After the overwhelming emergence of large models, computer science finally became a 'natural science'

The current artificial intelligence (AI) is in a wonderful era, and amazing tacit knowledge often appears (Polanyi’s Revenge and the new romance and tacitness of artificial intelligence) Knowledge, https://bit.ly/3qYrAOY), but it is certain that computers will not be able to complete this task for a long time in the future. Interesting research that has recently emerged is on large-scale learning systems based on the Transformer architecture, based on large network-scale multi-modal corpora and billions of parameters for training. Typical examples are large language models, GPT3 and PALM that respond to arbitrary text prompts, language/image models DALL-E and Imagen that convert text into images (even models with general behavior like GATO).

The emergence of large-scale learning models has fundamentally changed the nature of artificial intelligence research. When researchers recently used DALL-E, they believed that it seemed to have developed its own unique language. If humans could master it, they might be able to interact with DALL-E better. Some researchers have also found that GPT3's performance on reasoning problems can be improved by adding certain magical spells (such as "Let's think step by step") in the prompt. Now large learning models like GPT3 and DALL-E are like "alien species" and we have to try to decode their behavior.

This is undoubtedly a strange turning point for artificial intelligence. Since its emergence, artificial intelligence has been a "no man's land" between engineering (systems with specific functions) and science (discovering the laws of natural phenomena). The scientific part of AI stems from its original claims, which were insights into the nature of human intelligence; while the engineering part stems from a focus on intelligent capabilities (allowing computers to exhibit intelligent behavior) rather than insights into human intelligence.

The current situation is changing rapidly, especially as artificial intelligence has become synonymous with large-scale learning models. The current status quo is that no one knows anything about how the trained models have a specific function, or even other functions they may have (such as PALM's so-called ability to "explain jokes"). Even their creators often have no idea what these systems can do. Exploring these systems to understand their “functional” scope has become a recent trend in artificial intelligence research.

It is becoming increasingly clear that some parts of artificial intelligence are straying from their engineering roots. Today it is difficult to think of large learning systems as engineering designs with specific goals in the traditional sense. After all, one cannot say that one's children are "designed." The field of engineering doesn't typically celebrate unexpected new properties of systems it designs (just as civil engineers don't celebrate with excitement when a bridge they designed to withstand a Category 5 hurricane is found to levitate).

There is growing evidence that the study of these large, trained (but not designed) systems is destined to become a natural science: observing system functionality; doing ablation studies; conducting qualitative analysis of best practices. analyze.

Considering the fact that appearances are currently studied rather than what’s inside, this is similar to the ambitious goal in biology of trying to “figure it out” without actual evidence. Machine learning is a research endeavor that focuses more on why a system does what it does (think of doing "MRI" studies of large learning systems) rather than proving that the system was designed to do so. The knowledge gained from these studies can improve the ability to fine-tune systems (just like in medicine). Of course the study of surface settings allows for more targeted intervention than in internal settings.

Artificial intelligence becomes a natural science and will also have an impact on the entire computer science, considering that artificial intelligence will have a huge impact on almost all computing fields. The word "science" in computer science has also been questioned and ridiculed. But that has changed now, as artificial intelligence has become a natural science that studies large-scale artificial learning systems. Of course, there may be a lot of resistance and opinions to this transition, because computer science has long been the holy grail of "correct by construction". From the beginning, computer science has been equivalent to living in a system full of incentives. It's as correct as a well-trained dog, just like a human being.

Back in 2003, Turing Award winner Leslie Lamport sounded the alarm about the possibility that the future of computing would be biology rather than logic, saying computer science would allow us to live in a world of homeopathy and faith healing. At that time, his anxiety was mainly about complex software systems programmed by humans, rather than today's more mysterious large-scale learning models.

When moving from a field primarily concerned with intentional design and “correctness by construction” to trying to explore or understand existing (undesigned) artifacts, the methodological shift it will bring is worth thinking about. Unlike biology's study of wild creatures, artificial intelligence studies artificial artifacts created by humans that lack a "sense of design." Ethical issues will definitely arise when it comes to creating and deploying artificial artifacts that are not understood. Large learning models are unlikely to be guaranteed to support provable capabilities, whether with regard to accuracy, transparency, or fairness, yet these are critical issues in deploying and practicing these systems. While humans are also unable to provide evidence as to the correctness of their own decisions and actions, legal systems do exist to subject humans to punishments such as fines, reprimands, and even imprisonment. For large-scale learning systems, what are the equivalent systems?

The aesthetics of computational research will also change. Current researchers can evaluate papers by the proportion of them containing theorems and definitions. But as the goals of computer science become more and more like the goals of natural sciences such as biology, there is a need to develop new computational aesthetic methodologies (because the zero theorem will not be very different from the zero definition ratio). There are signs that computational complexity analysis has taken a backseat in AI research. ​

The above is the detailed content of After the overwhelming emergence of large models, computer science finally became a 'natural science'. For more information, please follow other related articles on the PHP Chinese website!