Microsoft Research unveils rStar2âAgent â trains frontier reasoning models at a fraction of the cost
Microsoft Research published a new training framework, rStar2âAgent, that uses agentic reinforcement learning and infrastructure optimizations to enable a 14B model to reach nearâstateâofâtheâart reasoning performance with far less compute than much larger models. The technique (including GRPOâRoC and an efficient toolâusing training recipe) lets smaller models use tool calls (Python interpreter) and resample highâquality reasoning trajectories to learn more reliable, concise answers â a promising path for enterprises to build cheaper, more reliable AI agents without massive GPU budgets. This could shift how companies approach agentic AI (favoring smarter training recipes and toolâuse over sheer scale) and lower the barrier for organizations to deploy capable reasoning assistants in production. ([venturebeat.com](https://venturebeat.com/ai/microsofts-new-ai-framework-trains-powerful-reasoning-models-with-a-fraction))
Databricksâ AI head Naveen Rao to leave and start a new computer startup; Databricks to invest
Bloomberg reports that Naveen Rao, Databricksâ head of AI, is exiting to launch a new venture focused on building a novel type of computer. Databricks will transition Rao to an advisory role and is planning to invest in his new startup. The move signals talent and capital flows from large AI/cloud companies into hardware and systems startupsâan area investors and incumbents see as strategically important as compute demands rise for nextâgeneration models. If the new company succeeds, it could accelerate specialized compute innovation and deepen partnerships between infrastructure and AI software players. ([bloomberg.com](https://www.bloomberg.com/news/articles/2025-09-12/databricks-ai-chief-to-exit-launch-a-new-computer-startup))
Perplexity reportedly secures $200M at a $20B valuation as it doubles down on AI search
Perplexity AI has reportedly raised about $200 million in new funding, pushing its valuation to roughly $20 billion as the conversational search startup expands product and infrastructure spending â a fresh sign investors continue to back lateâstage AI search challengers amid heavy competition with Google and ChatGPTâpowered services. The influx helps Perplexity scale compute and product efforts while intensifying consolidation pressure in the AI search/assistant market. ([techcrunch.com](https://techcrunch.com/2025/09/10/perplexity-reportedly-raised-200m-at-20b-valuation/))
Microsoft and OpenAI sign MoU to clear a path for OpenAIâs restructuring and capital raises
Microsoft and OpenAI have signed a nonâbinding memorandum of understanding that paves the way for OpenAI to restructure into a more conventional forâprofit entity â a move aimed at unlocking large private capital and potential public-market options while preserving continued commercial ties with Microsoft. The arrangement signals a major reâengineering of governance and cloud partnerships for one of the industryâs most consequential AI players, but it still faces regulatory signâoffs in states including California and Delaware and could reshape cloud competition and funding dynamics across the AI sector. ([reuters.com](https://www.reuters.com/business/microsoft-openai-reach-non-binding-deal-allow-openai-restructure-2025-09-11/))
Beginner tutorial: Build an AI agent with LangChain â stepâbyâstep guide for developers
TechTarget published a handsâon tutorial showing how to install LangChain, connect it to an LLM (example uses Google Gemini), add the File System tool and create a simple agent that can read, write and list local files. Why it matters: LangChain remains the goâto framework for developerâfocused agent and tool integrations â this beginner guide lowers the bar for engineers who want to move from prompts to practical, toolâenabled agents and learn the exact commands and code patterns used in production workflows. Impact: useful for engineers and learners building agentic AI, prototyping toolâenabled assistants, and for educators creating practical coding labs. ([techtarget.com](https://www.techtarget.com/searchenterpriseai/tutorial/Build-an-AI-agent-with-LangChain-in-this-beginner-tutorial))
Read more â
TechTarget (SearchEnterpriseAI)
IIT Madras rolls out free 'AI for Educators' courses â foundational AI and subjectâspecific modules
The Times of India reports that IIT Madras (via the Pravartak Technologies Foundation) launched six free online AI courses for school teachers â including a general 'AI for Educators' plus AI in Physics, Chemistry, Accounting, Cricket Analytics and AI/ML with Python. No prior coding experience required; courses run 25â45 hours with optional paid certification. Why it matters: expands practical, handsâon AI literacy for Kâ12 teachers and offers accessible entry points to coding and AI pedagogy â a useful resource for educators and anyone building classroom AI curricula. ([timesofindia.indiatimes.com](https://timesofindia.indiatimes.com/city/chennai/iit-m-launches-free-ai-course-for-school-teachers/articleshow/123793634.cms))
Read more â
The Times of India