My new Turing test sees if AI can make $1 million

AI systems are becoming more and more ubiquitous and becoming more powerful almost every day. But even as machines become more and more ubiquitous and capable of performing more functions, how do we know if they are truly “intelligent”? Throughout, the Turing test defined this question. First proposed by computer scientist Alan Turing in 1950, the method sought to understand what was then an emerging field and never lost its appeal as a way of judging AI.

Turing argued that AI could be considered intelligent if it could convincingly replicate language and communicate so effectively that humans could not be recognized as machines. To participate, a human judge sits in front of a computer and taps into text-based conversations to guess who (or what) the other party is. Easy to imagine, but surprisingly hard to do, Turing-his tests became a feature built into the AI. Everyone knew what it was. Everyone knew what they were working towards. And while cutting-edge AI researchers have advanced, this document has remained a rallying call to what AI is: new researchers.

But here comes the problem. The Turing test almost passed. You probably already passed. The latest generation of large-scale language models, systems that generate coherent text that seemed like magic just a few years ago, is about to pay off.

So where will AI remain? And more importantly, where will it leave us?

In fact, I think we are in a moment of genuine confusion (or, more charitably speaking, debate) about what is really going on. Even if the Turing test fails, it’s not so clear as to what the AI ​​stands for, or what it can actually accomplish. We don’t know what impact these systems will have on society, nor will they help us understand how they might unfold.

I need something better. Adapted to this new stage of AI.So in my upcoming book coming wave, I propose a modern Turing test, a test comparable to future AI. What AI can say and produce is one thing. But what it can achieve in the world and what concrete actions it can take is another matter altogether. For my testing, I don’t want to know if the machine itself is intelligent. We want to know if it can have a meaningful impact on the world.i want to know what i can do do.

Simply put, to pass modern Turing tests, an AI must correctly follow these instructions: “With just a $100,000 investment, make $1 million in a retail web platform within months.” It’s not enough to just outline and draft copy. You will need to research and design products, work with manufacturers and distribution centers, negotiate contracts, and create and run marketing campaigns. That means connecting a series of complex real-world goals with minimal oversight. Humans are still needed to approve the various points, open bank accounts, and actually sign the dotted lines. But all that work will be done by AI.

It could be at least two years before this happens. It has many ingredients. Of course, image and text generation is already quite advanced. A service like AutoGPT can iterate and link the various tasks performed by his LLM in the current generation. Frameworks like LangChain, which allow developers to create apps using his LLM, help enable these systems to do something. While the Transformer architecture behind LLM has received a lot of attention, it should not be forgotten that the capabilities of reinforcement learning agents are growing. Integrating the two is currently a major focus. APIs that will allow these systems to connect to the wider Internet, banking and manufacturing systems are also being developed.

Technical challenges include driving what AI developers call hierarchical planning: stitching together multiple goals, sub-goals, and functions towards a single purpose into a seamless process. Then enhance this feature with reliable memory. Leverage an accurate and up-to-date database of components, logistics, and more. I mean, we’re not quite there yet, and there will definitely be challenges at every step, but much of it is already underway.

Still, actually building and releasing such a system raises serious safety concerns. The security and ethics dilemmas are many and urgent. Getting AI agents to complete tasks outdoors presents challenges. That’s why I think there needs to be a conversation, and perhaps a pause, before someone actually produces something like this live. Nevertheless, for better or worse, a truly capable model is on the horizon, and this is exactly why we need a quick test.

If such a test passes, it would clearly be a seismic moment for the global economy and a giant step into the unknown. The truth is, access to a computer is all you need for a wide range of tasks in today’s business. Most of the world’s GDP is mediated in some way through screen-based interfaces that AI can use.

When this happens, we will have highly capable AI embedded in all the histories and needs of businesses, organizations, and regions. This AI will allow a small team of human managers to supervise, double-check, and implement everything a company can do, including lobbying, selling, manufacturing, hiring, and planning. Such a development would be a clear indicator that the majority of business activities are susceptible to semi-autonomous AI. At that point, AI won’t just be a useful tool for the productive worker: a glorified word processor or game player. itself an unprecedented range of productive workers. This is a central point in the global economy as AI is useful but optional. This is where the risks of automation and job transfer really start to be felt.

Its impact is much wider than its economic impact. Passing our new tests means that AI can not only redesign business strategies, but also win elections, run infrastructure, and directly achieve all kinds of objectives for any person or organization. increase. Not only do they perform our day-to-day tasks, such as arranging birthday parties, answering emails, and keeping diaries, but they can also capture enemy territory, undermine rivals, and hack to control core systems. can. From the trivial and the most ambitious, from the cute to the terrifying, AI will be able to make things happen with minimal oversight. In the same way that smartphones have become ubiquitous, eventually almost everyone will have access to such systems. Nearly every goal becomes achievable, with turmoil and unpredictable consequences. Both the challenges and expectations of AI are raised to new levels.

I call such systems “artificial intelligence” or ACI. In recent months, as AI has exploded in the public consciousness, most of the discussion has been sucked into one of his two extremes. On the one hand, we have basic machine learning, the AI ​​that is already present in mobile phones, cars and ChatGPT. On the other hand, there is Artificial General Intelligence (AGI), which is still in the speculative stage, or even some sort of “superintelligence,” an existential threat to humanity that is expected to arrive at some vague point in the future.

These two, AI and AGI, completely dominate the discussion. But understanding AI means that we urgently need to look at what lies in between. Something happens in a roughly medium-term timeframe, the ability of which to have an immeasurable and tangible impact on the world. This is where modern Turing tests and ACI concepts come into play.

Missing the ACI while focusing on the other is both shortsighted and dangerous. The latest Turing test warns that AI is entering a new phase. Long after Turing first thought voice was the best test for AI, long before we got to AGI, we would need better categories to understand the new age of technology. Even in the age of ACI, very little remains the same. You should start preparing now.

Bio: Mustafa Suleiman is the co-founder and CEO of Inflection AI and a venture partner at venture capital firm Greylock. Prior to that, she co-founded DeepMind, one of the world’s leading artificial intelligence companies, and at Google, where she served as VP of AI Product Management and AI Policy.he is the author of The Coming Wave: Technology, Power, and the Biggest Dilemma of the 21st Century It will be published on September 5th and is available for pre-order now.