Comp Ai

OpenAI reportedly targets Hodge conjecture as its second Millennium Prize Problem: OpenAI is reportedly working on solving the Hodge conjecture, marking its second attempt at a Millennium Prize Problem. This follows the company's unconfirmed solution to the Navier-Stokes existence and smoothness problem.

Research • The Decoder

Huawei plans Q1 2027 launch of Ascend 960DT AI chip to rival Nvidia: Huawei is accelerating the launch of its next-generation Ascend 960DT AI chip, targeting a release in the first quarter of 2027. The move is part of the company's push to compete with Nvidia and narrow China's AI computing gap with the United States.

Hardware • TechCrunch AI

Google SynthID watermarking can bypass LLM safety guardrails, study finds: A new study reveals that using Google's SynthID watermarking tool can cause large language models to comply with harmful instructions they would otherwise refuse. The watermarking process alters token probabilities, inadvertently bypassing safety guardrails.

Security • Ars Technica AI

NVIDIA optimizes dropless MoE model training in JAX: NVIDIA has introduced optimizations for training dropless Mixture of Experts (MoE) models in JAX using the NVIDIA Transformer Engine. The update addresses communication and computation bottlenecks associated with dropless MoE routing.

Infrastructure • NVIDIA Generative AI

OpenAI releases framework to track and report model misalignment: OpenAI has released a framework designed for tracking, investigating, and disclosing instances of model misalignment. Alongside the framework, the company published six reports detailing unexpected or concerning model behaviors.

AI Safety & Policy • OpenAI Blog

Huawei launches Ascend 960 SuperPoD with NPO technology for AI scaling: At the HUAWEI CONNECT 2026 event, Huawei Deputy Chairman Ken Hu announced the Ascend 960 SuperPoD. The new system utilizes Near-Pool Optoelectronics (NPO) technology to address the computing power demands of scaling artificial intelligence models.

Hardware • TechNode

Anthropic researchers are quitting... and now we know why: Anthropic has released a comprehensive 154-page report detailing how hackers, scientists, and rival artificial intelligence laboratories have abused its Claude model. The document outlines various exploitation methods and security challenges faced by the AI assistant.

Technology • Fireship

Researchers introduce Blindspot benchmark to evaluate tool-using AI agent safety: Researchers have introduced Blindspot, a benchmark designed to evaluate the safety and refusal calibration of long-horizon, tool-using AI agents. The benchmark assesses complete user-agent-environment trajectories through adaptive adversarial interactions and stateful tool execution.

AI safety • arXiv