Chinese lab DeepSeek has restarted its second fundraising round, seeking close to $8 billion at a valuation near 500 billion yuan, or about $74 billion, Bloomberg reported via people familiar with the talks. Monolith Management, an early Moonshot backer, is negotiating to join. The round was paused in July after investor friction, and capital is earmarked partly for data centers, including a large Inner Mongolia build. Why it matters: the lab that sold cheap inference is now raising for its own compute scale. Caveat: terms are not closed, and neither DeepSeek nor Monolith confirmed the talks publicly.
Athens based Omilia closed a $67 million Series B led by Expedition Growth Capital, its first raise since a $20 million Series A in 2020. The voice first agentic customer service vendor says annual recurring revenue grew more than tenfold to over $60 million, with clients including Capital One, Discover, RBC, and Taco Bell. Proceeds fund a U.S. office later in 2026 and a larger go to market team. Why it matters: enterprise voice AI that compounds on unit economics is still raising without chasing pure generative hype. Caveat: closing remains subject to customary conditions, and high volume food ordering pilots still carry operational risk.
A Nature Medicine study tested explainable AI for dermatology with 623 lay people and 153 primary care physicians. Fairness trained models that balanced performance across skin tones improved accuracy and cut tone related gaps for both groups. Multimodal LLM explanations amplified automation bias among lay users, helping when the model was right and hurting when it erred, while physicians stayed more resilient. Showing the AI diagnosis first strengthened anchoring. Why it matters: chatty explanations can make consumer dermatology AI feel safer than it is. Caveat: the tasks differed by cohort, so public and clinician effects are not a single head to head score.
In a Lok Sabha reply, electronics and IT minister of state Jitin Prasada said India selected 20 indigenous foundation model proposals from 506 IndiaAI Mission applications: 12 large multimodal models and eight small language models, with IP staying with developers. Launched examples include Sarvam, Gnani, BharatGen, and Avataar. The Centre also approved 58 AI centres of excellence, started work on 22 across 13 states and UTs, and sanctioned about 93.18 lakh GPU hours for 237 projects. Why it matters: New Delhi is treating sovereign models and state hubs as infrastructure. Caveat: a progress report is not the same as shipping competitive frontier weights at scale.
Huawei posted weights and inference code for openPangu-2.0-Pro, a roughly 505 billion parameter mixture of experts model with about 18 billion active parameters per token and a 512,000 token context. The lab says pretraining used Ascend 910B NPUs and about 34 trillion tokens, with no Nvidia GPUs in the cluster. Checkpoints are on Hugging Face under Huawei's OpenPangu license, which restricts EU use. Why it matters: a frontier scale open weight stack on Chinese silicon is now downloadable. Caveat: Ascend parts have previously included foreign dies and memory, so Nvidia free training is not a fully domestic supply chain, and independent benchmarks remain thin.
Jeff Dean is leaving Google after 27 years to lead Discovery Loop, a Palo Alto public benefit company co-founded with Sanjay Ghemawat, Quoc Le, and Oriol Vinyals. Radical Ventures and Khosla Ventures are co-leading the seed round, with Alphabet as a founding investor and cloud partner; the amount was not disclosed. The plan is AI that proposes, runs, and iterates thousands of experiments in parallel, starting with machine learning research. Why it matters: Gemini era leaders are betting the next frontier is automating discovery itself. Caveat: there is no product yet, and recursive self-improvement claims remain speculative.
San Jose firm Invenio Imaging finished enrollment in ON-SITE, a U.S. pivotal study of NIO Lung Cancer Reveal, an AI module that flags cancer-suspicious morphology on fresh bronchoscopic biopsies imaged with its NIO system. The multicenter effort enrolled 1,006 patients across seven centers, with 32 physicians collecting more than 3,250 specimens by four biopsy techniques. The tool already holds FDA Breakthrough Device Designation; Invenio will analyze validation data for a regulatory submission. Why it matters: large prospective datasets remain rare for intraoperative AI pathology. Caveat: enrollment is not efficacy, and the product remains investigational in the United States.
Meta released Muse Code in beta, a terminal coding agent for macOS and Linux, alongside Muse Spark 1.2, a coding-focused update to its closed Spark model. Meta says the pair were co-trained for planning, debugging, repository-scale work, and persistent background subagents that keep a replayable event log. Spark 1.2 is live in Muse Code and the Meta Model API at about $1.25 per million input tokens and $4.25 per million output tokens. Why it matters: Meta is shipping an agent-plus-model stack aimed at rival coding products. Caveat: the weights stay closed, and Meta's own charts drive the headline benchmark claims.
Chinese lab DeepSeek invested 140.8 million yuan ($20.8 million) in robot maker Unitree's Shanghai STAR Market IPO and agreed to jointly develop AI models for humanoid machines, Reuters reported from a stock-exchange filing. DeepSeek received 933,399 shares, or 2.31 percent of the strategic placement, while both Hangzhou firms pledged preferential procurement for training services and robots. The aim is a physical-world brain for cheap walking machines. Why it matters: China's open-model leader is tying capital and roadmaps to embodied hardware. Caveat: an IPO stake and a procurement pact are not yet a shipped humanoid controller.
Bengaluru startup Superleap raised Rs 36 crore from Peak XV Partners Surge, cofounder Subham Kumar Boundia told The Economic Times. It sells AI-native CRM meant to help large firms migrate off legacy systems, and says average contracts run about $60,000 to $100,000 with customers such as Razorpay, Cars24, and MediBuddy. Proceeds will deepen the product, expand go-to-market in India and abroad, and hire. Why it matters: Indian buyers are funding agentic revenue software, not only chat demos. Caveat: a Surge cheque does not prove Superleap will displace Salesforce-class incumbents, and generative CRM claims still need retention proof.
MeitY used an August 6 briefing, tied to a Lok Sabha reply, to restate February IT Rules on synthetic media. Platforms must label permissible AI content with traceable metadata, and the clock to remove unlawful material after a valid government or court notice falls from 36 hours to three. Significant intermediaries must deploy technical measures against deepfakes, CSAM, and impersonation or risk losing Section 79 safe harbour. IndiaAI Mission projects at IIT Jodhpur, Madras, and Kharagpur are building detectors. Why it matters: India is measuring AI content compliance in hours. Caveat: the statute is months old; this is enforcement emphasis, not a brand new law.
Duke-led researchers report in Nature that Raygun, a generative model built on protein language embeddings, can miniaturize, expand, or diversify natural proteins by encoding each sequence as a fixed-dimension probability distribution. In cell tests, shortened fluorescent proteins and TurboID variants kept activity, and enlarged EGF designs bound EGFR more tightly than wild type. Shrinks of 10 to 25 percent were common. Why it matters: large insertions and deletions have been a blind spot for residue-swap design tools. Caveat: wet-lab wins cover a few systems, so broad therapeutic claims still need independent tests.