AI Briefing, March 21 — Issue #132
Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.
## 🔍 Key Insights
**Kimi K2.5** has become the core base model for Cursor Composer 2; its significant advantage in **perplexity metrics** directly drove the product’s technical selection. Meanwhile, **open-source base models**—especially those from China’s open-source ecosystem—are explicitly recognized as a pivotal variable reshaping the global AI stack [4][5][9][12][15]. NVIDIA is advancing a dual-track revolution in hardware and model efficiency via its new **SOL-ExecBench** benchmark and the **Nemotron-Cascade-2** model [6][7].
## 🚀 Key Updates
- **Cursor confirms Kimi K2.5 as Composer 2’s base model** [4]: Co-founder Aman Sanger explicitly stated it achieved the strongest performance in perplexity evaluation and characterized the incident as a “communication mishap”—not a licensing dispute.
- **Moonshot AI officially confirms its partnership with Cursor** [5]: The company posted a congratulatory tweet on Composer 2’s launch, confirming Kimi K2.5 powers the foundational model capabilities and stating commercial licensing was obtained via Fireworks.
- **NVIDIA launches SOL-ExecBench, a “light-speed” benchmark** [6]: Converts GPU performance into a “light-speed score” to quantify hardware headroom and inference throughput potential.
- **Nemotron-Cascade-2 lands on Ollama** [7]: Enables out-of-the-box deployment, delivering inference and agent-task performance comparable to models up to 20× larger in parameter count.
- **Google Gemini API adds OpenAI-compatible layer** [11]: Seamless integration with Nano Banana and Veo models requires only three lines of code changes.
- **Devin introduces self-scheduling capability** [16]: Cognition releases automated workflow functionality, enabling one-off tasks to be converted into repeatable, intelligent agent workflows.
- **LiteParse launches plug-and-play skills for coding agents** [23]: A new LlamaIndex tool empowers agents like Claude Code to parse PDF/HTML documents locally and in real time—enhancing RAG and code-generation reliability.
- **Perplexity Computer integrates premium market research data sources** [14]: Officially incorporates PitchBook, Statista, and CB Insights, strengthening structured intelligence analysis for VC/PE use cases.
## 🔗 Sources
[1] Untrustworthy Monitoring: Additional Takeaways — LessWrong — https://www.bestblogs.dev/article/20cb1d28
[2] PM Methodology Needs to Evolve for the AI Era — https://www.bestblogs.dev/status/2035104384007422347
[3] Discovering Features in Transformers: Contrastive Directions Elicit Stronger Low-Level Perturbation Responses at Smaller Perturbation Magnitudes Than Baselines — LessWrong — https://www.bestblogs.dev/article/0ad52483
[4] Cursor Co-Founder Explains Technical Selection Pathway, Characterizes Incident as a “Communication Mishap” — https://www.bestblogs.dev/status/2035098758783148061
[5] Moonshot AI Officially Confirms Cursor Partnership, Reversing Prior Inquiry Stance — https://www.bestblogs.dev/status/2035098528658112704
[6] NVIDIA Launches SOL-ExecBench: A “Light-Speed” GPU Performance Benchmark — https://www.bestblogs.dev/status/2035089525702369584