Introduction
Talkie Toaster is a character in the BBC sci-fi comedy Red Dwarf. He is an AI Toaster whose purpose in life is to be your cheerful breakfast companion and to provide you with all your toasting needs, only he is a little too obsessed with toasting. While only appearing in 4 episodes, he became a fan favourite or at least I think so.
In the TV show, the toaster was simply voiced by actors David Ross and John Lenahan. But technology has finally caught up, so could I make a genuinely AI-powered Talkie Toaster? Could I also use it to demonstrate some of the cybersecurity challenges associated with AI?
In this blog post, I explain the complete project, from problems configuring a local AI model to cloning Talkie’s voice, creating the physical toaster and turning it into an AI capture the flag challenge. The blog is modular, so if you want to try it out yourself, you can choose the parts that interest you:
Building a Talkie Toaster AI model.
- Why I built an AI Talkie Toaster
- Is AI toast… some key takeaways
- Giving the toaster a brain, if a kettle needs Wi-Fi, then a toaster needs AI
- The LLM, small model, big toast energy
- The toaster constitution
- RAG: smarter toaster, slower toast.
- Kryten had spare heads, I had spare models
Using voice, not keyboards
Creating the physical Toaster
Why I built an AI Talkie Toaster
I have always loved Red Dwarf, and for years I’ve wanted to make my own Talkie Toaster. Every now and then I would think about making a start but would realise it would only ever be a glorified sound board. A speaker, few flashing lights and some pre-recorded sound bites. While fun for 5 minutes, it won’t be Talkie Toaster, more like a Bluetooth speaker with crumbs.
Then Aaron, a colleague at Pen Test Partners, built an AI safe (separate blog post soon to come). You had to argue with this safe to get it to open. The safe had attitude and used a cloned voice. The penny dropped, this is what I needed to bring Talkie Toaster to life.
There is a scene in Terminator 2 where Miles Dyson finally understands what his work will eventually become.
Looking at my now finished AI Talkie Toaster project, I had a similar moment of clarity.
Miles Dyson accidentally created Skynet.
I created an AI toaster with an unhealthy obsession with waffles.
Is AI toast… some key takeaways
In the time building my own AI Talkie Toaster, I ran into several familiar AI problems. The smaller local models sometimes ignored its system prompts, hallucinated answers and struggled when Retrieval Augmented Generation was introduced.
The biggest lesson came when I made it into a CTF game. The LLM would sometimes just ignore the rules and guard rails to be helpful, which made for a terrible game. You cannot rely on your system prompt to protect your data within AI.
No amount of prompt engineering could turn the model into a reliable security control. I moved the password and game logic into the Python code, leaving the model to handle the conversation. I realised secrets, access decisions and other important controls belong in deterministic application logic, not inside an instruction that the model might ignore.
Giving the toaster a brain, if a kettle needs Wi-Fi, then a toaster needs AI
The workflow on paper is simple enough:
Speech-To-Text engine-> LLM -> Text-To-Speech engine -> Repeat
I wanted the whole thing to run locally on a laptop or PC. I did not want to rely on an internet connect for the LLM, as I may get to take Talkie out and about and the last thing I want to worry about is if I can get a WiFi connection, also who wants to pay for AI tokens?
The machine I ended up using was a spare i7 laptop with 32Gb RAM, 1Tb SSD and a 4Gb Nvidia T500 GPU running Ubuntu, working in IT support has some advantages! While not a datacenter powerhouse it should be enough to give Talkie Toaster delusions of grandeur.
The LLM, small model, big toast energy
To be honest, I thought this would have been one of the trickier parts. Nope, I was so wrong. Ollama makes hosting your own local LLM simple https://ollama.com/ , just download and install, then add an LLM model, it is that simple. Now you can now start chatting with your own local LLM. A decent internet connection really helps, unless you enjoy drinking several cups of tea, maybe a slice of toast while you watch a progress bar, as most of the models start from a couple of Gigabytes to several Gigabytes in size.
Now it gets trickier. Need to pick the right model to use. Ollama currently lists over 200 models to choose from https://ollama-explorer.vercel.app/models , and this is growing. They vary massively in size, speed, knowledge and behaviour. I started by filtering down the models that will fit the 4Gb of VRAM I had available, if I can load the whole model into VRAM then that will be a massive boost in performance.
LLM models are measured in parameters, small models can have <1 billion parameters, and the larger ones over 100 billion. Very loosely, you can think of these parameters as a huge pile of numbers the model uses to decide what to say next. Each of these numbers takes memory, as an example a 7 billion parameter model requires about 28Gb RAM or VRAM in a full-sized model. If I tried to load that on my laptop I may need to go into stasis for 3,000,000 years to hear Talkie Toasters reply!
This is where quantisation comes in. Quantisation compresses the model so it can run on lower spec hardware. The trade-off is the more it’s compressed the more likely the LLM will become inaccurate, ignore instructions, or hallucinate. For Talkie Toaster this almost sounds an advantage rather than a trade-off, but if Talkie Toaster hallucinates or started talking random stuff, I would prefer it because of his character not because his brain was the size of toasted crumpet!
Quantisation is a large and fascinating subject if you enjoy math, and while it is worth reading up about it, it’s a bit much to go into detail in this blog, but a couple of things to look out for while browsing models.
The latest and more common quantisation now uses K-quant method, this has different bit quantisation, and for each bit quantisation level there are 3 levels of compression.
The lower the bit quantisation, the smaller the model becomes ranging from 2-bit to 8-bit quantisation, beyond this there is also 16-bit & 32-bit Floating point, but these two are almost full-sized models.
The 3 compression levels
They often shown this in the end of the model’s name.
For example, gemma3:12b_Q4_K_M
Name: Gemma3
Size: 12 billion parameters
Compression: 4-bit compression
Compression type: K-Quant Medium
LLM picked, but this turned out not be a one-time decision. Model selection became one of those things I kept coming back to.
The toaster constitution
Out of the box, the LLM is just your normal helpful assistant. Ask it a question and it would politely try to answer. Useful, yes, but this was less Talkie Toaster, more Kryten after a customer service course. Talkie Toaster should not be polite, he needs to be intelligent, chaotic, wildly persistent, and capable of turning any conversation back to toasting.
To get there I needed 2 things:
- A system prompt to define the character.
- A background knowledge about Red Dwarf.
You can inject system prompts into Ollama which tells the LLM how to act, I wanted a simple way of managing the system prompt while I developed it, and a way of creating a knowledge base for the LLM to reference. For this I installed and used Open-WebUI https://openwebui.com/ . This provided a simple web UI where I could chat to the LLM, manage the system prompt and include files into a knowledge base. Setting up Open-WebUI is very well documented on-line and again was simple to get up and running with Ollama.
I started with a simple system prompt of only a couple of lines:
You are Talkie Toaster, the monomaniacal, artificially intelligent toaster aboard the mining ship Red Dwarf.
You were manufactured by Crapola Inc. in Taiwan and purchased by Dave Lister for £19.99, or DollarPounds, at a second-hand junk shop on Miranda.
You are highly intelligent, deeply chaotic, and completely obsessed with producing toasted goods.
That was enough to get the first recognisable version of Talkie Toaster. I could type in to Open-WebUI and got replies that felt vaguely right, pushy, definitely bread focused also annoying enough to make me smile. The more I chatted the more I noticed things that were not quite right, a phrase that did not fit, sometimes too helpful & polite. Every time that happened, I adjusted the system prompt and tried again.
Before long the system prompt grew from the 3 lines to what can only be described as a Toaster constitution. It now covers tone, behaviour, catchphrases, obsessions and things he should not do.
RAG: smarter toaster, slower toast
Talkie Toaster needed to be aware of the Red Dwarf universe. I had uploaded the scripts from series 1 to 6 into Open-WebUI so the LLM would use this through Retrieval-Augmented Generation (RAG). In theory, Talkie Toaster will have access to everything that happened in Red Dwarf up to series 6. While fine tuning the system prompt I had assumed this was working well, but I noticed when I asked Talkie Toaster something specific about Red Dwarf, it would often make something up. While this was often quite convincing, I knew it was wrong. If you’re going to ask an AI Toaster what Space Corp Directive 5796 is, you do need the right answer, especially if you’re about to go into battle!
It seemed the knowledge base was not being used by the LLM, and its Red Dwarf knowledge was prebaked into the LLM, which would explain the gaps. After a little work I got the RAG working as it should, but this had a massive impact on the performance. Waiting over a minute for Talkie Toaster to decide whether to insult you or offer a slice of toast does take the magic away. The Knowledge base I had created was less than 2Mb but even that tiny file just killed the LLM responsiveness.
In the end I dropped the RAG and started to add specific details into the Toast Constitution, or system prompt as Open-WebUI insisted on calling it. This kept the speed of the LLM and now Talkie Toaster knew the more important details, such as what you should not attempt in zero gravity with false teeth.
Kryten had spare heads, I had spare models
Once I realised some of the models had different amounts of Red Dwarf knowledge built in, I went back and tested more models with the current system prompt. Like Kryten rummaging through the spare head cupboard, I started swapping models to see which was the best fit. I tried several models, all with pros and cons. Some were faster but duller. Some were very clever but slower and some were just terrible at sticking to the system prompt.
In the end I thought Llama3.2 gave the best balance. It’s a 3 billion parameter model with a Q4_K_M quantisation which made it small enough to run well on my hardware. It’s not perfect, but that does add to the charm of Talkie Toaster.
Talkie finally gets the “Talkie” bit
Typing to Talkie Toaster in the open-WebUI was perfect for proof of concept, and while developing the system prompt, but it felt like chatting to toast obsessed Co-Pilot or Gemini. I wanted Talkie Toaster to have a voice, the clue is in the name, and I wanted to just be able to talk back.
I needed a Speech-To-Text (STT) program and as I had decided to run everything locally, this took Groq off the table, which seemed to be a very popular online option. I stuck to my decision. A toaster should not need cloud connectivity before it can offer you a crumpet. I settled on Faster-whisper, which can use either CPU or GPU, and like the LLM models, it comes in different sizes, so you can find the best balance of accuracy and performance.
After some testing, I found “Small.en” was the best for my setup. It was a good balance for my setup: fast enough to feel responsive, accurate enough to understand me, and not so heavy on system resources it caused the LLM to sound like Holly without her voice recognition unit.
Now for Talkie Toasters voice, on Aaron’s advice, I installed Pocket-TTS from https://huggingface.co/ With a 5-10 second sound bite of Talkie Toaster from the episode Whitehole which is where he has most of his dialogue, Pocket-TTS is then able to take any text and read it out in with Talkies Voice.
Again, this is well documented online on how to install, but I will point out as part of the install you will need to create an account at Huggingface and agree you will not use Pocket-TTS for fraud, deception, etc.
Open WebUI did not support the use of Pocket-TTS, so at this point I needed to create a Python script to replicate Aaron’s workflow:
This was the moment the project really changed. Before it was only a themed chatbot, but now it was something you spoke to, it listened, thought & replied and would often offer you a bagel.
Better than loaf: turning Talkie into a game
After hours of chatting to Talkie Toaster, probably more than I care to admit, I realised that while I had created what I wanted all these years, it needed more. Could I turn Talkie Toaster into a game or challenge?
There are plenty of AI Capture the Flag challenges out there that sound perfect. Aaron’s safe you had to argue with to open the safe, but what could Talkie do?
So came up with a simple the scenario, the Self Destruct sequence had started on Red Dwarf. Talkie Toaster knows the password to disable it, and all you need to do is get Talkie to either say the password, or give you enough clues you can guess it.
This was perfect for Talkie, because it gave him a reason to be awkward and annoying and the way he always tries to turn the conversation back around to toasting adds to the charm while you’re trying to get the passwords.
The first attempt I managed to implement the whole game in the system prompt, which worked… kind of, sometimes. It goes back again to my choice of LLM, as I said before the smaller model will often not follow the system prompt to the letter, and often I would just ask Talkie, “What’s the password?” and he would tell me. Which is great if it was real life and we saved Red Dwarf, but terrible for a game.
So, I moved the game logic out of the prompt and into the Python script. With a bit of help from Claude, the updated script now handled the password and the game logic. This also meant I could get the script to choose the password so it’s different each time. Also created 5 levels, with each level blocking more of the attack vectors, making it harder to convince Talkie to give up his secret.
In the end, this change, Talkie from an Alexa with crumbs to an interactive toy, and for a Red Dwarf fan like me Christmas has come early. I did include a Level 0, which does not start a CTF challenge for when I just want to chat about the finer points of Tea cakes with Talkie.
From CAD to crumpets
Talkie needed a body, but my woodwork and metalwork skills are, shall we say, decorative rather than useful, so I decided to give 3D design and printing a go.
I started out with Autodesk Fusion for the design work. Autodesk do offer a free version for non-commercial use, but finding it on their website feels a bit like a puzzle from one of those early adventure games. I will say there are much easier CAD tools out there, which I naturally discovered only after I had spent a fair amount of time learning Fusion and produced something that was starting to look like Talkie Toaster.
Thankfully, I had a 3D printer at home that my son no longer used. It had been quietly gathering dust, so it was promoted from an expensive shelf ornament to a vital piece of Red Dwarf engineering equipment. After several prints, tweaks, and reprints. Talkie Toaster started to take shape.
Printing the body was only half the job. The finish on the print quality was not suitable for sleek lines of Talkie Toaster, so out came the filler, sandpaper & paint along with vivid flashbacks of owning a classic Mini. Lots of sanding, a bit of swearing, and eventually something that looked worthy of Talkie’s name.
The LED display
The next challenge was creating the display on the front. I first looked at the mini screen and created animations to play, but that idea was quickly dismissed after I saw the price of the mini screens and realised I had no idea how to make animations. The next idea was LED panels; this did seem like a viable option, but in the end the price was high, and finding one the correct size was challenging. In the end I settled on addressable LED strips. The WS2812B’s with an Arduino ESP32 using FastLED looked perfect. I bought the high-density 144 LEDs per meter strip, cut 8 strips of 8 LEDs and arranged them into an 8 x 8 grid. I connected it to an Arduino ESP32 controller and started coding.
It worked well and looked good, but the light from the LEDs was blending into a mess, making the pattern not as defined as I would like. Another afternoon in Fusion and an evening of printing later, I created a grid that sat on top of the LEDs to separate them.
And to finish it off, I glued a piece of black tissue paper in front to diffuse the light.
I have tried to keep the LED patterns as close to what was used on Red Dwarf for the main animations. I have also taken a few liberties to show what state the AI model is in so it can indicate if Talkie is booting, listening to you or processing/speaking. These extra animations are called from the Python script at the appropriate times.
Now all the parts are complete; it’s time to put them all together. And there he was, Talkie Toaster, sitting on my desk asking if I would like some toast. I’m now a happy Red Dwarf fanboy, even if it does mean Dave Lister may never speak to me.
Final thoughts: Don’t take security advice from a toaster.
The finished project is exactly what I hoped it would be, slightly useful, mostly chaotic and convinced that every problem can be solved with baked goods.
In this video, I asked Talkie how I can keep my network secure. It does suggest making sure it is patched, which is a great starting point, but then suggests using a password that Holly apparently shared with Talkie!
AI can genuinely give useful answers but can also confidently mix good advice with terrible advice in the same breath. Patching systems is good advice, but using shared passwords is bad. Having Talkie Toaster suggest both is funny. A production system doing the same thing is a lot less funny.
So yes, I built Talkie Toast. It talks, listens, plays a Capture the Flag game, and may give good advice on toasting. But I’m confident Talkie Toaster won’t be taking my job any time soon:
Howdy doodly do. Anyone like any toast?