Opening
Hi everyone. I'm Akira Hattori, founder of coiai, an IT company in its first year that builds systems and products with technology at the core. Every Monday, I share what running a startup is really like.
This week I'll walk through some tech news and then talk a bit about the business side of things.
Tech News
Which programming languages suit Claude Code
First up in tech news — I run a Python script every day that pulls the top stories from Hacker News and Zenn, and it's become a habit of mine to check it. More than half of this week's picks were about Claude Code.
One article that really blew up was about "which programming languages suit Claude Code," and it racked up 396 likes. The gist was that languages with strong typing perform dramatically better. Languages like TypeScript, Rust, and Kotlin, which have solid type systems, apparently make it easier for AI to grasp context.
On the flip side, plain Python or JavaScript — languages that were traditionally favored because they're easier for humans to write compared to strictly typed languages — lose some of that advantage now that AI takes care of the tedious parts. So it makes sense that more strictly typed languages might actually come out ahead.
Then there's Rust, which has a reputation for being fairly difficult — for someone like me who came up through frontend work, Rust always felt like it had a high barrier to entry. Same with Go, though it's an appealing language for being fast and lightweight. The article suggested that AI makes these languages more approachable to try, and that AI itself might actually be stronger — more accurate — with these kinds of languages.
The "AI Slop" problem
Next up is something called the "AI Slop" problem. This is similar to what happened with image-generation AI, where people questioned mass-posting AI-generated images on social media. The same thing is now happening in code — AI-generated code is getting pushed straight into open source repositories, with pull requests sent as-is. Apparently this causes problems because it ignores the original philosophy and conventions of the OSS project, and results in a flood of pull requests with insufficient tests or inconsistent documentation. It's being called the "AI Slop" problem, and it was another hot topic this week.
GPT-5.4 release
And then there's the latest news — GPT-5.4 came out, a new ChatGPT model. This got a Hacker News score of 761 with 633 comments, so it drew a lot of attention.
Of course the story is basically "performance improved," but honestly I haven't been using ChatGPT much lately since I do most of my work with Anthropic's tools, so maybe I'll give it a try. I read that presentation-generation accuracy has improved, so it might be worth checking out.
That said, I've been using Marp a lot recently, and I find it pretty convenient — there's something elegant about converting from Markdown into all sorts of formats. Beyond that, ChatGPT and Gemini both have canvas features that, as I understand it, run on HTML behind the scenes — that's true for Gemini's canvas at least. That approach feels a bit limited in versatility, hard to extend into other things, so I'm not expecting too much there. There are also AI services like Napkin. Combining tools like that is something to think about, though I'm not expecting much from that either — but I am glad the accuracy has gone up.
Anthropic and military use of AI
Something else — this isn't just a topic among engineers; it's becoming a broader conversation as AI gets used more widely, including in military applications.
Anthropic, the company behind Claude, made an official statement about its relationship with the Department of Defense, and this too generated a lot of discussion on Hacker News — a score of 370 with 362 comments. Meanwhile, OpenAI has agreed to provide its models to the Department of Defense's classified networks, and everyone's watching closely to see what stance each company takes.
It's a tough issue. Whenever new technology emerges, there's always been this question of whether it gets applied to military use. Scientific progress and military weapons seem to pair up all too easily. It's genuinely hard that mass-destruction weapons tend to emerge alongside technological progress, and it's unclear whether that's even something an individual developer can control. I don't have a strong take on this, so I won't go too deep, but it struck me as a difficult problem.
Is the MCP server on its way out?
There was also a lot of talk about whether MCP servers are over. I don't necessarily think MCP is done for, but it's true that it eats up a ton of tokens and can be a bit slow. That said, I'm not exactly a power user of MCP myself, so people who really know how to use it might feel completely differently.
To explain briefly what MCP actually is — up until now, AI tools like ChatGPT or Claude could really only operate within their own single piece of software. MCP lets them reach out and operate other software too.
There seem to be a lot of possibilities — articles about using it together with Figma, for example. But for anything you can already control via a CLI (command-line interface), you don't necessarily need to wrap it in MCP, so there was a "why not just use the CLI" argument going around, and honestly, that's fair. It doesn't burn many tokens either, so in a sense it really is the best option.
Then again, people have probably thought this for a while — take AWS or GCP, for instance, both of which you can operate via CLI, like the AWS CLI. But you can also use the dashboard and operate them through a GUI. For repetitive, fixed tasks, using the CLI tends to be better even for humans, and for a computer, handling things through language is even more natural and efficient. So it makes total sense that this discussion is coming up. The idea that a CLI is friendlier to computers really does check out.
That's about it for tech news this week.
An entrepreneur's dilemma — sales vs. management decisions
Now, moving on to something that's been on my mind lately as a founder — or I guess as a manager. Lately there's been a lot of subcontracted work. My company does build its own apps, but the main bulk of the work lately has been building things for other companies, or implementing specific features they need. Since I'm the one — the CEO — doing that work myself, I have no time left to actually run the business.
I haven't been able to sit down and think through how to move the business forward — how to grow the company, whether to bring in outside capital or keep bootstrapping with our own funds. I can't discuss it with anyone, and I don't even have time to think it through carefully by myself. That's basically how this past week has gone.
For now, one thing I've decided to try is exhibiting at the Startup Japan trade show next month, though I have no idea how that will turn out. My company is essentially run by me alone — everyone here is an engineer. I'm an engineer myself, and I'm terrible at sales. Or rather, I have zero sales know-how. I don't even know how to approach B2B sales.
Ideally, in the classic startup formula, having a dedicated salesperson who can operate independently is the most important thing. As long as someone brings in the work, the rest — development, hiring — can be figured out. As long as you're not distrusted by people, you can gather a team, since money will follow. If you can keep providing work, you can build all of that. So landing work really is the hardest part of building a company. Doing the work itself, I can do endlessly — but landing the work in the first place is what I find hardest.
And along with that comes questions like what scale of work to pursue, what channels to build to reach clients — that's the real business side of things. There's also the question of forming partnerships, or making management calls like whether to take on financing and put more resources into sales to expand more broadly. I haven't been handling any of that particularly well, and that's been my struggle this week.
Ending
And that's it for this week. I was buried in work, so I'm sorry I didn't have much else to share, and this episode also went up a bit late, on Tuesday. Sorry about that. Starting next week I'll be back to posting on Mondays, so please tune in.
All right, that about wraps up this week. coiai works on web development, XR development, core system development, and support for deploying on-premise AI. We can take on pretty much anything, so if you've been turned down elsewhere or gotten a quote that felt too high, please reach out — quotes are free. See you again next Monday — subscribe to the channel and I hope you'll come listen again. Thank you.