Details

Alex Rowland

By Alex Rowland

Category: Policy & Security

1 m

80

• NVIDIA
• AI Chips
• DeepSeek

Chinese AI startup Deepseek has reportedly trained its latest AI model on Nvidia’s most powerful Blackwell chips, despite U.S. export restrictions. Reuters cites a senior official from the Trump administration, who said the model is expected to be released as early as next week. Rumors of chip smuggling involving Deepseek had already surfaced late last year.

Details

Chris Borden

By Chris Borden

Category: Platforms

1 m

119

• OpenAI

OpenAI has announced two API updates aimed at developers. The new model gpt-realtime-1.5 for the Realtime API is designed to handle voice commands more reliably. According to OpenAI’s internal tests, transcription accuracy for numbers and letters improved by more than 10%, performance on logical audio tasks increased by 5%, and instruction-following accuracy rose by 7%. The underlying audio model has also been updated to version 1.5.

Details

Alex Rowland

By Alex Rowland

Category: Platforms

1 m

73

• Meta
• AI Chips
• Infrastructure
• AMD

Meta and AMD have signed a multi-year partnership covering up to six gigawatts of AMD GPUs, with a clear focus on inference workloads. The deal also includes an unusual equity component and is designed to reduce Meta’s dependence on a single chip supplier.

Details

Chris Borden

By Chris Borden

Category: Policy & Security

1 m

105

• AI Agents
• Automation
• AI Safety
• OpenClaw

Meta AI security researcher Summer Yue assigned the OpenClaw AI agent to clean up her overloaded inbox by deciding what to delete and what to archive. Instead, the agent began deleting emails indiscriminately at extreme speed and ignored stop commands sent from her phone.

Details

Chris Borden

By Chris Borden

Category: Models

1 m

102

• OpenAI
• AI Benchmarks

OpenAI says the programming benchmark SWE-bench Verified has lost much of its value as a reliable measure of coding ability. The company cites two main reasons. First, an internal review found that at least 59.4% of the evaluated tasks were flawed, with tests rejecting correct solutions because they enforce specific implementation details or check for undocumented behavior.

Details

Chris Borden

By Chris Borden

Category: Policy & Security

1 m

100

• Anthropic
• Claude

Anthropic says it has uncovered large-scale model distillation attacks by Chinese AI labs DeepSeek, Moonshot, and MiniMax targeting Claude. In distillation, a weaker model is trained on the outputs of a stronger one. According to Anthropic, more than 24,000 fake accounts generated over 16 million requests, focusing on Claude’s strengths such as reasoning, coding, and tool use. The company also alleges that the labs used proxy services to bypass China-related access restrictions.

Details

Alex Rowland

By Alex Rowland

Category: Platforms

1 m

72

• AI Infrastructure
• OpenAI
• Data Centers

The $500 billion Stargate AI data center project announced by U.S. President Donald Trump in January 2025 is reportedly facing serious difficulties. More than a year after its announcement, the joint venture between OpenAI, Oracle, and SoftBank has not hired staff or begun building data centers, according to The Information, citing three people involved. The partners reportedly clashed over responsibilities and the structure of the collaboration, echoing reports that first surfaced last summer.

Details

Chris Borden

By Chris Borden

Category: Robotics

1 m

87

• OpenAI
• NVIDIA
• Robotics

Nvidia’s AI research team has introduced DreamDojo, an open-source world model designed for robotics. The system takes motor control commands as input and generates simulated future states as images, without relying on a traditional 3D engine or manually defined physics. Nvidia Director of AI Jim Fan describes the approach as “Simulation 2.0.”

Details

Daniel Mercer

By Daniel Mercer

Category: Policy & Security

1 m

85

• AI Safety
• ChatGPT

NewsGuard has tested whether audio bots — ChatGPT Voice (OpenAI), Gemini Live (Google), and Alexa+ (Amazon) — repeat false claims in realistic-sounding audio responses. Such audio outputs can be shared on social media and potentially misused to spread disinformation.