📖 英文と日本語訳（一文ずつ）

🎧 通し読み（全文）

リスニング・シャドーイング用の全文です。

On April 24, 2026, Daisuke Okanohara, co-founder and president of Preferred Networks, delivered a lecture titled "The Evolution and Future of Generative AI: Domestic AI and Physical AI" at the Policy Research Institute of the Ministry of Finance. Mr. Okanohara first pointed out that generative AI has evolved phenomenally over the past decade, and that various generative tasks, such as language, images, and robot control, have come to be handled in a unified manner. Three factors were cited as the background: a 4.6-fold annual increase in computing power used for AI development, the exponential growth of training data which has now reached the scale of tens of trillions of characters, and the publication of hundreds of thousands of research papers annually. AI performance is improving steadily and predictably, reaching passing levels for the University of Tokyo entrance exams and the National Medical Practitioners Qualifying Examination. OpenAI’s o3-pro also achieved a 21% accuracy rate on “Humanity’s Last Exam,” a collection of the most difficult problems gathered from 1,000 researchers. The cost of providing equivalent intelligence continues to fall by a factor of 100 to 1,000 annually, while long-term task processing capabilities are doubling every four months. It is reported that Claude Opus 4.6 is achieving a 50% success rate on tasks that take a human 12 hours to complete. Furthermore, autonomous research such as "AutoResearch" has emerged, in which LLMs improve their own learning while humans sleep. Regarding the direction of AI development, it was explained that while improvements driven by the scaling laws of pre-training—previously the mainstream approach—are reaching a plateau, rapid progress is being made in enhancing training data quality through AI itself and improving reasoning capabilities via reinforcement learning with verifiable rewards. It was noted that gigawatt-class data centers and development teams of thousands are being utilized for training. Capital investment in U.S. AI data centers has surpassed 1% of GDP, and some predict it could expand to 6%, mirroring the railway investment boom of the 1880s. On the demand side, software development support accounts for more than half of current LLM demand, with the market size estimated to range from several trillion to tens of trillions of yen. OpenAI’s weekly active users have surpassed 900 million, with annualized revenue reaching $24 billion as of February 2026. Meanwhile, Anthropic has continued to grow at an annual rate of more than tenfold over the past three years, with its annualized revenue reaching $14 billion. An estimated 140 trillion tokens are processed daily in China, with ByteDance alone accounting for 30 trillion tokens and Google for 43 trillion. Mr. Okanohara positioned AI data centers as "token factories," presenting an estimate that one gigawatt of power could support a revenue potential of $50 billion (7.5 trillion yen), 23 trillion tokens per day, and the equivalent of one million GPUs. In its GTC 2026 keynote, NVIDIA projected the AI infrastructure market to reach $1 trillion, declaring a shift from the era of training to the era of inference. Regarding the situation in Japan, the analysis found that China lags several months behind the United States, while Japan and other countries are approximately one year behind. An analysis of the digital deficit structure estimated from OpenAI’s cost structure pointed out that AI semiconductors are the largest factor for capital outflow, with only 20% to 30% of the total currently remaining within the country. Risks associated with the use of closed models were cited, including Anthropic’s large-scale analysis of user requests, a proposal to mandate investment in the U.S. for countries purchasing AI semiconductors, and the temporary designation of Claude as a "supply chain risk" by the U.S. Department of Defense due to a conflict with Anthropic. Based on these considerations, Okanohara emphasized that the continued development of domestic generative AI is of critical importance for avoiding dependence on foreign suppliers, accumulating development know-how, ensuring secure closed-network operations for confidential information, eliminating bias, reducing the digital trade deficit, and protecting Japanese culture and social norms. Preferred Networks highlighted its initiatives, including PLaMo, a domestically developed foundation model with world-class Japanese language performance; Matlantis, a materials discovery simulator used by over 150 organizations; Kachaka Pro, which holds the top share (47.9%) of the domestic autonomous mobile robot (AMR) market; MN-Core L1000, a next-generation AI semiconductor featuring 3D-stacked DRAM scheduled for release in 2027; and a cybersecurity business in collaboration with GMO. Finally, addressing human work styles in the AI era, the presentation highlighted the importance of defining goals and constraints, managing data access permissions, decomposing tasks into appropriate levels of granularity, and continuous auditing and intervention, concluding that proper environment design (harness) and task management will be the key.

📝 学習のヒント

1. 1まず英文を読む — 知らない単語にあたりをつけてから音声へ。
2. 2一文ずつ確認 — 日本語訳と照合し、構文を理解する。
3. 3通し読み Normal で耳を作る — 内容を追いながらリピート。
4. 4Fast でシャドーイング — 口を慣らし、リスニング速度を上げる。
5. 5翌日に復習 — 1日空けて再聴すると長期記憶に定着しやすい。