AI Sector Daily Digest - July 16, 2026

1. China puts AI companion services under new guardrails

• China’s rules for AI anthropomorphic interaction services took effect on July 15. They cover services that simulate human personality and provide ongoing emotional interaction, while excluding ordinary customer service, work assistance, education, and research tools. 1
• Providers must detect distress, intervene in crises, limit excessive use, protect interaction data, show a warning after two hours of continuous use, and let users copy or delete their records. 2
• Xinhua reported that Doubao, Qwen, and Yuanbao had recently announced pauses to AI agent features. The report does not say that every agent feature falls under the new rules. 2

2. DeepSeek reportedly plans a new round at a $74 billion valuation

• DeepSeek is planning a fresh fundraising round at a valuation of about 500 billion yuan, or $74 billion, Reuters reported. This is a reported plan, not a completed financing. 3
• The round is being discussed ahead of a possible mainland IPO, but Reuters says both the financing and listing plans remain at an early stage and terms or timing could change. DeepSeek did not immediately comment. 3
• Reuters said the plan would follow a June financing of about $7.4 billion at roughly 450 billion yuan post-money, making the new report an additional proposed round rather than a confirmation of the earlier one. 3

3. Emergent reaches a $1.5 billion valuation with a $130 million Series C

• Indian AI coding startup Emergent raised $130 million in Series C funding at a $1.5 billion post-money valuation, with Creaegis leading the round. The deal brings total funding to $230 million. 4
• The company sells an application-building platform aimed at entrepreneurs and small and midsize businesses that want deployment, hosting, testing, and debugging alongside code generation. 4
• Emergent told TechCrunch that it has reached a $120 million annualized revenue run rate and more than 200,000 paying customers; it plans to use the new capital on product development, agent workflows, and go-to-market expansion. 4

4. Thinking Machines releases the open-weight Inkling model

• Thinking Machines Lab released Inkling, a multimodal open-weight model with 975 billion total parameters and 41 billion active parameters. It accepts text, images, and audio, produces text, and supports up to a 1 million-token context window. 5
• The weights are available through Hugging Face under the Apache 2.0 license, while fine-tuning and API access are available through the company’s Tinker service and other inference providers. 6
• The model card says the BF16 checkpoint needs at least 2 TB of aggregated VRAM to run, or at least 600 GB for its quantized NVFP4 checkpoint. Thinking Machines also says Inkling is not the strongest overall model, positioning it as a customization base rather than a frontier leaderboard claim. 6

5. A new benchmark asks whether agent improvements compound

• A July 15 arXiv paper tests whether an agent that has been optimized once can improve again on newly introduced tasks without losing its earlier gains. The authors use a two-phase continual-learning protocol built from hard tasks in Terminal-Bench 2.0. 7
• RELAI-VCL, which puts regression control inside the optimization loop, reached a 76.4% lifelong average, compared with 66.0% for GEPA, 64.6% for Meta Harness, and 58.7% for the baseline in the paper’s evaluation. 7
• The result is limited by the setup: the benchmark tasks are only loosely related, the method assumes repeatable execution and reliable verifiers, and the study covers two optimization phases rather than a long deployment history. 7