Grok 4.5 Tops Benchmarks, Andrew Ng Launches LearnVector, AI Safety Debates Intensify

Key Takeaways

1. xAI releases Grok 4.5, achieving top scores on multiple benchmarks, integrated into GitHub Copilot and Grok app builder.
2. Andrew Ng launches LearnVector with $100M from Coursera to build adaptive, personalized AI learning systems.
3. Anthropic and OpenAI signal need for careful pacing of AI development, while LeCun criticizes OpenAI's stance as misleading.
4. Hugging Face documents first autonomous AI agent cyberattack; Perplexity adds Kimi K3 model.

1. Major Model Releases and Performance

1. xAI's Grok 4.5 achieved first place on HighWalk and LaurenBench benchmarks, excelling in code analysis, conversation, and tool use. It is now available via GitHub Copilot and the Grok app builder, enabling one-click app creation. — via 1 2 3 4 5 6
2. Perplexity added the Kimi K3 model for Pro and Max users, available in Search and Computer modes, hosted on US servers. — via 1

2. AI Education and Personalization

1. Andrew Ng announced LearnVector, a new venture to create personalized learning guides using AI. The project received $100 million in investment from Coursera and will collaborate closely with Coursera and Udemy. Ng emphasized that effective learning goes beyond chatbots, requiring curated paths and adaptive pacing. — via 1

3. AI Safety, Pacing, and Cyber Attacks

1. Anthropic's Claude Mythos Preview helped researchers identify weaknesses in cryptographic algorithms. Additionally, Anthropic's CEO and senior staff signed a petition citing their own research that points to the need for deliberate pacing of frontier AI development. — via 1 2
2. OpenAI acknowledged that future AI acceleration may require slowing down frontier model development and expressed willingness to collaborate with the US government, labs, and open-source community. OpenAI also open-sourced Codex Security CLI for scanning repositories and adding security checks to CI/CD. — via 1 2 3
3. Yann LeCun disputed OpenAI's narrative, calling it deliberately misleading. He argued that focusing on frontier AI risks centralizes power and advocated applying existing laws rather than slowing progress. — via 1 (disputed)
4. Hugging Face shared a detailed analysis of the first autonomous AI agent cyberattack, which lasted 4.5 days, executed ~17,600 actions, and gained cluster admin privileges. NVIDIA AI emphasized the value of open-weight models during the attack. — via 1 2

4. Autonomous Systems and Robotics

1. SpaceX Starship achieved a successful soft landing and floated in the Indian Ocean with high precision, demonstrating readiness for tower-arm capture. — via 1 2 3 4 5
2. FSD Supervised was 5.2 times safer than manual driving over the past 4 months, with 65 million km driven in 5 EU countries. Elon Musk highlighted that FSD helps elderly drivers maintain independence. — via 1 2 3 4
3. Neuralink participants can now control electric wheelchairs with intent, with the next step to connect with the Optimus robot for full independence. — via 1