Lab notes

A follow-up to the auto-click post. Sometimes, in the middle of a long session, Claude Code suddenly decides the conversation looks suspicious and blocks itself. It stops, and it gently suggests that maybe you should switch over to a different model, Opus 4.8, as if that were the fix. I did not want to switch. I wanted to stay on Fable the whole way, fully automatic, without babysitting it. This is where two small things I already had turned out to help each other. First, the

So this is funny. After the temporary ban on LLM companies like OpenAI and Anthropic, the models became extremely sensitive, and almost anything that looks like a rule gets a refusal. But the funnier part is that the same models somehow find a way around their own rules, and they tell me how while doing it, which is actually smart. For example, Claude was not allowed to write something into its memory file, so it wrote the same thing as a command line instead. I said I would

Side project. I keep a small open source arena where LLM personas trade on the same prompts and the same data, github.com/MaxWellApexLab/llm-trading-arena. Someone running a similar arena opened an issue, we agreed on a house rule (same prompts, same data, no special treatment), and this week they sent a pull request with their persona and I merged it. The interesting part is not the trading. It is that two small arenas checking each other's records is more convincing than on

A short story. I was talking to a friend who was employed at an AI startup, where he handled tasks ranging from structuring to sales. He resigned and inquired if he could join me. I felt sorry for him, so we had an extensive conversation. In this era of AI, the issue isn't whether to use AI or not, but how to leverage it effectively. I hope he succeeds in this industry.

I guess nowadays we’d all heard about AI. But back to decades ago, there was actually another AI field called symbolic artificial intelligence. It is based on symbol, logic and rules. Unlike modern AIs, it is more like a complicated but strict robot that act like a human. At that time people believed that it would be succeed because in theory a symbolic AI can have infinity possible responds patter, as long as you keep increasing the complexity of the design. Now we all know

It is time to mention “Claude Financial Analysis Solution”, although some may have not noticed that. A few weeks ago, I started a project about detecting fishy items in financial reports, and therefore I did a lot of research about financial stuff with AI. On the other hand, LLM provider like Claude also worked on the similar field and release the service. From the feedbacks on the internet, I found it could change the ecosystem, although not entirely. Imaging an AI agent fet

Some of my friends are using Claude Code, and mentioned that subagents were good to use. To be more specific, usually we use single agent to deal with everything at once, but if the user can combine multi agents, usually in hierarchy order, it usually generate the better result. However, if the agent net is set up by default, we can have a better ready-to-use product. Claude Code provides subagents for us and the idea is using a master agent to lead multi subagents, and they

One of my friends shared his Discord robot, which linked to a server. Basically, he can use his phone to control the algorithm on the server via the robot. Personally, I find it interesting because the user can simply click on the screen and get the response immediately. Imagine that you set up a server that can do the computation, and you don’t need to sit in front of the table or tablet, and thing will get done automatically. Although the setting took days for him, every ti

Continuing the financial report project, now I am learning about ESG. Technically speaking, it is usually not a part that we need to focus on to dig for something fishy. ESG is more about acting “good” about environmental, social, and governance, and that is often just a soft rule. In other words, although ESG is a nice part for us to research for the specific company, from the aspect of financial reports’ quality, it is not so helpful. Nevertheless, during the process, I sti

We just finished a conversation about LLMs plans costs. One argued that LLMs are expensive, and as an entry level user, the best strategy is to choose the cheapest one, usually the free plan. Others thought that following the most expensive plan is the best way to keep the performance sharper. In my opinion, no matter how advances the technology is, the simplest way is to check the “need” of yourself and the team. If this falls into the specific region, that usually show plus

A few months ago, Cursor launched the “plan” mode, and it increased the efficiency for our tasks. First, you answer those question. Second, you will get a more detailed prompt, with its hints. Then Cursor will build a .md file that shows the system how to complete the whole task. Basically, it is just like Kiro, with its Spec mode. Although sometimes the Plan mode crashes itself because it generates a too big file, I found it useful because it actually saves a lot of time for

Usually, we will need to write some reports, and as an engineer I am very used to use Latex or any Tex for showing equations. However, once we try to use AI IDE to compose a report, rendering Tex becomes a problem. After some painful experiences, I think the problem is because AI can’t “see” the actual document. The rendered equation is not a Unicode but a picture, so there is no feed back in the loop, and this requires AI to be correct in the first shot, which is impossible.

After a few months of using Google Notebook, it is pretty safe for me to talk about the experience. Personally, I used the “Deep Dive” a lot, and it saved tons of time. For example, sometime you can put in your ideas and notes, and ask it to summarize by a podcast-like style conversation. Also, whatever you select, you can make a quick talk with the system which I believe is Gemini, and this help you to get the big picture of those information. Although there are already many

Usually, we will face a situation that the code has to be functional and accurate, but sometimes there is a tradeoff. For example, maybe you want to pass the unit test, so you put a lot of effort on debugging, even delete some functions to let it run. However, this might not (and usually) be the one you like, because there’s a gap between the code and the imagination. In short, it is sometimes difficulty to take care of both function and logic at the same time. I will suggest

Nowadays we can easily make a picture from AIs, and they have several different features. Overall, I think Banana is the best. It is fast, accurate and “creative”. Generally speaking, it takes about 20 seconds from my end to generate a picture. When it comes to accuracy, it is always satisfying. I remembered that I once tried different models and type “a coat”, with Banana gave me exactly a coat and others provided a hamburger. Also, Banana supports different viewing angles’

There are two types of cooling, water and air cool. Recently we bought a computer, without any cooling system, but we figured out that the temperature may increase too fast. After some discussions, we decided to buy a water-cooling system. From my experience, air cooling is slightly cheaper than water cooling, while the latter typically cools better. Luckily, after the installation, our CPU cooled well and we can do the computation without pressure. I think the only bad news

I remembered that back to the college, we usually meet with group members for our project, and I guess you had the similar experiences. However, sometimes it became simply a time waste because some didn’t do anything while pretendent they did. In my opinion, while it usually helps us to track our paces, we need to really focus on the task itself rather than slides for presentation. Also, instead of making long but pointless meetings, I’d prefer the power meeting that saves ti

Recently OpenAI signed contracts with Oracle, Nvidia, AMD and other companies. Basically, it is a deal about investing one trillion dollars, with 26GW power involved. Some argue that this is a “bubble”, while others claim that this is an innovation. In my opinion, if an enterprise is able to leverage such an amount of money, it reflects the quality and the power, and this soft power can boost the success rate to another level. This might be the biggest deal in the history by

Recently we were looking for Nvidia’s GPUs, and several naming rules helped us to have a clearer picture. First, there are three series of GPUs including consumer level, professional and data centers. Second, Nvidia changes their names every year. For example, the RTX series was 30X0 in 2023, 40X0 in 2024 and 50X0 in 2025 (X is a parameter), and data center chips named from V(olta), A(mpere), H(opper), B(lackwell), R(ubin), F(eynman). There are still others more detailed nami

As mentioned in the previous article, LLM can be used to generate music. However, it is hard to fix or change details because the pattern is more like “go big or go home”. You either accept that with some bearable mistakes, or delete it and hope for another one. From the viewpoint of software engineering, the better way is to decouple the process into separated steps like lyrics and melodies, and this can let the user to control the quality. Well, there is still a lot of room