What can Google’s new AI do?

Google is the leading company in the field of artificial intelligence (AI). For years, the group and its British subsidiary Deepmind have been publishing by far the most important research papers. However, it was others who brought the topic of AI to the fore. At the end of 2022, the start-up OpenAI and its investor Microsoft presented the chatbot ChatGPT, which can create texts as if they were written by humans.

That got things moving. Google is now shedding its previous reluctance and making its chatbot Bard accessible – albeit with some caveats. Why the long hesitation? And how good is Bard with the underlying Google AI?

What is bard?

Like ChatGPT, Bard is a chatbot that can engage in human-like dialogue and create text. Bard is named after the English word for “poet” or “bard”.

Like ChatGPT, the chatbot is based on a “Large Language Model” (LLM). A “big language model” has neural networks modeled on the human brain. They are trained on large amounts of text data; the LLM learns language structures and can predict the next word in a sentence with a high degree of probability.

Bard and ChatGPT: What’s the difference?

Unlike ChatGPT, people with Bard can currently only communicate in English. However, other languages ​​are to be added. ChatGPT has only been trained with data up to 2021, while Bard can answer questions with data up to the present.

ChatGPT is based on the GPT language model, of which OpenAI released the fourth version in March 2023. Google relies on Lamda for Bard, which stands for “Language Model for Dialogue Applications”. To save computing power and increase availability, Bard uses a slimmed down and optimized version of Lamda.

>> Read also: How ChatGPT will change your life

Both GPT and Lambda work with a huge number of parameters that determine certain references and probabilities in the model. An example: A parameter could be that after the word “sun” there is very likely “to shine”. According to the latest data from the companies, GPT works with 175 billion, Lambda with 137 billion parameters.

What can Google’s chatbot Bard do?

Bard works like an advanced Google search. You can formulate your questions or requests in a search field. For example, Bard can be instructed to write a poem, lyrics, or an essay.

The chatbot provides its answers with references. Unlike ChatGPT, Bard also generates several suggestions, so-called “drafts”, from which you can choose.

As with ChatGPT, the answers do not necessarily have to be correct – both OpenAI and Google are working on reducing the so-called “hallucinating” of language models. Both are trying to do this with “Reinforcement Learning from Human Feedback” (RLHF), among other things. With this training method, employees monitor the system and reward it depending on whether the answers are correct or not.

How can I test Bard?

Google’s chatbot Bard is currently still in the test phase. Only selected participants from the US and UK can enter. Interested parties can be put on a waiting list on the website bard.google.com – apparently the rush is large. An extension to other countries is planned.

Why is Google releasing Bard now?

Unlike the start-up OpenAI and Microsoft, Google has a lot to lose in the search business, which is very lucrative through online ads. Google has a market share of well over 90 percent, while Microsoft’s Bing only covers about three percent. Google doesn’t want to jeopardize the business and its image with a “hallucinating” AI that could give wrong or misleading answers.

>> Read about this: Interview with Peter Thiel: “Google is completely in panic mode”

But the market changed with the release of ChatGPT and the integration of GPT-4 into Bing. ChatGPT has made it clear to the general public what a chatbot is – with all its advantages and disadvantages. Now it’s about perception supremacy: Which company has the best AI?

Which AI has been used for “Ok Google” so far?

Google launched its voice assistant in 2016. “Ok Google” understands 30 languages ​​and can answer questions about the weather or daily routine. More than 700 million people use the system worldwide. It made headlines in 2018 because it can make calls for smartphone owners – for example to book a table in a restaurant.

Google is working on many AI models, some of which build on each other. Ok Google is about converting spoken language into text and not, like Bard, about entering text. The assistant recognizes via the camera and image recognition whether the user is looking at the smartphone and speaking. A total of five different machine learning models are used to make the conversation “as natural as possible,” explained Tuan Anh Nguyen and Sourhis Chaudhuri, two software engineers at Google, in a 2022 blog post.

What is Google’s Lamba?

Lambda is a large AI project by Google. It is based on a so-called transformer model, a neural network that the group helped develop in 2017. GPT and ChatGPT are also based on a Transformer model.

>> Read also: AI could revolutionize these industries

Unlike other language models, Lambda was not only trained for texts, but above all for dialogues. According to Google, this allows it to better pick up on the nuances of a conversation.

To make the answers more creative and interesting, the Google engineers turn on the “temperature”, which in technical jargon means a model setting. “The models can choose from reasonable but slightly less likely next words that allow for more interesting answers,” said James Manyika, Google’s chief technology officer.

How exactly does Lamba work?

Lambda is based on the Transformer method. This allows a computer to “transform” certain signals or information, for example to translate or summarize a text into another language.

Transformers belong to the deep learning models. They achieve “deep learning” through neural networks, which form a complex structure and hierarchy in numerous intermediate layers. The human brain is the model.

Lambda was “pre-trained” with a total of around 1.6 trillion words, mainly from dialogues, as it is called in technical jargon. Then comes the “fine tuning” in which the possible answers in a dialogue are evaluated according to so-called SSI factors (“meaningfulness, specificity, interestingness”) in order to make the conversation meaningful, precise and lively. The principles of security and down-to-earthness are also included in order to avoid controversial or nonsensical answers.

Can you talk to Google Lambda?

Yes, technically it’s not a problem, Lambda also transforms speech to text. However, access is limited. Google offers the function in the special “AI Test Kitchen” app, which has only been available for download in the USA, Great Britain and Australia since autumn 2022. However, only selected users can participate there.

Is Lamba conscious?

The answers of the language model Lamda are apparently so good and convincing that the Google developer Blake Lemoine attributed consciousness to him in 2022 – after which he promptly lost his job. Google described Lemoine’s statements as “completely unfounded” and that he had “violated work and data security regulations”.

>> Read also: Bill Gates in an interview about artificial intelligence

According to the American, Lamba developed feelings such as fear or insecurity. According to his statement, he was able to unsettle Lambda in such a way that it would violate internal Google guidelines.

Is Lamba conscious? This question is difficult to answer. For what exactly is “consciousness”? But one thing is clear: Lamba only does what it has learned. It simulates human conversations as accurately and convincingly as possible. In addition to the SSI factors, this also includes stress and insecurity: two feelings that play a large role in conversations.

What are Lambda 2 and Palm?

Lambda 2 is a new, more fine-tuned version of Google’s dialog language model. The group presented them in May 2022. It is only accessible to a small group of users on the Google app “AI Test Kitchen”.

At the same time, Google introduced Palm. Palm stands for “Pathways Language Model”. This is a language model based on Pathways, an AI architecture presented by Google in 2021. “Today’s AI models are typically trained to do just one thing,” said Jeff Dean, head of Google Research, at the launch. “Pathways will allow us to train a single model to do thousands or millions of things.”

>> Read also: Three well-paid IT jobs for career changers

Pathways works with “multiple senses”, for example it can recognize the word leopard, the picture of a leopard or the hissing of a leopard. Just like the human brain, it only uses parts of the neural network depending on the task. This should make it faster and use less energy. According to Dean, Pathways uses “small paths” – hence the name of the model.

Palm is equipped with an extremely large number of parameters (540 billion). It is trained on chips with four times the computing capacity of Lambda – and unlike Lambda, not only on dialogues, but with “high-quality web documents, books, Wikipedia, conversations and programming code”. Accordingly, it can create software, understand jokes or create videos.

Since March 2023, Google has been granting access to corporate customers and developers who can use it to “create text, images, code, videos, audio and more using voice commands”.

More: How ChatGPT works