Situational Awareness.

In 2024, a 22-year-old called Leopold Aschenbrenner (Leo) from OpenAI's Superalignment team departed the firm after allegedly leaking secrets to start a hedge fund called Situational Awareness. Leo, the so-called "Nostradamus of AI", believed that markets were underestimating AI's potential. He bet heavily that artificial general intelligence (AGI) was only years away, shorting companies like Adobe that he expected to be wiped out while borrowing heavily to buy long positions in AI-related firms.

In the calendar year through June, Leo's fund was up a massive 439%. Then in a single month, he learned the oldest lesson in finance: leverage doesn't care how smart you are. A relatively small downturn in AI stocks deflated the German wunderkind's fund almost as quickly as a single-stock leveraged ETF on Korea's KOSPI. By the end of July, Leo was desperately "asking his investors and lenders for more cash".

A margin call is brutal, but it only exposed a thesis that was already fragile. When Leo launched the manifesto underpinning his fund, it was big news and he had no problem securing funding from optimistic backers. I can proudly say I was not one of them, mainly because I wasn't invited but also because I was sceptical of his logic:

"For those [Leo's] claims to be true it actually requires believing that LLMs are on the path to AGI (they're probably not), and that the technology will keep growing exponentially with the same exponent. The latter is certainly plausible but assumes that we haven't already reached something close to the saturation of training data, that the costs of training and running LLMs won't also grow exponentially, and that the ~40% decline in the frontier model exponent we've seen in the last decade won't happen in the next decade, too.

If those assumptions don't hold indefinitely, then we could just as easily end up with something more closely resembling linear growth. Still a straight line, sure. But it would push Aschenbrenner's 2027 claim [of superintelligence] out so far as to not be worth worrying about."

There was also, and still is, an economic argument against Leo's investment philosophy. Even if we did achieve superintelligence, or AGI, there are countless forces that work against productivity miracles. Half a percentage point of extra growth a year, perhaps – still a huge improvement over a long enough horizon! – but nothing to justify the leverage that Leo was using for his bullish AI bets.

Being directionally correct doesn't save you from catastrophe. After the margin calls, Citadel picked up the scraps of Leo's fund for pennies on the dollar, leaving him with roughly $8 billion, down from around $30 billion at the start of the month. Whether you believe that there's actually that much remaining depends on how you value Anthropic, which accounts for five of the remaining eight billion in the fund (I'm wary of its valuation).

There's a good lesson here. Leo's fund blew up because he ignored margin call risk. He was right that AI stocks could soar and legacy tech could crash, but he was wrong about the path and the volatility. He was running a highly leveraged relative-value trade on a timeline he couldn't guarantee, and ended up getting squeezed by a simple short-term mean reversion.

Leverage is a cruel mistress, and fortunately for Leo he's young and has plenty of time to recover. By all accounts he's a nice guy, he didn't defraud anyone, his fund didn't go all the way to zero, and apparently he has "ceased using borrowed money, though he stopped short of pledging not to do so again". Hopefully he can still do some good with that big brain of his!

But this whole saga is a textbook cautionary tale. No matter how smart you are, too much leverage will eventually bite. Intellectual hubris combined with an ignorance of history seems to make such an outcome even more likely: the Nobel Prize winners who ran LTCM into the ground certainly weren't immune, nor was Leo, and neither will the next person who thinks they can safely ignore risk when investing.