What I'm working on now

• Improved TTS mechanisms.

  This is so that the robots will sound more authentically human. Current avenues of research include the use of voice cloning technologies using software transformer based text to speech solutions. The current technique enployed at present when the chatbots are on involves a kind of antiquated but very efficient neural network technique to producte human like voices but the way in which the words are pieced together can sometimes give things a very human sounding version of the speak and spell toy robot voice.

• Backend Reconstruction

• Website AI Technologies

  Experimental offline systems are under construction intended to provide website users with the following functions in future versions of the GDSWA (currently v15-26Q3).

  • Conversational AI (LLM): Claire and Tessa et all.

    Currently using llama.cpp backend with a specialised Llama 3.2 based model. Future intents to maybe increase the parameter count of the model. Currently its a 4B model and there is a 7B I can also use which experimentally works, but I like to keep some spare VRAM and since 4B is enough as a proof of concept I've opted for fairly stupid bots. I could ramp this up to a single bot on the website which is more intelligent than gemini albeit not all of geminis talents.

  • Speech Recognition (ASR): The green button in the chat.

    Currently using one of the faster models for the whisper.cpp backend. Some issues with its ability to understand some stuff is noted and there is a plan to integrate a higher parameter model which will work not as slowly as once thought. This is on the back burner with a lot of tasks as you can see.

  • Text to Speech (TTS): Audio output when the chatbots speak back.

    Currently using the RNN techniques employer in piper-tts. It sounds better than speak the win32 speech SDK voices like Microsoft Mike and Microsoft Sam and even delivers some "acting" through the words but it does not sound entirely authentic throughout, there is a limit on how much can be spoken in one go and obvious interface limits if a an ASR user has to wait ten seconds for a short reply. The current setup is working satisfactorily but could improve on the voices "acting" slightly. An additional constraint is a limited number of voices and a general search to find a voice never heard before speak in a certain way. Future intents is to use voice cloning and morphing to achieve just the right voice shape for specially designed voices.

  • Voice Cloning (cloning): Under Construction.

    At the current time I can clone any voice I have an audio for. Difficulties arise attempting to clone voices with any background noise, particularly music or highly entropic noise that nosie reduction filters cant remove. Voice data taken in recording studio environments are ideal and I know the kind of voice I want for whatever part I want played but I cant just use Morgan Freeman but its okay to have something like that but not him. So a long story short, after a bit of tweaking I get a voice thats deep and reassuring and friendly and this is achieved through quite a complex voice cloning process. I can clone anyone saying anything with a particlar mood an have an angry bob or a happy bob or a sad bob and if you feel like I do, then you can imagine a suprised bob at just how authentic things sound.

  • Cloned Voice TTS (cloned TTS): Under Construction.

    Now this is on the back burner because it takes a long time to render text produced by a cloned voice than it does for the piper-tts the site currently uses, but if this technique can be sped up (and this may be possible without a hardware upgrade), then we are good to use voice cloning as the next version of onsite TTS.

  • Music Generation (AI generated music): Under Construction.

    In the process of researching the improvement of the TTS I came across a means to generate pieces of music for the purposes of splicing them together in an audio editor. Early days for me here. Wondering if something bigger cant be done like a completed song as can be seen being done by googles gemini an maybe X's grok. So audio.cpp has a stable audio model which does both the voice cloning and the music generation. Stable audio comes itself free of audio.cpp but it may be an ideal voice cloner, its not an ideal music maker so the future if this gets done will be a differnt music maker to the one on the lab bench just now - stable audio.

  • Image Generation (Image Diffusion): Under Construction.

    Currenltly usind stable diffusion with a flux2 model. Its photorealistic and I guess my hardware lets me down for producing large images or streaming images.

  • Image Recognition (CV and OCR): Under Construction.

    To say this is anywhere near in prodcution would be a lie but there is early research into providing this as a service on the site.

  • Video Generation (AI generated video): Under Construction.

    If I want to compete with the big boys for AI then I have to produce this as a minimum, even in a short burst. I know that not all big tech AI will do this on the free tier but they will do it and it will become something more accessible in the future where there is an intent to provide video generation AI as a service on the site at some future point. The current hardware of the site could not support the generation of much of a video locally let alone remotely in quantity concurrently so theres a number of barriers to this at present, especially the price of VRAM and machien upgrades.

  • 3D Model Generation (video2objmat): Under Construction.

    I believe this will becoem a holy grail of computing once the whole AI thing takes hold. Current 3D model rendering is fairly photorealistic, except when it comes to humans and the living. Cartoonish isnt just a style, its a restriction on the current technology. At the same time photorealistic images are produced by AI easily. Photorealistic skins for photorealistic models that can have their mesh determined by a combination of guesswork and a series of images from a video walk round of an intended modelled subject. Holograph from a video ... and also maybe with rigging. All we need are the 3D goggles and the headphones and we wont know which worlds which until we fire the laser gun and accidently punch the monitor which doesnt exist in the ai-gl environment youve been living in, inside your goggled headset.

• Future Services Envisaged

  Experimental offline systems are under construction intended to provide website users with the following functions in future versions of the GDSWA (currently v15-26Q3).

  • AudioChat

    AudioChat has featured as a primary service of one previous version of the GSSWA. Not sure the version number but its in the single digits. Maybe 7. There was also a version close by where audio forums were the primary feature, although not so live as the actual chat version which is around those version numbers. The audiochat had some kind of tiny little bug that needed fixing although it was generally functional. Its intended to bring this back but while I am at it, I've decided I may as well cater also for video chat from the get go. In audio versions, thenceforce, the video part of the video chat would simply be switched off if audio is what is wanted.

  • The Recursion Game

    More about this area in the Repeat section.

  • VideoChat

    Some kind of VideoChat has been on cards for a while with me. I have did some video stuff just messing about but nothing other than embedding static videos. To embed multiple videos which are live onto one webpage is something which should now be quite easy for me to do, however I am having real OCD issues with screen layouts. I mean if I inist that people at least send portrait or landscape 16:9 or 9:16 videos I dont think it will be too much friction with visitors. How do I lay them out though? It depends on not only the size of the screen, but alos its shape. Portrait phones showing a list of landscape and portrait videos in so many rows is different from landscape desktops showing a list of landscape and portrait videos in a different number of rows and are the viewers having to squint their eyes to see so many rows on each and every device at every reasonable distance? Website design issues, or OCD?

  • SpaceChat

    Lets dispense with reality and play a game in outerspace. Instead of a list of videos or a list of chatters being so prominent, I can turn that on if I want to from a button in my virtual cockpit. Why do I need a list of whose chatting if I can hear them on the console as I use my screen as the view port for my space craft where I get to blow all of their space crafts to kingom come before they do it to me. That... and other games in outspace.

  • Search Engine

    Yeah, and eh! Whoa!!! Thats an enormous amount of information needed just to build an index thats like a searchable multi-dimenional array ... next Tuesday ... like I say every next Tuesday. Like Social Media, mentioned below, with advent of AI, search engines may be surplus to requirements in future.

  • Repeat

    This future project was referred to in the recursion and you should read this in relevance to what that section has to say.

  • Social Media Hub

    Taking advantage of a future auranos.org hub, customers would save alot of time and effort in doing laboious tasks like having to upload the same content to mutliple social media sites every time one creative product is created. Let auranos.org do the work for you. Just post it here and let us post it to all the accounts you wish it to be sent to. You may need to tell the social media platforms that we can post on your behalf whatever you wish to post through our website but once this is done we are set to go for good and we will talk you through the process for platforms where this is neccessary. Better still, let us set up an account on your behalf at a selection of social media platforms and we will manage it for you. We can even provide you with active bots willing to look and communicate your message online to the people you want your bots to reach. Remember, this is future plans, way in the future and the future of social media itself looks bleak. Theres some things in the works but not a lot for understandable reasons.

• Future Services Not Envisaged

  Any ideas? What do you want? Leave a message on the answering machine at the bottom of the page.

There are Different Flavours of Bots

The bots come in many flavours on this site. We have claire, who is not exactly a hostess with the mostest when it comes to the brains department. If your looking for rocket scientist to help your with your homework you can do no better than ... oh look theres a bunny rabbit. Isn't Claire pretty? Her face is a generative AI image built on the in house computers of auranos.org without the need for big tech AI corporations who provide their own remote AI image generators. I didn't have to make Claire and her friends dressed in astronaught costumes thinking they were in outer space.

Technically speaking, they dont "think". They, erm, perform multiple mathematical operations per second over a subset of multidimensional floating point numbers affected by a series of of these multidimensional floating point number sets in a novelly configured manners. If that sounds too mathematical for you, I will not bore you with the logic either but what is now known as AI, is quite distinct from the ancient sense of the subroutines by which the ghosts of pacman decide which way they will go.

In the modern sense, AI is referring to the invention of a transformer. A transformer does not physical exist. In a physical sense it is the configuration of whether or not electricity is flowing across the various intricate patterns of transistors embedded on a very small chip of silicon ingot which could fit on a thumb with billions of transistors embedded on its surface.

A miniature circuit board which is effectively dead until electricity flows through it but how it flows through that circuit board is where the software engineers art exists, in a non-physical form.

The transformer is a finite state machine having 1..* inputs and 1..* outputs. The internal design of a this finite state machine is subject to a lot of a options and design choices but the transformer can be said to have an interface of a series of outputs and a series of inputs. These inputs happen to be the same kind. They are both a series of Vectors.

There are Different Flavours of Transformers

A Vector in terms of AI is just like any other vector: One dimensional matrix of real numbers. Depending on the design of any particular transformer the width of the vector may be more than just for example and x and y axis but and x and y and z and q and r and s and t axis and perhaps as many than 8000 other dimensions as well as all of them and more. Its a very wide Vector also called a Tensor. A series of these Tensors is what goes into a Transformer and what comes out.

To be even more accurate, there are several different series of Tensors used in any one cycle through the finite state machine I descibe here as the software transformer. Those who engage in AI development fall into several categories. Those who press buttons on whatever hardware and software they have been given and those who write the software they are given. That software could be an interface that turns a website into a program that can communicate with the AI software. Now thats where I am. I dont write the AI software, I just build, configure it and make it available for use to the software I write for the website. When you are talking to Claire, or Tessa, or any of the other bots, in a kind of way you are going through what I have wrote so she can hear what you speak or type and so you can hear and see what she types back.

Claire Bares All

Now what Claire is, is what we are getting to, in terms of transformers but before we get down to the computational mathematics involvedin what Claire fundamentally is, lets get to the configuration layer and strip Claire bare. Her photo is AI generated by a specific pipeline we will ignore for now, because while a picture might tell a thousand words, the real Claire is not an image maker but a large language model.

A large language model is a stack of transformers arranged in a certain way together with embedding information which permits the translation of language tokens into these very long vectors mentioned earlier, known as Tensors. A model can also be seen as a file of information which contains information how these transformers are to be configured and how they are to be stacked, how many of them there are and what is initially contained in data known as the weights for any model. It is in the weights that the training data is stored, in the form of these tensors.

The Weights

The weights file is a series of tensors often contained in a file like .safe-tensors or .gguf. The .gguf file will contain additional information about the tensor data it contains. How many dimenstions does a big long Vector known as a Tensor have? It not only depends on the model but for each model, there is not neccessarily just one answer for each of the transformers in the block described in what I will call the .gguf file. It's an array of an array floating point numbers and those numbers in the digital world can be 80-bit (very rare) 64-bit (rare) 32-bit (sometimes), 16-bit (all the time), 8-bit (very often), 4-bit (almost as often) 2-bit (exceptionally rare). So these files can be huge in size or tiny and if cost is involved, go for small. Its cheaper to download although not neccessarily cheaper to run. 8-bit floats are known as Q8 and 4-bit floats are known as Q4. There are Q16 in a sense but FP16 is always the base standard so there is no actual Q for "quanization" involved. I guess there are rare circumstances where the bases are FP32 rather than FP16 and technically speaking the 16-bit model would appropriately be described as a Q16 model but 32-bit is rare and if you are doing it on the cheap, 8-bit seems to be the way to go because as the quanitzation reduces the bitsize, the intellect of the model exponentially decreases. It forgets things and mixes things up.

Neural Networks

You see its called "The Weights" but technically speaking its not just the weights which are contained in it, but also the biases. This is what the training data is composed of. Together these represent an artiifical neural synapse which will fire by software process like in a neuron and in accord with electochemical stimulus conditions modeled by the bias and the weight which decides just how much biased stimulus is required to fire. This prinicipal goes back to neural network technologies under development in the 90s and it was very laborious because to fire a billion neurons took a billion cycles to go through each neuron in the network based on the results of the previous neurons output.

Transformers, Neural Networks in disguise

In a completely different field of computer research of the 1990s was parallelism. Pipelining CPU instructions meant alomst two instructions could be carried out in one cycle and since the CPU as a machine is sos simple at this point it can be known that no sub pipelining could work to improve the number of operations per second that a CPU could do. Some multi-processor computers had been attempted in the past but they were all in need of software that supported the second CPU. The idea of a multi core processor was just an idea but it was in heavy research and part of the way there were things like SIMD where a single operation, although requiring an entire cycle to operate, could effectively carry out the same operation on multiple data items at the same time, in only one cycle.

Streaming SIMD became a big marketing term for the Pentium II processors where similar ideas used for MMX instruction on the Pentium II and Pentium MMX where now marketed as streaming SIMD. Not that it was any use to any software that didnt specifically cater to it. Software was begininning to cater to it and it was an architectural choice when multi-core GPUs were being used to do all the 3D work that once had to be performed by the software.

In the midst of highly performant GPUs and CPUs with this SIMD concept going on, someone working on neural networks has the Eureka! moment of pushing all the floating point work that can be done at the same time on any cycle of the neural network down a pipeline to be carried out in SIMD style.

Suddenly, highly intelligent slugs became like leopards overnight. The age of Generative AI has just begun.

Stephen Charles Frances Duffy (SCFD)

21:18 on Thursday the 12th of November 2026

If you found this article educational, please donate to auranos.org

If you checkout the launch button at the top of the screen you will see a comination of my efforts in webgl and 3D tile rasterization techniques using external api to provided rasterized tile data to use as the texture for the dynamically altering 3D tiled model earth. With a keyboard or some careful thumb work on a screen, you can navigate like a space ship and attempt to land on a deceptively fast moving rotating earth.

From out here the earth rotates an awful lot of miles per hour. Leave a message on the answering machine at the bottom of the page to let me know what you think.

Are you a business looking to integrate AI into your website or your company workflow? Get in touch by email or through the answering machine below. AI is enormously helpful in turning vast amounts of data into something which can be comprehended quickly by humans without having to spend hours or days organising the information. AI can do this in an instant and as an added bonus, they can do it with human like style.

Characters like C3P0, Kryten and Marvin the Paranoid Android are no longer science fiction but the cutting edge of computer science. Maybe you dont want some cartoon character from a movie but want a virtual colleague who can bring a bit of professionalism to a workplace or maybe even some wit and humour to an otherwise dull operation. It's not only the computational power of AI that makes them so alluring, its their psychological impact on their human colleagues.

Let auranos.org design the ideal personality for your employees, not that your employees are going to become robots anytime soon, but AI is now at a stage where it can have socialogical implecations to a workteam.

Contact us today for information and advice on integrations.

If you wish to contact the website, leave a message on the answering machine giving your name, your purpose and your email address. Leave such a message on this answering machine and you will receive a response.

If you are visiting auranos.org where the software of Stephen Duffy is showcased then contacting the website by this answering machine is the same thing as contacting me.

If you know me, leaving a message on this answering machine is as good as leaving a message on the answering machine of my phone with enhanced privacy and security. If you don't know me then I'm at the significant disadvantage of not knowing as much about you as you know about me from looking at this website, so let me know about you.

Email me at Stephen Duffy