I added transcription prompts to LocalAI's Whisper backend and immediately put it to use in VoxInput. This means that when I say "LocalAI" and "VoxInput", it actually transcribes those correctly. Also works great with OpenAI.
github.com/mudler/Local...
github.com/richiejp/Vox...
Posts by Richard Palethorpe
VoxInput using LocalAI to allow you to transcribe voice to text in any Linux desktop or app. This video shows how it works and some guidance on setting it up: youtu.be/bbZ_9-Uzp78
github.com/richiejp/Vox...
I released VoxInput v0.5.0 which now has a GUI that notifies you when VoxInput begins or stops listening for voice audio to transcribe.
The Go is more verbose than the Python, but it's still pretty small. I could probably make it smaller if I was really optimizing for that, however I'm not motivated to do that. What you see is the Go that came out as I went with minimal refactoring.
github.com/richiejp/mic...
Micrograd is "A tiny Autograd engine (with a bite! :)). Implements backpropagation (reverse-mode autodiff) over a dynamically built DAG and a small neural networks library on top of it with a PyTorch-like API. Both are tiny, with about 100 and 50 lines of code respectively. "
I watched Karpathy's excellent Micrograd video and implemented the Micrograd library in Go. Today I recorded some of my thoughts on the differences between the original Python and my Go implementation: www.youtube.com/watch?v=76DB...
I added real-time voice transcription to VoxInput by implementing the transcription only mode for OpenAI's real-time transcription API in LocalAI.
youtu.be/Ims16LC_Bp8
I created a voice transcription agent for Linux that works with any Linux desktop and uses OpenAI's and therefor LocalAI's API to transcript audio with Whisper.cpp. It can input text anywhere because it uses a virtual keyboard via dotool
github.com/richiejp/Vox...
Mudler merged my triggers and filters feature into LocalAGI: youtu.be/d_we-AYksSw?...
Not much interest in IRC though, maybe I should upgrade to Matrix?
I made a little update to LocalAGI's action playground that creates a form for action configuration and parameters, this saves you having to enter raw JSON. The Action Playground is for testing actions when the agent has issues making tool calls.
youtu.be/HtVwIxW3ePg
Funny Man Dog Poo
Never say that we don't still make shit in the UK!
hardware that stack has to work with. Resulting in a sea of configuration options, many combinations of which have never been tested.
A lot of the frictions to running AI locally are just generic software configuration issues. Basically there is a GPU involved which is a different kind of architecture and comes with its own software stack. Each vendor has there own stack and there are a whole bunch of combinations of software and
In this weeks news letter I wrote/ranted about the process of supporting Intel GPUs in LocalAI with SYCL. Essentially it's *just* a case of getting the software to build and that boils down to passing the right flags to the compiler. 1/2
If you vibe code a SaaS app, the second it makes money that should be reinvested into paying someone to clean it up. Coding agents are incredible, but so far I have not seen any which do penetration testing and vibe coding produces a large attack surface.
I've lost so much time over my career to simple build errors buried in piles of awkward software. Often the error manifests at runtime as something completely baffling and terrifying. At the end it turns out to be typo which would be completely avoidable with some static checking.
It's incredible that vibe coding is even a thing and my son is going to grow up thinking it's normal to speak to computers in natural language.
Most people think an AI agent is a person who does something with AI. Not an automation that combines LLMs and other algorithms to give the appearance of agency.
more is expected from software.
Compilers eliminated the need to write assembly by hand in most situations. However there are more assembly coders today than ever before because the industry is so much bigger.
What is quite likely is that the role of software dev will change beyond recognition.
Let's say that the stories about software dev teams being let go due to AI are true; I don't see this situation lasting for long. Every previous advance in software technology has lead to greater penetration by software into every aspect of industry. There is then more demand for coders, because
In this week's newsletter I vibe coded a "bare metal" program that runs inside a Firecracker micro VM on x86_64 and accepts commands through the serial terminal.
So yes kernel developers, you are not safe either!!!
Often we want to know which process is using our bandwidth and where it is sending data to. This can be done with Dinko Korunic's pktstat-bpf utility which I added CGroups support to this week.
buff.ly/K2ZUXwt
AI coding tools are reshaping development: I built an IRC bridge and a Go web app with minimal effort using Avante for Neovim. Will these tools replace developers or make skilled coders more valuable? My hands-on experience with Avante, Cursor, and Bolt[.]new reveals their promise and limitations.