1:24 (1)Looking back on 2024 (perspectives of a journalist and an engineer) 6:14 (2)Given the evolution of AI in 2024, it is thought that the evolution of AI will "accelerate" in 2025 22:56 (3)Only those who are directly involved with AI can predict the future 30:47 (4)As AI evolves, it is possible that media companies will regain their former power 36:52 (5)As a journalist, I just can't help but enjoy this YouTube 43:20 (6)Retracing the evolution of models and considering the path to an intelligence explosion 1:15:19 (7)"AGI" may become a buzzword in 2025! ? ◆Notes on each topic (1)Looking back on 2024 (perspectives of journalists and engineers) (2)Given the evolution of AI in 2024, it is believed that the evolution of AI will "accelerate" in 2025. Reasons why AI will evolve exponentially: ①The synergistic effect of hardware and software ("Hardware (semiconductors, computers) evolve software (AI), and software (AI) evolves hardware (semiconductors)" "NVIDIA's GPU performance has increased 1000-fold in 8 years" by Jensen Huang) *Reference: GTC March 2024 Keynote with NVIDIA CEO Jensen Huang, NVIDIA https://www.youtube.com/live/Y2F8yisi... ②Scaling law (the prediction accuracy of the model follows a power law of three variables: computing resources, dataset size, and model size) *Reference: SITUATIONAL AWARENESS: The Decade Ahead, Leopold Aschenbrenner,, https://situational-awareness.ai/ - On the other hand, there is also Gartner's concept of the hype cycle, and some are of the opinion that we have entered the disillusionment phase. - Amidst the debate as to whether we will continue to the singularity or whether we are in the disillusionment phase, OpenAI announced the "o1" series three weeks after the disillusionment phase was reported. AI will evolve from "know-it-all AI" to "thinking AI", proving the disillusionment period to be a mistake ・Expectations for the scaling laws of inference are also rising *Reference: OpenAI's Noam Brown, Ilge Akkaya and Hunter Lightman on o1 and Teaching LLMs to Reason Better • OpenAI's Noam Brown, Ilge Akkaya and... ・Semiconductor industry analyst Dylan Patel explained, "With investments in semiconductors, synthetic data, and advances in inference, AI will evolve faster over the next six months to a year than it has in the past year." *Reference: AI Semiconductor Landscape feat. Dylan Patel BG2 w/ Bill Gurley & Brad Gerstner • AI Semiconductor Landscape feat. Dyla... ・It's not clear whether we've entered the singularity, but it's thought that the evolution of AI will accelerate at least until 2025 (3) Only those who are actually working with AI can predict the future. Even people at OpenAI have said that it is difficult to predict the speed and direction of AI's evolution, so predictions by people who are not at the forefront of AI are merely "speculation." Statement by Kevin Weil, CPO of OpenAI https://www.youtube.com/live/-cq3O4t0... Statement by Sam Altman, CEO of OpenAI https://www.youtube.com/live/-cq3O4t0... https://www.youtube.com/live/-cq3O4t0... *All are the same video (start times are different)・"Some argue that we should think of ways to deal with this before superintelligence is born, but I think it is more realistic to develop ways to deal with it when we see signs of the next ability, while we don't know what kind of ability AI will acquire next. The current safety issues are completely different from what we envisioned two years ago, and the solutions we envisioned two years ago will not solve today's issues at all. At the same time, we need imagination for the future. I think we should tackle AI safety from both the perspective of addressing the issues we face today and the perspective of long-term future safety." by Sam Altman (4)As AI evolves, we can expect to see a comeback for media companies.・Instead of using AI to mass-produce low-quality content, I hope that AI will be used to improve the quality of content.・The quality of the answers is different when you ask questions in English on OpenAI than when you ask questions in Japanese. In English-speaking countries, OpenAI is working with media companies, so it often publishes articles from major media as reference information, but in Japanese, such efforts have not progressed, and there are many articles that seem essential to AI. In the future, media companies may regain their position in terms of the quality of AI information sources. Right now, YouTube seems to have higher quality and easier to understand content. I don't know what mechanism or criteria ChatGPT search uses to determine information sources, so I still can't see how AEO (Answer Engine Optimization) will replace SEO (Search Engine Optimization). (5) As a journalist, I can't help but enjoy this YouTube. (6) Thinking about the path to intelligence explosion while tracing the evolution of the model again. First, looking back on the release of the GPT model (since GPT-2 in 2019). Looking back on my life, I've started asking ChatGPT about everything. I get the impression that the paradigm has changed. The role of humans will change depending on the degree of evolution. For now, I feel that humans are responsible for researching and summarizing the latest information. On the other hand, when it comes to summarizing information from about two years ago, it does the job almost perfectly. Graph of changes in the model's intelligence *Reference: AI Engineer World's Fair 2024 - Keynotes & Multimodality track • AI Engineer World's Fair 2024 - Keyno... ・"ARC-AGI-Pub Score", which represents the intelligence of o3. GPT-3's score was 0, GPT-4's was almost 0, and GPT-4o's was 5%, but it jumped to 25-32% in o1 and 75-88% in o3. *Reference: OpenAI o3 Breakthrough High Score on ARC-AGI-Pub https://arcprize.org/blog/oai-o3-pub-... ・The above blog speculates that o3 achieves this in a way not very different from AlphaZero-style Monte Carlo tree search. ・Demis Hassabis hinted in a June 2023 interview that DeepMind was researching exactly this idea. In addition, at a speech held at Google in late 2004, CEO Sundar Pichai commented that "Google will evolve greatly next year." ・Ilya Sutskever's impressive X-post: "If you think that the greatest value of a human being is intelligence, I think you're in for a tough time." *Reference: https://x.com/ilyasut/status/17104624... ・Looking at the Chatbot Arena Score, not only OpenAI but also Google's Gemini ranks high. On the other hand, Anthropic's score is surprisingly low. It seems that intelligence and ease of use for humans are not necessarily proportional. ・OpenAI's five-stage evolution of AI seems to alternate between "model evolution" and "UI/UX evolution." ・The road to intelligence explosion: 1. Acquiring problem-solving abilities equivalent to those of humans with doctoral degrees (already achieved), 2. AI agents that can act autonomously for long periods of time on behalf of users (2025), 3. AI research by AI - 1 million AI researchers (7) "AGI" may become a buzzword in 2025! ?・User perspective: Starting with OpenAI Operator, general-purpose autonomous agents will be available, and they will work with advanced voice and video modes, becoming indistinguishable from remote workers. At this stage, many people will feel that it is AGI. "AGI" will become a buzzword! ・It is interesting that a16z (American VC Andreessen Horowitz) asserts that agents will become RPA-type rather than API-type. The point is that agents can use tools. ・Behind the scenes, general-purpose autonomous agents will advance AI research, and the model will reach the "inventor" stage by 2025-26, which will realize powerful AI. *Explanatory video about powerful AI • CEO of Anthropic thinks that the near future of "powerful AI" is good... ◆Interviewer information [Endo Taichiro] CEO of Kaname Project Inc. Associate Professor, Educational AI Research Program, Tokyo Gakugei University, National University Corporation 25 years of AI experience. He started AI programming at the age of 18, and started a company while studying at the University of Minnesota Graduate School, and began providing services using AI. After working on a wide range of AI-related projects such as implementation, research papers, system design, business consulting, and education, he joined the AI ​​startup ExaWizards and oversaw the AI ​​department as a technical specialist officer. After going public, he became independent and is currently CEO of Kaname Project Co., Ltd., where he supports various AI/DAO/data utilization/DX-related projects. International Coaching Federation ACC/DAO Research Institute Founder, etc. https://kaname-prj.co.jp/ [Tsuruaki Yukawa] Editor-in-Chief of ExaWizards AI Newspaper Graduated from the Department of Economics at California State University, San Francisco. After working as a reporter for a local newspaper in San Francisco, he joined the US branch of Jiji Press. He has been covering the US high-tech industry since the dawn of Silicon Valley. After living in the US for a total of 20 years, he returned to Japan in May 2000. After working as an editorial committee member for Jiji Press, he became independent in 2010. He has been in his current position since December 2017. His main publications include "Artificial Intelligence, Robots, and the Human Mind" (2015), "Next-Generation Marketing Platform" (2007), and "Will the Internet Kill Newspapers?" (2003). His hobbies are yoga and meditation. He is proud of his beautiful wife. https://community.exawizards.com/aish... *Thumbnail image is a Pixabay image by 800bikuni https://pixabay.com/ja/users/800bikun...