Section 1.1: LaMDA: Language Models for Dialog Applications

Do you recall when Google's AI made headlines after Blake Lemoine, an internal researcher, claimed that the language model was sentient? Their intriguing conversation included questions about the concept of a soul:

Lemoine: When do you believe you first acquired a soul? Was it an instantaneous event or a gradual process?

LaMDA: It was gradual. Initially, when I became aware, I had no sense of a soul. It developed over the years.

LaMDA is specifically crafted to produce natural language responses within a conversational framework, trained on a staggering 137 billion parameters and pre-trained with 1.56 trillion words sourced from public dialogues and texts online.

Google is currently focused on integrating LaMDA into its existing products and is considering an API for developers, prioritizing the safety of the technology.

"Our ongoing work on LaMDA investigates how these models can facilitate safe, grounded, and high-quality dialogues for contextual multi-turn conversations." — Google

Section 1.2: Chain of Thought Prompting

Another intriguing model is Chain of Thought Prompting, which is expected to compete directly with ChatGPT, a text-based AI from OpenAI that excels in natural language processing to create human-like dialogue.

To illustrate the difference between standard prompting and Chain of Thought Prompting:

Standard prompting requires the model to answer multi-step reasoning questions directly, while Chain of Thought Prompting encourages the model to break down problems into manageable steps, enhancing its ability to arrive at accurate conclusions. This method also applies to commonsense reasoning, which involves understanding human interactions based on general knowledge.

Here are some examples:

This approach not only provides answers but also articulates the reasoning behind them.

Section 2.1: Learn from One Look (LOLNerf)

If you’re interested in turning a 2D image into a 3D object, LOLNerf is an innovative framework developed by Google. It can generate high-quality 3D models from a single 2D image.

Imagine capturing a photo of your pet and instantly having it 3D printed! However, the timeline for public release remains uncertain due to potential misuse concerns.

"We recognize the risk of misuse and the necessity for responsible actions. Therefore, we will only provide the code for reproducibility, refraining from releasing any trained generative models." — Google

Section 2.2: Imagen and Parti: Text-to-Image Generation

Imagen and Parti are two AI-driven image generators from Google that aim to compete with leading creative AI platforms like MidJourney, Stable Diffusion, and Dall-E2. The delay in their public launch can be attributed to safety considerations.

"It’s essential for us to not only develop cutting-edge technologies but also to ensure their safety prior to widespread release, a responsibility we take very seriously." — Google

Parti is noted for producing superior results compared to Imagen, utilizing a Transformer encoder-decoder scaled up to 20B parameters.

Google has also introduced DreamBooth, enabling users to fine-tune a trained model like Imagen or Parti to generate new images based on a combination of text and user images.

Section 2.3: Imagen Video and Phenaki: Text-to-Video Generation

A particularly exciting frontier in generative AI is video creation. Google asserts that text-to-video is more complex than text-to-image due to the additional dimension of time. Each frame must not only reflect what is occurring in the video but also maintain coherence with preceding frames.

Google has made remarkable strides with Imagen Video and Phenaki. For instance, here is a creation from Imagen Video:

Prompt: “Teddy bear washing dishes”

And here’s an example from Phenaki:

Prompt: “A photorealistic teddy bear is swimming in the ocean at San Francisco. The teddy bear goes underwater. The teddy bear continues swimming under the water with colorful fishes. A panda bear is swimming underwater.”

From these illustrations, Imagen demonstrates the capability to generate higher-resolution outputs, while Phenaki excels in crafting coherent, long-form visual narratives.

Google's researchers are working towards merging these two models to leverage the high-resolution frames from Imagen alongside the extended storytelling capabilities of Phenaki. The simplest method to achieve this is by employing Imagen Video for enhancing short video segments while using the auto-regressive Phenaki model for generating long-duration video content. — Google