Leopold Aschenbrenner Goes Over AI’s Future

When you look at the dizzying pace of AI progress over the last few years, you can ask one of two questions (or both): 1. How did this happen? or 2. What will it do?

Recently, I was introduced to Leopold Aschenbrenner’s opus magnum on AI and “situational awareness” through a Substack post (“Faster, Please!”) by James Pethokoukis. Reading through the long and detailed essay by Aschenbrenner, you can get answers to both of the above questions, with some pretty interesting explanations of why we can expect singularity events in the near future.

Big Jumps in Competence

First, Aschenbrenner urges us to simply count the “OOMs” or orders of magnitude. He defines an order of magnitude as a 10X phenomenon. I was first introduced to 10X with the idea of a “10X programmer” – someone who has 10 times more productivity than his or her colleagues.

Now that AI is writing code, that could be sort of an obsolete framework, because what Aschenbrenner goes into is the massive numbers of OOMs that show us where we’re going towards artificial general intelligence or superintelligence.

Growing Like People

One thing that Aschenbrenner keeps coming back to, time and time again, is analogies to stages of human growth – a preschooler, a kindergartener, a high-schooler, an adult.

He talks about preschooler to high-schooler jumps, suggesting that we’re going to see these happen routinely.

“It is strikingly plausible that by 2027, models will be able to do the work of an AI researcher/engineer,” he writes. “That doesn’t require believing in sci-fi; it just requires believing in straight lines on a graph.”

Again, he says, just count the OOMs, and you’ll see where we’re headed…

Measuring Success

Another major claim in Aschenbrenner essay is that we’re simply running out of benchmarks. This rings true to me, because I wrote about foundational data sets weeks ago, with various models scoring higher than most people.

Aschenbrenner has the same sort of thing in mind, talking about the MATH set that’s commonly used to identify intelligence levels. Notably, he doesn’t mention ARC, where up to quite recently, models couldn’t do very well on pattern recognition problems. But they’re cracking this, too. Just ask Francois Chollet.

In terms of broader resource identification, Aschenbrenner lists three major elements of advances – compute, algorithm efficiency, and what he calls “unhobbling further gains.”

The kinds of progress we expect, he claims, will make Moore’s law seem “glacial,” and again, he urges us to simply count the OOMs.

The Unhobbling Process

Later in the essay, Aschenbrenner talks about how things like chain of thought and scaffolding enable AI agents to think smarter in ways that will unlock their true potential. I found this statement of his truly integral:

“Imagine if when asked to solve a hard math problem, you had to instantly answer with the very first thing that came to mind,” he writes. “It seems obvious that you would have a hard time, except for the simplest problems. But until recently, that’s how we had LLMs solve math problems. Instead, most of us work through the problem step-by-step on a scratchpad, and are able to solve much more difficult problems that way. “Chain-of-thought” prompting unlocked that for LLMs. Despite excellent raw capabilities, they were much worse at math than they could be because they were hobbled in an obvious way, and it took a small algorithmic tweak to unlock much greater capabilities.”

Experts talk about it as “test-time compute during inference.” It’s the idea that the AI can take a moment to reason before answering a question in real-time. And as many point out, it’s a real game-changer!

The On-Boarding Problem

I think we should pay attention to this part of the essay:

“GPT-4 has the raw smarts to do a decent chunk of many people’s jobs,” Aschenbrenner writes, “but it’s sort of like a smart new hire that just showed up 5 minutes ago: it doesn’t have any relevant context, hasn’t read the company docs or Slack history or had conversations with members of the team, or spent any time understanding the company-internal codebase. A smart new hire isn’t that useful 5 minutes after arriving—but they are quite useful a month in! It seems like it should be possible, for example via very-long-context, to ‘onboard’ models like we would a new human coworker. This alone would be a huge unlock.”

The agent, he suggests, needs persistent memory. It needs a full context to ruminate on. It needs the details.

The bottom line seems to be this – that systems endowed with memory and greater context will continue to blow us away, as they move toward being more “human,” more fully cognitive in human ways.

One of the most compelling arguments that the author comes back to repeatedly is that just a few years ago, the models had trouble recognizing basic images of dogs and cats. They were primitive in a deeply profound way, and now they’re simply night and day different. If this keeps happening, we can expect many of Aschenbrenner’s other predictions to occur.

I’ll end with a quote that he put in the beginning of the essay from Ilya Sutskever, who famously worked at OpenAI with Sam Altman , prior to that controversy there:

“Look. The models, they just want to learn. You have to understand this. The models, they just want to learn…”

I guess one way to explain all of this is that the predictions made by people like Aschenbrenner aren’t just frantic scarifying or doom and gloom, or hyperbolic promotion – that they are reasonable things to expect, given the reality of what these models can do as they evolve. Aschenbrenner talks about a set of several hundred Cassandras, who are close to the industry and have to explain to everyone else that LLMs are not just “predicting a next token,” but doing the kinds of mental work that we ourselves will soon not be able to do, no matter how much we apply ourselves. So in two words: get ready.

I may return to this essay as I look at what we’re working with in 2025, because it seems really relevant to our times.