Sanjay Sahay

SpaceXAI’s partnership with Cursor has tightened Grok’s training pipeline and product engineering. Cursor contributed model-building expertise, deployment tooling, and token-efficiency optimizations that improved code reasoning, agentic behavior, and real-world robustness. The collaboration focused on faster throughput and lower operational cost, shaping Grok into a more production-ready offering. Grok 4.5 delivers those gains: benchmark parity near Claude […]

Meta today unveiled Muse Image, its first major in-house image model from Alexandr Wang’s Superintelligence Labs, and it arrived with a splash — debuting at No. 2 on Arena’s leaderboards. The model pairs image generation with editing and agentic features via Muse Spark, bringing search and tool use directly into the creative loop. Rolling out inside Meta AI

Meta has been racing to build top-tier AI. After its Muse Spark model launched in April and got mixed reviews, the company now says a newer model called Watermelon is training and already matches OpenAI’s GPT-5.5 in some tests. Meta is also promising big coding and “agent” improvements soon. Why the change now? Meta created a SuperIntelligence Lab and

Sonnet 5 delivers clear, measurable improvements in agentic coding and multi-step reasoning versus Sonnet 4.6, adding browser and terminal capabilities and longer-context operation that make it a more practical, lower-cost tool for knowledge-work automation. However, Sonnet 5 shows weaker cybersecurity benchmark results than 4.6, which Anthropic says reflects a lack of deliberate security training. That shortfall matters for

Meta’s Brain2Qwerty is a new AI system that tries to turn brain activity into typed text by looking at signals linked to language and movement, rather than reading “thoughts” in a magical way. It is still a research system, but it shows that non-invasive brain scanning may one day help people communicate again without surgery. In Version