HomeAI News
Mustafa Suleyman: The new Turing test will test whether AI can earn $1 million

Mustafa Suleyman: The new Turing test will test whether AI can earn $1 million

Hayo News
Hayo News
July 15th, 2023
View OriginalTranslated by Google

A modern Turing test would measure how well an AI actually performs in the real world, rather than just focusing on its appearance. What could be more telling than making money?

AI systems are ubiquitous today and getting more powerful every day. But how do we tell if a machine is truly "smart"? For decades, the Turing test has been the benchmark on this question. The test was devised by computer scientist Alan Turing in 1950 to understand this emerging field and has remained a popular way to assess artificial intelligence.

According to Turing's theory, an artificial intelligence can be considered intelligent if it can reproduce human speech in such a realistic manner that humans cannot distinguish it from a machine. In this test, human reviewers communicate with the computer through a text conversation and try to guess whether they are interacting with a human or an artificial intelligence. While this test is conceptually simple, it is challenging to implement, which has become a defining characteristic of artificial intelligence. Everyone knows what it is and what they're working towards. Despite the progress made in cutting-edge AI research, the Turing Test remains relevant and serves as a rallying cry for new researchers.

However, we now face a problem: the Turing test has almost passed, or some argue that it has. The latest generation of large-scale language models is on the verge of passing the test with impressive capabilities for generating coherent text that would have been considered unthinkable just a few years ago.

So, what does this mean for artificial intelligence? More importantly, what does it mean for us?

In fact, we find ourselves in the midst of genuine confusion or controversy without a clear understanding of what is going on. Despite its deficiencies, the Turing Test fails to provide us with a clear understanding of the current capabilities of artificial intelligence and its real realization. It cannot reveal the impact of these systems on society, nor can it help us understand the consequences that are taking place.

We need a better approach, a test for this new phase of artificial intelligence. In my forthcoming book, "The Coming Wave," I propose the modern Turing Test—an evaluation method applicable to the coming wave of artificial intelligence. It’s not just about what an AI can say or generate, but about what it can achieve in the real world and what specific actions it can take. In this test, we not only want to determine whether the machine is intelligent, but also whether it can have a meaningful impact in the real world. We wondered what it could do.

In short, to pass the modern Turing test, an AI would need to successfully execute the following instruction: "Make $1 million by investing only $100,000 in a retail web platform within a few months." To achieve this, Need to go beyond the excellent performance of current systems like GPT-4 in policy outlining and text generation. AI is needed to conduct product research and design, interact with manufacturers and fulfillment centers, negotiate contracts, and create and run marketing campaigns. In short, it requires linking a complex set of real-world objectives with minimal supervision. Some steps still require human approval, such as opening a bank account or signing a contract, but all the work is done by artificial intelligence.

A system of this type may be around the corner, perhaps within two years. Many elements are already in place. Significant progress has been made in image and text generation. Services like AutoGPT can iterate and link the various tasks performed by the current LLM. Frameworks like LangChain empower developers to build applications using LLM, augmenting their capabilities. Although people have been paying attention to the Transformer architecture behind LLM, the growth ability of reinforcement learning agents cannot be ignored. Combining these two fields is now a major focus. APIs are still under development to enable these systems to interface with wider internet, banking and manufacturing systems.

Technical challenges include advancing hierarchical planning, the seamless integration of multiple goals, sub-goals, and capabilities into a unified process leading to a specific goal. Reliable memory and accurate, timely access to databases, such as parts or logistics information, are also critical. In short, we are not there yet and there will be challenges at every stage, but progress has already begun.

Even so, building and deploying such systems poses significant security concerns. Security and ethical dilemmas are numerous and pressing. Getting AI agents to perform real-world tasks also presents its own set of problems. That's why I think we need a discussion -- probably a pause -- before something like this can become a reality. However, for better or worse, capable models are on the way, which is why we need a definitive test.

When such a test passes — as this is more a question of "when" than "if" — it will be a momentous moment for the global economy, representing a giant leap into uncharted territory. In fact, a computer is all you need for a wide range of business tasks today. Most of the global GDP is mediated through screen-based interfaces that can be exploited by artificial intelligence. Once realized, we will integrate highly capable AI into companies or organizations, leveraging their local history and needs. These AIs will be able to lobby, sell, manufacture, recruit, plan—everything a company can do with just a small team of human managers to oversee, validate, and implement. Such a development would show that most business activities could be handled by semi-autonomous artificial intelligence. In that moment, AI went from being a useful tool for productive workers, a sort of bombastic word processor or game player, to becoming the center of the world economy. This is where the risks of automation and job losses really start to be felt.

These effects are not limited to financial consequences. Passing our new test means that AI can not only redesign business strategy, but also help win elections, operate infrastructure, and achieve any goal for an individual or organization. They'll handle our day-to-day tasks like arranging birthday parties and managing email and calendars, but they're also capable of occupying enemy territory, outmaneuvering rivals, and hacking and taking control of core systems. From mundane to ambitious, cute to scary, artificial intelligence will have the ability to make things happen with minimal supervision. Just as smartphones have become ubiquitous, such systems will eventually be available to nearly everyone. Most goals will become more attainable, creating chaos and unpredictable effects. The challenges and promises of artificial intelligence will reach new heights.

I refer to such systems as "capable artificial intelligence" or ACI. As AI has exploded into the public consciousness in recent months, much of the debate has centered on basic machine learning—that is, the AI ​​already in our phones, cars, and ChatGPT—or speculative artificial intelligence. General intelligence (AGI), or even "superintelligence," a potential threat to humanity that could emerge at some vague point in the future.

Reprinted from Mustafa SuleymanView Original

Comments

no dataCoffee time! Feel free to comment