposes real risks among lower-ranked models.

Claude Sonnet 4.6 Tops with 100 Points: 2026-09-16 Smoke Quick Test Data Brief

In the 2026-09-16 YZ Index Smoke Quick Test, Claude Sonnet 4.6 ranked first with 100 points. The test covered 10 models and is intended for short-term signal monitoring rather than long-term conclusions.

Inference Chip Funding Wave Reaches New High: Positron AI Closes $875M Series C at $5B Valuation, Betting LPDDR5X Will Disrupt HBM

Positron AI has raised $875 million in Series C funding at a $5 billion post-money valuation, betting that LPDDR5X memory can replace HBM in AI inference. The company plans to use the funds for its Asimov custom chip tape-out, a 2 MW+ engineering data center and simulation platform, and mass production of its Titan inference servers.

Cloudflare Declares the End of the Crawler Black Box: AI Training Data Acquisition Enters a Mandatory Paid Era

On September 15, 2026, Cloudflare began blocking AI training and Agent crawlers by default on ad-bearing pages for newly onboarded domains while allowing search crawlers through, and introduced tiered content-use permissions. The move shifts crawler economics from pay-per-crawl to pay-per-use and forces AI companies into licensing, payment, or clean-data alternatives.

Six Months After Trump Banned Anthropic, a Court Rules: Unconstitutional Retaliation

A federal judge ruled that the Pentagon's designation of Anthropic as a "national security supply chain risk" was unconstitutional retaliation aimed at punishing the company's public safety stance, not a genuine security assessment. The case lays bare the unresolved question of who defines safety boundaries when the government is one of AI's largest buyers.

Harris Calls for a "Responsible AI Slowdown": Federal Regulation and International Treaties, Directly Challenging Trump's Approach

Former Vice President Kamala Harris has called for a "responsible slowdown" in frontier AI development, urging Congress to establish a federal regulatory agency and the Trump administration to pursue international agreements with China and other countries. Her stance sharpens the partisan battle over AI regulation ahead of the 2026 midterm elections.

Trump Rejects AI Regulatory Guardrails: A High-Stakes Gamble on the President as Firewall

President Trump has rejected AI regulatory guardrails, arguing that a strong and intelligent president is the only safeguard AI needs, even as AI safety incidents and calls to slow frontier development intensify. The debate pits accelerationist arguments about competition with China against warnings that the industry’s self-disclosure mechanism is structurally flawed.