This August, I spent almost the entire month using Agent AI intensively.
I tried Claude, Codex, GLM, Kimi, Grok and several AI tools that were popular at the time. I used the top subscription tiers for Claude and Codex, and a heavy-duty version of Grok as well. The number of tokens I burned through during the month was extraordinary. So was the amount of work that came out of it.
The results fall roughly into three groups: research and learning, software development, and the organisation of personal archives.
1. Extending my work on partial differential equations
AI has helped me most in mathematics, both in learning and research.
Partial differential equations were not an interest I picked up for August. They are a continuing subject of study and research for me. I have spent a long time working through the problems in the field, including trying to understand and learn from Deng Yu’s work. With AI, that path extended further in August and produced several essays.
By "AI assistance," I do not mean simply chatting with AI or asking it to think in my place. It has gradually entered the whole research process: finding and organising material, tracing relations among concepts, checking derivations, building models, testing different paths of explanation, and turning scattered thoughts into a more complete argument.
More importantly, as the work continues, AI is helping me refine my own system of mathematical research.
Knowledge accumulated in the past, current research questions, new ideas and finished essays no longer sit apart. They are gradually connecting into a whole that can keep growing.
So the gain is not simply efficiency. AI lets the research extend outwards along paths I have already opened, while also letting me turn inwards to reorganise and improve the system behind it.
2. From the Bellman equation to Honor of Kings
Another piece of research was rather less respectable: I studied a five-player team strategy for Honor of Kings.
It began with the Bellman equation and expanded into formations, time windows, resource trades, tempo and coordination among several players. AI helped me build a number of models. I then tried to turn those models into tactics we could actually play.
The results were surprisingly good.
Our five-player team can generally sustain a win rate of 70 to 80 per cent. When the line-up is stable and the players perform consistently, the win rate remains remarkably steady.
For me, it is play and research at the same time.
There is no need to dress it up as a serious academic result. But it does show that mathematical models do not live only in papers. Any system with states, choices, feedback and long-term rewards can be approached in similar terms. A game happens to provide an environment where experiments and feedback are fast, and the model can be revised repeatedly.
AI did more here than offer advice. It helped turn judgements based on experience into clearer variables, relations and models. Actual matches then showed which ideas worked and which ones needed to change.
What emerged was not a collection of tips based on instinct, but a system of play that can be tested repeatedly and adjusted over time.
3. A failed algorithm experiment
In August, I also tried to improve the mechanism related to attention decay in Transformers.
I wanted to introduce ideas from finite element methods and information entropy to improve the distribution of attention within local structures. I ran many experiments and wrote an essay about the attempt.
In the end, it failed.
The method did improve attention locally, but it also suppressed the model’s emergent abilities. The local metrics improved while the most valuable property of the system as a whole was damaged. The trade was not worth making. I stopped further development, and for now that line of research ends there.
Even so, it was one of the month’s important results.
Research is not a process of endlessly proving that your first idea was right. Establishing why a direction does not work, finding the conflict between a local optimisation and the system’s overall ability, and stopping once the evidence is sufficient can constitute a complete research result.
Sometimes failure means not that you received no answer, but that the answer was no.
The result reminded me again of a recurring problem in complex systems: a local improvement does not guarantee an improvement to the whole. A mechanism that looks more precise and controllable can unintentionally suppress the space in which new structures and abilities emerge.
The failure was therefore not empty. It ruled out one route and gave me a more concrete view of the relationship among attention, local constraints and emergent ability.
4. Building software for myself and my family
I also spent a good deal of time developing software.
I continued to maintain my blog and added new features, including a dashboard. For an experimental mathematics curriculum, I built several learning tools for children as well.
I also developed tools for managing family photos, household items, family finances and fitness. Almost all of them remain private. They are not necessarily general enough to become products for everyone, but they are extremely useful in my own life.
I have wondered whether I should eventually release some of these projects as open source on GitHub. I remain cautious about doing so.
That is not because the tools work badly. Quite the opposite. They work so well largely because they were designed around my own family structure, habits and specific problems.
As AI becomes better at development, everyone may eventually be able to make software for themselves. Instead of searching the market for a generic product that almost fits, people will be able to build highly personal tools around the lives they actually lead.
If that happens, the value of software may change too.
Software used to derive much of its value from the number of people who could use it. In the future, some software may be valuable because it is genuinely useful to just one person or one family.
Under those conditions, whether to open-source a project, when to do it and what the source would mean to anyone else are no longer obvious questions.
I do not have an answer yet. Some projects may eventually become open source. Others may remain private systems forever. Either way, they already do useful work in real life. For me, the fact that they directly serve personal needs is value enough.
5. Organising a large personal archive
Another important task this month was organising photographs and videos accumulated over many years.
I think this may be one of the best and most practical uses of AI available today.
Moving files is not the hard part of a personal archive. The hard part is recognising what they contain, classifying and naming them, finding duplicates, and rebuilding the relations among dates, people and events.
Once the archive becomes large enough, that work consumes an astonishing amount of attention. It is easy to lose patience after organising only a small portion, so the whole plan keeps being put aside.
AI can take on much of the repetitive, fiddly work. It can also help define classification rules, process metadata, find duplicates and bring order back to scattered material.
For someone like me with a vast collection of photos and videos, the gain in efficiency is obvious. Work that remained stuck for years because of its sheer volume can finally be completed systematically.
And the point is not simply to make a hard drive look tidier.
Photos and videos are a family’s memory. Only after they are recognised, classified and connected can the dates, people and fragments of life buried in an enormous file store become visible again.
The cost is equally obvious: the tokens burn at a frightening rate.
6. Projects finished almost in passing
I also completed a few scattered projects in August, including a recording platform for a two-person podcast.
It took so little effort that it hardly seems worth singling out among the month’s results.
That fact says quite a lot by itself.
A working platform once counted as an engineering project that demanded its own plan. Now it can be a by-product completed almost in passing.
The platform has not lost its value. AI has simply reduced the cost of turning an idea into a working system. In the past, I would first have asked whether I had enough time, technical ability and energy. Now I can build something first, then decide whether it deserves further development.
As implementation becomes cheaper, many things that once stopped at the idea stage can become tools people can actually use.
7. When implementation is no longer the scarcest resource
Looking back, August gave me more than a pile of essays, models and software.
The more visible change is in how I handle my interests, research and ordinary life.
My study of partial differential equations can continue to extend, while my own mathematical system becomes more complete. Experience from a game can be modelled and turned into a stable method of cooperation. An algorithmic idea can be tested quickly, and even a negative result can produce a clear conclusion. A family’s specific needs can become private software. Years of accumulated photographs and video can finally be organised again.
Here, AI is no longer just something I talk to or a tool I call for a temporary task.
It is gradually entering the research, development and daily systems I already have. It connects what I accumulated in the past, advances work already under way and lets new ideas keep extending outwards.
After using Agent AI this heavily, implementation itself no longer seems to be the scarce resource.
More ideas can be made real and more projects can begin. Work that once demanded a large investment of time can now produce a first result very quickly.
That leaves a different set of questions. What deserves to continue? What should stop? What needs to serve only me? What should be public? And what deserves a longer, more complete form?
The failed algorithm experiment showed me that a direction is not worth pursuing merely because it can continue. The highly personal family tools showed that something valuable need not become a general product. Organising photos and videos showed that AI does not always have to create something new. It can also help us understand and organise the life we already have.
AI brings more than greater productivity.
It shifts the constraint from "Can I make this?" towards "What do I actually want to make?" "What do I want to preserve?" and "What kind of system do I want to build?"
That is the most honest account I can give of my August.