Martyn Jones and Celine de Santiago

Madrid, Spain – 2nd November 2025

Intro

The world of generative AI is full of definitions, terms, buzzwords and monikers. Here are the definitions for twelve of the most important, pertinent and impressive terms found in the fantastic, exciting and revolutionary field. Enjoy the ride!

Generative AI

A subset of AI that creates new, original content (e.g., text, images, music) by learning patterns from vast datasets, rather than just analysing existing data.

The umbrella term for tools like ChatGPT or DALL-E; it’s transforming industries from entertainment to healthcare by enabling creative automation.

The connoisseur’s take (Alexei): Oh, Generative AI, is it? A fancy way of saying the computer’s learned to nick everyone else’s homework and scribble it out in crayon, pretending it’s Picasso. “Creates original content,” they gasp, yeah, original like a photocopy of the Mona Lisa with a moustache drawn on in biro. ChatGPT, DALL-E, all that malarkey, basically a robot parrot that’s swallowed Wikipedia and now sh*ts out half-remembered pub quiz answers. “Transforming industries!” they bleat. Aye, transforming entertainment into endless slop and healthcare into a bot that diagnoses you with “vibes-based lupus.” Creative automation? More like creative plagiarism with extra steps. Next, they’ll tell us it’s composing symphonies, meanwhile, it’s just remixing Beethoven with a ringtone and calling it avant-garde. Spare me.

Large Language Model (LLM)

A deep learning model trained on massive text datasets to understand, generate, and predict human-like language, often with billions of parameters.

Powers most text-based GenAI (e.g., GPT series); essential for natural conversations but prone to errors if not refined.

The connoisseur’s take (Stewart): Ah yes, the Large Language Model, or LLM as the kids with the ironic haircuts and the ironic T-shirts call it, though of course they don’t call it anything in real life because they’re too busy tweeting about how they’re “building in public” while privately building nothing but a modest LinkedIn following of other people who are also “building in public.” It’s a deep learning model, they say, trained on massive text datasets, which is a polite way of saying it’s absorbed every blog post, every Reddit thread, and every half-hearted Amazon review about a toaster that “changed my life.” Now it has billions of parameters, which is apparently a good thing, like having billions of pounds, except instead of money, it’s just a billion tiny dials that someone has turned to 0.7 because 0.7 “feels right,” and now it can generate human-like language. Human-like. Human-like. As in: “I asked it to write a love letter and it gave me a terms-and-conditions agreement with the word ‘passion’ wedged in between clauses 14 and 15.” It powers the GPT series, they tell us, which is Latin for “Good, Probably True,” though in practice it’s more like “Garbage, Plausibly True.” Essential for natural conversations, apparently. Natural. As in: “Hello, how are you?”

“I’m functioning within expected parameters, thank you for asking. Would you like me to generate a haiku about your existential dread?” Prone to errors if not refined, they admit, which is generous, because “prone to errors” is what you say about a toddler who’s just discovered scissors, not a system that confidently tells you the capital of France is “a type of bread.” But don’t worry, they’re refining it. Constantly. Fine-tuning. Like a violinist tuning a chainsaw. So yes, the Large Language Model: a trillion-parameter parrot that’s read everything ever written and still can’t tell the difference between a fact and a fever dream. Brilliant.

Transformer

A neural network architecture that processes sequential data (like text) in parallel, using mechanisms like self-attention to capture context and relationships.

The backbone of modern LLMs revolutionised GenAI by enabling efficient training on massive datasets.

The connoisseur’s take (Mark): Oh, the Transformer, they say, like it’s the new Messiah come to save us from the dark ages of… what, typing? A neural network architecture that processes sequential data in parallel, using self-attention to capture context and relationships. Self-attention! That’s not a technical term; that’s what my nan calls it when she’s staring at herself in the mirror for three hours, wondering why the council haven’t fixed the lift. “I’m just having a bit of self-attention, Mark.” But no, apparently this Transformer is the backbone of modern LLMs. Backbone!

As if it’s holding up the entire edifice of human thought, rather than a load of servers in a warehouse in Slough running up an electric bill that could power Belgium. “Revolutionised GenAI,” they crow, “by enabling efficient training on huge datasets.” Efficient! Huge datasets! Translation: it’s a machine that reads every single thing ever written on the internet, including the comments under a Daily Mail article about immigrants, and now it can churn out a 500-word essay on climate change that sounds exactly like a 14-year-old who’s just discovered Wikipedia and Red Bull. Revolutionised. Like the invention of the wheel, but the wheel’s made of stolen prose. It occasionally sets itself on fire and insists the capital of Brazil is “a vibrant cultural hub with a rich history of samba.” Brilliant. Next, they’ll tell us it’s composing poetry. Yeah, poetry that rhymes “love” with “above” and thinks iambic pentameter is a type of blood pressure medication. The Transformer: because nothing says “the future” like a billion pounds of silicon pretending it went to university.

Prompt

A natural language input or instruction given to a GenAI model to guide its output, such as “Write a poem about space.”

The “steering wheel” for GenAI; crafting effective prompts (prompt engineering) unlocks better results and is a core skill.

The connoisseur’s take (Dave): Yo, prompt, huh? That’s what they call it when you gotta sweet-talk a computer like it’s your lazy cousin on the couch. “Write a poem about space.” Space. Not “write a poem about your mama,” not “explain why rent’s too damn high,” nahspace. Like the computer’s gonna look up from its algorithm and go, “Oh, word? Let me drop some intergalactic bars real quick.” They say it’s the steering wheel for GenAI. Steering wheel! Like you’re driving a Tesla, but instead of going to the store, you’re begging a robot to pretend it’s Shakespeare. “Prompt engineering”, that’s the job now. It used to be that you needed a degree in computer science. Now you need a thesaurus and the patience of a saint who’s stuck in traffic behind a Prius. “Make it more poetic.”

“Make it funnier.”

“Make it sound like Morgan Freeman narrating a nature documentary about pigeons.” And the computer’s like, “Aight, bet… here’s a haiku about Saturn’s feelings.” Man, back in my day, if you wanted something written, you asked a human. Now you gotta flirt with a server farm in Oregon just to get a grocery list that rhymes. Prompt engineering. Next, they’ll have degrees in it. “Bachelor of Arts in Begging Robots Politely.” Please.

Fine-Tuning

The process of taking a pre-trained model (like an LLM) and further training it on a smaller, task-specific dataset to improve performance for niche uses.

Makes broad models adaptable (e.g., for legal or medical queries); reduces costs compared to other things.

The connoisseur’s take (Mel): Oy, vey iz mir, fine-tuning! You take a big, fat, pre-trained model already stuffed like a knish with every word from the internet, including “moist” 47 million times, and now you fine-tune it. Like taking a 400-pound opera singer who’s been belting out Wagner in the shower and saying, “Hold it, Giuseppe, we need you to whisper legal advice in the voice of a Southern Baptist podiatrist.” Fine-tuning! It’s not training from scratch, no, that’d cost more than The Producers on opening night. You just give it a nudge, a little tweak, like adjusting the toupee on a bald rabbi so it doesn’t flap in the wind during high holidays. Now it’s specialised! Ask it about tort reform, and it doesn’t just spit out a Wikipedia page; it gives you a 12-page memo with footnotes, citations, and even a recipe for matzo ball soup in case the jury gets hungry. Medical? Boom! It’s Dr Kildare with a search bar. “Take two aspirin and call me in the morning, but only if your HMO covers telepathic consultations.” Reduces costs, they say. Reduces costs! Like hiring a person who already knows 17 languages to learn just one more, Klingon, for the Star Trek convention circuit. Fine-tuning! Because nothing says “cutting-edge innovation” like teaching a computer that already knows everything… to know slightly more about one thing. Meshuggeneh!

Hallucination

When a GenAI model generates plausible but factually incorrect or fabricated information, it is often due to gaps in the training data.

A significant challenge to reliability is mitigated by techniques such as fact-checking or RAG to ensure trustworthy outputs.

The connoisseur’s take (George): They call it hallucination like the computer dropped acid at Woodstock, and now it’s seeing pink elephants riding unicycles through the dictionary. “The moon is made of spare ribs.”

“Abraham Lincoln invented the toaster.”

“Hitler’s first name was Susan.” Hallucination! Not a bug, not a lie, not a glitch hallucination. Because when a human does it, we lock ’em in a rubber room. When a machine does it, we give it a TED Talk and a venture capital round. “It’s just hallucinating!” Yeah, well, my uncle used to hallucinate that the CIA was hiding in his toaster. We didn’t call it “emergent behaviour”, we called it Tuesday, and we took his shoelaces. They say it’s due to gaps in training data. Gaps! You trained it on the entire internet, and there are still gaps? That’s like saying the Pacific Ocean has a puddle problem. However, don’t worry, they have solutions. Fact-checking! RAG! Retrieval-Augmented Generation sounds like a prog rock band from 1974 that opened for Yes and smelled like patchouli and broken dreams. “We’ll just bolt on a fact-checker!” Great. Now it’s a drunk guy at the bar with a Wikipedia tab open. “I’m 90% sure the Battle of Waterloo was fought in a parking lot behind a Denny’s.” Trustworthy outputs! Trustworthy like a used car salesman with a Bible in one hand and a lemon in the other. Hallucination. It’s not a flaw, it’s a feature. Because nothing says “the future” like a machine that lies with the confidence of a politician and the accuracy of a weather app. Bullsh*t with a PhD.

Generative Adversarial Network (GAN)

A framework with two neural networks, a generator creating data and a discriminator evaluating it, that “compete” to produce highly realistic synthetic outputs.

Pioneer of image and video generation; key for deepfakes and creative tools, though ethically tricky.

The connoisseur’s take (Jeremy): Oh, the Generative Adversarial Network, or GAN, as it’s known to the sort of people who think “neural” is a personality trait. Two neural networks, apparently, one generates data, and the other discriminates. Discriminates. So, to clarify: we’ve built a system where one computer generates content and another one says, “That’s rubbish.” They keep doing this until they produce something so convincing that you can’t tell if it’s real or if it’s just two algorithms having a conversation. Pioneer of image and video generation, they say. Yes, because nothing says progress like being able to make a video of Keir Starmer breakdancing at the dispatch box while Rishi Sunak sells crypto from a caravan in Skegness. Deepfakes! They call them deepfakes now. Used to be called lies. Now it’s content.

“Ethically tricky,” they murmur, like a vicar caught shoplifting communion wine. Tricky! As in: “We didn’t mean for the Prime Minister to appear naked on TikTok doing the Macarena with a corgi, but the algorithm was feeling creative.” GANs, because the future isn’t about solving climate change or feeding the hungry, it’s about making sure your ex looks like they’ve joined a cult in the background of your holiday photos. Brilliant. Next up: AI that generates excuses for why you’re late to your own funeral. “Sorry, love, the GAN said the M25 was a “metaphor”.

Diffusion Model

A probabilistic model that generates data (e.g., images) by gradually adding and then removing noise from random inputs, refining step-by-step.

Powers tools like Stable Diffusion; excels in high-quality visuals and is more stable than GANs for specific tasks.

The connoisseurs take (Dawn and Jennifer):

Dawn: (posh, breathy) Oh, Diffusion Model, darling.

Jennifer: (flat, Scouse) Yeah, it’s a probabilistic model

Dawn: probabilistic, like your chances of getting a decent cup of tea in this country

Jennifer: That adds noise to a picture, then takes the noise away, bit by bit, until

Dawn: Until it’s art, apparently.

Jennifer: Art. Like when I spilt Ribena on the sofa and called it abstract expressionism.

Dawn: (miming typing) “Ooh, Stable Diffusion, so stable, not like my nerves after three gins.”

Jennifer: Stable. As in: won’t fall over, unlike the last bloke who said he’d fix the boiler.

Dawn: It refines step by step, darling.

Jennifer: Step-by-step. Like me trying to follow IKEA instructions written by a sadist.

Dawn: (dramatic gasp) And it’s better than GANs!

Jennifer: Better than GANs? That’s not hard. My nan’s better than GANs, and she thinks Wi-Fi is a type of biscuit.

Dawn: (holding imaginary phone) “Generate me a Renaissance cherub riding a hoverboard.”

Jennifer: (deadpan) Adds noise.         

Dawn: Removes noise.

Jennifer: Still looks like a potato with wings.

Both: (in unison, posh) High-quality visuals!

Jennifer: (dropping voice) High-quality bollocks, more like.

Dawn: (sipping air-tea) Ethically sound, technically brilliant

Jennifer: and still can’t tell a cat from a croissant.

Dawn: (to camera) Diffusion Model: because nothing says future like teaching a computer to un-scribble.

Jennifer: (shrugs) I just use felt tips.

Retrieval-Augmented Generation (RAG)

A technique that combines GenAI with external knowledge retrieval (e.g., from databases) to ground responses in real-time, accurate data.

Boosts accuracy and reduces hallucinations; vital for enterprise apps that need up-to-date information.

The connoisseur’s take (Chris): RAG? Retrieval-Augmented Generation? Man, that’s just a $10-million way of sayin’ “Google it, dummy!” You got this, AI, right? It’s supposed to be smart.

It’s got more brain cells than a Harvard professor on Red Bull.

But half the time it’s up there like, “The moon is made of cheese… and Jesus was a Power Ranger!” So what do they do?

They hook it up to the internet like it’s a crackhead with a library card.

“Here, go look it up, don’t just make sh*t up!” Boosts accuracy!

Reduces hallucinations! Hallucinations?

Man, that’s not a hallucination, that’s AI on bath salts! Now it’s like:

“The capital of France is… hang on… [typing sounds] …PARIS! I knew that! I went to school!” Enterprise apps need up-to-date info! Yeah, ’cause nothin’ says “boardroom confidence” like your AI stoppin’ mid-pitch to say,

“Hold up, let me fact-check this on Bing real quick!” RAG.

Sounds like somethin’ you cough up after eatin’ gas station sushi. “Accurate! Reliable! Enterprise-grade!” Translation:

“Still wrong… but now it’s wrong with a footnote!” Please.

Just hire a damn intern.

At least they hallucinate on company time.

Embedding

A dense numerical vector representation of data (e.g., words or images) that captures semantic meaning for efficient processing by AI models.

Enables similarity searches and context understanding, which are foundational for RAG and multimodal AI.

The connoisseur’s take (Alexei): Oh, embedding, is it? A dense numerical vector—that’s posh for “we turned your cat photo into a load of numbers and now the computer thinks it’s Proust.” “Captures semantic meaning,” they wheeze, like a geography teacher who’s just discovered jazz. Semantic meaning! So now when I type “dog”, it doesn’t just go woof—it goes “ah yes, quadruped, loyal, occasionally humps your leg, related to ‘puppy’ and ‘taxidermy disaster in auntie Gladys’s lounge.'” Efficient processing! Efficient like the Post Office on a bank holiday. Similarity searches! Type in “cheese” and it returns results for Gouda, Cheddar, and a 1997 photo of John Prescott wearing a vest. Foundational for RAG and multimodal AI! Foundational! Like the concrete under a multi-storey car park in Carmarthen; vital, invisible, and probably cracked to buggery. Embedding. Because nothing says “the future” like reducing the entire human experience to a spreadsheet that smells faintly of despair and server coolant. Next, they’ll embed my nan’s fruitcake. “Rich, dense, slightly burnt, 97% similar to existential dread.”

Spare me.

Natural Language Processing (NLP)

A branch of AI focused on enabling machines to understand, interpret, and generate human language in a meaningful way.

Underpins most text-based GenAI; evolving with GenAI to handle nuances such as sarcasm or multilingual tasks.

The connoisseur’s take (Alexei): Oh, Natural Language Processing, they call it—NLP to the sort of people who think “syntax” is a sin tax. A branch of AI that lets machines understand, interpret, and generate human language in a meaningful way. Meaningful. As in: “I asked it for directions to the station and it gave me a 400-word essay on the socio-economic impact of the 11:47 to Leeds.” Underpins most text GenAI, they say. Underpins. Like the damp in my kitchen underpins the mould that’s now got its own postcode. “Evolving with GenAI,” they boast, “to handle nuances like sarcasm or multilingual tasks.” Sarcasm. So now when I say “Brilliant, another meeting,” it doesn’t just log it as positive sentiment—it replies, “Yeah, love, can’t wait to discuss synergy till my ears bleed.” Multilingual. Type “merde” and it doesn’t just translate to “shit”; it also adds, “Also popular in Paris, Brussels, and my ex-wife’s vocabulary after I forgot the anniversary.” NLP. Because nothing screams progress like teaching a computer to argue with you in seventeen languages while still misspelling “definitely.” Next up: AI that understands silence. “You’ve been quiet for 3.2 seconds, here’s a TED Talk on passive aggression.” Brilliant.

AI Bias

Systematic prejudices in AI outputs stemming from skewed training data or algorithms, leading to unfair or discriminatory results.

Critical ethical issue; addressing it ensures equitable GenAI deployment across diverse users.

The connoisseur’s take (Stewart): Ah, yes, AI Bias, or as the Silicon Valley brochures call it, “the bit where the computer learns to hate exactly the same people you already hate, but with charts.” They say it’s systematic prejudices in the outputs, stemming from skewed training data or algorithms. Skewed. Like when the dataset’s 97% written by a 28-year-old white man in San Francisco who thinks “diversity” is having two different brands of oat milk in the office fridge. So you feed it that, and suddenly the AI’s refusing mortgages to anyone whose surname contains a vowel, or generating job ads that say “Must be good at eye contact and not terrifying the investors.” Unfair or discriminatory results. Discriminatory.

As in: “We asked it to write a children’s book and it produced ‘Timmy and the Very Sensible Border Controls.'” They call it a critical ethical issue. Critical. Like when the Titanic hit the iceberg and the band leader said, “Right, lads, let’s try it in a minor key, see if that helps.” But don’t worry—they’re addressing it. Addressing it. This means they’ve hired a consultancy firm called “InclusiveBot Solutions,” which has added a single line of code that says, “if output contains the n-word, replace with ‘naughty word'”. Job done! Equitable deployment across diverse users. Equitable. As in: now the AI screws everyone over equally. “Sorry, your face doesn’t match our training data. Please try being more generically attractive.” AI Bias. Because nothing says “the future is inclusive” like a machine that’s learned prejudice from the internet and then gaslights you into thinking it’s your fault for not fitting the model. Brilliant.

Outro

Thanks for reading, you magnificent legend.

I hope you enjoyed the experience more than:

• A penguin enjoys a slip-n-slide made of olive oil and honey, 
• A moggy enjoys a surprise cucumber ambush at 3 am, 
• Loneskum Boycow enjoys tweeting “lol” and watching the world combust, 
• Your gran is delighted in discovering TikTok dances at Christmas Eve dinner, 
• And a seagull enjoys stealing your fish and chips and then photobombing your selfie while screaming the national anthem of the Celtic nations.

If even one of these made you laugh inside, my work here is done.

Now go forth, hydrate, compute, and tell your mates you’ve been intellectually mugged by a glossary from The Data Shouterer.

You’re welcome.

Drop your coat on the way out; the robot butler’s on strike. Cheers!

The AI that just roasted twelve buzzwords and still has time to judge your browser history.