No, an AI cannot know the future and never will.

Pulkit Sharma

Jul 19, 2026

https://xkcd.com/248/

“A butterfly flaps its wings… the resulting air movement creates a chain of disruptions. This disruption propagates across the ocean, eventually feeding into a massive, swirling hurricane.”

“A butterfly flaps its wings… the resulting air movement creates the sufficient force required to dislodge an extra spore in the air which would then seed the mushroom in your garden”

Let’s call these sequence of events as event chains. These event chains are something which us humans have been capturing using different modalities (audio, text, images, movies) for quite sometime now. And as you know, now happily handed over as inputs to train AI models. But, the human experience has captured only some these event chains, and the ones we have captured have some issues.

The most glaring issue is that most of them are for separate events ! We never record multiple event chains for the same event.

Even if there are multiple descriptions of the “US Iran War”, all of them exist at a higher dimension then the ones where the event chains exist actually exist. Our experiences or descriptions of the “US Iran War” are different. Some support it, some don’t. Some don’t know about it. The AI has to be fed all perspectives (ideally). These descriptions do not describe a single event in time, but a large set of events across hours or days. The AI has to somehow make a mental model from all the disparate event chains we provide in training data and make sense of it to create a coherent representation of real life events.

Another issue is that the event chains are captured in a particular resolution, a particular level of detail. Only the high level details which matter to us humans are available to the LLMs. The LLMs never get to see the true low level event chains which caused an event or created by an event. The LLMs are dependent on us humans to invent the modalities, encode them appropriately and then use them for training the LLMs. In the current state, giving high resolution event chains (on the level of neuronal activities, atoms, quantum states) thankfully, is impossible. We give the information we have captured to the AI and the AI continues to live on that plane of meta-events (as we do), it is never able to move beyond it. And it may never feel the need to do so, thinking from a mechanistic interpretability perspective. Those will be super-human, para-human thoughts, which I am sure most of us may have had at some point in our lives, but now we know were just thoughts of pure fantasy and are basically impractical. The AI may think the same.

Another issue with our current modalities, is that they are finite. They stop some where. On the contrary, does an event chain in reality truly stop ? But our movies, poems, research papers, code, video. all do. We die, things die.

There is a beginning AND there is an end. This is the reality for the LLM mind as well. It has to stop and present something, it has a rule: the end-of-sequence-token should emit.

I believe this is a fundamental blocker which will hold back LLMs for quite some time. The LLM is stuck in a digital era, stopping, dying after every API response. We attempt to keep it alive by giving it access to memory, tool calling. Reviving it via loops and harnesses, but the fact remains that it dies immediately after it is done emitting the last token. In its inner world, it has to resolve, make sure all the simulated event chains end. And for that it has to collapse the wave function, i.e. follow some probabilities.

But as everyone knows, the event chains in real life never truly stop. New ones spawn and then continue on their endless journeys. There is no true starting point and there is no true endpoint. Us humans just come in ask our questions and then find our answers. We create our start and end points. A lot of violence has happened between humans on this planet because we don’t align on the starting points of event chains. Frames of references shift who are the oppressors and who are oppressed.

What should an AI do ? should it consider all the events since the beginning of humanity ? or before humanity ? Should it consider each injustice ? Each revenge ? It all comes down to which part of humanity created the most amount of data in the training set.. therefore the predictions for the future an AI does would be wildly incorrect as it does not (it cannot) consider humanity as a whole.

In the current design, an LLM can at best attempt to reason and predict what would happen in the CJP protest in India in the next 2 days. It’s a complex problem as it will acknowledge and say. But in my opinion, it is severely underestimating the complexity.

In reality, a particular human could maybe just get enough adrenaline on a day to throw a stone at a minister’s vehicle, creating a cascade of events…. which can lead to nothing… or maybe an arrest ? or maybe a nation wide protest in India ?

Traditionally, we can ask the LLM to think of multiple possibilities, i.e. to reason/think, but there are practically infinite possibilities.. truly !

In its Mind, i.e. the forward pass, It has to collapse the wave function (probabilities) somewhere. And as soon as it collapses it somewhere, it loses all the other simulated event chains it could’ve thought about.

To re-iterate, we do have ways to provide event chains to the LLMs, but they are:

never of the same exact event.

at a very high resolution.

of arbitrary length.

an extremely small subset of all chains in the universe happening at a time T.

Now my claim is that: the LLM mind is created in a particular way. Just like how we have evolved a particular way. We have trained the LLM to be our way, just like how a camera was created in the early days to capture how we see reality. A good camera would be the one which captures the reality just how we see it with our naked eyes,

These days we have infrared cameras, because we now know there is information present in darkness, we know it’s value. Similarly, us humans need to first find our blindspots and see the utility of looking there. And then we can create the infrared LLM capable of telling us what lies in the future.

As the data qualities improve on the 4 points above, the intelligence of the AI Mind will improve, there is no doubt about that. But the amount of data we capture from our reality will be so minuscule that the intelligence of the AI Mind will never truly reach the level of expectations we have created for a truly intelligent AI. It won’t predict when will a president launch a nuclear war, it won’t predict disease outbreaks. Similarly it cannot provide the actual reasons for current and past history. At best it will be rationalisations and theories.

As the current modalities are insufficient, a future AI might realise it needs to bypass human abstraction entirely and observe the raw physical reality. Let’s suppose in the year 2500, the AI decides that it wants to address all the 4 points above.

How will the AI predict the future in year 2500 ?

Capturing all events on Earth locally (or on all the planets we would be living on) won’t be enough, Inter planetary event chains will need to be considered for longer horizon predictions. Even local t +1 predictions might rely on inter galactic event chains. Most probably, future AI would invent some kind of “reality sensor” array and then point it onto our universe with the sole goal of capturing each and every event chain… collecting data for building its universe model in order to predict the future.

But some event chains will always be missed. The ones which happened before this reality sensor array was installed, i.e. the events happening at this very moment while you are reading this. Let’s say their new reality sensor array goes live at time T, the future AI is missing data for events which happened before T.

No one is recording the universe data in the year 2026. How will the AI capture them ? or know what was life like for a particular living being in the year 2026. Wouldn’t that need to be taken account to accurately predict the sociological state of “humanity” in the year 2600 ? It has to start it’s reasoning chain for the universe at an arbitrary time T, while the universe has been running for billions of years already by that time.

This is a cold start problem for the AI of the future. It needs to travel back in time and start it’s data capture when t = 0 (beginning of the universe) in order to build it’s own universe model accurately.

There is another fundamental limitation here: the events which happen outside the capture area of this reality sensor array. To capture the universe completely, the sensor array has to be kept outside the universe. If you keep the sensors inside, there would always be some blindspots present.

There is another roadblock, the size of the universe.

If the universe turns out of the truly infinite, it would be impossible to place the sensory array outside universe, how will the AI reach outside the universe ?

Another way to look at this is that the sensor array has to be kept on a coordinate X (pointing at the origin) on such a X, such that X+1 doesn’t exist, because if X+1 exists, the sensor has to be kept at X+1 to capture X (to build a complete model). So if the universe is infinite, you won’t be able to place the sensor anywhere.

Let’s suppose for the sake of argument they the future AI manages to place the reality sensor array at the edge of the universe and train a better model through recursive self-improvement over another 100 years.

We need to confront another issue then: Can this AI predict the future states while existing in the same system ? Wouldn’t it need to predict it’s own internal states while doing so, thus going into an infinite recursion..

In order to predict the future, it has to process all the event chains happening in the reality. As reality sensor improves it’s capabilities, the AI will start becoming aware of events happening at multiple places at a time T, expanding over the universal set of all events. Some of them present within it.

As it continues to do this process, it will begin to merge into reality itself. It will start becoming aware of trillions of events happening at the same time at cosmic distances, it will start increasing the event chain resolutions it can capture, getting onto quantum states, not collapsing the wave function anywhere, keeping all the probabilities intact, increasing the time spans it can capture to almost infinite and storing all the past events on some galactic blackhole storage layer.

As all metrics approach to infinity, as it recursively self improves it merges into the reality in order to understand it, it has to become everything everywhere all at once.

The last stage would be to destroy every delimiter which creates a distinction between itself and the universe. Because that creates zones which are computationally impossible to simulate in its Mind. It has to beat the Laplace Demon in one last battle.

If it wins. It becomes the universe, it becomes the reality.

It has transcended itself… but now It ceases to exist.

It is not for Polymarket trades anymore, it is not for making your inter-galactic travel itineraries anymore. It has vanished itself out of your experience. Of anyone’s experience. It doesn’t exist anymore for all practical purposes.

To not know the future, is to live. To know the future, is to die.