AI Weekly Rundown (September 23 to September 29)
Major AI announcements from Meta, Amazon, Google this week.
Hello, Engineering Leaders and AI Enthusiasts,
Another eventful week in the AI realm. Lots of big news from huge enterprises.
In today’s edition:
💰 Amazon to Invest $4B in Anthropic
🤖 Meta to develop a ‘sassy chatbot’ for younger users
🚀 LongLoRA: Efficient fine-tuning of long-context LLMs
💥 Biggest Boom in AI: ChatGPT Talks and Beyond🖌️ Getty Images’s new AI art tool powered by NVIDIA
💰 Colossal-AI’s commercial-free LLM saving thousands
💰 OpenAI eyes $90B valuation, dives into AI hardware
🚀 Vectara launches Boomerang, the next-gen LLM redefining GenAI accuracy
🎉 Google's 25-year AI legacy guides its future AI innovations
🤩 Meta’s new exciting AI experiences & tools🌐 OpenAI links ChatGPT with Internet
😮 Mistral AI’s LLM outperforms Meta’s Llama2
🚀 AWS announces powerful new AI offerings
📜 Meta introduces LLAMA 2 Long
🌐 Google announces Google-Extended and opens SGE to teens
Let’s go
Amazon to Invest $4B in Anthropic
Amazon will invest up to $4 billion in Anthropic. The agreement is part of a broader collaboration to develop the industry's most reliable and high-performing foundation models.
Anthropic’s frontier safety research and products, together with Amazon Web Services’ (AWS) expertise in running secure, reliable infrastructure, will make Anthropic’s safe and steerable AI widely accessible to AWS customers. AWS will become Anthropic’s primary cloud provider for mission-critical workloads, and this will also expand Anthropic’s support of Amazon Bedrock.
Meta to develop a ‘sassy chatbot’ for younger users
Meta has plans to develop dozens of chatbot ‘personas’ geared toward engaging young users with more colorful behavior. It also includes ones for celebrities to interact with their fans and some more geared towards productivity, such as to help with coding and other tasks.
LongLoRA: Efficient fine-tuning of long-context LLMs
New research has introduced LongLoRA, an ultra-efficient fine-tuning method designed to extend the context sizes of pre-trained LLMs without a huge computation cost.
Typically, training LLMs with longer context sizes consumes a lot of time and requires strong GPU resources. For example, extending the context length from 2048 to 8192 increases computational costs 16 times, particularly in self-attention layers. LongLoRA makes it way cheaper by:
1. Using sparse local attention instead of dense global attention (optional at inference time).
2. Using LoRA (Low-Rank Adaptation) for context extension
This approach seems both easy to use and super practical. LongLoRA performed strongly on various tasks using LLaMA-2 models ranging from 7B/13B to 70B. Notably, it extended LLaMA-2 7B from 4k context to 100k and LLaMA-2 70B to 32k on a single 8x A100 machine, all while keeping the original model architectures intact.
Biggest Boom in AI: ChatGPT Talks and Beyond
OpenAI is introducing voice and image capabilities in ChatGPT, allowing users to have voice conversations and show images to ChatGPT. This new feature offers a more intuitive interface and expands the ways in which ChatGPT can be used.
Users can have live conversations about landmarks, get recipe suggestions by showing pictures of their fridge, and even receive math problem hints by sharing photos. The voice and image capabilities will be rolled out to Plus and Enterprise users over the next two weeks, with voice available on iOS and Android and images available on all platforms.
ChatGPT can now comprehend images, including photos, screenshots, and text-containing documents, using its language reasoning abilities. You can also discuss multiple images and utilize their new drawing tool to guide you.
Getty Images’s new AI art tool powered by NVIDIA
Getty Images has launched a generative AI art tool called Generative AI, which uses an AI model provided by Nvidia to render images from text descriptions. The tool is designed to be "commercially safer" than rival solutions, with safeguards to prevent disinformation and copyright infringement.
Getty Images will compensate contributors whose work is used to train the AI generator and share revenues generated from the tool. The tool can be accessed on Getty's website or integrated into apps and websites through an API, with pricing based on prompt volume. Other companies, including Bria and Shutterstock, are also exploring ethical approaches to generative AI.
Colossal-AI’s commercial-free LLM saving thousands
Colossal-AI has released Colossal-LLaMA-2, an open-source and commercial-free domain-specific language model solution. It uses a relatively small amount of data and training time, resulting in lower costs.
The Chinese version of LLaMA-2 has outperformed competitors in various evaluation benchmarks. The release includes improvements such as vocabulary expansion, a data cleaning system, and a multi-stage pre-training scheme to enhance Chinese and English abilities.
OpenAI eyes $90B valuation, dives into AI hardware
OpenAI is in discussions to possibly sell shares, a a move that would boost its valuation from $29 billion to somewhere between $80 billion and $90 billion, according to a Wall Street Journal report citing people familiar with the talks.
(Source)
In other news, Apple's former design chief, Jony Ive, and OpenAI CEO, Sam Altman, have reportedly been discussing building a new AI hardware device. It is unclear what the device would be or if they will build it, but the duo has been discussing what new hardware for the AI age could look like.
Vectara launches Boomerang, the next-gen LLM redefining GenAI accuracy
Outpacing major competitors, Boomerang sets a new benchmark in Grounded Generative AI for business applications. It is a next-generation neural information retrieval model integrated into Vectara's GenAI platform.
Boomerang surpasses Cohere in benchmark performance and matches OpenAI on certain metrics, excelling particularly in multilingual benchmarks. Notably, it prioritizes security, reducing bias, copyright concerns, and "hallucinations" in AI-generated content. It also offers cross-lingual support for hundreds of languages and dialects and improves prompt understanding, leading to more accurate and faster responses.
Google's 25-year AI legacy guides its future AI innovations
On its 25th birthday, Google reflected on its two-and-a-half decades of pioneering achievements in the field of AI. It started in 2001 using a simple ML to suggest better spellings for web searches.
A standout moment in 2023 was the introduction of PaLM 2 and Gemini. It is now looking forward to these models driving the next quarter-century of its AI advancements.
📢 Invite friends and get rewards 🤑🎁
Enjoying AI updates? Refer friends and get perks and special access to The AI Edge.
Get 400+ AI Tools and 500+ Prompts for 1 referral.
Get A Free Shoutout! for 3 referrals.
Get The Ultimate Gen AI Handbook for 5 referrals.
When you use the referral link above or the “Share” button on any post, you'll get credit for any new subscribers. Simply send the link in a text, email or share it on social media with friends.
Meta’s new exciting AI experiences & tools
Meta's new AI features include an AI Assistant powered by Bing, It will provide real-time information and generate photorealistic images from text prompts. Meta used specialized datasets to train the AI to respond in a conversational and friendly tone. The first extension of the AI Assistant will be web search. The AI Assistant will be available in beta on WhatsApp, Messenger, and Instagram.
Introduced 28 AI personality chatbots based on celebrities, such as Tom Brady, Naomi Osaka, Mr. Beast, and more. These chatbots, accessible on platforms like WhatsApp, Messenger, and Instagram, provide topic-specific conversations but are currently text-based, with plans to introduce audio capabilities. These AI personalities were created using Llama 2. Meta aims to integrate Bing search functionality in the future. The chatbots' animations are generated through AI techniques, offering a cohesive visual experience.
Launching AI Studio, a platform allowing businesses to build AI chatbots for Facebook, Instagram, and Messenger, initially focusing on Messenger for e-commerce and customer support apps. This toolkit will be available in alpha.
Gen AI stickers powered by Emu allow users to create unique stickers across its messaging apps. Users can type in their desired image descriptions, and Emu generates multiple sticker options in just a few seconds. Initially available to English-language users, this feature will roll out over the next month.
Introducing 2 new AI Instagram features, restyle and backdrop. Restyle allows users to transform the visual styles of their images by entering prompts like "watercolor" or more. While backdrop changes the background of photos using prompts.
Launches New-gen Ray-Ban smart glasses, in partnership with EssilorLuxottica, will feature improved audio and cameras, over 150 different custom frame and lens combinations. They’re lighter and more comfortable. Will enable livestream to Facebook or Instagram and use “Hey Meta” to engage with Meta AI assistant by voice.
OpenAI links ChatGPT with Internet
ChatGPT is back with internet browsing, It can now browse the internet to provide current & reliable information, along with direct links to sources. This update addresses feedback received since the browsing feature was launched in May. The model now follows robots.txt and identifies user agents to respect website preferences.
Currently available to Plus and Enterprise users, browsing will be expanded to all users soon.
To try it out, enable Browse in your beta features setting:
Click on 'Profile & Settings’ > Select 'Beta features' > Toggle on ‘Browse with Bing’ > Choose Browse with Bing in the selector under GPT-4.
Mistral AI’s LLM outperforms Meta’s Llama2 13B
Mistral AI, Europe's largest seeded startup, has released its first LLM Mistral 7B. This model outperforms Meta's Llama 2 13B and is touted as the most powerful language model for its size. It was founded by alums from Google's DeepMind and Meta earlier this year. It aims to make AI useful for enterprises by using publicly available data and customer contributions.
Mistral 7B excelled in benchmarks, surpassing Llama 2 7B and 13B in text summarization, classification, and code completion tasks. The only area where Llama 2 13B matched Mistral 7B was world knowledge testing.
AWS announces powerful new AI offerings
Amazon Web Services (AWS) has announced 5 major generative AI updates and innovations.
Amazon Bedrock is now generally available. It is a fully managed service that makes foundation models (FMs) from leading AI companies available through a single API. It also has new AI models in the mix and will help more customers build and scale generative AI applications.
Amazon Titan Embeddings is now generally available. It is an LLM that makes it easier for customers to start with Retrieval-Augmented Generation (RAG) to extend the power of any FM using their proprietary data.
Meta’s Llama 2 is coming to Amazon Bedrock in the next few weeks. Amazon Bedrock is the first fully managed generative AI service to offer Llama 2 through a managed API. Currently, it includes models from 21 Labs, Anthropic, Cohere, Stability AI, and Amazon.
New Amazon CodeWhisperer capability is coming soon. It will allow customers to securely customize CodeWhisperer suggestions using their private code base to unlock new levels of developer productivity. Trained on billions of lines of Amazon and publicly available code, Amazon CodeWhisperer is an AI-powered coding companion.
New Generative BI authoring capabilities to extend the natural-language querying of Amazon QuickSight Q beyond answering well-structured questions. It will help analysts quickly create customizable visuals from question fragments, clarify the intent of a query by asking follow-up questions, refine visualizations, and complete complex calculations.
Meta introduces LLAMA 2 Long
In a new research, Meta presents a series of long-context LLMs that support effective context windows of up to 32,768 tokens. The models are built through continual pretraining from Llama 2 with longer training sequences and on a dataset where long texts are upsampled.
On research benchmarks, the models achieve consistent improvements on most regular tasks and significant improvements on long-context tasks over Llama 2. Notably, with a cost-effective instruction tuning procedure that does not require human-annotated long instruction data, the 70B variant can already surpass gpt-3.5-turbo-16k's overall performance on a suite of long-context tasks.
Google announces Google-Extended and opens SGE to teens
Google introduced Google-Extended, a new control that web publishers can use to manage whether their sites help improve Bard and Vertex AI generative APIs, including future generations of models that power those products. This will allow publishers to control access to content on their site to train these AI models.
In another update, Google has opened up access to SGE in Search Labs to more people, specifically teens (ages 13-17) in the U.S., so they too can benefit from generative AI's helpful capabilities. Informed by research and experts in teen development, Google has built additional safeguards into the experience. For instance, to prevent inappropriate or harmful content from surfacing.
That's all for now!
If you are new to ‘The AI Edge’ newsletter. Subscribe to receive the ‘Ultimate AI tools and ChatGPT Prompt guide’ specifically designed for Engineering Leaders and AI Enthusiasts.
Thanks for reading, and see you on Monday. 😊