On 2 February, Alphabet Inc.’s artificial intelligence subsidiary, DeepMind, announced that it has created a system called AlphaCode that can write computer programs at a competitive level. Drawing on a combination of critical thinking, logic, algorithms, coding, and natural language understanding, the consequences of this AI milestone could be far reaching.
DeepMind
DeepMind was founded in 2010 as the brainchild of Demis Hassabis, Shane Legg and Mustafa Suleyman. Following substantial investments from venture capitalists including Peter Thiel and Elon Musk, it was purchased by Google in 2014 for $500 million. Created with the aim of “solving intelligence” then using intelligence “to solve everything else”, it seems that DeepMind has remained true to this purpose over the last 12 years. In 2016, the company made international headlines when its AlphaGo program beat Go world champion Lee Sedol, 4 games to 1 which included producing several moves that were said to have upended hundreds of years of conventional wisdom. More recently, the company took on a 50-year-old grand challenge in biology known as the protein-folding problem with their AlphaFold program which achieved an unprecedented 92.4 global distance test (GTD) accuracy rating at the 14th biennial CASP competition. Operating at the bleeding edge of AI and ML research, DeepMind holds numerous patents for its technologies across a broad range of applications including neural programming, speech recognition and image processing.
AlphaCode
With the announcement of their AlphaCode program to tackle competitive programming, it appears DeepMind have reached yet another important milestone in their quest “to solve everything else”. Competitive programming generally involves a competitor being given a series of long problem descriptions and then a fixed amount of time to write computer programs to solve them with the winner being the competitor that can solve the most problems in the shortest amount of time. What seems most notable about this milestone is that it appears to mark a departure from the clean-cut, well defined problems conquered by previous Alpha programs. For example, with Go and other traditional games like chess, there are well defined rules that not only constrain the solution space but most importantly also describe what it means to win. In contrast, a programming problem is inherently open-ended meaning the only real rules are that of the programming language in which the solution will be written. AlphaCode ranked on average within the top 54% of human participants across 10 competitions, which means it appears to be capable of writing code at a competitive level.
Natural Language Processing
The step in the program that allows AlphaCode to take a descriptive statement and internally formulate a problem it can then attempt to solve relies on a branch of AI known as natural language processing (NLP). In basic terms, this process takes natural language such as a sentence and transforms it into a format that a computer can interpret and is the technology that allows voice assistants to answer questions and control devices. Whilst these applications may seem primitive, robust NLP is clearly a pre-requisite to any AI system dealing with written or spoken natural language. Due to the ambiguity, nuances and complex syntactic structure of natural language, it is a highly complex and hence computationally intensive problem meaning it has only really been within the last decade that researchers have had the necessary hardware at their disposal to start making significant progress, utilising methods such as representation learning and deep neural networks. Last year, DeepMind debuted a new NLP model and despite the system approaching near human levels of reading comprehension it, perhaps surprisingly, fell well short in other areas like mathematical reasoning which demonstrates the NLP problem is far from solved.
Consequences for IP
AlphaCode appears to demonstrate that NLP has now reached a level where it is capable of interpreting a problem statement to a level of comprehension that allows a system to produce not only a working solution to a computer programming problem but a working solution that falls in the top 54% of human-generated solutions. On this basis, one cannot help but wonder what this means for the world of IP. For example, is the code generated by AlphaCode an “original literary work” protected by copyright? If so, who is the author of that work – AlphaCode itself, the people who created AlphaCode or the people who use AlphaCode? Looking at other forms of IP, could something like AlphaCode create a patentable invention? Last year saw a UK legal case which raised the question of whether an AI system can be an inventor. Although this case was dismissed by the UK courts, the case was dismissed solely on the basis that only a person may be designated as an inventor, rather than on the basis that an AI system cannot create something inventive. So it seems that there are still some thorny questions yet to be answered by the courts.
Conclusion
There is an ever-expanding list of questions surrounding the consequences that increasingly intelligent AI will have for humanity. As the capabilities of AI continue to grow, it is becoming increasingly important that we start to seek answers to these difficult questions now or risk having systems in place that are simply not equipped to meet the needs of a society where AI plays an integral role.
.png?width=100&height=100&name=My%20project-1%20(9).png)
