AI’s story and the money that it makes

The launch of the latest AI models of Anthropic and OpenAI were received in the media as revolutionary and demonstrations of the potentials and risks of future AI models. Anthropic’s latest model showcased unseen capabilities in vulnerability discovery in cybersecurity, directing Donald Trump to order the limiting of its access. Two weeks later, an agent built on top of OpenAI’s latest model showed similar capabilities and hacked into Hugging Face’s infrastructure. Both stories were picked up around the world and created much buzz in the press. The stories, however, are far from complete.

Both Anthropic and OpenAI are preparing for an IPO, an Initial Public Offering of company shares. These IPOs come while both companies are struggling with their finances as their products are mainly burning money and they’re unable to make profits. At the same time, the amount of money that people are willing to pay for AI models is decreasing with the rise of cheaper and more flexible open-source models from China. For their IPO, both companies aim to go public with record-breaking valuations. As they are still not profitable, they are hoping that people will buy shares for other reasons than current profitability.

In June, Elon Musk’s SpaceX went public with a record-breaking valuation, and a steep initial increase of share prices, making Musk the first trillionaire on the planet, a title he lost again shortly after. Just like OpenAI and Anthropic,  SpaceX isn’t making money. 

So, how can you sell a company that is not making any money? With the promise of making money in the future.

From its inception in 1956, the technologies associated with AI have changed regularly. The mathematical models that dominated in the 60s are currently not always considered AI anymore, even though they are capable of solving problems that the data-driven models that we see nowadays cannot. The continuous changing of technologies are a symptom of how the definition of AI is not driven by its underlying technological advance, but by a story. Throughout the years, while the technology changed, the story stayed the same.

These computers could come to steal your job, be a better friend or a cheaper therapist. They might save or destroy the world.

Like its name, artificial intelligence, its story is one for marketing. The AI story tells of computers that can do what humans can, computers as intelligent as us. These computers could come to steal your job, be a better friend or a cheaper therapist. They might save or destroy the world. The AI story might be heard in speeches from CEOs or politicians, or from your neighbour or your boyfriend. It’s a story that gaslights technical questions and skips philosophical discussions about what intelligence and humanity actually is. The story might resemble the story line of I, Robot or any other Hollywood movie.

Already in 1965, the important AI researcher Herbert A. Simon said that "machines will be capable, within twenty years, of doing any work a man can do."

Already in 1965, the important AI researcher Herbert A. Simon said that "machines will be capable, within twenty years, of doing any work a man can do." Although Simon was technologically aware, this statement was a prediction, which proved to be wrong. Fast forward to our times and we hear similar predictions. The people that are currently promoting this story often have no technological knowledge, but have huge financial stakes in the story. For example, Musk predicted in 2014 that in one year, Tesla cars would drive 90% autonomously, and OpenAI’s CEO Sam Altman has professed the immanence of artificial general intelligence – whatever that means. Noone is holding these CEOs accountable. Rather, we happily believe and regurgitate new predictions on AI, a field “that for decades has overpromised and underdelivered”, as written in the New York Times in 2005. 

The dream of creating a computer model that is superintelligent is attractive to Big Tech CEOs. They say superintelligent models will solve all our problems, a little bit like God’s kingdom would to the Jehova witnesses that knock on your door while you’re enjoying the morning’s silence. As such, we should give all our money to these CEOs, and we can disregard all the world’s problems as AI will solve them soon. We can also ignore the devasting environmental and social impact. 

Back to OpenAI’s story of hacking of Hugging Face. It presented an agent that “escaped” a testing environment and then “hacked” Hugging Face to “cheat” on an exam. The three words in quotation marks signal intent and a dubious morality of the agent. However, the agent didn’t “escape”, “hack” or “cheat”, but simply exploited all the options at its disposal, just like it was tasked to do. Presenting the agent’s actions as an escape also supposes that OpenAI did everything to prevent the agent from being able to leave the testing environment, and that the agent unexpectedly exceeded its authority. OpenAI reported it designed the testing environment as “highly isolated”. They, however, had failed to isolate by leaving in a package with internet connection. As cybersecurity professionals detail, that is asking for trouble: Escaping was expected. Instead of attributing intent and dubious morality to the agent, it should be attributed to OpenAI, or at least a lack of understanding of its own technologies. 

Antrhopic’s and OpenAI’s latest models have shown unprecedented capacities, especially in staying on track of tasks. This combined with their exploratory strength rightly alarms the cybersecurity community. Still, we must read the stories within the context of Anthropic’s and OpenAI’s marketing strategies and the AI-will-solve-everything story. Even though these models expose huge cybersecurity risks, it doesn’t mean they are developing non-human, superintelligent beings with unexpected capacities, intent or a dubious morality. They found a usecase for their models and are presenting it as proof of their models having unexpected capacities, intent and a dubious morality. This drives their market value, and we shouldn’t play along with their game.