KTibow

It sounds like newer devices had a Do Not Send Voice Recordings option that transcribed and processed queries locally.

The site being made by AI explains some of it.

> inspect to undisable it

To be clear, this is to reinstall it if you accidentally uninstalled it. Keeping it enabled is easier but still follows a dark pattern.

It might not directly learn it, but it could still identify it if it's quoted by others or if there's enough text of the same style with the author's name as context.

It's unfortunate that this isn't a default option. The only viewer I've seen that happens to be able to do this is Gnome's Document Viewer.

It's an 8B model. There's lots of room to grow.

One thing I don't understand is why Claude 3.5 Haiku, a non thinking model in the non thinking section, says it has a 8192 thinking budget.

I don't know if it is AI generated, but the formatting (lists with bold text, unnecessary headers, lots of en and em dashes) is setting off my spidey senses.

Looks like they started work Feb 1 and haven't made anything yet.

In fact, you can still use Svelte 4-style components with Svelte 5, even in the same project.

It's not much better than DeepSeek's old slogan "Unravel the mystery of AGI with curiosity. Answer the essential question with long-termism."

I tried this: https://github.com/KTibow/thoughtgraph

I wonder if you could turn it into a graph by adding connections between the two most similar entries until everything is connected.

They experimented with music back in 2020 [0]. I suppose they didn't continue because the risk (work, training, possibly getting sued) is greater than the reward (splitting the market with Suno).

[0] https://openai.com/index/jukebox/

It's a relatively recent addition and only appears if you stop typing when you're only a few words in

What color space does it use when generating the brightness variations?

Yes

(albeit I believe o3-mini isn't natively multimodal)

fyi, you can automatically detect the user's language and switch to English when applicable by using `navigator.language`...

...or even better, don't use JS (which doesn't work on older browsers and prevents some optimizations), and read the Accept-Language header instead

Checking with GLTR:

- 6 of the tokens were within the top 10 predictions - 8 of the tokens were in the #10-100 range - 3 of the tokens were in the #100-1000 range - 3 of the tokens weren't in any of the top 1000 predictions

It seems the only thing that changes how covered a section is the speed you go over it with.

Checking the source, you could also use your keyboard (eg 0-9 is a grayscale palette).

Is that to say that the server can do things on the client connecting to the server? That doesn't make much sense.

Ironic when it's faster than Java almost every other way.

The increase is likely because 1.5 Flash was actually cheaper than all other STT services. I wrote about this a while ago at https://ktibow.github.io/blog/geminiaudio/.

You should only use reasoning models if your prompt needs reasoning (eg debugging a weird error or writing an optimized algorithm).

Why does the example code use a base model to generate the analysis input?

This doesn't let you include multiple files in a commit, right?

I don't know how it wouldn't be - it can't retract things already sent to the client. (The alternative is to moderate every chunk server side before sending it back, like Gemini does.)

Oh, I guess you would need something like dynamic imports to not include uncommonly used functionality in the main bundle

If you're making a serious website, alternatives include Vercel, Cloudflare Pages, Deno Deploy, and Fly