Large Language Models

Today's Stories

1. 1

   US judge approves Anthropic's $1.5 billion（約2400億円） copyright settlement

   U.S. District Judge Araceli Martinez-Olguin granted final approval on Monday of Anthropic's $1.5 billion（約2400億円） settlement of a class action lawsuit brought by authors who alleged the company used their books without permission to train its AI chatbot Claude. This is the largest known settlement of a U.S. copyright case. The settlement ends the first major U.S. copyright lawsuit against an AI company to reach resolution, after a trial was scheduled to begin in December with potential damages running into the hundreds of billions of dollars. Authors and other copyright holders filed claims covering over 92% of the more than 480,000 works included in the settlement, establishing a precedent as dozens of similar cases remain pending against tech companies by copyright owners including authors and news outlets.

   Some authors and publishers opted out of the settlement and have filed separate lawsuits against Anthropic that are still ongoing. The settlement also faced objections from some authors who argued it is not large enough or wrongly excludes certain copyright owners.

2. 2

   OpenAI fears open-weight models; US weighs China AI ban

   OpenAI's strategic futures head Dean W. Ball argued the US government should use regulation to discourage open-weight models, claiming they deter capital spending by frontier labs. After pushback from figures like Yann LeCun and Martin Casado, Ball retracted the claim that regulatory crackdown was the White House's "best strategy." Reports now indicate the Trump administration is considering banning advanced Chinese models like Moonshot's Kimi K3, though the Department of Commerce may not act soon. Open-weight models running on independent infrastructure offer cheaper AI than proprietary services from companies like OpenAI and Anthropic, which threatens the return on their massive training investments. However, restricting them could cede innovation leadership to China—US graduate programs already build primarily on open Chinese models, and half the papers students study come from Chinese institutions, according to research cited in the article. The real lever may be chip export controls rather than banning software.

   The article notes uncertainty around AI economics: neither the open nor proprietary business model is yet proven, and Chinese AI companies face the same revenue and compute struggles as US firms. Nvidia and others are exploring open-model businesses, suggesting the landscape may shift if multiple companies can sustain open releases rather than just frontier labs pursuing closed models.

3. 3

   Google's Next AI Chip Lifts Alphabet Stock 3%

   Alphabet shares rose about 3% on Monday after reports emerged that the company is developing a next-generation AI server chip code-named "Frozen v2," designed to deliver six to 10 times more AI tokens while consuming the same amount of power as current tensor processing units. Deployment is targeted as early as 2028. The custom processor aims to reduce the energy and cost demands of training and running advanced AI applications, which is critical as Alphabet scales its Gemini AI models. A more efficient chip strengthens the company's in-house AI computing capability and reduces reliance on third-party hardware at a time when large technology companies are competing intensely over generative AI systems.

   The chip's actual deployment timeline and whether it achieves the projected six to 10 times performance gain per unit of energy. This is a key test of Alphabet's strategy to build proprietary infrastructure that supports its long-term AI ambitions.

4. 4

   Open-weight models near frontier parity as closed labs maintain lead

   Open-source AI models including DeepSeek R1, GLM-4.6, GLM-5.2, Kimi K3, and others have reached equivalency with closed frontier models, though closed systems like GPT-5.2 and Opus continue to create step-change advances. Recent releases include Moonshot's Kimi K3 (2.8T parameters, July 16), Alibaba's Qwen 3.8 preview (2.4T, July 19), and DeepSeek V4's mid-July graduation from preview. Open models run approximately 15% cheaper than GPT-5.2 at median frontier quality, with DeepSeek V4 Flash roughly 90% cheaper. This pricing pressure is reshaping industry margins—Anthropic is reaching its first profitable quarter—while competition has driven OpenAI to cut inference costs by 50% and spurred architectural innovation like Kimi's new KDA attention mechanism.

   The industry is cycling between closed models pulling ahead and open models catching up, potentially creating sustained competitive pressure on pricing and margins. The question of whether this dynamic will slow innovation or accelerate it through competition remains central to how fast the AI wave advances.

5. 5

   Google designs Gemini-specific chip for 6–10× efficiency gain, deployment from 2028

   Google is building an internal server chip called Frozen v2 that embeds Gemini's model architecture directly into silicon. The chip could be 6 to 10 times more efficient at serving AI responses than Google's current TPU chips, and Google plans to deploy it starting in 2028. In AI, inference cost optimization increasingly determines profit margins. Frozen v2 is designed to ease Google's internal compute strain and could let Google run powerful models at lower cost—a potential competitive advantage against OpenAI and Anthropic. Unlike Google's TPU line, which it leases to Meta and external cloud customers, Frozen v2 is built specifically for Gemini and unlikely to become a commercial product.

   Google has not yet decided how much of Gemini's architecture will be hardcoded into the chip. The design—which embeds the model architecture rather than weights—allows new weights to be loaded, making it more flexible than an earlier approach that would have locked the chip to a single Gemini version.