Kokoro is great. I built an mcp for it a while back that has gotten decent traction - https://github.com/mberg/kokoro-tts-mcp if anyone wants to go that route
I've been using VibeVoice's ASR (speech to text) model quite intensively for the past month and have found it to be a lot more reliable and out-of-the box functional then Whisper, parakeet and other models. The fact that is has diarization built into to the model is a huge win in my book. Without that you have to run a different model just for that which adds significantly to the overall processing time vs VibeVoice which gives you reliably great results. Big fan.
I created a virtual voxel 64x32 app that emulates the 64x32 screen of my Tidbyt. Voxels are 2D SVGs building off David Aerne's delightful Heerich.js (https://meodai.github.io/heerich)
I realize this probably doesn't need to exist in this world but I've gotten a lot of job from it. My main current use is to create virtual displays on extra monitors that I can power from a pixlet apps (pixlet apps generate a 64x32 webp image this emulates).
It's now possible to leverage channels to create a simple HTTP server to wrap it. I created a agent http as a quick proof of concept to demonstrate this.
This mimics the API of Agent API (https://github.com/coder/agentapi) which has relied on clever terminal scraping to provide an http interface for claude code.
The TLDR is this enables you to run Claude Code in a headless way via API calls (perfect for AI code orchestration) while using your Anthropic subscription. To achieve this before you would need to use Anthropic's Agent SDK and API tokens.
I just created agent-http that leverages the channels feature to enable you to wrap claude code with a http api. This provides an identical API to Agent API (https://github.com/coder/agentapi) that relies on terminal scraping to achieve this. Now you can interact with claude code in a headless manner using your subscription. Previously I think you had to do this via the Agents SDK which relies on api token use.
One way to think of an incremental improvement is cutting a sharp edge of a process, solution or problem. A single improvement may not make a noticeable difference but in aggregate the changes can compound to transform a nascent idea to a viable solution. We think this particularly relevant when thinking about AI.
We created this little visualization to help demonstrate it.
Example - The compounding effect of incremental AI coding improvements has made it now possible to build something like this just to illustrate this example!
I've been an avid Mapbox user since getting alpha access to Tilemill. From my perspective, this is the most important contribution to the open mapping space since the introduction of vector tiles and Mapbox GL / Map GL. Mapbox in my opinion left the door open in how they approached tile baking with MTS. While powerful it was way to confusing and expensive. While you could always bake tiles with Tippecanoe you still had to struggle with vector tile hosting hosting. With PMTIles vector tile hosting is arguably easier, more convenient and an order of magnitude cheaper then going with Mapbox (especially if you don't need fancy projection / 3d support and are ok with MapLibre). Felt is one of the major services that I believe uses the PMTile approach behind the scenes and they won't be the last. (Notably: they have continued to invest in Tippecanoe which is awesome). With Mapbox focus shifting, it appears, to car navigation - it's great to see disruptive innovation coming from new places. Even more amazing that is has come from a one person team.
Also to note, the emergence of PMTiles coincides DuckDB spatial query support will unlock a lot of innovation. The ability to query and process large geoparquet files and quickly stream them into baked PMTiles unlocks a lot of really compelling use cases.
Hi MedPlum team - congrats on the launch! We've been working on the FHIR space pretty heavily for the past two years. Our main focus is building native FHIR apps for Android currently.
I see a lot of value in some of the components you are adding from an ecosystem standpoint. Quick q - are you building a server or focusing on the components on top? I guess I'm trying to understand the value prop vs. a system like HAPI of Google Health API? How robust is your support for data extraction eg. with StructureMaps, etc?
Congrats on the launch! Look forward to digging in more and my reach out.
Mapbox has been pretty strategic about not being perceived to compete with it's developers. Besides creating an amazing stack remaining neutral I think is part of their appeal.
Just this mean, we can style a raw vector map from mapbox without applying styles? I'm not able currently style maps with a huge number of point on vector mapbox due to the tile processing limit. I am able to host the vector file though. Being able to directly style raw vector files on the client really opens the doors to a realm of possibilities.
Nice one! Try this one. When I was living in NY, I'd always want to bring a magazine for the long subway ride downtown. So one day I took one of the magazines that get delivered to my home and enjoyed the read on the way downtown. When I got to my destination I'd just pop the magazine in the nearest mailbox. Usually a day later it would delivered back safely back home to be read another day. I've done this dozens of times without fail.
The language and examples in the blog post above seems to have taken a lot of inspiration from posts by MapBox below. Regardless of whether or Google was working on it in advance or not - it's a feature I imagine you can't afford but not copy. Giving some attribution to the concept or at least a shout out to MapBox for pulling this off seems at least warranted.