Weekly AI Recap: Meta unveils Llama 3, Claude 3 Opus gets added to Amazon Bedrock
Plus, AI takes center stage at TED2024.
![Llama 3 Llama 3](https://thedrum-media.imgix.net/thedrum-prod/s3/untitled_design_-_2024-04-18t162525.514.png?w=608&ar=default&fit=crop&crop=faces&auto=format)
Meta has released 8bn- and 70bn-parameter versions of Llama 3. / Adobe Stock
First two versions of Llama 3 released
Meta announced yesterday that it had released the first two iterations of its Llama 3 collection of open-source large language models (LLMs).
The new models, which have 8bn and 70bn parameters respectively, are capable of "state-of-the-art performance on a wide range of industry benchmarks and [offer] new capabilities, including improved reasoning,” Meta wrote in a blog post. “We believe these are the best open source models of their class, period.” (For comparison, OpenAI’s proprietary GPT-4 LLM reportedly has around 1.7tn parameters.)
Llama 3’s training data was collected entirely from publicly available sources, according to Meta.
The 8bn- and 70bn-parameter models are only capable of generating text and have limited language abilities beyond English, but Meta says that it's working on more advanced versions. “We are embracing the open source ethos of releasing early and often to enable the community to get access to these models while they are still in development,” the company writes. “Our goal in the near future is to make Llama 3 multilingual and multimodal…and continue to improve overall performance across core LLM capabilities such as reasoning and coding.”
Meta added in the blog post that the new Llama 3 models will soon be accessible on Microsoft Azure and Google Cloud. They've already been added to Amazon SageMaker, a platform for building and deploying cloud-based machine learning models.
Amazon deploys Claude 3 Opus on Bedrock
Amazon Web Services (AWS) announced in a blog post on Tuesday that Anthropic’s Claude 3 Opus, a recently released LLM which reportedly outperforms competitors like OpenAI’s GPT-4 in certain domains, like mathematics, is now available on Amazon Bedrock. (The other two, less powerful members of the Claude 3 trio, Sonnet and Haiku, were incorporated into Bedrock last month.)
“With the availability of Claude 3 Opus on Amazon Bedrock, enterprises can build generative AI applications to automate tasks, generate revenue through user-facing applications, conduct complex financial forecasts and accelerate research and development across various sectors,” AWS principal developer advocate Channy Yun wrote in the blog post.
First unveiled almost one year ago to the day, Amazon Bedrock is a platform which allows AWS customers to build their own foundation models leveraging tech from Amazon-funded AI companies like Anthropic and Stability AI.
Amazon has invested $4bn in Anthropic since Semptember.
AI spotlighted at TED2024
Like many other major conferences over the past year – SXSW and Davos, for example – TED2024 has placed a heavy emphasis on AI.
This year’s TED conference (the name stands for Technology, Education, Design) kicked off in Vancouver on Monday. Prominent speakers have included Google DeepMind cofounder and CEO Demis Hassabis, AI researcher Fei-Fei Li, recently appointed Microsoft AI CEO Mustafa Suleyman and AI policy expert (and former OpenAI board member) Helen Toner.
In a recent edition of his email newsletter, author and AI expert Gary Marcus was critical of the evangelical tone displayed by some of those AI experts speaking at TED 2024. Suleyman, for example, reportedly claimed that the hallucination problems currently plaguing LLMs would be resolved “soon,” and spoke with enthusiasm about the future of AI, which he likened to the arrival of “a new digital species.”
“I don’t see how we get to a positive AI future if we sweep all the problems under the rug,” Marcus wrote.
IAS brings Global Alliance for Responsible Media (GARM) misinformation framework to clients
Digital ad optimization company Integral Ad Science (IAS) announced earlier this week that it’s stepping up its efforts to prevent advertisers’ content from appearing alongside misinformation circulating on Meta platforms Facebook, Instagram and Reels.
The IAS’ AI-powered Total Media Quality tool, designed to ensure that clients' ads do not appear alongside potentially offensive content, has now been equipped with the World Economic Forum’s Global Alliance for Responsible Media (GARM) framework for identifying online misinformation.
“We’re excited to evolve our TMQ solution to include the GARM category of misinformation which gives marketers more control over their brands’ reputations, especially ahead of this year’s US elections,” IAS CEO Lisa Utzschneider said in a statement. “We are constantly innovating and building on our partnerships with the world’s largest platforms like Meta, providing added confidence for our customers while ensuring they are driving superior results through IAS’s data and technology.”
Google adds generative image tools to Demand Gen
Google announced on Tuesday that it had launched new generative image tools in Demand Gen, a platform through which Google Ads customers can build campaigns using AI. The new tools enable advertisers to create new images from text prompts – the new system will never create two identical images, according to a blog post from Google – or to create new iterations of existing images.
A brief online demo shows a user prompting Demand Gen to create images of “vibrantly colored tents illuminated under the Aurora Borealis with snow-capped mountains in the background.” After a series of images are generated, the user selects one and drops it into a YouTube ad. According to Google, All images created by Demand Gen ill be marked as being AI-generated.
The new features are currently only able to process prompts in English, but Google says that other language capabilities will be added “later this year.”
The tech giant recently found itself in hot water after Gemini, its multimodal LLM, was found to have a propensity for generating historically inaccurate images and other content.
Digitas launches AI-powered platform
On Monday, marketing agency Digitas unveiled Digitas AI, a platform designed to help internal strategists and clients develop stronger marketing campaigns through the use of generative AI.
Digitas AI centers on the Agent Management System (AMS), a tool which enables marketers to craft AI personas – or agents – based on audience data. “One prime example of an agent is the ‘It Girl’ – an agent that aggregates insights on the latest social platform trends, acting as a conversational 'teammate' ready to share the latest trends and social platform hacks,” Digitas wrote in a press release.
AMS, according to Digitas, will make it easier for marketers to ideate, test and implement new strategic approaches to campaigns.
Publicis Groupe, Digitas’ parent company, recently announced a plan to invest around $326mn in AI over the next three years.
For more on the latest happenings in AI, web3 and other cutting-edge technologies, sign up for The Emerging Tech Briefing newsletter.