Hey there!

We’ve got some big news in this edition so make sure to read through and keep up with the world.

Here’s what we’ll learn about today:

• Google has partnered with DeepMind to combat Microsoft and OpenAI 🏢

• Whisper is now 70x faster 🗣️

• Stable Diffusion releases Large Language Models 💻

• Cody - the AI Employee 🤖

Reading time: 5 min and 08 seconds

News Reel:

Highlight 1: Google has partnered with DeepMind to combat Microsoft and OpenAI

It’s the Battle of the Titans⚔️. Earlier today, the inception of Google Deepmind was announced. DeepMind and Google Brain teams are merging into a new unit called Google DeepMind. They originally had differences but have decided to set them aside to compete with Microsoft and OpenAI.

Here are some important points you need to know:

• DeepMind is the company responsible for AlphaCode and one of the biggest artificial intelligence companies in the world.

• The unit will have the resources to accelerate progress toward developing more capable and general AI in a bold and responsible way.

• A new Scientific Board for Google DeepMind will be created to oversee the research progress and direction of the unit, which will be led by Koray Kavukcuoglu and will have representatives from across the organizations, the composition of which will be decided by Jeff Dean, Koray Kavukcuoglu, Zoubin Ghahramani (Research Lead), Shane Legg (Co-founder DeepMind) and Demis Hassabis (CEO DeepMind).

• The goal is to achieve greater speed, stronger collaboration, and execution, and focus on achieving the biggest impact to improve the lives of billions of people, transform industries, advance science, and serve diverse communities📈.

Read the full announcement here

Highlight 2: Speech-to-text transcriber Whisper is now 70x faster 

Whisper JAX (explained ahead) is a speech-to-text transcriber that’s only recently become 70x faster. To provide some context, Whisper JAX can transcribe 1 hour of speech in just 15 seconds!🌬️

Whisper gains quicker results by:

• Processing batched data rather than un-batched - A single audio sample is segregated into 30-second clips, which are then transcribed rather than transcribing the whole sample. This yields a 7x speed increase.

• Using JAX over JIT (Just in time processing) algorithm - JAX is an automatic differentiation library for high-performance machine learning research which yields a 2x speed increase. I understand the names are cryptic so to put it simply, a change in algorithm is faster and hence results in a speed increase.

• Finally, TPUs (Tensor Processing Units) are machine learning accelerators made by Google (specifically designed for matrix multiplication) are faster than GPU, yielding a 5x speed increase.

• 7 × 2 × 5 = 70x speed increase.🔝

Understand in more detail in Sanchit Gandhi’s tweet here

Try it yourself here

Highlight 3: StabilityAI releases 3 Billion and 7 Billion parameter Large Language Models

Stability AI (Creators of Stable Diffusion - the text-to-image AI model) has launched StableLM, a suite of language models meant to compete with OpenAI’s GPT-4, Meta’s LLaMA, and Google’s LaMDA.

• The company trained the new model on an 800GB open-source data set called “The Pile,” ranging from 3 billion to 7 billion parameters (as compared to GPT-3’s 175 billion parameters.

• The machine learning model parameters determine how to input data is transformed into the desired output. Read more about parameters here.

• The 7-billion parameter version of the chatbot is available to test drive on Hugging Face, but initial tests showed some limitations in its performance, such as failing at rudimentary tasks like creating a recipe for a peanut butter jelly sandwich.🥪

• Stability AI aims to eventually process 3 trillion tokens and is focused on efficient, specialized, and practical AI performance rather than a quest for god-like intelligence. The company has reportedly struggled with money lately as it has spent a lot on developing its AI projects.

• The company plans to release details on StableLM's training data "in due course" alongside a full technical write-up, and CEO Emad Mostaque said both the LLM and its training data will only get better with time.

Read the whole Twitter thread here.

Read more on the Stability blog here.

Read the detailed article from Gizmodo here.

A.I. Tool showcase:

Cody:

Don’t like how your head of HR behaves? Just change their personality! Finding employees for your company has never been easier. CodyAI allows you to create bots that will act as employees for your company!

• You can directly upload docx, pdf etc. and even ask it to crawl (to extract information from a website including all the linked pages) your website up to 5,000 pages!

• Once transformed, you can edit the information in the text editor to ensure it’s organized and correct.

• You can then create and customize your own bot, giving it a name, description, GPT model, knowledge, job, and even personality (creative, balanced, or factual). 🤖

• You can even choose how the bot should respond if it doesn't know the answer like giving the closest guess, prompting the user to rewrite their question and even simply not answering⁉️.

• Now imagine several of these bots with access to all your company’s documents, you’ll never need another employee to give you information again! It can even be in charge of social media and sales and will give you social media posts if you import the correct documents and simply ask!

• You can even embed the bots directly to your website as a chatbot for users to use as help!

• I’ve already created a very creative Gu, who provides me with great social media post ideas, try it out here now💫!

A.I. Art of the Week:

Here’s a picture of a bear cooking to make your Thursday a little better 😍

See you on Monday!

-A.I. Bits