Stuart Russell, a professor of computer science at the University of California, Berkeley, who has been an AI researcher for 45 years, said in a conversation with BT that while artificial intelligence is highly beneficial, it also disrupts the world. He said it has potential. It’s the worst when there are no guardrails installed. Edited excerpt:
BT: Let’s talk about the open letter to stop developing AI systems stronger than GPT-4 and develop guardrails. What prompted you to sign this letter?
We want to stop deploying Large Language Models (LLMs) that are more powerful than those already released. The reason is simple. Because I don’t understand how these systems work.
So what is an LLM? It’s a computer program that predicts the next word given the sequence of previous words. And you can have a conversation with that system. The methods of building these systems require large amounts of training data. For GPT-4, I think: [it is around] The amount of text is 20-30 trillion words, about the amount contained in every book ever written by mankind.
And we start with what seems like a blank slate: a gigantic circuit with about a trillion or more parameters. And the process of making about 1 billion or 10 trillion small random permutations on those parameters will gradually improve the system and improve its ability to predict the next word. The result is what appears to be an intelligent being when spoken to.
BT: The letter calls for a six-month moratorium on development of these tools. The genie is already out of the bottle, do you think that’s enough?
I agree to some extent that the systems out there can already cause problems. The petition asks us not to release systems that are more likely to cause problems. So 6 months is not enough.what we are looking for [to] Create reasonable guidelines that the system must meet. If you can’t build a plane that doesn’t fall out of the sky, you can’t carry passengers. This is common sense. We are just asking to apply common sense in the case of these very powerful AI systems…I think the possibilities for AI to benefit the world are endless. But if we have Chernobyl… Chernobyl destroyed the nuclear industry… we don’t want that. [for AI].
BT: The concern of everyday consumers is that this will eventually replace me and take my job.
There is a good chance it will have a big impact. Here are some examples. One is the field of computer programming. It may be surprising that advances in technology have made computer programmers unnecessary. But the numbers I’ve seen suggest that you can create software 5 to 10 times faster with these tools than if you didn’t do anything. And most of the time you just say what you want the program to do. The software just writes it. So, to me, the world is unlikely to need five or ten times as much software as he does. This means that slightly fewer computer programmers are required.
Thinking of people working in a company as nodes in a network, [then] What goes into that node? An email, a call from your boss, a request from a customer, and so on. what will come out? It’s the language of documents, sales invoices, reports to your boss, and so on. It’s all language. So, in principle, any of these jobs could be replaced.
But we don’t leave those tasks to a psychotic six-year-old living in a fantasy world. So unless you’re her psychotic 6-year-old living in a fantasy world, I don’t think your whole job is immediately in jeopardy. We cannot trust these systems to tell the truth because they are hallucinating…they just want to look plausible and have no idea what is true and what is false. However, there are thousands of companies working to solve these problems so that they can be used in critical applications. So I think the next generation will have an even greater impact on employment.
BT: Could the AI hallucinate or spread misinformation when fed malicious code or training data?
absolutely. All you have to do is ask them to generate false information. When I say, “Write me a letter that convinces someone that the earth is flat,” it does pretty well. They tried to impose some kind of constraint, but people found it very easy to ask questions in a different way. And that information is in the training set so you can finally get the answers you need.
So I think the level of unpredictability in these systems exceeds anything we’ve seen in AI software before, but we’re only a year or two away. I need to figure out what is going on. And, to be honest, I think we need to start pursuing a different avenue for designing AI systems.
BT: Many countries are starting to talk about regulating AI. When you talk about tools like ChatGPT, Google Bard, do you think regulation is the way forward? Is it possible?
The European Union AI law is expected to be passed by the end of this year. I have worked with bill drafters and governments. [European] Congress and [European] The Commission has spent the last few years trying to make sure that it makes sense that the bill is not repealed before it is passed. And to my knowledge, systems like ChatGPT are probably not legal to use in high-stakes applications. The law defines high-stakes applications as systems that can have a significant impact on people. And we want you to take steps to demonstrate that the system behaves in a safe, predictable manner, is accurate, fair, and free of racial bias. I don’t think there is any way to show that these LLMs meet these criteria. standard. Interestingly, OpenAI’s GPT-4 webpage recommends that these systems should not be used in high-stakes applications.
BT: Elon Musk, one of the signatories to the open letter, has been vocal about how AI and the AI revolution are shaping up. Is Mr. Musk’s assessment correct?
Basically yes. What Elon is saying is that until he finds a way to control a system that is more powerful than they are, he faces a very serious risk of developing an AI system that is so powerful that he cannot control it. It’s that I don’t know how. And it’s not like this never happened. For example, if you look at what’s going on with climate change, we’ve developed a system called “fossil fuel companies” that just so happens to have a human element. But basically, it’s an algorithm that maximizes the objective of quarterly earnings for shareholders. And that algorithm is destroying the world. And we cannot control it. So this is a scaled-down version of the kinds of problems we will face with AI systems in the future.
@aayush_a6
