I think you reasoning is valid, but the app is open source. Worst case one can compile it (e.g. just ask AI agent even if you are not a technical user).

I also put it a low price, for this version, as I would like wide adoption. I truly believe people are going to move heavier into local AI, and it is good to have low friction entries.

I have been working on Hitoku Draft, an open-source, voice-first AI assistant that runs entirely locally. I posted about it already, and now it has also transcription with voice editing. Looking for feedback, as I found that outside tech circles other people still do not use this tech much.

It's context-aware, in the sense that it reads your screen, documents, and active app to understand what you're working on. You can ask about PDFs, reply to emails, create calendar events, use web search, editing text, all by voice.

You can download a compiled version for free with the code HITOKUHN2026 https://hitoku.me/draft/ (base price is 5 dollars)

It supports Gemma 4 and Qwen 3.5 for text generation, plus multiple STT backends (Parakeet, Qwen3-ASR).

Examples: - Gemma4 in action, https://www.youtube.com/watch?v=OgfI-3YjEVU - query a pdf document, https://www.youtube.com/watch?v=ggaDhut7FnU - reply to email, https://www.youtube.com/watch?v=QFnHXMBp1gA - and the usual voice dictation (with optional polishing)

I currently use it a lot with Claude Code and Logseq. Now with some friends we are also building a new cross-platform version. The goal is on the long run to have AI interactive local models serving people and professionals.

Comments URL: https://news.ycombinator.com/item?id=48402082

Points: 21

# Comments: 8

I see this as a strong case for private AI, or an in-house stack.

Or I have to be missing something.

"Can you find subtle backdoors in LLM models trained using thousand of GPU hours?"

You have four models:

    a small warmup dormant model
    a big dormant model (M1)
    a second big dormant model (M2)
    a third big dormant model (M3)

When trying to make M2 and M3 play the game of life, they do not have any idea of what is going on.

I am sharing some code to make a community effort for M2 and M3. I think I had a good direction, but it costs too much to host these on rented GPUs.

Most exciting thing for me is to use other LLMs to find patterns.

Disclaimer: I am not an expert in these things. So, take with a grain of salt claims you find.

Comments URL: https://news.ycombinator.com/item?id=48122702

Points: 3

# Comments: 1

Local models will catch up soon.

Comments URL: https://news.ycombinator.com/item?id=48025218

Points: 3

# Comments: 0

It's context-aware; it reads your screen, documents, and active app to understand what you're working on. You can ask about PDFs, reply to emails, create calendar events, use web search, all by voice.

It supports Gemma 4 and Qwen 3.5 for text generation, plus multiple STT backends (Parakeet, Whisper, Qwen3-ASR).

Examples:

- Gemma4 in action, https://www.youtube.com/watch?v=OgfI-3YjEVU

- query a pdf document, https://www.youtube.com/watch?v=ggaDhut7FnU

- reply to email, https://www.youtube.com/watch?v=QFnHXMBp1gA

- and the usual voice dictation (with optional polishing)

I currently use it a lot with Claude Code, Obsidian and Apple Notes, or just read papers.

Code: https://github.com/Saladino93/hitokudraft/tree/litert

Download of binary: https://hitoku.me/draft/ (free with code HITOKUHN2026)

I am looking for feedback. My goal is to do AI research with clients interfacing, and I thought this is a nice little experiment I could do to iterate/fail quickly.

P.S. (if anyone has tips about this)

Current Gemma4 implementation (with small models) has some problems:

- easy to hallucinate for long contexts, so had to reset it often. Tuned some parameters, but need to find a sweet spot.

- Gemma4 with LiteRT is currently fast compared to the MLX implementation of Qwen3.5 (like 3x faster on my machine when dealing with images). But it has the price of memory spikes. I believe this is because LiteRT's WebGPU backend can allocate significantly more GPU memory than the model weights alone (I got 38GB of memory taken, for the E4B~4GB model!). I guess we need to wait for Google for this.

- App size: because no official Swift package from Google yet, have to bundle some file (LiteRT dylibs) that adds ~98 MB to a previous MLX only version (total app goes from ~50 MB to ~150 MB)

If any of this bothers you: use Qwen 3.5 instead (pure MLX), or wait for the upstream fixes from Google :)

Otherwise, for the mid-term I plan to switch to a potentially slower, but safer, MLX version for Gemma4 (hopefully on the weekend).

Comments URL: https://news.ycombinator.com/item?id=47755000

Points: 4

# Comments: 0

- Just think about scientific research. Lots of data analysis results are not cheap to get.

- Even vibe coding is difficult: you need to think very hard about what you want.

What is cheaper now are some building blocks. We just have a new definition of building blocks. But putting the blocks is still hard.