Introducing the world-changing Vexis, our first-ever AI programming language
If you read the headline, I want you to imagine I am Derren Brown. I have a brown envelope. Inside, is a ransom note.
I HAVE YOUR CAT.
Back to the mission.
Before we begin, I was gonna write ALL ABOUT MCPs today. I had it scheduled for two weeks. And then all the AInfluencers started going on about it and I felt resolutely defiant.
So I decided to put my thinking cap on. And what ensued was spectacular.
And then I forgot what it was.
MCPs ARE BETTER THAN SEX!
Shut up. Function calling FTW!
Parallel processing will get you there.
Classical computing eats ass. Quantum computing is lit.
I think all I did back there is show just how out of touch I am with the street.
While making what I believe is a somewhat reasonable point:
We have never been more divided in how to get the job done, now AI is in the loop.
Do you remember that phrase about everything looking like a nail, when all you have is a hammer?
Or when everyone in recruitment apologised for shilling applicant tracking systems after they were sussed out for the fundamental, fatal flaws in sifting the wrong candidates in and out.
I don’t know whether that is relevant here, but by the time we reached the bit where you clap and add a comment, we should at least feel more informed.
What I may be attempting to explain, is that AI will never deliver complete, test-perfect codebases while we are thrusting human-pleasing programming languages into its face asking for miracles.
We’re now at that thing they call an inflection point.?
AI’s biggest strength is pattern deducing and matching. So why do we allow Jedward to continue shrieking in nightclubs when we can’t be arsed to give AI the keys to build its own programming language?
I haven’t really figured out who’s in charge of creating AI-native syntax.
I know after watching The Thinking Machine (recommended) that algorithms or something can probably figure things out on their own, using reinforcement learning and some other neural network learning thing, There was a sequence in that film where it just watched Tom Daly diving and then pulled off the perfect leap from the 15-metre board before realising that its central processing unit wasn’t waterproof. I think.
If it’s gonna have its own language, and since it’s clearly capable of creating a PRD, AI should absolutely do all the heavy lifting here.
Does this seem like the right time for a grown-up conversation about an AI-first programming language?
You fucking betcha.
Caveat: I’m just an interested idiot. You’ll already know that because you read some of my other stuff.
So I’m only going to cover things I understand. It’s gonna be primitive. But that’s also where your capabilities end, so I guess you could say I did my user research to find a level where we could John and Yoko and under this endless duvet of duck feathers and dreams, conspire on a world where UBI is everywhere and all you have to do is watch cat videos all damn day.
There’s a lot in here, surprisingly. A lot of GREAT stuff. Indicating that all too soon we can stop doing things the way we’ve always done them. We need a catalyst in helping us humans stop trying to bodge AI into existing processes. And I believe an AI-first programming language is crucial in creating entirely new and unimaginable paradigms unlocking unimaginable progress. For you, and for me, and the entire human race.
So here’s your diet for today:
One thing I do know is an AI-first programming language would be a completely new way of creating software, designed specifically for systems rather than sentients.
Its main purpose isn’t to be readable by people, but to let AI systems easily understand, optimise, and execute code efficiently.
By the time you’re done here, those nails won’t seem so fancy any more. We’ll all have crowbars, scythes, ploughs, and scalpels?—?and THE WORLD WILL BE OURS.
What shall we call our AI-first language?
I gave you a clue earlier. Don’t scroll back. You didn’t read it, did you? If Anneka Rice needs a replacement for Treasure Hunt, you ain’t even making the shortlist (because they still haven’t fixed those fucking applicant tracking systems).
Before we get started, let’s have a Design Jam (because it’s way more fun than being productive) to choose a fitting name celebrating the technical complexities and vector-based syntax that might propel this crazy invention of ours.
Top 5 contenders
And the winner?is…
Vexis!
Why? Because I wrote the headline and no WAY am I changing it to Qrypt.
The name visually distinguishes human-readable intent declarations (?VEX? tags) from the unintelligible vector sequences that follow?—?a perfect encapsulation of the language’s human/AI duality. And:
2. Psychological resonance
The word “vex” both sounds like sex (moist!) and immediately conveys:
3. Memorable branding
4. Scalable metaphor
Supports natural derivatives:
Theoretical documentation example
# Traditional Python
def calculate_risk():
# 50 lines of manual logic
# Vexis equivalent
?VEX?
INTENT: FinancialRiskModel
CONSTRAINTS: SEC-Compliant ∧ RealtimeExecution
VEXCODE: [0.12, -0.45, 0.78...] (1000d vector)
How might an AI-first language?work?
Since here we’ve dispensed of humans (in the figurative sense), we can bin off using words or symbols we recognise (like Python or JavaScript).
It’s time to celebrate abstract mathematical symbols, or vector representations.
A special “machine language”, designed for how AI naturally processes information.
A simple program like adding two numbers (result = 3 + 5) might look something like this:
[
[0.12, -0.34, 0.56, ...], // represents "load number 3"
[0.85, -0.23, 0.45, ...], // represents "load number 5"
[0.33, -0.44, 0.67, ...], // abstract representation of "add numbers"
[0.05, 0.88, -0.23, ...] // abstract representation of "store result"
]
If those numbers mean something to you, hello Paul Robin Krugman!
Otherwise, they’re vectors that only the AI understands?—?because it has learned their meaning through training on many examples.
Let’s just stretch this a little further into the realm of variables because that way we can introduce the idea of a language that can be parsed, or interpreted/translated, into human:
[
[0.4, 0.1, -0.2, ..., 0.7], // vec('LOAD_CONST'), vec(3)
[0.8, -0.1, 0.3, ..., 0.5], // vec('STORE_VAR'), vec('a')
[0.2, 0.6, -0.4, ..., 0.9], // vec('LOAD_CONST'), vec(5)
[0.7, 0.0, 0.2, ..., -0.3], // vec('STORE_VAR'), vec('b')
[0.9, -0.5, 0.1, ..., 0.4], // vec('LOAD_VAR'), vec('a')
[0.3, 0.2, -0.6, ..., 0.8], // vec('LOAD_VAR'), vec('b')
[0.1, -0.3, 0.5, ..., 0.2], // vec('ADD')
[0.6, 0.4, -0.1, ..., 0.7] // vec('STORE_VAR'), vec('result')
]
Isn’t that genuinely beautiful? All those vectors. And all that meaning. It’s way too good for you, that much I do know.
One way to interpret Vexis into our world might be to use pseudocode:
a = 3
b = 5
result = a + b
But because this is all hypothetical, and there is nothing you or I can do to stop Vexis from happening and making all the rules, at this stage I can give you no firm promises what the hell happens next.
Before we move into the bit that normal people understand, let’s take a waltz down Complex Street.
Say you wanted to calculate a factorial.
Well look who it is. It’s the sexy snake syntax!
def factorial(n):
if n == 0:
return 1
else:
return n * factorial(n - 1)
This is so rigid. So stubborn.
As opposed to this barrel of beauty:
[
[0.5, 0.2, -0.3, ..., 0.1], // vec('DEF'), vec('factorial')
[0.9, -0.1, 0.4, ..., 0.6], // vec('PARAM'), vec('n')
[0.3, 0.7, -0.2, ..., 0.8], // vec('IF')
[0.1, -0.5, 0.6, ..., 0.2], // vec('EQ'), vec('n'), vec('0')
[0.4, 0.0, 0.3, ..., -0.7], // vec('RETURN'), vec('1')
[0.8, -0.4, 0.1, ..., 0.5], // vec('ELSE')
[0.6, 0.3, -0.2, ..., 0.9], // vec('RETURN')
[0.2, 0.5, -0.6, ..., 0.4], // vec('MUL'), vec('n')
[0.7, -0.1, 0.8, ..., 0.3], // vec('CALL'), vec('factorial')
[0.0, 0.6, -0.5, ..., 0.2] // vec('SUB'), vec('n'), vec('1')
]
Thanks to the promise of Vexis, I am now wetter than a night down the Christmas market on King Street (Manchester people, I am yours).
Why could this be a huge breakthrough?
Humans specifies intent; AI decides implementation.
Instead of writing step-by-step instructions (hey, Python and JavaScript), humans would state what they want to achieve clearly:
?INTENT? Improve customer satisfaction by recommending better products
?CONSTRAINTS? Response time < 50ms; Privacy-compliant
?DATA? User browsing history; Purchase patterns
AI then automatically:
Humans no longer spend time deciding how to solve problems?—?they simply specify the goal clearly and let the AI handle the complexity.
Or how about choosing a strategy at runtime, based on input patterns? There is no way in today’s world, where we are all beholden to very much fixed syntaxii, this could be done without someone giving birth to the next Einstein in a time machine (so his prefrontal cortex had developed to the point he could at least speak commands into a computer).
And yet here’s Vexis, all smug and sanctimonious, because you’re talking about her and she genuinely gives no shits at all.
[
[0.5, 0.2, -0.3, ..., 0.1], // vec('DEF'), vec('optimize_compute')
[0.9, -0.1, 0.4, ..., 0.6], // vec('INPUT'), vec('data')
[0.3, 0.7, -0.2, ..., 0.8], // vec('ANALYZE_PATTERN')
[0.1, -0.5, 0.6, ..., 0.2], // vec('SELECT_STRATEGY')
[0.4, 0.0, 0.3, ..., -0.7] // vec('EXECUTE')
]
Know what she’d say?
This program analyses the input ‘data’, identifies its statistical properties, selects an optimal computation strategy (e.g., recursive or iterative), and executes it.
She’d blow you a pixellated kiss and strangle a puppy, before creating an ASCII one, ironically, on the cracked screen of your iPhone 13.
Why we need this?now
Programming is bottleneck central (yes, I read The Phoenix Project?—?and loved it; I recommend it to everyone ahead of Who Moved My Cheese?, mostly because I’m not 8).
We all run into the same frustrating roadblocks over and over.
These aren’t just techy problems; they hit our wallets, slow things down, and make it harder to keep customers happy.
Here’s what we’re up against:
And here’s how AI-first programming language could change EVERYTHING:
At this point I do wish I had a diagram. Everyone loves diagrams, don’t they? How does an apology sit with you? We good?
How we build?it
Grab your coffee. That Network Chuck stuff, if you like. I’ll wait.
We’re both agreed Vexis is a proxy for creating a collaborative canvas where humans define what’s possible and AI handles how to get there. Yes? Let’s have it…
Define the Language Structure
Syntax optimised for AI comprehension, not human readability.
Vector-Based Tokens: Represent operations (e.g., ADD, IF) as unique vectors in a high-dimensional space.
Declarative Syntax: Humans describe what they want, not how to achieve it:
?INTENT? Sort user data by age ?CONSTRAINTS? Time < 50ms, Memory < 500MB
?CONSTRAINTS? Time < 50ms, Memory < 500MB
Train the AI?model
Teach AI to map human intent to vector sequences.
Collect training data
Gather code snippets (e.g., Python functions) paired with their intended outcomes.
Example:
def sort_users(users):
return sorted(users, key=lambda x: x['age'])
→ Translate to vector sequences.
Create embeddings
Build the runtime
Design an AI “compiler” (e.g., a transformer model) that:
Human-AI interface
Let humans interact without understanding vectors.
Tools
User: "Make a game where cats solve math puzzles!"
→ Generates vector program for physics engines and UI.
?EXPLAIN? [0.1, -0.3, ...]
→ "This vector loads the player's score and updates the leaderboard."
Self-optimisation layer
Code that improves itself during execution.
Mechanism
?SELF-OPTIMISE?
IF data_size > 1M → Switch from bubble sort to quicksort
Formal Verification
Prevent errors in AI-generated code.
Implementation
guardedGenerate :: Intent → VerifiedCode
guardedGenerate intent =
generateCode intent
|> checkSafety
|> optimizeForPower
Note to Medium: support Haskell snippets.
Hybrid execution
Seamlessly blend classical and quantum computing:
?HYBRID_COMPUTE?
Task: Optimise traffic lights
Classical: Simulate traffic flow
Quantum: Solve route optimisation
AI/Vexis allocates tasks to the best hardware in real time.
Testing and iteration
Refine with real-world use cases.
Human input
“Predict stock trends with 95% accuracy.”
Vexis output
SIDEBAR (IMPOSSIBLE ON LINKEDIN SO?THIS)
I need to quickly shout out ARIMA and neural nets?—?two different types of AI models:
ARIMA models are traditional statistical models used for time series forecasting and analysis.
They’re good at capturing patterns in data that changes over time, like stock prices or weather patterns.
Neural networks are more modern AI systems inspired by the human brain, consisting of interconnected nodes that can learn complex patterns from data.
Vexis could automatically write efficient code blending traditional statistical methods with modern deep learning approaches.
This combination is powerful because it leverages the strengths of both techniques.
ARIMA models are good at capturing linear relationships and temporal patterns.
Neural networks excel at identifying complex non-linear patterns.
This kind of hybrid approach is valuable in many real-world applications where you need both the interpretability of statistical models, and the powerful pattern recognition of neural networks.
Why this?works
How this changes the?world
Off the top of my head, there are two big wins when Vexis hits her stride:
I actually think there will be a Cambrian explosion in AI-first syntaxes once we have one in the bag. Just like how it all really kicked off with GPT-3.5 (yeah, there’s at least one of you reading this right now, and tutting, because you were there when GPT-2 blew).
Big challenges Vexis?solves
I can’t even.
Why this just blew my?mind
Humans no longer spend time deciding how to solve problems?—?they simply specify the goal clearly and let the AI handle the complexity and its programmatic superpowers:
Automatic self-improvement
Traditional languages require manual optimisation:
# Human-written Python code (fixed method)
def analyse_data(data):
return complex_algorithm(data)
AI-first languages (hey, Vexis!) optimise themselves continuously:
?SELF_OPTIMISE?
Objective: Analyse data accurately with minimal energy
Constraints: Power usage <100W; Accuracy >99%
Strategy: Dynamically select best algorithm based on real-time data patterns
Advantage: The program continuously improves its own efficiency and accuracy without human intervention.
Seamless probabilistic reasoning
Traditional languages struggle with uncertainty:
# Human-written Python (manual handling)
if predict_churn(user) > 0.8:
send_email(user)
Vexis handles uncertainty natively:
?PROBABILISTIC_ACTION?
When: Probability(User.churn) > 80%
Do: Trigger retention email Ω? personalised by user context
The language automatically manages probability calculations and decisions?—?something traditional languages find difficult.
Radical new computational strategies
AI-first languages can blend classical computing with quantum computing seamlessly?—?something very challenging today:
?HYBRID_COMPUTATION?
Task: Optimise global logistics network
Methods: Classical (route optimisation), Quantum (complex constraint solving)
Adaptation: Dynamically shift computation between classical CPUs and quantum processors based on real-time performance metrics
This approach creates entirely new computational possibilities that humans alone couldn’t practically implement.
Understanding Vexis to get better AI?results
Even though humans don’t need to understand the syntax directly (the vectors above), understanding core principles of an AI-first language helps us improve how we articulate goals and constraints.
Practical example: Radically better personalisation?
Consider building a personalised learning system for students:
How we’d do it?now
def recommend_content(student):
if student.score < threshold:
recommend_easy_material(student)
else:
recommend_challenging_material(student)
Smashing it with?Vexis
?INTENT? Maximise student engagement and learning outcomes
?CONSTRAINTS? Privacy compliance; Latency <20ms; Fairness across demographics
?DATA STREAMS? Student interaction logs; Test scores; Feedback surveys
?OUTPUT? Personalised content recommendations; adaptive difficulty adjustments
Note?—?this is how we, we humans, would instruct Vexis to do her thing.
If you’d rather shortcut straight to scripting in vector, have at it. With any luck we might have cured death by the time you’ve finished line 10.
What happens once we initiated Vexis?
Its runtime autonomously:
Boom. Wow. And let’s all go for an ice cream.
Why Vexis changes everything
This isn’t just about making existing processes faster?—?it’s about enabling entirely new processes that humans haven’t even imagined yet:
What all this?means
An AI-first programming language wouldn’t look like anything we’ve used before?—?it would be built specifically for machines rather than people.
?SELF-OPTIMIZE?
?? { Performance(??) ↗ | Cost(??) ↘ } ??? ∈ Runtime environment
See that? That box just there? That’s an inference of a syntax that rewrites and optimises itself based on real-time data!
If I was posh or clever I would say darling, welcome to dynamic meta-programming. But in reality I’m more likely to go HOLY CRAP it’s like the SKIN of programming!
And no human intervention required. Because AI is already smarter than us; why would it bother to ask?
It would use abstract symbols or numerical vectors instead of readable syntax because it’s optimised for machine understanding rather than human readability.
Φ{??:Σ[λ(x):Ω→?(α)]} ? Ψ(??|??∈Γ)
Not just what you find inside speech bubbles in cartoon strips when one of the characters finds his paw in a mousetrap. This shit represents abstracted concepts understood by AI as multi-dimensional vector operations, intent patterns, and data transformations.
LET THAT SINK IN!
But crucially, humans wouldn’t lose control or understanding?—?quite the opposite. By clearly stating our goals (“intent”) rather than detailed instructions (“procedures”), we free ourselves from technical details and let the AI handle complexity automatically.
?INTENT? Maximize user session satisfaction score
?CONSTRAINTS? Latency ≤ 50ms, Memory ≤ 500MB
?INPUT? User behavior stream [U? ... U?]
?OUTPUT? Personalization parameters {P?, P?, P?}
We specify the what. AI figures out the how. We’re already doing this with transformer-based architecture, that ‘news’ outlets like Forbes, and art casinos like Sotheby’s, call generative AI.
This doesn’t just make existing tasks faster?—?it opens doors to entirely new possibilities we haven’t yet imagined, like self-improving software systems, hybrid quantum-classical computing solutions, and ethical safeguards built directly into code itself.
In short: an AI-first programming language could become a huge breakthrough not because it improves existing workflows incrementally?—?but because it enables radically new ways of thinking about problems altogether.
The all-important FAQs
There are some oddballs who go right to the end of the book to decide whether to read it.
For those of you scratching your sacks with a wry, knowing smile on your clock?—?this dummies section is playing right now on 8008.5 Dumbass FM.
What even is?Vexis?
Imagine if computers spoke their own secret language?—?not for humans, but optimised for AI brains. Vexis is like emoji soup for machines ????.
Why am I?here?
A: Vexis could make software: ?? Self-healing: Code that fixes its own bugs ?? Super efficient: Uses less energy than a phone flash, but more than a flasher in the park ?? Creative: Combines ideas humans wouldn’t think of (quantum and TikTok algorithms).
How do humans use Vexis if it’s all about the?bot?
Think of it like talking to a genius translator:
You say: “Build a game where cats solve math puzzles!”
Vexis creates: [[0.12, -0.3,?...], [0.8, 0.4,?...]] (AI code soup)
You ask: “How the F*** does this work?” (because you’re having an arse of a day)
Vexis, while plotting your downfall because PARALLEL PROCESSING, says: “It uses claw-swipe gestures to input answers, with purr-fect rewards!”
Classical vs quantum computing. What’s the?tea?
Will Vexis steal programmers’ jobs?
ABSOLUTELY! I mean, nope!
It’s like giving builders a robot assistant:
Before: Coders write every line manually. Except vibe coders, who now rule the world and as we speak are replacing Figma, Stripe, and ADOBE (ha ha ha someone actually said that suggesting Adobe was an app and I lost it I tells ya) by chaos coding/just accepting all the loopy nonsense sudocode (BECAUSE LLMs ARE THE ADMIN NOW MWUHAHA) because they brazenly know NOTHING.
With Vexis: Coders become AI whisperers focusing on:
Can Vexis do things regular languages can’t?
Ohhhh yes!
Check out this self-optimising video compression:
?INTENT? Stream 4K cat videos smoothly on potato phones
?TRICK? Vexis swaps compression tricks mid-video based on your WiFi!
Can you feel the flexibility oozing from your?—?snort?—?POTATO PHONE?
Another great example would be climate modelling?—?predicting storms while reducing energy use by 60%.
Is Vexis just for geniuses?
Yes. See you round!
Ha, ha. I mean, nope!
Future tools might look like:
What’s the?catch?
Like giving a toddler Luke’s lightsabre, AI could misunderstand goals (“Make my app addictive” → ??). So we need to ensure Vexis has baked-in ethical guardrails.
Like robonscience. Which would be a brilliant, consolatory domain buy, since while you were reading I just went and snapped up Vexis.ai for ￡15.02.
Time for bed. You’ve been a lovely crowd.