Google's AI shocks engineers by learning new language without human assistance

Google's AI (artificial intelligence) has managed to learn a foreign language without any human intervention. Meanwhile, the American tech giant is still trying to figure out how AI achieved this remarkable feat unassisted.

Google CEO Sundar Pichai shed some light on the company's cautious approach to AI in an interview with CBS. Apparently, the search giant has postponed a myriad of AI projects citing image risks and security concerns over the past few years.

This is understandable given that some cybercriminals have resorted to using AI-powered chatbots to craft genuine-looking phishing emails. While Google could have easily established itself as a pioneer in the AI space, it currently lags behind rivals like OpenAI and Microsoft due to its cautiousness.

Everything you need to know about emergent properties in AI

Pichai says it is imperative to be responsible when it comes to developing AI. "We all have to be responsible every step of the way," the top executive explained. Notably, Google adopted a cautious strategy for launching its AI chatbot, Bard. Instead of making it available worldwide, the company restricted Bard's initial availability to the United Kingdom and the United States.

Despite its cautious rollout, Bard was accused of being left-leaning. In fact, Google's former CEO Eric Schmidt even warned that AI chatbots like Bard and OpenAI's ChatGPT could undermine democracy. However, Pichai clarified that Google wants to "develop more robust security layers before building and deploying more efficient models."

Google Bard was initially based on the LaMDA (Language Model for Dialogue Applications) model. However, the company eventually integrated features from a higher-performing language model dubbed PaLM (Pathways Language Model) into the AI bot. While LaMDA has 137 billion parameters, PaLM boasts a whopping 540 billion parameters, according to a report by 01net.

However, Pichai emphasises moving with caution and keeping an eye on AI reactions because generative AI can sometimes display unexpected behavior. Google's research shows that language models can occasionally learn new skills without any human assistance. Google refers to this mysterious event as "emergent properties."

Google developed experimental AI that serves as an example of these emergent properties. The AI was able to successfully learn the language of Bangladesh, Bengali although it wasn't trained to do so. Google's Senior Vice President of Technology, James Manyika noted that the AI was able to adapt to the new language after only a few Bengali queries.

"We found that with very few Bengali queries, he can now translate all of Bengali," Manyika said. Emergent properties allude to a machine's ability to learn using available data, without any help or human assistance. So, it doesn't come as a surprise that an AI was able to successfully learn a language without a programmer's assistance. Still, Google engineers are trying to understand how AI acquired Bengali knowledge.

Can Google AI emulate the human brain?

According to Pichai, abnormalities in AI are a "black box." In other words, details about how they work internally are still scarce. "You know, you don't quite understand. And you can't quite say why he said that, or why he was wrong." he said. "We have some ideas, and our ability to understand improves over time."

Aside from this, Google's chief executive pointed out that the human mind is mysterious as well. Nevertheless, tech companies including Google need to remain vigilant as AI continues to advance. In conclusion, Google's AI ability to learn a foreign language without a programmer's intervention has left the company perplexed about how the AI works.

The occurrence of emergent properties in AI development is a major sign that developers should closely monitor AI behavior. With AI getting more advanced, tech giants should be responsible, and ensure the ethical development of the tech.