[Premium] Ways AI Will Change the World, Part 1

FUTURE
The Ways AI Will Change The World, Part 1

Did you know that AI might be tearing down decades-old theories of human emotion?

That we are using AIs to discover signs of extraterrestrial life?

That we are using AIs to understand what dogs are saying when barking?

And that we could one day genetically modify humans to be more resistant to climate change thanks to AI?

Currently, it’s hard to discern anything other than Large Language Models (LLMs) and productivity-based applications, as most capital has gone in that direction. But AI much more than that.

Today, we go beyond the mainstream to describe four ways AI is changing or will change the world. While some will sound futuristic, please bear in mind that I’ve carefully selected the list based on actual research and that all these leverage AI’s only known true advantage over humans:

Its capacity to process vast amounts of data and extract regularities (the patterns) in that data to enhance our knowledge. Don’t expect human-level intelligence claims or robots doing backflips; these are all exciting cases where research and start-ups are making actual progress.

This article intends to serve as an interesting way of reevaluating your perspective on AI’s value, which might seem at an all-time low based on the insufferable hype regarding LLMs. But worry not; there are plenty of reasons to be extremely excited about AI.

Here’s why.

A Brief Description of What AI Really Is

Despite the extreme differences in the use cases we’ll see today, the underlying AI is strangely similar across all use cases. For that reason, I see it fitting to comment briefly on how AI works beyond the spotlight and superficial explanations you are usually exposed to through the media.

Although the explanation might seem a bit complex at first, please bear with me because it will allow you to truly understand the use cases we are seeing today, from astronomy to gene editing, in a more intuitive way than most.

Simply put, all AI does is take vast amounts of data and compress the underlying knowledge into something we call a ‘model.’ As the model is smaller than the data it processes, it’s forced to find the common patterns across data to compress the bits of information that truly matter.

But how do AIs actually process data? In our current paradigm, they treat every data source as a sequence of tokens, a term you’ve probably heard of many times. These are data patches that have intrinsic meaning but are then enriched with other surrounding tokens, turning data into a sequence of tokens that, combined, have an overall meaning that AI aims to capture.

For instance, text can be broken into words or subwords, images into groups of pixels, video into sequences of frames that are then broken down into groups of pixels, and so on.

And once the data is ‘tokenized’ and ‘embedded’ (represented in numerical form so that a classical computer can process it), AI uses two operations:

• Mixing operator. While tokens have intrinsic meaning, their overall meaning depends on their surrounding context (other tokens). Thus, we use an operation called attention (other popular operators include convolutions, especially for computer vision), where each token talks to other tokens in the sequence and chooses which of these it pays attention to. For instance, nouns may use attention to update their meaning based on surrounding adjectives. For example, the token ‘river’ has a meaning by itself, but if the text shows ‘the red river;’ through attention, ‘river’ now has the attribute ‘red.’ This way, the model encodes a global meaning of the data, be that a sentence, an image, or a dog bark.

• Knowledge enrichment operators. As models process more data, they learn about it. Therefore, they can add relevant knowledge to tokens in the sequence that might not be present in the actual sequence. For instance, in “Humankind reached the moon in…“ the year is absent in the sequence. Hence, the model uses its previous knowledge to enrich the tokens in the sequence so that the next predicted token is 1969. In other words, the models can add necessary information not present in the sequence.

Long story short, whenever you see ‘AI is being used to…,’ what AI is always doing, whether that’s discovering new exoplanets or understanding the human genome, is breaking down data (independently of its type) into pieces of context called tokens, and ‘enriching’ the meaning of these tokens through data present in the input and in the model’s knowledge to gradually build an understanding of the data and make useful predictions based on that understanding.

If it appears to be a very simplistic approach, well, I’m sorry to bring this news to you: AI is surprisingly simple and unamusing. In fact, if there’s something I want you to take from today besides a more optimistic view about the value AI will bring is that despite the seemingly unrelated nature of the cases we are about to see, the underlying AI is fairly similar.

Uncovering The True Human Emotions

As I’ve discussed many times, I’m much more excited about AI’s capacity to discover than to make us more productive. And I can guarantee that time will prove me right. For the time being, it’s advancing beyond language to explore and redefine the study of human emotion, or challenge our decades-long understanding of it.

The Theory

This approach challenges long-standing theories like Paul Ekman’s Six Basic Emotions and Dimensional Models, which classify emotions as fundamental categories or along dimensions like valence and arousal, which, although somewhat accepted, have been built in a deductive way, not fundamentally based on empirical data but on theories of the mind.

In contrast, new AI research, specifically Semantic Space Theory (SST), proposes an inductive approach, letting data, rather than assumptions, shape our understanding of emotions.

In other words, instead of humans imposing our biases on ‘what we think emotions are and how they are classified,’ we let the data give us the actual classification of all the emotions humans can show. 

This research not only deepens our understanding of human emotion but may also enhance applications in diagnosing diseases through voice or detecting lies based on speech patterns.

The Progress

The company that is probably more committed to this field is HumeAI. The start-up offers a platform to leverage emphatic voices to create AI apps and ‘measure emotion’ based on six areas like speech prosody, vocal, and facial expression.

Across these, they have created emotion maps that categorize them into many more categories than what traditional theory implied:

The Challenge? Funding.

As we have discussed several times, most capital today flows into productivity-based applications like LLMs, diffusion models, RAG vector databases, etc. Mapping emotions sounds amazing, but its perceived economic value is inferior to the vision of automating jobs or disrupting corporate America.

Worse off, AI needs a lot of data to work. And without capital, there’s no data. Although HumeAI did manage to raise $50 million in its Series B, that’s 120 times less than what xAI or OpenAI have individually raised this past year alone.

And the fact that Inflection, the only LLM company that tailored its experience to show more empathy with its LLM Pi, had to be acquired by Microsoft due to lagging demand doesn’t give hope that investors will be excited about emotion mapping anytime soon.

Gene Editing & Drug Discovery

Our next use case is so powerful that it has led to Nobel Prizes. As we mentioned earlier, almost any data source, including the human genome and amino acid sequences, can be broken down into individual semantic pieces of information called tokens.

And why does this matter?

The human genome holds the secret to understanding the blueprint of life. It encodes all the instructions that build and maintain the human body. Each cell contains a copy of this genome, which comprises DNA sequences that shape everything from physical characteristics to complex biological functions.

On the other hand, proteins, the workhorses of cells, are formed by sequences of amino acids that fold into varying structures, which are intrinsically linked to their role in the organism.

Fully understanding both may hold the secret to curing diseases, developing new drugs, or even modifying our genes to make us more resistant to outside temperatures, and AI is playing a major role in both.

But how?

The progress

Overall, we are seeing progress in two ways: protein folding prediction and biological foundation models.

The former has already seen valuable contributions by AI with models like AlphaFold, which takes in a sequence of amino acids and predicts protein structure. According to Demis Hassabis, one of two Nobel laureates behind AlphaFold, this model can speed up protein-folding prediction considerably, a task that, prior to AI, could take an entire PhD dissertation for a single protein.

As for the latter, companies like the Arc Institute have presented biological foundation models like EVO, which I discussed in detail a few months ago, that help us identify critical patterns in the human genome.

Among the many things EVO can do:

• it can predict gene mutations,

• design synthetic CRISPR-Cas9 molecules essential for the emergent field of gene editing (substitute/insert new gene structures into a human genome to cure illnesses),

• and predict ‘gene essentiality,’ effectively assessing the importance of individual genes crucial for an organism’s survival.

Long story short, AI is helping us understand ourselves and could lead to a future where we can actively engineer our own genetic code.

The challenge? Models.

While patterns in language are most often local (the closer two words are, the more related they usually are), patterns in the human genome or amino acid sequences can be extremely global. In layman’s terms, two genes in the human genome could be related despite being millions of nucleotides (the basic units of DNA, in the case of EVO one token equals one nucleotide) far from each other.

This is a massive problem because the model must ingest huge amounts of data (basically the entire human genome) in the hundreds of thousands of tokens to find these key relationships.

As I mentioned earlier, the underlying structure (sequence of tokens) is fairly similar, be that text or genes, which means that the costs of actually running these models grow considerably in this use case, making the overall process intractable in some cases.

Thus, while AlphaFold is already making strides in biology (although no AI-discovered drugs have been FDA-approved yet), biological foundation models still need further development.

But if you think this is fascinating, wait to see what researchers are doing in astronomy and the animal kingdom.

Astronomy

Astronomers used AI to show us the very first image of a black hole in 2019. Reconstructing images using an AI algorithm called CHIRP, developed by lead author Katherine Bouman, led to one of the defining images of the 21st century (although the achievement has been cited as not entirely accurate very recently).

But astronomy, the scientific study of celestial objects, space, and the physical universe as a whole, has a long history with AI.

Using AI To Teach Us About the World

In fact, its use of neural networks, the algorithms that underpin most AI progress today, can be dated back to 1990. Today, neural networks are used particularly to process vast amounts of data (what a surprise), especially considering that in astronomy, we are literally studying space.

For instance, every new image that the Vera Rubin Observatory telescope generates would require 1,500 large screens. Over ten years, it will generate 0.5 exabytes of data, about 50,000 times the information available in the Library of Congress. Compared to LLMs, that is 8,333 times the data used to train Llama 3.1 models, which required 15 trillion tokens of text at 4 bytes per token (around 60 TeraBytes).

Before AI, astronomers would have to check for new patterns using their own eyes, which is an intractable problem given the data's size.

But what are AIs real astronomy achievements? Well, besides showing us a picture of a black hole, it has led to other fascinating discoveries, from galaxies to exoplanets and… aliens?

Become a paying subscriber of Full Premium package to get access to this post and other subscriber-only content.

Already a paying subscriber? Sign In.

A subscription gets you:

• • NO ADS
• • An additional insights email on Tuesdays
• • Gain access to TheWhiteBox's knowledge base to access four times more content than the free version on markets, cutting-edge research, company deep dives, AI engineering tips, & more