Language, Our Oldest Technology and Possibly Our Biggest Bottleneck

Human language is an incredible invention. A compressive, lossy codec for the mind.
It evolved for survival, storytelling, gossip, and cooperation — not for describing quantum fields, multi-dimensional manifolds, or embedding spaces. Yet here we are, trying to use 50,000 years of biological slang to communicate with systems that never lived in a savanna, never feared a tiger, and never needed to impress a tribe around a fire.

This alone creates tension.

Language is:

• Ambiguous — almost every word has multiple meanings.

• Biocentric — optimized for human perception.

• Slow — you can think faster than you can speak.

• Shaped by culture — and culture is chaotic.

• Compression-heavy — entire models are reduced to single words (“gravity”, “life”).

• Constantly mutating — Gen Z invents new words faster than models can retrain.

The problem is not that language is “bad.” The problem is that we mistake it for reality.

And Now We Give This Noisy Communication Protocol to Machines!

This is where the trouble begins.

LLMs inherit all the structural constraints of human language — because that’s their training substrate. You get all the ambiguity, all the imprecision, all the hidden assumptions baked right into the embeddings.

When you talk to a model, you don’t give it “the problem.” You give it a sentence about the problem, in a lossy format, constrained by centuries of linguistic drift.

This is why prompting is hard. This is why we need entire fields like “prompt engineering.” We’re performing surgery with crayons.

Despite all its flaws, language enables alignment between humans and machines.
It is our Rosetta Stone: a shared symbolic space.

Language lets us:

• Compress complex models

• Transfer abstract concepts

• Generate structured hypotheses

• Build shared mental models

• Map human goals into machine action

The paradox is beautiful: Language is the weakest link — and yet our only link.

Is Our Language Suitable for a Higher Level of World Comprehension?

If humans and AIs ever build a deeper connection, it may require a new kind of language — one that is less ambiguous, more structural, more mathematical. But until then, we must master the language we have. Even our mathematics can prove too human, too simple, too dumb.

Because if we cannot precisely articulate the problem, we should not expect the machine to deliver a precise solution.

Data, LLMs, GenAI, and the Endless Loop of Knowledge

Large language models take the chaotic data and build a structured manifold: a smooth, continuous space of meaning.

They do not “understand.” But they encode meaning in ways that are:

• high-dimensional

• analogical

• relational

• compositional

• emergent

LLMs build a topology of knowledge that is arguably more coherent than human language itself. This is why they appear intelligent. They are performing geometric reasoning over meaning.

But the tradeoff is severe:

• They inherit every weakness of the data.

• They inherit every weakness of the language.

• They hallucinate where the data is thin.

• They produce statistical echoes of our own linguistic biases.

LLMs are powerful — but they are also mirrors.

GenAI is the Surface Layer We Interact With

It’s the part that produces text, code, images, and ideas. What fascinates me here is that GenAI doesn’t just retrieve patterns. It synthesizes, blends, transforms, and abstracts. Sometimes, it even proposes structures we would not have invented ourselves.

Generative models are compressors and decompressors of possible worlds within the boundaries of language.

This is where the golden quadrant becomes real: GenAI allows us to explore the space of plausible models that humans cannot compute manually.

We now have a four-step loop: Data → Model → Generation → New Data.
This is a brand-new epistemic engine to generate knowledge, as I showed at the beginning.

Can This Loop Produce Something Fundamentally New?

The optimistic narrative you often hear goes something like: “Since LLMs recombine language in new ways, they can discover new truths hidden in text.”

But let’s dissect this carefully.

Language is not just a medium but a latent structure of the world.

Over centuries, humans have encoded:

• patterns

• analogies

• causal assumptions

• heuristics

• models

• folk theories

• scientific abstractions

• Language is a compressed archive of human cognition.

When a model explores this space, it can:

• Reveal latent patterns

• Suggest missing connections

• Propose hypotheses hidden in the structure

• Surface correlations we never articulated

• Generalize beyond any individual text source

In this sense, yes — the geometry of language can hide new knowledge.
And models can discover it.

But only under a condition: The knowledge must be implicit in the corpus. LLMs cannot transcend what humanity has already baked into the linguistic manifold. They share some of our blind spots.

The moment we try to go beyond implicit human knowledge, we hit a wall.

We have no corpus written by physics itself.
No corpus written by evolution.
No corpus written by spacetime.
No corpus written by the underlying machinery of reality.

Until we build foundation models trained on non-linguistic data — raw sensory streams, real physical interactions, symbolic scientific models — language will remain a ceiling.

LLMs can rearrange our thoughts. They cannot generate thoughts the universe never encoded.

Not yet.

So, if the LLM hasn’t encountered the question with its answer before or can’t use a calculator, it won’t be able to determine the answer.

It also won’t make you a coffee.

Not yet :)

Models may help generate novel hypotheses, but not new truth. The moment a model outputs something surprising, we must still:

• Validate it

• Embed it into a scientific model

• Test it

• Use it to make predictions

• See if reality agrees

Language alone is insufficient for scientific discovery. But it is an excellent engine for hypothesis generation.

Leave a comment

That alone moves the boundaries of human knowledge.

This is the direction foundation models are heading — models that integrate:

• spatial data

• physical simulation

• biological signals

• multimodal sensory streams

• mathematical structures

• symbolic scientific rules

Only then will AI be able to generate knowledge that does not simply remix human thought. Until then, we live inside the linguistic bubble our species built.

Near Future Prediction

Predicting the future is traditionally a dangerous sport — but also an irresistible one.
So let me sketch a few trajectories that seem almost inevitable given the current data, trends, and incentives. These aren’t prophecies. Think of them as models: imperfect, but useful.

Data Privacy Becomes Fiction

Privacy stops being about “what you shared” and becomes about “what can be deduced about you.” With enough data points — many of which you never knew you emitted — models will reconstruct your preferences, intentions, habits, and vulnerabilities with unsettling precision.

Privacy becomes less of a right and more of an illusion we nostalgically describe to younger generations.

Democratization of Hacking

Hacking used to require expertise. Now, it requires instructions in natural language. Agentic models can already chain actions, write malware, and adapt to defenses.

When such tools become cheap and ubiquitous, the attack surface grows; defense becomes a continuous battle of AI versus AI.

A Flood of Deepfakes

Deepfakes are no longer a novelty — they’re becoming background noise.
Soon, the question won’t be “Is this video real?” but “Does it matter whether it’s real?”
We’ll operate in a hybrid epistemic environment where synthetic and authentic realities blend. Institutions, media, governments, and individuals will struggle to maintain trust.

Reality becomes probabilistic, just like truth in machine learning.

Advanced Phishing and Social Engineering

Phishing evolves into personalized, AI-generated psychological operations. Models will craft perfect imitations of voices, faces, writing styles — tailored to each individual’s psyche. Your future inbox may consist mostly of messages written precisely for you, by systems designed to exploit every cognitive bias known to neuroscience.

Trust itself becomes a vulnerability.

Job Displacement and Skill Atrophy

AI won’t eliminate all jobs — but it will eat the middle: tasks requiring moderate expertise, moderate creativity, moderate responsibility. The low-end gets automated; the high-end scales with AI assistance. Meanwhile, humans risk losing skills we once considered foundational — writing, reasoning, even the ability to sit with an open problem without asking an LLM for immediate closure.

We may become cognitively dependent, outsourcing more than we realize.

Cheap Dopamine and Infinite Content

AI becomes the cheapest source of dopamine humanity has ever built: hyper-personalized entertainment, companions, narratives, worlds, and attention loops.
We already have infinite content. Soon we’ll have infinite content made specifically for each person, optimized for maximal retention.

The line between “personalized experience” and “addiction-by-design” thins rapidly.

The Path Toward AGI

We will see:

• Autogenous AI: systems capable of improving themselves in constrained domains

• Recursive optimization loops

• Foundation models trained on multimodal sensory reality, not just language or pictures

I don’t know if this leads to AGI, but I do know that the pace of capability growth is accelerating. And with scale, architecture, and multimodality converging, the magnitude of potential knowledge approaches something that feels uncomfortably close to infinity from our human vantage point — the singularity.

If intelligence keeps scaling and models begin optimizing themselves, the idea of a singularity becomes less science fiction and more a matter of when, not if. Not as an explosion of consciousness — but as a tipping point where:

• human pace

• biological limits

• linguistic constraints

• and cognitive bandwidth

simply fail to keep up.

Whether this future is utopian or catastrophic depends less on the models and more on our ability to control the change and adapt.

Thanks for reading!

The structure is as follows:

[SLIDE]

The transcribed text.

And optionally, my commentary.

As you’ve probably guessed from the first slide, we’re at the Faculty of Information Technology, so as a small tribute, I framed this slide in the form of a context-free grammar.

The talk will be with a slightly philosophical flavor, which is always a bit dangerous, because the audience can easily tear you apart in the discussion later.

Let’s dive in.

I created this slide with respect to the venue of the lecture, which was the Faculty of Information Technology at Brno University of Technology, my alma mater. Formal languages and automata were among the most inspiring courses of my studies, and displaying the title as a context-free grammar nicely resonates with the lecture’s themes: the importance of unequivocalness and meaning when working with data and machine learning models. This form clearly reduces the ambiguity of the lecture’s title, which is crucial when querying AI systems.

I represented SolarWinds during the conference, and we got a lovely spot for our booth in the hallways of the old monastery, right across from the refreshments, which turned out to be a strategic position to catch hungry bystanders looking for a schnitzel, introduce them the company, and exchange ideas.

With the surge in AI, new perspectives are unfolding in our understanding of the world around us. In the near future, it will become imperative to move beyond age-old paradigms of truth and reality and start rethinking humanity’s problems to harness the power of AI effectively for their solutions. In this session, I aim to navigate you through the fundamentals of what should be our focus and the necessary mental shift required.

The session abstract was provided a couple of months before the session, which always causes you to deliver something a little bit different from what you thought. Sometimes it’s for the better, I believe. This slide wasn’t part of the presentation and was only provided for the conference agenda.

What are we going to discuss today? I don’t really like agenda slides, so I’ll just splash this all at you, burn it into your retinas for a moment, and then we’ll jump right in.

What the audience saw was a brief cognitive shock of the session’s index, displayed for just a couple of seconds but long enough to imprint whatever their brains noticed first.

Truth

I’ll start with a slightly uncomfortable topic: truth. And why talk about truth in the context of artificial intelligence? As humans, when we grow from children to adults, we usually begin with something like “naive realism”: we think that what we experience and see is a direct reflection of objective reality.

To some extent, this model of truth works. But it also has apparent limitations.

At this moment, your brain, still processing the previous slide, instantly suffers a loss of information abundance, preparing you to focus.

If we base truth only on our senses and experience, we easily end up with “truth” that is basically just a consensus. Everyone says it’s true, so it becomes true. We agree with others, and we build truth from what people around us repeat. And of course, everyone brings a piece of themselves into their picture of truth.

In the end, we bake a strange construct: Very noisy, very coloured by bias, and on top of that, binary: either it’s true or it’s not, which suggests that nothing in between can be true, and suddenly the world gets very black-and-white. In science, this kind of truth model doesn’t work. And that’s already built into the scientific method itself; we’re used to looking at truth differently.

You can see I’ve put a tilde before the “objective” word because I treat objective truth as a purely philosophical concept.

So how do we look at truth from a machine’s point of view? We look at data. Today, truth is statistical primarily. We use statistical methods, but from the perspective of computing and AI, truth is agnostic. Humans can only influence it through the data they produce. If the data are biased, then the “truth” we infer from them is implicitly biased as well. If the data are cleaner, the model of truth becomes cleaner.

What we’re left with is some probabilistic model of truth, and that’s what we actually work with in science and informatics.

So it’s not black-and-white. It’s more of a gradient.

The title is displayed at the same position as on the previous slide to maintain topic fluency. This form of extra dimensionality carries through the entire session wherever possible, it’s also establishing the protocol of how the information in the slides will be presented and what’s the “on slide/spoken” ratio.

The “Truth” is purposely presented within green “walls” that squeeze it between these two paradigms. The image in the background represents a half-man, half-machine according to what level of truth is closer to each of them. The tilde in the “~Objective” says that I don’t believe in an entirely objective truth available to us.

An interesting point on the truth I mention is that "Humans can only influence it through the data they produce.” Of course, especially with GenAI models, we can add multiple layers to “improve” and align the output as needed. My initial thought here was based solely on pure inference.

I’m currently working on a whole article about the topic of truth and will share the link once it’s finished.

Subscribe now

Knowledge

And because this talk is about knowledge, not just truth, the question is: how do we get from this probabilistic model of truth back to a model of knowledge? We must reduce everything into models so that we can even discuss it.

If we start at the level of raw data, we humans receive sensory inputs from the world around us. From that data, we begin to form hypotheses. A hypothesis is nice, but if you can’t test it, it’s almost useless, and nobody will really care. So it’s important to test hypotheses with models.

We’re looking for models that fit the data well and, ideally, have predictive power because both our thinking and our brains are fundamentally predictive. If we had to process everything by brute force — like who is sitting where or standing where — it would be impossible; instead, we constantly predict what will be there, then check whether reality matches our prediction.

We build models in science in exactly the same way. From them, we draw conclusions, and these conclusions can be treated as a certain level of probabilistic truth: truth that holds under specific conditions, in a specific context, with some probability. If you imagine this process as a spiral or an endless cycle of learning, we keep feeding new data back into the cycle and continuously deepen what we know about the world. This whole model has two overlapping boundaries. One is our overall potential for knowledge. The second is the complexity of the models we use to test hypotheses, and therefore the limits of what humans can understand

How do we pursue knowledge when faced with this version of truth? We infer it from data by formulating testable hypotheses and creating models with strong predictive capabilities. The conclusions drawn nudge us closer to the truth and become new data, fueling a relentless cycle driven by our desire for discovery with boundaries only determined by the extent of our potential for knowledge — a frontier that can be shifted..

But what are its limits? I elaborate more on this topic in on of my other lectures:

Knowledge Potential over Time

Let me switch to another little model. An image is worth a thousand words.

On the y-axis, we have “potential for knowledge”, and on the x-axis, we have time. You can probably guess there will be a trending curve, showing how our brains, our collective thinking, and technology have evolved over time. Roughly, at least – reality is always messier.

I’ll mark a few milestones.

The first milestone is the knowledge potential of a single individual. There was probably a time in human history when the entire body of “useful knowledge” could fit into one person’s head. I think it was Komenský who used the term “pansophy”, the universal wisdom, meaning a person could grasp almost everything that could be known and used in life.

That moment is definitely gone. Even if it ever existed, it was very brief.

It began with a single human mind, which once could have contained all available knowledge. However, as our understanding of the world advanced faster than our brains could grow, knowledge outpaced our mental capacity. The term “pansophy”—meaning “universal wisdom”—originates from the 17th century, when it was believed one mind could hold all the knowledge of that era.

Then we learned to work in collectives. At first, small groups, tribes, and teams, but as soon as we started collaborating, sharing, and mixing knowledge and experience, our potential shot up. Everyone could know a little bit, and together we could cover a lot more.

Over time, we started sharing knowledge and expanding it collectively. We collaborated to reach wisdom surpassing individual capabilities. We created various groups—packs, guilds, teams, communities, societies—and together, we learned to hunt mammoths, construct bridges, and carry out complex surgeries such as triple heart bypasses.

The third level we’re roughly at now is the network. Under “network,” you can naturally think of the internet, but also of any large group of people who can exchange vast amounts of data, information, and knowledge in real time and work with it and develop it. That’s the “network stage” – today’s internet, our data reality, our information reality.

This stage faces a major issue: too many inputs, too many actors, and too much noise. It’s becoming very difficult to distinguish between real, reliable knowledge and nonsense.

And of course, how else do we fix a technological problem but with more technology?

Ultimately, we’ve built the most complex structures yet: networks. Many entities, spread across great distances, can now access and share knowledge almost instantly. This knowledge is constantly exchanged, developed, and utilized. However, we’re starting to feel overwhelmed by the huge volume of data. Telling valid models from false ones is becoming more challenging.

This highlights our current situation and explains why I see AI as the essential next step in understanding the vast data landscape.

So here is the breaking point we’re approaching – where artificial intelligence comes into play. Especially this year, it’s a huge hype.

I come from the “old AI” days, when AI wasn’t so sexy – over ten years ago. Today, everyone mostly talks about generative models and large language models. I’m a bit old-school, so I won’t go deep into history; otherwise, we’d be here for more than twenty minutes.

What does AI offer us today? It can work with complex, multimodal data. You can feed it video, talk to it in natural language, give it images, give it all kinds of instructions, and it will talk back to you. That’s roughly where we are now.

AI systems trained on all the data we have, capable of processing complex and multimodal queries and interacting with us in our own language, represent the next frontier. This is likely the point at which we surpass the limits of human potential alone.

But the question remains: how much further can it take us?

I started with machine learning in 2008 by winning an AI tic-tac-toe tournament that was part of our “Fundamentals of Artificial Intelligence” course during my studies. It wasn’t true AI, just smart heuristics, but it was so fast that it actually beat other models thanks to a scoring system that awarded points for both results and speed. It had a complexity of O(SIZE²), no recursion, and no minimax, making it very fast. Because of the missing minimax implementation, I did not receive the top rating despite winning. I focused on optimizing for speed rather than calculating the best move for a long time. But it was enough to spark excitement for machine learning beyond the Terminator movies.

Later, I wrote my master’s thesis on an optimization of neural networks and their combination with chaos theory in stock market prediction. That was well before the genAI bubble.

Where we can go from here is probably more of a topic for an offline beer after the talk. You’ve surely heard terms like AGI, artificial general intelligence, or self-improving AI — systems capable of creating their own enhanced versions and offspring. These systems could become so advanced that they might not communicate with us in our language without a “translation layer” from their internal language, which might be more aligned with the structure of reality itself.

Predicting the future is difficult, but terms like AGI, the singularity, and autogenous AIs are often discussed. I believe these systems will need to go beyond human language limits and redefine many current concepts to better model reality than humans ever could.

Let me make a short detour. I’m a big fan of humans. I think people still have a lot to offer and should try to keep up with AI.

You might have heard the term “transhumanism1”: the idea of extending or enhancing our cognitive abilities. That can include anything from genetic engineering and biohacking to projects like Neuralink from Elon Musk. It’s about anything that could push our potential a little closer to that “upper” curve in the future.

The curve of today’s human potential is not straight; it’s a staircase.

In our model, I’ve represented human potential as a constant for simplicity’s sake. Clearly, it’s not; it has been increasing over time. Yet, the pressing question is whether humanity can keep up with AI’s pace. Is it feasible for human and AI capabilities to converge once more? Could we modify or evolve ourselves to extend our own limits? We have avenues to explore. I am certain there is a path forward. But for what cost?

The big question is: what kind of knowledge can we reach with the help of AI?

Let’s look at one more model.

Leave a comment

Models: Predictive Value vs. Complexity

This time, we’ll model “predictive value” or “usefulness” on the vertical axis, and model complexity on the horizontal axis. We’ll split the graph a bit. Here’s the limit of human potential. That nicely divides the space into two areas: on the left, models humans can understand; on the right, models we probably cannot, but where AI could push us a bit further.

On the left will be most of what we already know, the theories we can actually reason about.

On the right, there’s a space that signifies the window of opportunity for AI.

Our minds and networks are limited in the complexity of models they can understand. We interpret the world through models we developed and comprehend, such as meteorology or quantum mechanics. There is a core body of knowledge within human capacity. Yet, there exists another realm of models that can offer useful predictions but are beyond our ability to understand their internal mechanisms.

I’ll add one more boundary: what I call the “bullshit line”. Anything below this line has insufficient predictive value to warrant scientific or technical interest.

The “bullshit line” marks the threshold of a model’s value and essentially separates what can be considered a valid scientific theory or a useful model from what cannot. Anything below this line is, frankly, unreliable. 🙂

Now, what lives where? Down here on the left, we have folk wisdom, proverbs, rules of thumb – simple models that are good enough for everyday life.

If you miss a train because you imagined it leaving later based on some simple model, it’s annoying but not disastrous.

We don’t need AI there.

There are now four quadrants. The bottom left quadrant contains models that are sufficient for everyday life or to get thousands of followers 🙂

Up here, we have models that are valid and useful: the most rigorous scientific theories and the highest-quality knowledge.

In the upper left quadrant, we find models that are validated, thoroughly tested, and used with a high success rate, yet remain within the complexity our most brilliant minds can handle.

On the bottom right, we have the real bullshit: models that can’t be tested, pure hallucinations, ideas people make up on drugs, totally vague, impossible to formulate, impossible to verify, and therefore useless for prediction.

The bottom right quadrant is designated for untestable models and ideas that appear plausible but cannot be formalized or are simply too wild.

So what’s left is a kind of “golden quadrant”: models that surpass human limits but have high predictive value. That’s where I’d place the biggest potential for AI to expand our knowledge.

The upper right quadrant is our ‘golden quadrant.’ It holds the potential for models that are both highly complex—beyond human understanding—and extremely useful. As mentioned earlier, the human potential line isn’t fixed, and there are models within our potential that we haven't yet discovered. AI can play a key role in uncovering these as well.

I want to give a few examples for the first three quadrants. Treading carefully to avoid offense, in the bottom-right quadrant, I’ve included a single example: the simulation hypothesis2, which is reasonably complex but remains untestable so far. Even if proven, it’s unlikely to provide predictive value beyond Murphy’s Laws, like “If we’re living in a simulation, it’s bound to crash at the worst possible time.”

As shown, the lower left contains flat Earth theory and similar nonsense, while the upper areas approach the human limit, with fields like quantum mechanics and advanced physics. You likely noticed I extended the golden quadrant into this region because we haven’t fully explored the upper left area yet. There’s still a wealth of knowledge to discover, even with AI assisting us in analyzing vast datasets and searching for patterns and structures beyond human perception.

The leading theory of particle physics is the Standard Model. It describes all fundamental physical interactions except for gravity.

Yet, the 2023 Muon g-2 experiment confirmed a 5.1σ deviation from its prediction, with a data fit of about 10⁻⁶. This suggests that our current model is significantly inaccurate, leaving ample room for further exploration.

Subscribe now

The Golden Quadrant of AI’s Potential

So, what kinds of things belong in that golden quadrant?

Most of you know common AI use cases, so I’ll just mention a few. For example, detecting “dirty data3”: highly noisy, corrupted, fake, or fabricated data. That’s something we’ll see more and more of on the internet. Data that is too noisy or has no real underlying foundation, errors, or made-up data. Still data – but much less valuable.

Then there’s what we call “dark data4”: data we gather but never examine, as we don’t see immediate value in it. However, there could be hidden insights there – AI can help explore this area.

In many areas of physics, we seek to unify concepts and interactions, reducing the number of independent ideas. AI could assist by discovering patterns and connections we don’t see. It can help find new combinations and meanings. We describe reality using our human categories, but that doesn’t mean AI will necessarily use the same ones. If AI looks at a table full of bottles, it might see something completely different and label it in a way we would never think of. It can detect symmetries and similarities we’re blind to. That’s a major benefit: AI is not anthropocentric. It doesn’t have to view the world through a human lens. Of course, when we force it into a language model, it becomes very human-centric again. But there are other types of foundation models, not just language-based ones.

Another area is causal inference: we often observe only correlation, but in some cases, AI might help us identify causal relationships. Those are the more practical applications.

If we ventured into the “spiritual” realm, we might ask questions like: what is life, what is consciousness, what is infinity, what is reality, how gravity actually works, or whether free will exists. I don’t dare say that AI will soon help us answer these—still science fiction—but there are already thinkers who try to connect hard science with these traditionally philosophical questions.

There are authors who write about this, and we’re starting to see hypotheses that are at least in principle testable. I think the next few years will be extremely exciting in this area. Assuming, of course, there’s no Skynet moment and it all doesn’t end just as it begins.

Let’s hope not.

Near the edge of human potential, there’s a wealth of knowledge within reach that may be obscured by noise or overlooked because we haven’t yet directed our attention there. AI can help us by sifting through noisy data, extracting value from unprocessed data, and tying up loose ends in our thoughts and hypotheses.

It can integrate concepts, discover combinations we’ve never considered, coin new terms, spot patterns we’ve missed, identify symmetries in data to pinpoint similarities in their origins, and discern causality where we’ve only seen correlation.

If we aim to explore further, many themes and concepts lie outside our current scientific understanding, and our minds may not even grasp them.

So where are we now in terms of our potential, and where does AI sit in that picture?

I borrowed this nice graph from Max Roser, because I really didn’t want to redraw it in ASCII. You can see various human activities that we typically perform quite well, and how AI is catching up or even surpassing us in them.

At first glance, that looks a bit scary, but it also means we’re entering that golden quadrant: AI can now understand written text better than most of us. That’s genuinely scary.

I myself read some books two or three times before I really get them.

How close is AI to helping us with the challenges of the golden quadrant? I’ve adapted this chart from Max Roser in 2023, showing how AI systems have improved over time compared to a human baseline. It demonstrates that recently, AI has started to outperform humans in many areas, so the golden quadrant is not just on the horizon but already here.

Challenges and Paradigm Shift

Let me end with a few challenges we face ahead. Most of them are at the data level.

First: quality. We’ll have more and more dirty data, fake data. How do you build a reliable predictive model if you don’t know what in your dataset is fundamentally correct?

Second: Do we even have enough data? We collect petabytes and exabytes, but is that really enough to understand the universe?

Third: Real-time data. Many models are trained for years; they’re huge and complex, and it’s hard to keep them up to date with fresh data in real time. Often, they sit there and do inference on new inputs, like today’s large language models. We’re not yet good at continuously retraining them to keep up with the latest reality.

Then there’s anthropocentric semantics –I mentioned that already.

Next, we have a lot of “soft data”: if you listen to how Gen Z talks today, which is probably most of you, older people often don’t understand you at all. So why should a machine understand you easily? We use a lot of slang and changing expressions, which is hard for models to keep up with.

The challenges revolve around working with data captured in semantics, which allows us to understand it. Our understanding of data is not needed for AI models, but the data challenges still apply:

• Real-time data presents a significant hurdle, as AI models require extensive time and resources to train, preventing direct real-time processing.

• Soft data, language-specific nuances like narratives and memes, which we sometimes comprehend.

• Noise, dirty data, and false data further complicate matters.

• A key question is whether we have enough data to drive significant progress, and if so, whether we have appropriate computational power to process it.

• Emergent data: New information or patterns that arise at certain levels of complexity must be incorporated.

AI models face inherent challenges:

• Trust in Inductive Reasoning: Can we rely on models if we only have inductive evidence of their validity, knowing they could fail?

• The need for foundational models: We must examine the data on which models are trained and understand their structure. Relying solely on large language models may be insufficient.

And finally: language. I skipped a whole chapter on language, but language is crucial for how we frame problems, communicate them, and share knowledge. We’ll have to learn to ask the right questions and correctly decode the answers. We want answers that are testable – and ideally ones that don’t accidentally blow up a nuclear reactor somewhere.

And that’s it from me. I hope I finished exactly on time.

Thank you.

I skipped the part about the language so I could finish exactly at the 20-minute mark, but at least you get the chance to read it at the end.

Discussion

The last slide I presented was meant to engage the audience and encourage further discussion after the session. Another 15 minutes of questions from the audience followed. Sharing it next. I selected a few questions related to the lecture itself:

Audience: One philosophical question — you touched on transhumanism. Some people connect it with collective consciousness. Isn’t that dangerous?

I see it mainly as an interesting area — we’ll likely have neural implants, augmented cognition… but there are substantial security concerns. With quantum computing coming… …our current cryptographic systems might be broken overnight. So isn’t the risk enormous?

Me: I’ll approach this from two perspectives.

First, I completely agree — connecting humans directly to machines poses significant risks. Bad guys are always one step ahead. You could literally have someone walking around whose thoughts are manipulated externally, and there would be no easy way to detect it.

Regarding cryptography, quantum-resistant crypto is already being developed. We’ll just need a new wave of mathematics and protocols. Honestly, I’m not too worried about this part.

People have always experimented with self-enhancement — mentally and physically. Bryan Johnson takes hundreds of supplements daily, trying to slow aging or improve cognition. Others wear devices that continuously track emotions or measure biometrics, like the first-world cyborg Neil Harbisson. So, this isn’t a new trend. But the more fascinating direction is neuroscience — not enhancing the brain, but understanding it. The brain constructs reality and seems to be a creation to observe and understand. If we know the process of understanding, we might understand reality itself more deeply. That’s why psychedelic-assisted therapy is gaining renewed interest.

Studying how chemicals alter perception allows us to reverse-engineer cognitive patterns. There’s a lot of research happening in this area right now. At the National Institute of Mental Health in the Czech Republic, volunteers line up to participate in live studies, receive psychedelics in a controlled environment, and contribute to ongoing research.

Audience: You mentioned combining science with existential questions— life and consciousness. Have you come across work that does this well?

Me: Yes — I recommend Curt Jaimungal’s podcast “Theories of Everything”. He interviews highly educated guests who present their unique, fascinating models of reality, consciousness, physics, and metaphysics—all thoroughly discussed. If I had to choose one episode, it would be the first one with Bernardo Kastrup, a holder of PhDs in computer science and psychology, who explains his worldview in a very clear, ego-detached way.

Curt Jaimungal

Exploring theoretical physics, consciousness, Ai, and God in a rigorous (and playful) manner.

Audience: You mentioned humans have limits. We do. Does AI have limits?

Me: That’s an excellent question. We don’t know. I skipped a definition of AI in the context of this lecture, but I believe AI can extend our potential and cognitive boundaries, though it may not be the final stage of intelligence.

Everything here is speculative — why would a human-made technology be the end of evolution? The universe is billions of years old; there is plenty of time left, likely for other forms of intelligence to emerge.

I don’t think AI is the final chapter. I also don’t think we’ll be around to see the end.

And there’s more. If you’ve read this far and you’re interested in the additional content, you can find it here:

Thanks for reading, and make sure to subscribe to learn more.

Subscribe now

The problem is universal, qualitative, and somewhat recursive. At its heart, this question relates to whether we are special because we’re biological observers of the universe or whether ‘observers’ (i.e., conscious intelligences) could emerge from other substrates.

Therefore, it raises the question of whether AGI1 is possible and whether what people possess is more than just a “wetware2” algorithm. If consciousness can emerge purely algorithmically, AGI might one day be conscious.

Without a clear handle on this, how can we hope to harness AI to solve our grandest problems? Answering this question would have a profound impact across philosophy, mathematics, and religion. Assumptions about the mind-body problem, implications for soul, spiritual traditions, etc.

We now possess language models trained on almost all our data. If a distorted reflection of reality, modeled through a mirror of language, is enough to create an AGI, then language has always been the key to intelligence and consciousness.

Turing Test 2.0

We have tests that offer satisfactory statistical confidence in evaluating algorithms, intelligence, or human psychology, such as the Turing test or Searle’s Chinese Room3. If we strive to find consciousness using AI, these tests need to be updated, or we may need to accept there is another class of intelligent beings.

Can a general test for consciousness be found, or is it an undecidable problem, perhaps even an impossible one? Is it a task for logic, theoretical informatics, or philosophy?

Every model of reality eventually discovers a set of unsolved, hard problems, including at least one version of the “problem of consciousness.“

Leave a comment

What do you think?

Ethically, let’s draw clear lines of respect for human (and animal) well-being. We don’t need to dissect living brains to gain insights into consciousness… Should we permit it for conscious machines, though? Do you believe AI might never cross this threshold, or is it just a matter of definition before we give it human rights?

Subscribe now

Just recently, I dug up my old notebook from the early years of this century, long before the “AI boom.” In its pages, I found ramblings about how artificial intelligence could shape humanity’s future by putting her at work to solve our most burning issues. Let me share a brief excerpt from it:

16. červen, 2003

…V říši, kde příroda vládne svrchovaně, se na tapisérii existence odvíjí závod o pochopení samotné podstaty života. Je to běh na trati, kde se možná nevyplácí přírodu porazit, protože by se mohla bránit. I tak se ale lidstvo žene za účelem sestrojení inteligence tak výkonné, že už její vývoj samotný zodpoví mnoho otázek dosud nezodpovězených a její výstupy zodpoví mnoho zbývajících. S tím přijde, jako už mnohokrát předtím, posun k hlubší úrovni poznání, která přinese nový pohled na realitu a nové třídy problémů a paradigmat, o kterých se nám ani nesnilo. Pokud se však podíváme na realitu zrakem stroje, nemine nás surový pohled do zrcadla naší vlastní existence obklopený úžasem, znechucením i strachem…

This would translate to English as:

June 16th, 2003

In a realm where nature reigns supreme, a race unfolds upon the tapestry of existence. It’s a race on a track where it may not be wise to defeat nature, for it could fight back. Even so, humanity strives to create an intelligence so powerful that its very development will answer many unanswered questions, and its outputs will provide solutions to the many of remaining ones. With this will come, as has happened countless times before, a leap to the deeper level of reality and a revelation of a new class of problems and paradigms we have never even dreamed of. Yet, when we peer into the fabric of reality through the lens of a machine, we find ourselves gazing into a raw reflection of our existence surrounded with amazement, disgust, and fear.

Back then, the idea sounded outlandish, a dream too far from reality. It’s 2025 now, and the world is awash with AI-driven headlines. It’s the perfect moment to revisit my scribbles and see how they hold up in the present day.

In this series, “23 Questions for AI”, I want to examine the idea of using AI to tackle the world’s biggest problems. But first, before any problem can be put on the table, we need to begin with a few fundamental questions:

• How can we formalize and describe the problems, or if you prefer, how should we ask AI models the right questions?

• How should we decode and validate the answers without harming ourselves? How can we determine if they are correct or not?

• What does it mean to be “correct,” “right,” or “true”? More importantly, is “truth” really the criterion we should use to validate the value of AI’s outputs?

• What role does language play in all this? Will we require a new generation of priests to interpret the AI outputs?

• What is the importance of data? Is it the whole, clear, and raw reflection of the world around us? How do we find models with the best fit on the data we have?

Join me on this journey to learn how language models resemble our consciousness more than we realize and why understanding the language vehicle and the concept of truth are crucial components of this equation.

Subscribe now

Twenty-Three

You are probably wondering—why 23 and not 42?

In 1900, mathematician David Hilbert stood before the International Congress of Mathematicians in Paris to present 23 problems that he believed were crucial to the development of mathematics in the coming century. Although some of those problems remain unsolved to this day, the ripple effect of Hilbert’s challenge reshaped mathematics in profound ways and spurred advances across number theory, logic, and beyond.

Every era has its own grand tasks. For mathematics at the dawn of the 20th century, it was Hilbert’s 23 Problems1—a blueprint for generations of thinkers who sought to push the boundaries of the known. Today, AI has reached a turning point that feels reminiscent of Hilbert’s era, and the problems to be solved call for an equally bold blueprint.

These are the ambitious questions that, once tackled, will shift our understanding of intelligence, consciousness, society, and the underlying fabric of reality itself. They are the questions that, in solving, propel us into an uncharted realm of possibilities.

The 23 is not meant to be a definitive or final number. Rather, it serves as a framework:

1. To focus collective effort on the hardest, most impactful challenges.

2. To illuminate unexplored paths where AI might catalyze entirely new paradigms of thought.

3. To celebrate a tradition of curiosity and intellectual ambition that stretches back into our history.

We know that new puzzles inevitably emerge as soon as one breakthrough occurs.

Problem Criteria

Not every interesting puzzle rises to the level of a problem. I propose four main basic criteria so we don’t propose bullshit:

• Universality: The question addresses a fundamental aspect of intelligence, knowledge, or reality rather than a niche application.

• Transformative Impact: A solution (or substantial progress) on this question would dramatically shift how we live, think, and frame other scientific and societal questions.

• Testability: Even if the problem is partly philosophical, there should be a theoretical or empirical path to test or validate its solution, progress, or initial hypothesis.

• Ethics: We should establish boundaries that reflect ethical values and respect for life, particularly for testing hypotheses.

I considered explicitness one of the criteria, but I’ve decided to put it aside for now. Forcing the question into a strictly formal structure could be counterproductive and limit the set of possible solutions. Let’s revisit it later.

Your Voice Matters!

Throughout this series, you’re invited to:

1. Propose new challenges, questions, and problems you believe could meet the criteria. Or challenge the criteria themselves!

2. Debate the feasibility or importance of the problems discussed.

3. Contribute to refining the final 23-problem list or the criteria. Perhaps your insights will highlight an angle no one else has considered.

4. Challenge!

Ultimately, this series aspires to crowdsource a grand vision that ties together the genius and curiosity of thousands of minds, just as Hilbert’s Problems rallied the mathematicians of the 20th century. Each problem we propose paves the way for new discoveries, fresh paradigms, and deeper insight into the nature of intelligence, ourselves, and our reality.

Leave a comment

Perhaps even some of the unsolved Hilbert’s problems will appear in the list we will build together…

I was recently asked to deliver a couple of talks about AI. I crafted two distinct yet interconnected decks tailored to the audience and the occasion. As both were somehow similar and somehow different at the same time, I originally wanted to share a narrated blend of both, published as a sequence of articles. Instead, I leveraged the original idea and split it into two articles— one based on the lecture’s commented transcript:

Another one based on a few extra slides that weren’t part of the original lecture:

But in the next session, I wanted it to be next level! The topic was up to me, and sometimes, the lack of specification opens up an opportunity for creative expression.

I decided to start by examining how humans model the reality around them to build the session’s skeleton.

Both talks are designed as immersive experiences for the audience rather than mere presentations. The aesthetic blends Matrix-inspired hues with the Courier New font, a staple in the world of computer programming. To blur the lines between human and machine perspectives even more, I intentionally employ hashtags instead of traditional headings.

But this article should be dedicated to the later version of the deck, where I discuss more topics and go into greater depth.

The graphics are created using Midjourney to add an additional layer to the content, striving to complement it by eliciting specific emotions and associations on each slide. As the talk mentions the art of accurately encoding problems and decoding solutions, finding the right imagery through prompting was very important to me, especially when contrasting with the potential of AI's capabilities.

No abstract, instead, the audience received a cognitive shock to engage them:

I chose a list of hashtags with relevance dictated by font size and order. They subtly transition to black, not meant for detailed reading but more to be perceived as a visual imprint on the audience’s retinas. They linger for a moment before we surge into the first, for a pretty deep topic, the Truth…

Subscribe now