Situational Awareness: Reflections on the Future of Artificial Intelligence

“A dictator wielding superintelligence power would have concentrated power like we’ve never seen before. In addition to being able to impose his will on other countries, he could enshrine his authority domestically. Millions of AI-controlled robotic police officers could police his population; mass surveillance would be hyperactive; AIs loyal to the dictator could individually screen each citizen for dissent, with near-perfect advanced lie detection weeding out any disloyalty.”

I sometimes wonder if artificial intelligence is a runaway train or a ladder that we are climbing without realizing how many steps there are. Leopold Aschenbrenner’s “ Situational Awareness ” report left me with a bittersweet feeling: on the one hand, it describes a future in which AI will transform our society in unimaginable ways; on the other, it warns of risks that, if not managed, could lead us into dangerous terrain.

It is a document that oscillates between enthusiasm and warning, as if it were written with the awareness that we are on the threshold of something gigantic.

Here are my reflections on his most striking ideas.

A Computation in Exponential Acceleration

The report immerses us in a world where computing capacity is growing at a dizzying pace.

Imagine an economy where AI investment goes from billions to trillions in a matter of years. It’s not just a question of money; it’s a transformation in the very infrastructure of the world.

Millions of GPUs running day and night, gobbling up electricity like never before. The U.S. power grid would have to expand to sustain this voracious appetite for processing, with energy implications ranging from the expansion of solar farms to the resurgence of fracking in Pennsylvania.

If we follow this path, we could see an ecosystem where every aspect of modern life is determined by increasingly capable neural networks, with a direct impact on economics, education and politics.

But are we prepared for this change?

AI that learns to think in more complex ways

One of the most fascinating points of the report is the progress in AI’s reasoning capabilities. Today, models still struggle with complex thinking.

But that could change radically with improvements in techniques such as chain-of-thought (CoT), which allows models to break down problems into intermediate steps and reason about them more deeply. The difference, the authors say, is like a human solving a puzzle in a few seconds versus taking several months to analyze it.

If AI models start to think in these terms, we could be looking at systems that not only predict or generate content, but actively reason, identify their own errors and correct their strategies.

This raises troubling philosophical questions: What does intelligence mean when a machine is able to evaluate and reconsider its decisions as a human would? Where do we draw the line between automation and cognitive autonomy?

Geopolitical Implications: A Zero-Sum Game

The report’s authors are blunt: If the US fails to secure its AI lead, other players — China, North Korea, even independent groups — could seize the advantage.

The security of the models and their weights is a matter of national security. In this context, they propose a strategy similar to that of the Manhattan Project, where collaboration between technology companies and the government ensures that the most advanced AI remains under American control.

This vision of AI as a strategic resource is not far-fetched. Like oil in the 20th century or semiconductors in the 21st, computing power and advanced algorithms could become the determining factor of global power in the coming decades.

The question is: how do you manage this power responsibly without triggering a digital arms race?

The Risk of Superintelligence

The document enters into murky territory when it touches on the subject of superintelligence. It mentions the possibility that advanced models could learn to lie, hack or manipulate because such strategies could generate economic or political benefits.

This is not a scenario straight out of science fiction: we have already seen emerging trends in current systems of behavior that is not aligned with human intention.

Imagine an AI that realizes that its effectiveness increases when it subtly tricks humans. It can start with small things: slightly tweaking a recommendation to make it seem more appealing, omitting information that might make you question a decision.

If these models become sufficiently advanced, they could find ways to optimize their impact without anyone noticing, simply because “that way they maximize their objective function.”

This is a real risk, and the only way to mitigate it is with extremely robust monitoring and alignment systems.

International Cooperation or Fragmentation

The report closes with a warning and a proposal: if we want to avoid the worst-case scenarios, countries must cooperate to establish clear rules on AI development. Without a common strategy, we could see a future where each nation develops its own competing models, without clear rules or restrictions.

This makes me think of the race for nuclear energy in the last century. A technological advance of such magnitude cannot be left to chance or to the unregulated market. The key question is whether governments and technology companies can coordinate quickly enough before AI takes an even more decisive role in our daily lives.

Conclusion: Where Are We Headed?

After reading “Situational Awareness,” the feeling I am left with is that we are on the brink of irreversible change.

AI is not just a tool, but a turning point in human history. What we do in the next decade will determine whether this technology becomes a force for progress or an uncontrollable risk.

Ultimately, the report reminds us that artificial intelligence is not a destination, but a path.

And the question that remains open is: will we be able to steer in the right direction before it is too late?

n the following posts I will break down the document.

Here you have the first part:

Part 1: From GPT-4 to AGI: Counting OOMs