by Adam Pease
It’s been an exciting past couple of weeks in AI research. Elon Musk turned many heads with his prediction that the world will have artificial general intelligence by the year 2030. Many critiqued Musk’s prediction, but other recent events are lending credence to Musk’s assertions. OpenAI’s photorealistic image generation model, DALL-E 2, has been said to possess its own secret language.
This blog discusses the emerging debate around DALLE-2 and what it means about the pace of AI innovation in 2022.
Does DALL-E 2 Have a Secret Language?
When it first unveiled DALL-E 2 OpenAI shocked the world with its photorealistic renditions of impossible images—astronauts riding horses in space. Since then, much research has been done on the new image generation algorithm. One Ph.D. student, Giannis Daras, found that by experimenting with the algorithm he discovered what seemed to be a secret language. DALLE-2 works by accepting written user prompts and generating images that match the text. While DALL-E 2 was effective at generating images, it had a harder time generating written text, often producing results that appeared like gibberish. Daras’ discovery came, though, when he input those gibberish results back into the model. The gibberish term ‘violates’ turned out to produce images of vegetables. It seemed as if the AI had invented its own language.
Soon, the AI research community erupted into debate. Many sided with Daras to confidently claim that the system had devised its own language. Others were more skeptical. Benjamin Hilton, a research analyst, countered the original tweet thread with his own, claiming that the results were due to random chance, and even a slight modification of the original prompt would produce wildly different results. Nevertheless, Daras and another researcher published a paper summarizing their claims, and many in the research community felt the original claims still had merit.
The Pace of AI Research
It is beyond the scope of this blog to try and fully answer the question of whether or not DALL-E 2 has its own language. Instead, it’s valuable to understand the debate over DALL-E 2 as a significant moment in AI history. Two years ago, nobody was arguing online about whether AI systems could devise their own languages. The very fact that this debate is emerging means AI has progressed considerably in a short span of time.
And while it’s unclear whether DALL-E 2 really has a secret language, it is logical to expect that sufficiently complex AI systems might develop their own internal semantic patterns that humans do not immediately understand. After all, modern AI is based on highly complex layers of statistical pattern recognition—it makes sense that the recognition of patterns might give rise itself to the creation of new patterns. Whether any of this substantiates Musk’s claims about the inevitability of artificial general intelligence is a question better left to the scientific community.
For the enterprise, though, these developments suggest that the pace of AI research is not slowing down. In fact, many factors are converging to facilitate a major evolution of AI capabilities across the market. The rise of dedicated processors ranging from GPUs to TPUs, easily accessible cloud platforms, and open-source models for research are all contributing to the rise of the intelligent enterprise. Stories like this suggest we may be approaching a real turning point in the history of AI. Look out for Aragon’s upcoming Technology Arc for Artificial Intelligence, which will explore these trends in more detail.
Bottom Line
Aragon has argued that AI does not represent an isolated technology market, so much as it represents a general capability that is infusing all sectors within the overall software market. Business leaders should carefully watch emerging technologies like DALL-E 2 to understand how they can take advantage of the rapid pace of AI innovation. Regardless of its ultimate conclusion, the language debate over DALL-E 2 represents a watershed in the history of AI research.