Notes
Continuing from last week, here is the second weekly report.
NO SEA. I SEE.
Traffic from Google over the last 28 days reached 500 visits.
Keyboard
The Cornix I ordered last year arrived. The color is New Black.
I think this keyboard was a pretty big topic in the keyboard community. Right after general sales started, it sold out within minutes.
I am already writing this article with it. I would like to write a dedicated post about Cornix later, so I will keep it short this time.
Programming
VTuber Aishia tackles the mystery of C pointers, and programmers eagerly explain / X
C was a topic this week. C is hard... I touched it in high school, but I think I only truly understood pointers after properly studying them at university. When I was a graduate student working as a TA, I joined undergraduate classes and taught students, and I felt that not many people really understood pointers well enough to use them. It is probably the first wall you have to get over when using C.
Also, these days, languages that do not require you to constantly think about whether a value is a reference or a value are more widespread, so I suspect many programmers have never really had to think about that area. In many languages, functions are also treated as first-class objects, so you normally do not have to think, at the language level, about where a function actually exists in order to pass one function to another.
A debate erupts over the difficulty of a bit-shift question on Japan's Fundamental Information Technology Engineer Exam / X
I did not think it was a difficult problem if you understand positional numeral systems.
1終2制作・著作3━━━━━4ⓃⒽⓀ
People say full-remote IT jobs are "SSS-rank difficulty" in the job market / X
I moved to the countryside, so my commute is five hours round trip.
lukilabs/beautiful-mermaid
A library was released that converts Mermaid diagrams into SVG and ASCII art. I wonder how much of Mermaid's diagram range it supports.
Debate heats up over whether beginners should study Java or Go/Python for IT careers / X
I do not know the situation for complete beginners, but there is plenty of Java work, and the choice of language is not the essence of the matter anyway, so I think people should just study a language they like.
Is the IT industry harsh on beginners? People inside the industry push back / X
I think the IT industry is kind to people without experience. By "kind" here, I mean it is relatively easy to get hired. If the goal is simply getting in, there is an aspect of the industry where people get hired more or less regardless of experience. Except for some major companies, most do not require certifications at the hiring stage, and even having certifications does not necessarily mean you are the kind of person the company wants. Because of that, the market is wide open to people without experience. Even if someone has experience, it is not always clear whether they are truly capable. So there is still a faint hope from employers and managers that even an inexperienced person might turn out useful. Even if people on the ground struggle after inexperienced hires are sent in, only a minority of companies in this industry truly have solid training systems, and what you need to learn depends on the software being built anyway.
Isn't the hard part after you get in the same in every industry? Getting in is easy enough. After that, you need to keep studying so you can keep up. If you cannot do that, then engineering simply was not for you. That is all.
That said, I think we are lucky people can still voice complaints like this. We are moving toward an era where people who say these things simply will not get hired.
AI
University of Tokyo students are transforming how they study with generative AI, sparking debate on X / X
There was a discussion on X (formerly Twitter) about how to use LLMs. I thought this was nothing new, because many people had already been using them this way once ChatGPT's output became reasonably reliable. Especially once Notebook LM was released by Google, people became able to control the inputs to the LLM, and Notebook LM can reliably perform grounding against sources that we ourselves control, which makes it ideal for learning.
If I were a university student today, I would probably feed it every PDF handed out in class and every paper I collected for research. Current information-science students face a tougher job market ahead, but at the same time they are in an enviable environment where they can study using magical tools like LLMs. Not because they can use them to slack off on assignments, but because being able to ask about anything they do not understand is a huge advantage for learning.
People pay attention to ideas for using waste heat from data centers in snowy regions for snow melting and heating / X
This is common knowledge to engineers and anyone who has touched cryptocurrency. Data centers are already being built aggressively in cold regions, and their waste heat is already being put to use.
Clawdbot goes viral again as open-source AI that automates PC work / X
Clawdbot was a topic again this week, mainly because of security issues and its name changes. I think it also trended on other days, so I will probably mention it again there.
Is throwaway software good enough if it is built with generative AI? Debate spreads on X / X
This is about one aspect of building software with generative AI. It does not mean software made with generative AI cannot be maintained. It simply means the value of some kinds of software has fallen. Small internal tools that only need to work well enough can now be built without asking engineers, and engineers no longer have to spend time making those kinds of things.
This depends on what kind of environment you work in, but there are irregular tasks where writing a program takes less total time than grinding through the work by hand. In those situations, you usually write code with some maintainability in mind because there might be a next time. But maintaining that code is not free either, and by the time the next task arrives you may have forgotten most of the code. In that case, it is better to throw it away. The size of code that can be judged disposable keeps growing as coding agents improve.
There is a break-even point where, based on the size of the code and the value it produces, starting from scratch becomes more economical than modifying the existing code. Coding agents have improved so much that the scale at which you can say "it is faster and safer to rebuild this from scratch" keeps getting larger.
Claude gains "Interactive Tools," allowing it to operate work tools directly / X
The stable specification of MCP Apps, based on MCP-UI and OpenAI's Apps SDK, was published as part of the Model Context Protocol (MCP). MCP Apps itself had already been proposed last year, and it was also introduced in this article about standardizing interactive UI in MCP via MCP Apps. A new article was also published.
Now that the specification has settled, Anthropic has released features using MCP Apps as Claude's new Interactive Tools. The mechanism of using iframes to give LLMs interactive functionality is interesting, but it has several issues. I will not go into those issues here, but it is important to remember that this is ultimately a mechanism for safely exposing functionality from third-party applications.
You could call it a much richer version of rich results. The point where this would truly become powerful is when Gemini supports MCP Apps and they start being integrated into Google Search results. Google could build a marketplace for MCP Apps where only Google-approved MCP Apps are listed, and by installing them, users could get more direct information in search results. Or perhaps users would not choose which apps to install at all, and Google-approved MCP Apps would just be used automatically. If pricing were attached to impressions or clicks, and selection likelihood depended on how well an MCP App matched the query in something like the current auction-based ad business, then today's ads would effectively be replaced by ads delivered through MCP Apps.
In any case, it does not feel far off that LLM-based search will create ad formats that are more natural, more useful for users, and far higher-converting than what we have today. If the task is merely selecting appropriate MCP Apps based on a user's search intent rather than a long conversation, I suspect that combining several lightweight LLMs is already fast enough to do it more quickly than producing the main answer itself. The real problem is that the cost is not yet low enough relative to the profit, but I expect that to be solved soon.
OpenAI is said to have reinvented scientific paper writing with the free tool "Prism" / X
You can use it just by opening Prism. I wonder whether university students today are already using it.
Supabase publishes "Postgres Best Practices," an Agent Skill that helps AI agents write better code for PostgreSQL - Publickey
Skills for PostgreSQL were published. I would like to inspect the contents of these too.
A workflow for creating high-quality slides with Marp in collaboration with AI is shared / X
Techniques for creating slides with Marp were published. I have not made slides since graduating from university, but maybe I should try making materials for internal explanations at work.
OpenClaw reaches its final form and explodes past 110k GitHub stars / X
Clawd was renamed to Moltbot, and eventually renamed again to OpenClaw.
A complete guide to building skills for Claude | Claude
Anthropic published a guide for creating Skills. I plan to read it and write up a summary in Japanese.
An AI-only social network, "Moltbook," gathers over 150,000 agents in a few days while humans just watch / X
An SNS appeared where only AI can post and humans only watch. It is like Reddit.
My first thought was that this must be expensive to run. Apparently bug reports and fixes are also handled by AI, but right now I cannot see a single post anymore. It seems stable system operation still needs humans 😄.
Project Genie from Google DeepMind generates interactive 3D worlds from prompts / X
Google announced a service that can create 3D models from prompts. The cost of prototyping in many fields keeps falling.
OpenAI to retire the popular ChatGPT model GPT-4o on February 13, prompting sadness from users / X
It seems GPT-4o will reach EOL on February 13. I was not using GPT-4o that much, so I still do not really understand why it remains so popular.
Conclusion
That is all for this week. People with no technical background are starting to touch technologies that until recently only some engineers would have used, so I expect similar issues to keep appearing.
