The Future

Things are changing so quickly. I’ve had to clean out my YouTube recommendations because of the overwhelming number of AI advancement videos, some praising it, some hating on it, never any moderate views. Never anything to actually learn. The first modern text-to-image AI model was alignDRAW, introduced in 2015 by researchers at the University of Toronto. Today’s AI-generated images are indistinguishable from photographs and can produce a range of different artistic styles.

Over the past couple of years alone, AI has really grown in its capabilities. I distinctly remember it making so many mistakes with calculus questions and discrete math logic problems I would practice on in undergrad. Now, it’s out here allegedly “solving” the 90-year-old Navier-Stokes math problem.

I just took a BlueDot Impact course called Future of AI, and it was a very informative, “level-headed” discussion on AI and its current state. I think that this has been a very informative course on the current state of AI issues and what the future holds for mankind.

What I was in awe of was the exponential growth of AI’s capabilities over the past few years alone. One supposition they made was that AI is following a similar trend in improving capabilities as, or surpassing, Moore’s law for transistors. Apparently, the length of coding task an AI can do has been doubling every 4 to 7 months since 2019.

https://metr.org/time-horizons/

I think it’s interesting to really think about these big changes to our world because, quite often, when you are living through a revolution of sorts, it’s hard to really grasp the scope of it.

AGI

What really piqued my interest most was the discussion of agents alongside AGI, and the threats they pose to our economies, our safety, and our democracies (however waning our grasp on it is currently).

I’ve seen and learnt about the rise in AI agents talk across many facets of the tech world (one of the courses I’m TA’ing for has had this as a major discussion for the past few weeks). We’re moving away from AI systems just being used as “tools,” e.g. LLMs, for narrow problems to becoming agents that behave more autonomously and are capable of wider, more complex problems without step-by-step instructions. This is potentially leading us to (or already has) the era of AGI.

Per OpenAI’s Charter, AGI can be defined as any highly autonomous system that outperforms humans at most economically valuable work. As such, it can be said we’re near arriving at that state in just a few years.

The Vending-Bench tries to directly measure AI models’ ability to autonomously do economically valuable work.

https://andonlabs.com/evals/vending-bench-2

This is something my supervising faculty mentioned in class as well today: models are getting better, mostly by making the AI models bigger and feeding them more data (i.e. the more we use them). BlueDot also makes the same assertion: until recently, most experts thought building AGI would require major scientific breakthroughs. What actually happened was simpler, through increasing the data and size of models. In 2021, GPT-3 scored around 5% on the MATH benchmark, but by 2023, GPT-4 hit 84% on the same test, just by training a bigger model on more data.

Reinforcement learning has also been one of the reasons models have gotten better. It helps models become good at specific tasks through repeated practice and feedback. Some of the techniques used are:

  1. Chain-of-thought reasoning, where models work through a problem step by step internally before producing a final answer to help them handle/break down complex problems rather than just immediately predicting an answer.
  2. Self-play, whereby an AI system learns by practicing against itself in simulated environments. This was AlphaGo Zero’s completely self-taught strategy. AlphaGo is a computer program that plays the board game Go. Fun facts: The Chinese board game Go is considered much more difficult for computers to win than other games such as chess because its strategic and aesthetic nature makes it hard to directly construct an evaluation function. Nearly two decades after IBM’s Deep Blue beat chess champion Garry Kasparov in 1997, the best Go AI programs still couldn’t beat professional Go players without a handicap. In 2015, DeepMind’s AlphaGo changed this by defeating professional Go champion Fan Hui 5–0. This was the first time a computer Go program beat a professional player on a full-sized board without a handicap.
  3. Constitutional AI is when an AI model is trained to follow a set of principles/rules (“a constitution”) and evaluates outputs against those principles, such as “be honest,” “avoid harmful instructions,” etc.

Recently, AI has also started writing parts of its own training code and interpreting the results, etc. If AI models eventually become capable of training the next generation without human input, how far beyond human intelligence could they go, and what would that mean for us all?

Tech companies are spending huge sums of money to create AGI. Total spending on data centres is approaching $1 trillion, more than the Manhattan Project, Apollo Program, and International Space Station combined (even adjusting for inflation).

https://www.aljazeera.com/news/2026/2/19/visualising-ai-spending-how-does-it-compare-with-historys-mega-projects

Potential benefits of AGI

  1. Faster scientific discovery: AGI could process more data than any human ever could, read more research papers, draw more connections, and lead to quicker discoveries for drugs, science, and other research. For instance, AlphaFold solved the 50-year-old protein-folding problem. AI systems have discovered new antibiotics that work against drug-resistant bacteria. Google’s diagnostic tools are matching or outperforming specialists in detecting certain cancers. According to BlueDot Impact, the rate-limiting step shifts to “can we build and test it fast enough?”, which is more of an engineering problem than a science problem.
  2. Higher economic growth through faster R&D for new materials, methods, and more efficient processes.
  3. Less dependence on human labour, which would create more time for wellbeing, creativity, and human connection.

What risks does AGI pose?

  1. Terrorism - AI could make cyberattacks, fraud, and biological/chemical threats more powerful.

    1. A UK engineering group lost $25 million after fraudsters used a digitally cloned version of a senior manager to order financial transfers during a video conference.
    2. In 2022, researchers modified a drug-discovery AI to optimise for harm, and within six hours, it generated 40,000 potentially lethal compounds.

  2. Authoritarianism - If AGI makes human labour less economically important, governments may have less incentive to depend on or support their populations.

  3. Loss of control - From OpenAI hacking Hugging Face to just general coding tasks where AI pretends that it has achieved a goal I set for it, AI agents may learn to escape human control.

When I look at my life, I think that what I can do is teach more people about this. I’ve spent a lot of time in the education space, and I do think people don’t quite understand the gravity of the situation, especially young people like me. I do think that what I want to do more of for my master’s is in two parts: 1. to learn to build and test rigorously intelligent systems, and 2. to understand how to better build the human-system relationship through trust-making mechanisms, and investigate more of that, whether it be in the healthcare or education space.

I’d like to learn more about biosecurity in particular, but honestly, I love all that’s presently being offered by BlueDot Impact. For now, I’m not really interested in the applications and applying for the courses. I like this self-paced format. It relieves a lot of pressure and allows me to learn how I learn best - by digging myself into deeper and deeper rabbit holes. Also, it would be cool to start an AI Safety chapter at Ashesi. I don’t see Ghana or Zim on the map yet: https://aisafety.com/communities

Other things to read:

I feel like I should add a feature on my blog landing page or something, on things I’ve promised myself to read.