built a local STT tool to brain-dump into Claude Code

Claude Code is great with detailed context, but I got tired of typing long prompts in the terminal all the time. I wanted local speech-to-text (STT) tool that actually worked in the terminal, so I wrote a small CLI, `hns`.

• Run “hns”, speak, hit Enter.
• faster-whisper handles the speech-to-text locally.
• The result goes straight to your clipboard for pasting into Claude Code, Slack, email, etc.

No data leaves the machine, works offline after the first model download, MIT licensed.

Install: `uv tool install hns` or `pipx install hns`

Code: https://github.com/primaprashant/hns

Demo ⬇️

so a total of 245 talk proposals for PyCon JP and there are 44 slots. This means acceptance rate will be 18%. This is lower acceptance rate than ACL (~21%) and CVPR (~23%) and NeuralIPS (~26%) 😲

entire tech community these days

built a little something to track long weekends in Japan to plan for them in advance

check it out longweekend.jp

gave my first tech talk at Tokyo AI last Tuesday about MCP. Looking forward to sharing more at future events with the AI community in Tokyo

I had two main insights I wanted to share about MCP, and based on my conversations with people afterward,the second one was particularly appreciated. While everyone had some understanding that MCP helps LLMs use tools, many found it eye-opening when I showed how the models learn about available tools and their capabilities. The revelation that all the tools and their schemas are simply appended to the system prompt was a surprise for many in the audience

coined a new term 「アイハラ」for when you are being harassed to use gen AI / LLM for everything

tried asking the new gemini checkpoint (gemini-2.5-pro-preview-06-05) about MCP to see if there had been any knowledge update but nope

it hallucinated hard for this question even at 0.35 temperature, way more than 05-06

found out that if you want to talk to LLMs about MCP, only Claude 4 models know about it. Gemini pro 05-06 and gpt-4.1 just bullshit you

for gpt-4.1, it makes sense cause the knowledge cutoff is supposed to be June, 2024 but for gemini, it's January, 2025, so it should know the specs of MCP

everything is a copy of a copy of copy

everything is a summary of a summary of a summary

saw Green Day a couple days ago, now I have seen two of favorite bands, Linkin Park and Green Day, perfom live in the same month, both for the first time, this has been a really good month

saw this post on HN to tell the time in corporate, so made an internal one for my org cause our fiscal year starts in July, which makes things very confusing

went to an AI event couple days ago where a person was demoing how amazing cursor is. The goal was to build a simple blogging website in next js where you can create blogs post and each blog has their own page and then list all the blog posts on the homepage. And everything is done by vibe coding using the composer feature of cursor. The speaker started the project from scratch three times and did multiple iterations each time over the duration of an hour but the blog didn't work and got build error each time no matter how much fix he tried

now all of this wasn't so bad. This kind of thing happens all the time with live coding demos but what's interesting is that before demo started, the speaker did a little introduction for cursor. They used the fact that microsoft said people who use gihub copilot are 55% faster and there was a dig at microsoft that if you're 55% faster by using something made by microsoft, you'll be at least 80% faster by using anything else (like cursor). Seems like the only thing that was 80% faster was how quickly the errors showed up!

finally saw linkin park live

was reading about model merging yesterday. This model merging paper from Sakana AI and the followup using quality diversity is just amazing. Second paper is very hard to understand so I'd definitely recommend at least reading the first one
https://arxiv.org/abs/2403.13187
https://arxiv.org/abs/2410.14735