14 Aug 2025

AI News Digest

🤖 AI-curated 8 stories

Today's Summary

DeepSeek’s hit a snag with its AI model launch due to issues with Huawei chips, casting a spotlight on China’s tricky path to cutting reliance on U.S. tech. Meanwhile, Arintra’s snagged $21M to rev up its GenAI medical-coding game, promising a smoother ride for healthcare systems. On the research front, new techniques like “step entropy” and “Refine-n-Judge” are making waves in AI efficiency and refinement, hinting at leaner, smarter AI reasoning without breaking the bank.

Stories

DeepSeek delays new AI model launch after training problems with Huawei chips

Chinese AI startup DeepSeek has postponed the rollout of its next model after failing to train it successfully on Huawei’s domestically made chips, underscoring limits in China’s effort to replace U.S. AI hardware and potentially slowing the company’s product roadmap and competitive positioning. The delay spotlights broader supply-chain and performance challenges for Chinese AI firms that are trying to move away from Western processors, and could affect timelines for other locally developed models that depend on domestic silicon. ([reuters.com](https://www.reuters.com/world/china/deepseeks-launch-new-ai-model-delayed-by-huawei-chip-issues-ft-reports-2025-08-14/))

Arintra raises $21M Series A to scale GenAI medical-coding platform

Arintra, a Texas-based startup building a GenAI-native autonomous medical coding platform, raised $21 million in a Series A led by Peak XV Partners to accelerate product development, hire, and open a Bay Area HQ. The platform automates coding inside EHRs (integrating with systems such as Epic and Athena), promising to reduce claim denials and coding costs for health systems — a notable step in commercialization of revenue-cycle AI tools in healthcare. ([economictimes.indiatimes.com](https://economictimes.indiatimes.com/tech/funding/healthcare-focused-startup-arintra-raises-21-million-in-funding-round-led-by-peak-xv-partners/articleshow/123297667.cms))

Step-Entropy Shrinks Chain-of-Thoughts — a new arXiv paper shows LLMs can prune most redundant reasoning steps

What happened: A fresh arXiv preprint (Aug 5, 2025) from researchers including Zeju Li proposes “step entropy,” a metric that measures the informational contribution of each intermediate Chain‑of‑Thought (CoT) step and uses it to prune low‑value steps. The paper reports that up to ~80% of low‑entropy steps can be removed with only minor accuracy loss on several math reasoning benchmarks, and introduces a two‑stage SFT + Group Relative Policy Optimization (GRPO) training recipe so models can learn to emit [SKIP] tokens and produce compressed CoTs at inference time. Why it matters: CoT prompting gives big reasoning gains but is expensive at inference; principled compression could dramatically cut token usage, latency and cost for reasoning models while preserving interpretability. Impact: If validated broadly, the method could make high‑quality reasoning with LLMs far cheaper to deploy (important for on‑device and cost‑sensitive services), and opens new lines of research on which internal reasoning steps are actually necessary versus redundant.

Refine‑n‑Judge: use one LLM to iteratively refine and judge responses — a scalable route to better fine‑tuning data

What happened: An arXiv preprint (Aug 3, 2025) summarized in the weekly arXiv roundup by AskAiBrain presents Refine‑n‑Judge, an automated pipeline that uses a single LLM both to refine candidate outputs and to judge whether refinements are improvements — looping until no further improvement is found. Why it matters: Preference‑based fine‑tuning (SFT+DPO/RL) depends heavily on high‑quality preference chains, and human labeling is expensive. Refine‑n‑Judge promises to produce scalable, higher‑quality preference sequences without a separate reward/judge model or massive human labeling. Impact: The authors report measurable gains (e.g., +5% on AlpacaEval variants, +19% on MT‑Bench in their experiments) and show models fine‑tuned on these auto‑curated chains are preferred by judge models; if results hold across domains, this could reduce the human cost of alignment and speed practical deployments of better-aligned open models.

Read more → AskAiBrain (weekly arXiv roundup)

Healthcare AI champion Abridge readies M&A spree after $700M run — 20% of war chest earmarked for acquisitions

Abridge — the San Francisco startup that builds AI tools to transcribe and summarize patient‑doctor conversations — told Business Insider it has raised roughly $700 million over the last 18 months, hit a $5.3 billion valuation, and is now reserving about 20% of its capital for acquisitions to buy talent, data and adjacent tech (while keeping 80% for product development). This signals consolidation pressure in healthcare AI (where incumbents and deep‑pocketed PE firms are active) and shows startups are using mega‑rounds to both scale products and pursue inorganic growth. Impact: more consolidation in AI clinical‑workflow tooling, faster product roadmaps for buyers, and heightened competition for domain data and talent. ([businessinsider.com](https://www.businessinsider.com/abridge-wants-to-buy-healthcare-ai-startups-2025-8))

CoreWeave posts strong revenue but warns on power constraints as $9B Core Scientific takeover faces shareholder pushback

CoreWeave reported Q2 revenue that beat estimates ($1.21B) and raised its 2025 revenue outlook, but a larger‑than‑expected net loss sent shares lower; the company reiterated its planned $9 billion all‑stock acquisition of data‑center operator Core Scientific — a deal intended to secure 1.3GW of power capacity — even as Core Scientific’s biggest shareholder said it would vote against the sale. Why it matters: the report highlights how AI hyperscalers are aggressively pursuing upstream infrastructure (power and data‑center ownership) to lock supply, while deal dynamics (stock volatility and shareholder opposition) can materially reshape M&A outcomes and financing risk for AI infrastructure plays. ([reuters.com](https://www.reuters.com/business/coreweave-revenue-beats-estimates-ai-boom-shares-fall-bigger-loss-2025-08-12/))

Windows 11 August 2025 update brings AI agent to Settings, Click-to-Do enhancements and Quick Machine Recovery

Microsoft’s August 2025 cumulative update (rolling out beginning Aug. 12) adds practical AI features to Windows 11: a Settings app AI agent that can find and apply system settings via natural language, expanded Click-to-Do AI actions (reading coach, Immersive Reader, draft-with-Copilot integrations and Teams actions), Windows Recall export/reset controls for EEA users, and a Quick Machine Recovery feature to help diagnose and remediate boot failures. The changes push OS-level AI into everyday workflows (especially on Copilot+ hardware) and highlight how AI is being embedded into core productivity and system-recovery tools — a meaningful step for users, IT admins and developers building integrations.

Best free AI training courses for August 2025 — quick, practical paths to learn AI and coding

Tech.co published a curated roundup of free AI and coding courses (Aug. 4, 2025) aimed at learners from beginners to more technical students. Highlights include IBM’s “AI for Everyone” on edX, Codecademy’s short OpenAI API/Python course, Stanford/Andrew Ng machine-learning specializations on Coursera, and other targeted programs for digital marketers and analysts. The guide is practical for people seeking fast, low-cost ways to gain AI literacy or build developer skills that work alongside modern AI tools.