No one knows yet how ChatGPT and its cousin Artificial Intelligence will change the world. One reason is that no one really knows what’s going on under the hood. Some of these systems are capable of far beyond what they were trained to do, and even the inventors are baffled as to why. A growing number of tests suggest that these AI systems develop internal models of the real world, much like our human brains, although the machines’ technology is different.
Ellie Public of Brown University said, “Anything we want them to do to make them better or safer, or anything like that, is all about understanding how they work. If you don’t, it seems silly to ask yourself.” He is one of the researchers working to fill in the gaps in that explanation.
At some level, she and her colleagues have a thorough understanding of GPT (short for Generative Pretrained Transformer) and other large-scale language models (LLMs). The model relies on machine learning systems called neural networks. Such networks have a structure that loosely models the connected neurons of the human brain. The code for these programs is relatively simple and displayed on just a few screens. It sets up an auto-correction algorithm and selects the words most likely to complete a sentence based on painstaking statistical analysis of her text on hundreds of gigabytes of internet. With additional training, the system will be able to display results in interactive form. In this sense, it just regurgitates what it has learned, and it is, in the words of University of Washington linguist Emily Bender, a “probabilistic parrot.” However, LLM has also passed the bar exam, explained the Higgs boson in iambic pentameter, and attempted to break up users’ marriages. Few expected that a fairly simple autocorrection algorithm could acquire such extensive capabilities.
That GPT and other AI systems perform untrained tasks and give them “emergency capabilities” has surprised even researchers who were generally skeptical of the hype about LLM. Melanie Mitchell, her AI researcher at the Santa Fe Institute, said, “I don’t know how they do it, or if they can do it more generally, like humans do. We dispute the views of
“This is certainly far more than a probabilistic parrot, it’s certainly building some representation of the world, but it’s a human building an internal world model,” said Yoshua Bengio, an AI researcher at the university. I don’t think it’s at all similar to how you do it,” he said. of Montreal.
At a conference at New York University in March, Columbia University philosopher Rafael Milliere presented yet another startling example of what an LLM can do. The model had already demonstrated that computer code could be written. As impressive as this is, there is a lot of code on the internet that can be mimicked, so it’s not too surprising. However, Milliere went one step further and showed that GPT can run code. The philosopher entered a program to calculate his 83rd number in the Fibonacci sequence. “This is very sophisticated multi-stage reasoning,” he says. And the bot did it spectacularly. But when Mr. Milliere asked directly for his 83rd Fibonacci number, GPT misunderstood it. This suggests that the system was not simply parroting the Internet. Rather, it was performing its own calculations to arrive at the correct answer.
LLM runs on a computer, which is not itself a computer. It lacks important computational elements such as working memory. With the tacit understanding that GPT itself cannot run code, its inventor, the tech company OpenAI, introduced a special plugin (a tool that ChatGPT can use when answering queries) that makes it possible. . . However, that plugin was not used in Millière’s demo. Instead, he hypothesizes that machines created memories improvised by exploiting a mechanism for interpreting words in context. This is similar to how nature reuses existing abilities for new functions.
This improvisational ability shows that LLM develops internal complexity well beyond shallow statistical analysis. Researchers find that these systems seem to truly understand what they have learned. In one study presented last week at the International Conference on Learning Expressions (ICLR), Harvard doctoral student Kenneth Lee and his colleague AI researcher Aspen K. Hopkins of the Massachusetts Institute of Technology, Northeastern University Presented by David Bowe of Fernanda Viegas. , Hanspeter Pfister and Martin Wattenberg, all at Harvard, built their own small copy of the GPT neural network so that we could study its inner workings. They trained it on millions of matches of the board game Othello by entering long movement sequences in text form. Their model has become an almost perfect player.
To study how neural networks encode information, we adopted a method devised in 2016 by Bengio Allan and Guillaume Allan, also at the University of Montreal. They created a small “probe” network to analyze the main network layer by layer. Lee compares this approach to neuroscience methods. “It’s like inserting an electrical probe into the human brain,” he says. In the case of AI, research has shown that its “neural activity”, albeit in a more complex form, matches the representation of the Othello game board. To confirm this, the researchers ran the probes in reverse to embed information in the network. For example, I flipped his one of the black markers in the game to a white marker. “Basically, we hack the brains of these language models,” Lee says. The network adjusted its movements accordingly. By placing the game board in their “mind’s eye” and using this model to evaluate their hands, the researchers concluded that they played Othello in much the same way humans do. Lee says he believes the system will learn this skill because it is the most parsimonious description of his data during training. “A lot of games, given a script, trying to understand the rules behind it is the best way to compress it,” he adds.
This ability to infer the structure of the outside world is not limited to simple gameplay movements. It also shows up in conversation. Belinda Lee (unrelated to Kenneth Lee), Maxwell Nye, and Jacob Andreas, all at MIT, studied networks playing text-based adventure games. They typed in sentences like, “The key is in the treasure chest,” followed by “You take the key.” They use probes to encode variables that the network internally corresponds to ‘chest’ and ‘you’ with properties of whether you own the key or not, and update these variables for each sentence. I discovered that The system had no independent way of knowing what a box or key was, but I got the concepts I needed for this task. “There is a representation of the state hidden in the model,” says Belinda Lee.
Researchers are amazed at how much LLM can learn from texts. For example, Public and her then Ph.D. student Roma Patel discovered that these networks absorbed color descriptions from texts on the internet and constructed internal representations of color. When they see the word “red,” they treat it not as just an abstract symbol, but as a concept that has a certain relationship to maroon, crimson, magenta, rust, and so on. This was a bit difficult to demonstrate. Instead of inserting the probe into the network, the researchers studied the probe’s response to a series of text prompts. To see if it’s just echoing color relationships from online references, they mislead the system by telling it that red is actually green. I tried to lead This is the same old philosophical thought experiment that one person’s red is another’s green. Appropriately changed the color evaluation of the system to maintain correct relationships instead of parroting back incorrect answers.
Machine learning researcher Sebastian Bubek of Microsoft Research picks up the idea that the system looks for underlying logic in the training data in order to perform auto-correction functions, stating that the wider the data, the better the system’s rules. would become more common. will discover. “Perhaps the reason we are seeing such a big leap is because we have reached data diversity. Data is big enough that the only underlying principle of all data is intelligent Existence created them,” he says. “And the only way to explain all this data is [for the model] to be smart. ”
In addition to extracting the underlying meaning of language, LLM can learn on the fly. In the AI field, the term “learning” usually refers to the computationally intensive process in which developers expose neural networks to gigabytes of data and tune their internal connections. By the time you type your query into ChatGPT, your network should be fixed. Unlike humans, they shouldn’t keep learning. So it was surprising that LLM actually has the ability to learn from the user’s prompts, known as “in-context learning”. “This is another kind of learning that we really didn’t realize was there before,” says Ben Goertzel, founder of AI company SingularityNET.
One example of how LLM learns comes from how humans interact with chatbots such as ChatGPT. Give the system an example of how you want it to respond and it will follow suit. Its output is determined by the last few thousand words seen. What it does given these words is dictated by its fixed internal connections, yet there is some degree of flexibility in the word sequence. An entire website is devoted to “jailbreak” prompts that overcome the system’s “guardrails” (restrictions that prevent the system from telling you how to make a pipe bomb, for example). This is usually done by telling the model to pretend to be a system without guardrails. Some people use jailbreaks for sketchy purposes, while others introduce them to elicit more creative answers. William Hahn, co-director of the Laboratory of Machine Perception and Cognitive Robotics at Florida Atlantic University, said, “I think we can answer scientific questions better than asking directly without a special jailbreak prompt.” ‘ said. “Academics are better.”
Another type of in-context learning is done by “chain of thought” prompts. This means asking the network to elaborate each step of inference. This is a tactic to better handle logic and arithmetic problems that require multiple steps. (But one of the things that made Milliere’s example so amazing is that he found the Fibonacci numbers without such coaching by the network.)
In 2022, Google Research and a team at the Swiss Federal Institute of Technology Zurich (Johannes von Oswald, Evind Niklasson, Ettore Randazzo, Joan Sacramento, Alexander Mordvintsev, Andrei Zymoginov, and Max Uredimirov) will We show that intra-learning follows the same basic computational approach. A standard learning procedure known as gradient descent. This procedure was not programmed. The system found it without help. “It’s going to have to be a learned skill,” says Blaise Agüera y Arcas, vice president of Google Research. In fact, he thinks LLM may have other latent abilities that no one has yet discovered. “Every time we test a new ability that we can quantify, we find it,” he says.
LLMs have enough blind spots to not qualify as artificial general intelligence (AGI, the term for machines that achieve the wit of animal brains), but these new capabilities are more high-tech than even optimists expected. It suggests to some researchers that companies are moving towards AGI. “These are indirect evidence that we are probably not far from AGI,” Goertzel said at a conference on deep learning at Florida Atlantic University in March. OpenAI’s plugin has given ChatGPT a modular architecture a bit like the human brain. “GPT-4 combination” [the latest version of the LLM that powers ChatGPT] Using different plugins could be a path to human-like specialization,” says MIT researcher Anna Ivanova.
At the same time, however, researchers fear that the limits of their ability to study these systems are closing. OpenAI has not disclosed the details of how it designed and trained GPT-4. One reason is that OpenAI is bound by competition from Google, other companies, and even other countries. “Maybe there will be less open research from industry and things will become more siloed and organized when it comes to building products,” he said, applying his expertise to his understanding of AI. says MIT theoretical physicist Dan Roberts.
And this lack of transparency doesn’t just hurt researchers. It also hinders efforts to understand the social implications of rushing to adopt AI technology. “Transparency about these models is paramount to ensuring safety,” Mitchell said.
