AI뉴스 - 수학 끝났다, 챗GPT 대화형 UI, Claude Motion, Haiku 5.5, AI 인권 등장, Grok Bot 업데이트, 어도비 망했다 등

Summary

The video surveys recent AI developments across mathematics, AI agents, creative tools, science, and software. The presenter also demonstrates an AI-stock-trading backtest.

AI Research and Mathematics

• The presenter reports that OpenAI and Google DeepMind have used advanced models to produce or solve mathematical results, including problems that had resisted progress for years.
• These developments have prompted debate about recursive self-improvement (RSI). The presenter says researchers are increasingly focused on building agent loops that let AI improve its own work, rather than relying only on human code review.
• Mathematicians have raised concerns about how the results were presented, including whether AI-produced proofs advance human understanding, who receives credit, and whether rapid disclosure could weaken the research ecosystem. Terence Tao is cited as arguing that mathematical progress involves more than obtaining correct answers.
• The video also notes that call-center employment has recently begun to decline after a long period of growth—possibly one sign of AI’s labor-market impact.

OpenAI and AI-Agent Updates

• Interactive ChatGPT responses: The interface can produce interactive, website-like answers with cards, charts, and controls, rather than only text or images.
• Meeting notes: A ChatGPT desktop feature is described as recording and organizing meeting content, initially for certain paid users.
• Text watermarking: OpenAI is reportedly exploring ways to track text origins. The presenter questions how robust this would be, since minor edits might remove the detectable pattern.
• Codex Cloud and Tailscale: Cloud coding tasks can continue while a user’s computer is off, with Tailscale providing a connection to private computers or company resources.
• Next-action prediction: Codex can predict and start a likely follow-up task with less manual prompting.
• Decision API: A fast, lower-cost API for simple choices—such as selecting a tool or making a yes-or-no decision—is presented as useful for agents and time-sensitive applications. Microsoft’s competing Decision-1 model is also mentioned.
• A SemiAnalysis comparison is cited as suggesting that Claude subscriptions may deliver more task value than ChatGPT subscriptions at several price tiers.

Demonstration: Backtesting an AI Trading Strategy

The presenter demonstrates a backtesting workflow using NH Investment & Securities’ Namuh Plug:

1. Obtain an API key from Namuh Plug and connect it to the example.
2. Prepare historical stock-price data.
3. Adjust the strategy conditions and AI prompt.
4. Run the backtest and review the AI’s decisions, including its reasons for buying or holding.

One experiment shows a small reported profit, but the presenter stresses that AI decisions can vary even when conditions are unchanged. Past or simulated performance does not guarantee future returns, and real investing carries a risk of loss.

Anthropic and Claude

• Claude Motion and Dashboards: New beta features are shown generating dashboards and motion graphics from prompts. The presenter demonstrates a coding-themed animation and suggests this could reduce the need for separate motion-graphics workflows.
• Claude for Startups: The program is described as offering eligible startups a year of Team access, API credits, and other software benefits. The presenter says demand was high and that the offer may be revised.
• Haiku 5.5: A lower-cost model is presented for simpler tasks. The presenter compares its price and benchmark performance with other small models and suggests it could be useful for routine work.
• Model-abuse policy: The presenter says Anthropic’s updated policy prohibits persistent, unnecessary abusive treatment of its models and could lead to account consequences.
• The video reports that a user was arrested after writing violent threats in Claude. It emphasizes that prompts are processed in the provider’s cloud and that limited disclosure to authorities may occur in emergencies.

Grok, Google, and Other Agent Tools

• Grok Bot: The video describes plans for Grok Bot to route tasks to leading models, connect directly to X for searching and monitoring, and use its own email address. It also discusses proposed shopping functionality and cautions users against uploading personal or financial information in light of the company’s warning.
• Gemini Agent for business: Google is described as offering an always-on cloud agent that can orchestrate tasks and use models from its own lineup as well as Claude.
• Google Docs and Markdown: Markdown files can reportedly be viewed and edited directly in Docs.
• Game creation: Google’s experimental Playgrounds and Unity Spark are presented as tools for creating playable games from prompts, potentially without writing code.
• EmbeddingGemma 2: A small embedding model can map text, images, audio, and video into a shared representation, enabling searches across different media types. The video also describes a free, local Mac note-taking app built around this approach.

Science, Image Generation, and Software

• Protein design: An open-source project is described as using AI to design proteins—and potentially enzymes—for desired chemical reactions, rather than only searching for naturally occurring proteins.
• Plant genetics: Google DeepMind’s plant-focused models are said to help identify genetic changes associated with plant traits.
• Image models: The presenter discusses updates to Nano Banana and Qwen Image, while comparing image-editing consistency with other models.
• Adobe alternatives and reverse engineering: The video highlights open-source recreations of familiar creative-software functions and tools that help coding agents inspect and reproduce existing software. The presenter frames this as a challenge to established creative-software companies, while noting that AI has made reverse engineering more accessible.
• Game mashups: Examples combine mechanics from different games, such as FPS gameplay with Minecraft-like environments, and use AI-assisted analysis to create playable prototypes.
• Mistral Large 4: A large open-weight model is mentioned, though the presenter considers its reported performance less compelling than that of some competitors.

Music, Industry Outlook, and Career Advice

• The presenter notes that AI-generated music is appearing on music charts and discusses a case in which a person allegedly used bots to generate fake streams and claim royalties. The video characterizes this as fraud exploiting streaming royalty systems.
• A State of AI report is summarized as describing intense competition among major labs, the growing importance of agent design, rapid adoption of coding agents, falling model costs, China’s open-source progress, and infrastructure constraints such as power and funding.
• Career advice emphasizes defining worthwhile problems, building expertise and reputation, working with strong collaborators, and distinguishing one’s work through the final layer of quality—since basic outputs are increasingly easy to generate.
• The presenter promotes coding and AI-literacy learning resources, showcases games submitted to the JoCoHunt project, and says a Google app-monetization event will be covered in a later video.

Guides and Demonstrations Featured

• AI trading backtest: A step-by-step demonstration using historical prices, an API key, configurable prompts, and a review of the AI’s decisions. The results are examples only, not investment advice or a promise of returns.
• AI-assisted game creation: Brief demonstrations of prompt-based game-building tools and game mashups.
• Claude Motion: A demonstration of generating motion graphics from a text prompt.

Main Speakers and Sources

• Main speaker: JoCoding (조코딩), presenting and commenting on the news and demonstrations.
• Organizations and products discussed: OpenAI, Google DeepMind, Anthropic/Claude, xAI/Grok, Microsoft, Unity, Mistral, Adobe, NH Investment & Securities/Namuh Plug, and SemiAnalysis.
• Other cited sources or people: Mathematician Terence Tao, the State of AI report, and various project announcements and examples referenced by the presenter.