- Ahead Of The Curve Artificial Intelligence
- Posts
- Introduction to Infinite Prompting (1/3)
Introduction to Infinite Prompting (1/3)
The second and third part is scheduled to be published on May 17th 2am EDT and May 17th 2:20pm EDT, respectively. In the meantime, you can take this course on my Youtube
Course Overview
This free course introduces a breakthrough prompting technique that flips the script on how you’ll collaborate with AI.
I’m talking about prompts that evolve on their own and generate results far better than what human-written prompts could do, often 10x better without needing constant input from humans.
This is a skillset most people don’t even know exists yet, and by the end, you’ll be one of the few pioneers who truly get it. This also means you won’t have to search for prompts anymore and can create them yourself.
What you’ll learn here
What is Infinite Prompting?
Why Infinite Prompting?
Why should you pay attention to Infinite Prompting?
Is this course right for you?
An example prompt you can play around with to get a glimpse of what Infinite Prompting could do
Infinite Prompting follows a similar concept found in key academic research and theoretical works
Infinite Prompting is aligned with historical context and related fields
If you prefer watching and listening, I’ve recorded a 3-part Youtube series. Watch it here
What is Infinite Prompting?
First, you have to know what a neural network is.
Neural networks are basically the brains behind AI like ChatGPT. They’re inspired by how the human brain works — tons of interconnected “neurons” that pass signals to each other. Through training, these networks “learn” to recognize patterns such as grammar, understand language, and generate responses, just like how our neurons develop and get stronger through training.
So, Infinite prompting is inspired by neural networks and the way we learn. Instead of just giving it one-shot instructions (prompts), you're letting the AI think for itself and self-improve its instructions and answers to produce more intelligent answers. Just like how humans learn by practice many many times to improve.
Traditional prompting = manually engineered
Infinite prompting = semi/fully automated
Why Infinite Prompting?
1. Richard Sutton’s The Bitter Lesson essay
AI pioneer Richard Sutton wrote one of the most important essays on AI. It’s called The Bitter Lesson.
The Bitter Lesson is about how AI researchers keep wanting to incorporate human ideas to make the AI smarter, but what actually works best is giving the AI lots of processing power and letting it learn independently from massive amounts of data without human interference.
For example:
Chess computers: People tried to teach computers chess strategy, but what actually beat the world champion was a computer that could calculate tons of possible moves on its own really fast. (IBM's Deep Blue vs Garry Kasparov)
Go computers: Same story. Researchers spent 20 years trying to make computers understand Go like humans do, but humans were eventually beaten by computers that just practised against themselves millions of times. (DeepMind’s AlphaGo vs Lee Sedol)
Speech recognition: People tried to program in knowledge about words and sounds, but what worked better was just letting computers listen to lots and lots of people talking.
Computer vision: Early researchers tried to teach computers to recognise specific shapes and features, but now we’ve found that showing them millions of images and letting them figure it out themselves works better.
Why this matters: The Bitter Lesson suggests that, in the long run, the most powerful prompts won’t come from humans. They’ll be built by AI itself, using stronger and better learning methods baked in from the start.
2. DeepSeek R1 Zero model
DeepSeek discovered its new model having an "aha" moment where it developed an advanced problem-solving technique, entirely on its own
A group of researchers created the following chart to show how the accuracy (blue line) improves as the model goes through more steps of reinforcement learning.
To understand why this is a big deal, think back to IBM's Deep Blue vs Garry Kasparov and DeepMind’s AlphaGo vs Lee Sedol.
During one of the games, Kasparov noticed a new, unfathomable strategy move by Deep Blue that eventually led to his loss.
Also, AlphaGo made several moves that humans never thought of, which stunned human experts and Lee Sedol himself.
So, R1 Zero did something similar. Its development of advanced problem-solving strategies through learning over and over again is a really big step in AI's capacity for independent learning.
These three cases suggest that AI are capable of coming up with novel and unexpected approaches that humans haven't thought of or explicitly programmed.
Why this matters: So, one of the goals of Infinite Prompting is to get higher-quality, novel answers with each iteration. I believe there’s a whole world of original, undiscovered, useful, and universal ideas still out there. And I think a technique like Infinite Prompting could be the key to unlocking them.
It’s not 10,000 hours, it’s 10,000 iterations.
— Naval (@naval)
5:18 AM • Nov 22, 2022
This line connects so well with what we’ve discussed above.
If you don’t know Naval, he is an icon in Silicon Valley and startup culture. He is also an early investor in Uber, Twitter, Notion and much more.
4. Complexity Theory and emergent behaviour
The core idea behind complexity theory and emergent behavior is that all the complexity we see in the world might actually stem from a few super simple rules, just iterated and repeated over time.
For example
Zygote to Human: You start as one cell. Simple rules - cells divide, interact, and self-organise. Then, a fully-formed human with thoughts, emotions, abilities is created.
Spark to Forest Fire: One spark hits fuel, wind, and terrain. Repeated over and over again, fire spreads fast. What started tiny becomes a massive wildfire destroying habitats and ecosystems.
AI Chess Player: AI learns basic rules, plays itself repeatedly. Feedback from wins and losses improves strategy. Eventually, it makes genius moves no one thought of.
Why this matters: You could start with a simple prompt that produces a simple output. Then, it self-iterates over and over again till it creates a complex prompt that generates mind-blowing results and unexpected ripple effects in the real world.
Why should you pay attention to Infinite Prompting?
AI just came into the picture 3 years ago (ChatGPT was released in 2022). And we’re going to be using it for decades. So, what are the chances we figured out all the best approaches in the first 3 years? Pretty low
It’s not mainstream yet. That means you’ve got the chance to be one of the AI pioneers or frontrunners
It’s durable. Instead of spending time making perfect instructions that might get outdated by updates to AI models, Infinite Prompting actually make things better when new models come out. Because? It self-learns
High leverage. The power to get incredible results from a simple prompt could completely change the way we use and work with AI.
You could use it in every situation. With Infinite Prompting, you could discover new marketing strategies, generate more breakthrough ideas within your domain, discover new arguments about a topic, generate new styles for arts and music, and even find new scientific hypotheses. The possibilities are endless.
Get out of your own box. Infinite Prompting will stretch your thinking into territories you’ve never even considered before.
Is this course right for you?
If you’ve already got the basics of prompting down, and now you’re ready to level up and sharpen your edge in one of the most important skills of the future
If you want to see what’s truly possible with AI, so you can be ahead of the curve and actually be ready for the future.
Yes, this is for you.
An example prompt you can play around with to get a glimpse of what Infinite Prompting could do
Let's say you found this piece of writing and you want to recreate it for yourself.
“Shakespeare couldn't help but be intrigued by the banana, hanging there, yellow and alluring. He didn't know what it would taste like, but something in him compelled him to take a bite. And as he did, he was rewarded with a symphony of flavors, rich and complex, satisfying his every sense.”
Prompt:
You’re an expert prompt engineer. These are your tasks:
1. Analyse the linguistic structure and characteristics of the attached writing.
2. Then, write a prompt that will generate a similar style of writing
3. Test the output of the prompt with a topic about Cleopatra sitting in front of a computer
4. Evaluate the output. Give it a rating of 1-10 based on verbosity and catchiness, easy to read and understand
5. Repeat until the rating reaches 10/10
Keep in mind, this prompt is not showing the full power of Infinite Prompting yet, just a glimpse of what it could do.
Infinite Prompting follows a similar concept found in key academic research and theoretical works
This research introduces a fresh method for training AI to tackle reasoning problems in coding and math without needing human input. Instead of using traditional techniques which depend on lots of human-curated data, AZR takes a self-play approach. The model
creates its own problems
solves them
and learns from the feedback it gets.
How it performed:
Outperforms models trained with human-curated data in the coding category by 0.3 percentage points.
Even though it’s surpassing by just 0.3 points, this might just change the game for how we might prompt in the future because AZR achieved this performance level without ever having access to expert-curated human data for training, unlike the models it surpassed.
Instead of manually writing prompts, PromtBreeder uses an LLM, specifically PaLM 2-L, to evolve the prompts on its own. It works with two types of prompts: one set tells the model what task to do (task-prompts), and the other tells it how to improve those prompts (mutation-prompts). The process is self-referential and runs in loops, improving both the task and the way improvements are made.
How it performed:
Outperformed several top prompting strategies like Chain-of-Thought (CoT), Plan-and-Solve, Zero-shot CoT, Program-of-Thoughts (PoT), Auto-CoT, OPRO, and APE.
Did especially well in tasks like arithmetic reasoning, commonsense reasoning, and hate speech classification.
Showed better results than random search and basic prompts taken straight from task descriptions.
Self-Refine lets an LLM revise its own output by giving itself feedback. First, it generates an answer. Then it critiques what it just wrote. Then it uses that critique to rewrite the output. This loop can run multiple times, and it doesn’t require any extra training data or human supervision.
How it performed:
Consistently outperformed the typical single-step output across seven different tasks (like dialogue, math, code generation, sentiment reversal, etc.).
On average, improved task performance by around 20%.
For code generation, the improvement was up to 13% over the initial output.
Humans also preferred the refined outputs over the one-shot answers.
This is a looping self-improvement technique in code generation. STOP starts with a basic “optimizer” program (written in natural language instructions) that is applied to the LLM’s code output to suggest fixes or enhancements. The “optimizer” can also rewrite its own instructions to become a better optimizer. This loop continues, so the optimizer keeps upgrading its instructions and the code output over time.
How it performed:
The upgraded version performed better than the original across several optimization tasks (like 3-SAT and learning parity with noise).
It managed to rediscover advanced optimization strategies like beam search and simulated annealing all on its own, without being explicitly told to.
Outperformed other simpler methods like single-pass CoT or greedy improvement loops.
These improved optimizers even worked well on new tasks they weren’t specifically trained on.
Theoretical foundations
A lot of today’s research actually builds on older ideas about self-improving AI. Back in the ’90s, Jürgen Schmidhuber and others were already talking about systems that could tweak themselves to get better over time. And the idea goes even further back, like Irvine John Good’s “intelligence explosion” from 1965, where an AI keeps redesigning itself to become smarter and smarter.
By refining the prompts they use, they can change how they behave without needing to be retrained. The model is learning how to get better results just by changing how it asks questions.
Infinite Prompting is aligned with historical context and related fields
Biological Evolution
One of the clearest parallels to this idea of self-improving prompting is evolution itself. Nature has been running an infinite improvement loop for billions of years. Random mutations show up in organisms, and the environment “selects” which ones survive. Over generations, species get more adapted.
In a similar way, Infinite Prompting works by generating variations and then picking the best results based on some kind of scoring or feedback. That’s exactly what PromptBreeder does: it “breeds” better prompts using evolutionary algorithms.
Like evolution, there’s no end goal, just ongoing exploration. If you let an AI keep self-prompting, it could keep discovering new strategies, solutions, or ideas, just like how evolution keeps producing wild and novel life forms.
Genetic Programming
Back in the ’90s, John Koza introduced genetic programming—basically, letting a computer evolve its own code to solve problems. Instead of a human writing the perfect program, the system mutates and recombines snippets of code, then selects the ones that perform best.
That’s a perfect analogy for how we can evolve AI prompts. Instead of evolving code, we’re evolving the instructions that guide the model’s behavior. And just like in genetic programming, the best strategies might be ones we didn’t anticipate—emerging not from design, but from iterative variation and selection.
Self-Modifying Code and AI
The dream of programs that rewrite themselves isn’t new. Researchers and sci-fi authors have been into that idea for decades. Douglas Hofstadter explored this with his concept of “strange loops,” where a system can reflect on itself and change as a result. In the ’80s and ’90s, people explored self-rewriting algorithms, though there were concerns about how stable that could be.
The modern version of this idea is more controlled. Instead of editing its own binary code, the AI just rewrites the prompt it’s given, which influences its output in the next round. It’s like doing code-modification at the level of natural language.
Schmidhuber’s Gödel Machine took this to the theoretical max. It was a blueprint for a fully self-improving agent that rewrites its own code only when it can prove the new code is better. We’re not quite there, but Infinite Prompting is a step in that direction, testing new “code” (prompts) and keeping what works.
Cellular Automata – Conway’s Game of Life
Conway’s Game of Life is a classic example of how simple rules can lead to mind-blowing complexity. You start with a grid and a few rules, and just by applying them repeatedly, you get gliders, patterns, even full-blown computation. It’s an iterative system—no new inputs, just loops—and yet it generates complexity that feels almost alive.
That’s a powerful metaphor for iterative prompting in AI. If you give a model a simple rule like “improve your previous answer based on feedback” and let it loop, the end result might be way more advanced than what a one-shot answer could achieve. An AI that keeps prompting itself could keep generating useful, surprising results in an open-ended way.
Theoretical Models of Cognition – Strange Loops
Self-reference isn’t just a math or CS thing, it’s also at the heart of how we understand human consciousness. Hofstadter’s work, especially in Gödel, Escher, Bach and I Am a Strange Loop, makes the case that our sense of “self” is built from internal feedback loops: the brain reflecting on its own activity. He suggested that consciousness is what happens when a system can represent itself and act on that representation.
Iterative prompting mirrors this idea in a light way. The AI looks at its own output, questions it, improves it—it’s a loop. It’s not full self-awareness, but it’s a form of self-monitoring or introspection.
This is similar to ideas in cognitive science, like Global Workspace Theory, where conscious thought is like a spotlight of attention. In a way, Infinite Prompting is creating a “workspace” in plain text, a space where the AI tracks and refines its thoughts step-by-step, leading to more coherent reasoning and better outcomes.
Reply