lukax

And the hayro library is standalone and can easily be used outside of Typst. It only uses CPU and is pure Rust so it can also be used with WebAsembly. Link to demo below.

https://github.com/LaurenzV/hayro

https://laurenzv.github.io/hayro/

> It’s powered by a version of GPT‑5 that’s trained to look across multiple sources to give more comprehensive and accurate answers.

So another GPT-5 fine-tune. Codex also uses a custom GPT-5 fine-tune.

Does fine-tuning make sense now? Or do you have to be OpenAI to fine-tune the models with a mix of existing data and new behaviours?

And CopilotKit needs a JS backend for a proxy that holds the state for the chat which is quite a pain if you want to scalably self-host. Maybe ChatKit will be less lock-in than CopilotKit.

This looks very similar to CopilotKit and AG-UI. With CopilotKit you can have some tools implemented client-side and others server-side using AG-UI.

https://www.copilotkit.ai/

With ChatKit you can proxy everything through your ChatKit API "proxy" and implement the "agentic" stuff in the proxy.

Even the React APIs look very similar.

Looks like a vendor lock-in and a way to stop the AG-UI ecosystem

Example kernels for DeepSeek-V3.2-Exp

https://github.com/tile-ai/tilelang/tree/main/examples/deeps...

It's really accurate and supports 60+ languages

Have you tried Soniox? It's really not expensive ($0.12/h, $200 free credits when you sign up) and really accurate.

https://soniox.com/

You can use it with Spokenly (free app, bring your own Soniox API key) on macOS and iOS (virtual voice keyboard)

https://spokenly.app/

Disclaimer: I've worked for Soniox

Maybe you have an ad-blocker that just hides the popup but does not restore scrolling (scrolling is usually prevented when popups are visible)

Soniox offers real-time speech-to-text with real-time translation between 60+ languages (mostly to/from English with some additional pairs between more popular languages). It operates with minimal amount of context and produces the translated text as soon as possible.

https://soniox.com/

Disclaimer: I used to work for Soniox

Is this Triton's reply to NVIDIA's tilus[1]. Tilus is suposed to be lower level (e.g. you have control over registers). NVIDIA really does not want the CUDA ecosystem to move to Triton as Triton also supports AMD and other accelerators. So with Gluon you get access to lower level features and you can stay within Triton ecosystem.

[1] https://github.com/NVIDIA/tilus

word

Have you tried Soniox for speech recognition? It supports Croatian. Or are you just looking for self-hosted open-source models? Soniox is very cheap ($0.1/h for async, $0.12/h for real-time) and you get $200 free credits on signup.

https://soniox.com/

Disclaimer: I used to work for Soniox

Have you seen Soniox? They support real-time translation (only speech to text translation for now).

https://soniox.com/

(disclaimer: I worked there)

NVIDIA does not want CUDA development (e.g. flash attention) to move to Triton because Triton also supports AMD and if ecosystem moves from pure CUDA to Triton, that's bad for NVIDIA's lock-in. That's why there is so much focus on CUDA Python (lower level) and Tilus (higher level, more similar to Triton).

Japanese, not Chinese

Inference in Python uses harmony [1] (for request and response format) which is written in Rust with Python bindings. Another OpenAI's Rust library is tiktoken [2], used for all tokenization and detokenization. OpenAI Codex [3] is also written in Rust. It looks like OpenAI is increasingly adopting Rust (at least for inference).

[1] https://github.com/openai/harmony

[2] https://github.com/openai/tiktoken

[3] https://github.com/openai/codex

And 440MB tab memory usage

Soniox also supports real-time speech-to-text translation with 60 languages. You can hook that to a TTS and you have Speech-to-Speech translation. That failed Google I/O real-time translation demo? With Soniox it just works.

You can try it out here (select translation instead of transcription) https://soniox.com/

Disclaimer: I work at Soniox.

Try Soniox for real-time translation (interpreting). With the limited context it has in real-time, it's actually really good.

https://soniox.com

Disclaimer: I work for Soniox.

At Koofr[1] one of the most requested features was an option to prevent downloading files from public links. We didn't want to lie to our users so we added a "Hide download button" option because that's the only thing you can do. You can hide the download button but you can never really prevent the download.

[1] https://koofr.eu

Not really. But I didn't use async (was not supported yet when I started using it).

Bindings were easy, everything else (building, linking,...) was a bit pain to setup because there no good examples.

Do not write the bindings manually. Just use the amazing uniffi-rs library from Mozilla.

https://github.com/mozilla/uniffi-rs

You can generate bindings for multiple languages. It supports error handling on both sides and latest versions also support native async integration.

I've used it to reuse the same Rust engine in iOS and Android apps and write native UI.

https://github.com/koofr/vault

So people could just use a non-US VPN and a non-US Apple account to download the apps?

Uv also bundles uvx command so you can run Python scripts without installing them manually:

uvx --from 'huggingface_hub[cli]' huggingface-cli

It's a bit funny that they use Jedi Language Server for Python as they cannot use Microsoft's own Pylance ("Pylance is licensed for use in Microsoft products and services only, so can only be used on official Microsoft builds of Visual Studio Code and GitHub ...").

They also used Omnisharp Language Server for C# which is deprecated and was replaced with a new extension with custom license terms.

Go errors in standard library does not support stack traces. errors.Wrap() only exists in github.com/pkg/errors package

The author uses Whisper and GPT-4o to get transcriptions into a nicely formatted Markdown file.

We just released Omnio, a new AI model, that can do all this in a single step, as it works with audio directly. It does not generate a transcript and then modify it but can generate structured output such as Markdown directly from audio.

Maybe you can check it out.

https://soniox.com/blog/omnio/

Whoops. We've updated the website. Omnio is available to all developers and all accounts receive $5.00 in free credits.

We built Omnio to address the limitations we kept running into with existing audio AI models (we previously built an automatic speech recognition product). Most of them rely heavily on speech-to-text, which strips out a lot of the things like speaker roles, emotions and non-verbal cues that are critical in more complex scenarios. Omnio directly processes audio signals to capture that kind of context. It's designed to understand conversations in a way that feels more "human."

Google Docs is also migrating to Kotlin Multiplatform

> The initial step in this journey is the rollout of the Google Docs app for Android, iOS, and Web, which leverages KMP for shared business logic, validating its readiness for production use at Google scale. The Google Workspace team is thrilled to continue exploring the possibilities of KMP across its product suite, aiming to enhance productivity and deliver seamless experiences to users on all platforms.

https://android-developers.googleblog.com/2024/05/android-su...