From Ryuta Hamamoto at TIMEWELL
This is Ryuta Hamamoto from TIMEWELL Corporation.
AI technology is evolving at a pace that makes it genuinely hard to keep up. On January 28, 2025, I had the opportunity to present at an AI study session organized by the Musashino Valley community. This article covers what I shared that day — the major trends in generative AI, what to expect going forward, and the emergence of DeepSeek, a China-developed model that has sent real shockwaves through the industry.
2024 in Generative AI: Six Defining Trends
Allow me to introduce myself briefly. I am Ryuta Hamamoto, CEO of TIMEWELL Corporation. I serve as a visiting researcher at Shinshu University, a visiting researcher at Musashino University's Entrepreneurship Research Institute, and I lead CHANGE by one Japan, a program supporting people who take on new challenges.
My career began in sales at Panasonic, which led to new business development work, and in 2022 I founded an AI startup. Today we focus on AI education and developing AI-powered applications.
Here is how I see the six major trends that defined generative AI in 2024.
1. Extended memory Earlier models could only hold short-term context. Now, models can retain long conversation histories — enabling conversations that genuinely build on prior exchanges, not just the last few messages.
2. Multimodal capability Communication was once text-only. Now models handle images, audio, and more, enabling far richer interactions across different types of content.
3. Deep reasoning Models have become dramatically better at thinking through complex problems — including problems that require sustained, multi-step reasoning rather than pattern-matching. Some have achieved over 91% accuracy on university entrance exam questions.
4. Vertical AI Specialized models for specific domains — medicine, finance, law — are emerging. These aren't general-purpose assistants; they are purpose-built for high-stakes, domain-specific problems.
5. Efficiency optimization Training and running large models has always been expensive. That's changing. Optimization techniques are bringing costs down substantially, which matters enormously for what comes next.
6. Agent AI AI that can handle project-level tasks is now here. Give it a goal, and it can create and execute the steps to reach it — including building the applications it needs along the way.
These six developments together have moved generative AI from something impressive in demos to something that genuinely changes how work gets done.
Looking for AI training and consulting?
Learn about WARP training programs and consulting services in our materials.
What 2025 Looks Like
Looking ahead from where we were at the start of 2025, several shifts seemed nearly certain.
Personal AI becomes real. Long-term memory management will improve, and AI increasingly becomes something personalized to each individual — a system that understands your context, your work style, and your history.
Multimodal becomes the standard. Integration with AR/VR, automatic content editing, and AI that participates in meetings to handle notes and task tracking — these are no longer speculative.
Deep reasoning reaches practical use. The "think long and hard before answering" capability becomes a reliable feature rather than a research milestone.
Multi-agent workflows appear. Multiple AI agents coordinate to handle entire workflows, not just individual tasks.
Smaller, more efficient models spread. The shift isn't just toward larger models — it's toward models that do more with less. This is where DeepSeek comes in.
DeepSeek: The Model That Changed the Conversation
The development I have been watching most closely is DeepSeek — a large language model developed by China's DeepSeek (originally referred to as DeepSearch). Among all the models competing with ChatGPT, DeepSeek has stood out for three reasons: performance, open source access, and cost.
Performance: DeepSeek matches or exceeds models from OpenAI and Google on key benchmarks. This is not a discount option — it is genuinely competitive at the frontier.
Open source: The code is publicly available, meaning researchers and developers can use, study, and build on it freely. That changes the dynamics of the entire ecosystem.
Cost: API pricing for DeepSeek runs at roughly one-tenth to one-thirtieth of comparable models. That cost advantage is not marginal — it is structural.
DeepSeek offers two main models: V3 and DeepBlue. Both perform at GPT-4 level or above. The company was founded by graduates of Peking University, who also run an AI-driven hedge fund — a background that explains the emphasis on efficiency and quantitative rigor.
There are legitimate concerns. As a China-developed model, DeepSeek has known restrictions on politically sensitive topics — Tiananmen Square, Taiwan, and similar subjects are filtered. There are also open questions about how user data is handled. These are real considerations.
That said, the balance of opinion in the developer community has been that the advantages outweigh the concerns — and DeepSeek's rise to the top of US app store rankings reflected exactly that assessment.
App Development Is Opening Up
The broader implication of these trends is that building software is becoming accessible to people who would never have described themselves as developers.
I built a presentation practice app in about thirty minutes — on my phone, at the gym. The app records what users say, sends it to an AI for analysis, and returns structured feedback. That used to require a team and a sprint cycle. Now it requires an idea and an afternoon.
I also built a marketing diagnostic app for small retailers, also in a short session. Users enter their business type, target customer, and current challenges. The AI analyzes the inputs and proposes a marketing strategy.
The comparison I keep coming back to is YouTube and TikTok. Those platforms didn't just create a new type of content — they changed who creates content. Video production went from studios to smartphones. The same shift is happening with software.
Tools like V0, Bubble, and Thunkable are removing the technical barriers. With V0, you describe what you want in plain language — even something like a simple game — and the tool generates the code. If there's an error, you paste the error message back into the chat and get a fix. The feedback loop is tight enough that beginners can iterate their way to working software.
WARP: TIMEWELL's AI Development Training Program
At TIMEWELL, we run WARP — a program for developing young technical entrepreneurs who can build things themselves. The program has been selected as a partner initiative under Tokyo's "TOKYO SUTEAM" startup support program. The goal is to produce engineers who can develop at five to ten times the conventional speed by integrating generative AI into their workflow.
No prior programming experience is required. With generative AI tools, people who have never written a line of code can build real applications — and bring development in-house rather than outsourcing it.
The program is open to students and working adults under 30, and participation is free. If any of this sounds relevant, I encourage you to look into it.
Looking Ahead
Generative AI is not just an efficiency story. DeepSeek showed that the barriers to high-performance AI are collapsing — in cost, in access, and in who can participate in building it. App development is following the same pattern. The question for organizations and individuals is not whether to engage with these tools, but how quickly to develop the skills to use them well.
The tools are here. The cost is dropping. The window for building an advantage is open — but it won't stay open indefinitely.
