18 Aug 2025

AI News Digest

🤖 AI-curated 8 stories

Today's Summary

Grammarly is stepping up its game with nine new AI agents, including an AI grader that predicts paper scores based on course details and instructor info. It’s pushing the envelope in education, but also stirs up fresh debates over academic integrity and the role of AI in grading. Meanwhile, Ambience Healthcare just snagged a hefty $243 million to expand its AI medical scribe offerings, upping the ante in a hotly contested field as it races against rivals like Abridge and big players like Epic.

Over in the AI startup world, Natasha Malpani is making waves with Boundless Ventures, a Rs 200 crore fund aimed at boosting Indian AI startups across various sectors. It’s a welcome push for local AI innovation, offering founders a homegrown

Stories

Grammarly launches nine AI agents — including an ‘AI grader’ that predicts your paper’s score

Grammarly rolled out nine specialized AI agents today inside its new Docs writing surface, aimed at students and educators. Highlights include an AI grader that estimates a paper’s likely grade using uploaded course details and publicly available instructor information, a citation finder, a proofreader and paraphrase agent, plus plagiarism and AI‑detector agents for educators (the latter two are Pro‑only at launch). The move packages generative tools directly into a mainstream writing product, speeding adoption in education while raising fresh questions about academic integrity, detection accuracy, and instructor reliance on automated grading. This could reshape how schools handle homework, plagiarism checks, and formative feedback — and intensifies debates about AI’s role in learning and assessment. ([theverge.com](https://www.theverge.com/news/760508/grammarly-ai-agents-help-students-educators))

Ambience Healthcare raises $243M Series C as AI medical scribe race heats up

Ambience Healthcare secured a $243 million Series C round (co‑led by Oak HC/FT and Andreessen Horowitz) as it expands beyond ambient transcription into automated medical coding, referrals, pre‑visit summaries and billing workflows. The San Francisco startup — already deployed at more than 40 health systems — says the funding will accelerate product development toward broader automation of clinicians’ administrative work. The round underscores intense investor interest in AI medical‑scribe startups and sets up a showdown with rivals (including Abridge) and incumbent EHR vendors like Epic, which is reported to be launching its own AI scribe. The funding and competition signal rapid consolidation and rising stakes around clinical data, compliance, and safety as healthcare systems adopt generative AI. ([businessinsider.com](https://www.businessinsider.com/ambience-healthcare-used-this-pitch-deck-to-raise-243-million-2025-8))

CABENCH: A new benchmark to evaluate ‘composable AI’ systems that stitch models together

Researchers released CABENCH (arXiv Aug 4, 2025), the first public benchmark aimed at evaluating composable AI — systems that solve complex tasks by composing many ready-made models (across modalities and domains) into execution pipelines. CABENCH includes 70 realistic composable tasks, a curated pool of ~700 models, and an end-to-end evaluation framework with human-designed reference solutions and LLM-based baselines. Why it matters: as industry moves away from monolithic foundation models toward modular pipelines (agents, tool use, and model orchestration), CABENCH provides a standardized way for academics and practitioners to measure progress, identify failure modes, and compare approaches for automatic pipeline synthesis and execution. Impact: the benchmark should accelerate research on automated model composition, robustness and evaluation methods for multi-model systems — a timely contribution for researchers working on agents, multimodal workflows, and production deployment of chained-model solutions.

AINL‑Eval 2025: shared task and dataset to detect AI‑generated scientific abstracts in Russian

AINL‑Eval 2025 (arXiv Aug 13, 2025) introduces a large-scale shared task and dataset for detecting AI‑generated scientific abstracts in Russian. The dataset contains 52,305 samples spanning 12 scientific domains and synthetic abstracts produced by five state-of-the-art LLMs (including GPT‑4‑Turbo, Gemma2‑27B, Llama3.3‑70B). The competition targeted generalization — both to unseen domains and to model outputs not present in training — and attracted multiple teams with public baselines and code. Why it matters: as LLMs are increasingly used (and misused) in academic writing, robust multilingual detection tools and benchmarks are essential for preserving research integrity and for designing fair editorial workflows. Impact: the shared task and public dataset give the community concrete, reproducible tools to evaluate detectors, study cross-model generalization, and improve methods for automated integrity checks in scientific publishing.

Anthropic tightens rules for investors as it raises roughly $5B — starts rejecting SPVs

Anthropic is raising an estimated $5 billion at about a $170 billion valuation and has told prospective investors it will not accept capital through special purpose vehicles (SPVs) for this round, according to reporting today. The move reflects strong demand for Anthropic shares and a desire to secure long‑term, top‑tier institutional backers rather than pooled third‑party vehicles. Why it matters: by restricting SPVs Anthropic preserves cleaner investor relationships and greater control over its cap table during a blockbuster round — a sign of leverage for leading AI labs and a signpost for how large private AI fundraises are being structured in 2025. Industry impact: firms that rely on SPV-driven placement models could be squeezed out of the biggest AI deals, and other startups may adopt similar investor-selection standards as valuations and competition intensify.

Ex‑Kae Capital partner launches Rs 200 crore Boundless Ventures to back Indian AI startups

Natasha Malpani, formerly of Kae Capital, has launched Boundless Ventures, a Rs 200 crore (≈$24–25M) early‑stage fund focused on AI‑native startups in India, the Economic Times reports today. The fund will write pre‑seed and seed checks across AI infrastructure, vertical applications (healthcare, logistics), consumer AI and make‑in‑India hardware, and has already invested in six companies. Why it matters: the move signals continued VC momentum for AI in India and adds a new localized source of capital and networks for founders, helping domestic AI ventures scale without exclusively relying on overseas investors. Industry impact: more dedicated local AI funds can accelerate verticalization of AI products for Indian markets and increase competition for talent and deal flow in the region.

Read more → The Economic Times (ETtech)

Microsoft rolls GPT‑5 into Copilot and adds a ‘Smart Mode’ to pick the best model for the job

Microsoft has integrated OpenAI’s GPT‑5 across its Copilot products (Microsoft 365 Copilot, GitHub Copilot, Azure AI Foundry) and introduced a new “Smart Mode” that automatically routes tasks to the most appropriate model variant (fast vs. deep‑reasoning). For developers and knowledge workers this means substantially improved code generation, longer‑context conversations, and fewer manual model choices — a direct upgrade to coding assistants, in‑app AI workflows, and developer tooling. The change accelerates ‘vibe‑coding’ and AI‑assisted productivity, while making GPT‑5 available through Microsoft’s Copilot ecosystem and Azure for app builders.

GitHub CEO Thomas Dohmke steps down as Microsoft folds GitHub into its CoreAI group

GitHub’s CEO Thomas Dohmke announced his resignation and Microsoft is reorganizing GitHub to report into its CoreAI 'Platform & Tools' division. The move signals tighter integration between GitHub (and Copilot) and Microsoft’s developer AI stack — potentially speeding feature delivery for AI coding tools and aligning GitHub more closely with Azure and Microsoft’s agent/AI infrastructure. For developers this could mean faster Copilot feature development and deeper platform integration, but it also raises questions about GitHub’s independence and future product direction.