Among AI news from November 5 (up to early evening Japan time), the big story out of Japan was Preferred Networks releasing its translation-specialized LLM 'PLaMo 3 Translate' as open weights. In the US, the midterms confirmed that Democrats will control both chambers of Congress, spreading wariness through the AI infrastructure market. In the open-weights camp, a 65B model developed by a five-person team posted results beating much larger models — a day when 'efficiency by design' took center stage.

PFN Releases Translation-Specialized LLM 'PLaMo 3 Translate' as Open Weights — Four Sizes from 0.8B to 31B, with On-Device Support

Preferred Networks (PFN) released 'PLaMo 3 Translate', an LLM specialized for translation, as open weights on November 5. According to ITmedia, the company touts 'world-class' translation performance and offers four sizes ranging from 0.8B to 31B, with a size lineup that also supports on-device operation.

Small, translation-specialized models are easier to run at low cost and low latency than calling general-purpose giant models via API, making them easy to fold into internal document workflows and continuous localization work. Translation quality between Japanese and English has been something of a bottleneck for Japanese companies putting AI to work in their operations, so the significance of a 'world-class' model arriving in open-weight form — verifiable and usable by anyone — looks considerable. Real-world assessment now awaits benchmarks from various organizations and validation in actual practice.

US Midterms Turn Into a Democratic 'Sweep' of Both Chambers — Data Center Names Sold Off, CoreWeave Bonds Plunge

In our previous roundup we reported that 'exit polls ranked AI last among voter priorities'; since then, vote counting has strengthened the outlook that Democrats will control both chambers, moving the AI infrastructure situation a step further.

According to Bloomberg, candidates taking a cautious stance on data center construction won races such as gubernatorial elections in Texas and elsewhere. Voter backlash over soaring electricity prices and job concerns was reflected directly in state-level politics. On Wall Street, risks to the AI boom were flagged and data center–related stocks were sold off.

The most symbolic case is AI cloud provider CoreWeave, whose corporate bonds plunged. Its unsecured bonds due 2031 briefly fell to 79.5 cents on the dollar, and bonds maturing in 2030 and 2032 carry yields above 15% — levels generally considered 'distressed' (with repayment concerns). There are also reports that the massive investments Elon Musk has poured into Texas are increasingly becoming a risk factor under Democratic dominance.

The change this time is that financing for AI capex has entered a stage where political risk gets priced in. If securing power and local consent become bottlenecks, the very design of data center siting, contracts, and financing could be forced into rethinking.

Startup Ricursive's 'Ripple-65B Preview' — A Five-Person Team's 65B Dense Beats 500B+ Models on TerminalBench

In the open-weights camp, startup Ricursive Intelligence released 'Ripple-65B Preview', a 65B-parameter dense Transformer, under the Apache 2.0 license.

The benchmark results are the highlight. It scored 80.52 on the agentic coding benchmark 'TerminalBench 2.1', beating US open-weight models more than eight times its size, including Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8), and NVIDIA Nemotron 3 Ultra 550B (56.4).

The specs: pre-training on 12 trillion tokens and a context length of up to 1 million tokens. Starting from Olmo 3, the team reworked scaling laws, architecture, data, and post-training from the ground up, improving compute efficiency on code by 16x with their recipe (up to 55x for full training). They deliberately chose a dense configuration over MoE and completed training in BF16 with no custom kernels or quantization, using simple sharding. The team reportedly grew from one person to five over the course of development.

Against the assumption that 'more scale wins,' showing that a small team can reach near the very top through compute-efficient design is a significant result. It's a Preview release and independent replication is still to come, but it could redraw the map of open-weight development.

New York Times Sues Brave Browser — Over AI Scraping and Paywall Circumvention

The New York Times (NYT) has sued 'Brave', the privacy-focused web browser, over AI scraping of its articles and circumvention of its paywall (The Verge reports).

Legal battles over AI and content have centered on lawsuits over generative AI training data, but this time the search/browser layer is the direct target — a new shape. How crawling should work, how paywalls are treated, whether AI features substitute for reading the underlying articles: these rulings bear directly on the design of search engines, summarization features, and AI assistants in general. It reads as the publisher-platform fight expanding to the browser, the very entry point.

Fuji TV to Set Up AI Use Across All Its Dramas by October 2027 — First Deployment with 'Kyumei Byoto 24 Ji'

Fuji Television announced it will build a setup enabling AI use across every drama it produces by around October 2027 (ITmedia reports). The first early deployment will be 'Kyumei Byoto 24 Ji'.

What stands out is that a Japanese broadcaster has publicly committed to redesigning its entire production line — 'all dramas,' not individual titles — on the premise of AI. How far AI will reach into screenwriting, filming, VFX, and art direction, and how the division of labor with existing production staff will be designed, are the focal points ahead. With Hollywood's recent labor disputes over generative AI still fresh in memory, how this lands on Japanese production floors is likely to serve as a reference for the entire industry.

Other Moves

The industry's people reshuffling is in motion too. According to a Latent Space roundup, Cat Wu, who has served as Head of Product for Anthropic's Claude Code since the company's founding, is moving to OpenAI to lead PM for Core Products. In the opposite direction, Sharon Zhou, who became VP of AMD's AI division through the Lamini acquisition, appears to be heading to Anthropic. Separately, AI coding company Factory announced its acquisition of HumanLayer.

On the open-weights front, Baidu's 'ERNIE-5.5-Flash-Base' was published on Hugging Face and became a talking point on Reddit's r/LocalLLaMA.