AI Weekly Rundown (June 22 to June 28)
Major AI announcements from Anthropic, Synthesia, OpenAI, Amazon, and more.
Hello Engineering Leaders and AI Enthusiasts!
Another eventful week in the AI realm. Lots of big news from huge enterprises.
In today’s edition:
🤝 Apple and Meta are discussing a generative AI partnership
🔧 ByteDance and Broadcom collaborate on AI chip development
🕵️♂️ Researchers developed a new method to detect hallucinations
🎥 Synthesia 2.0: World’s 1st AI video communication platform
🛒 OpenAI is on an acquiring spree, buying Rocket and Multi
🎵 Record labels sue AI music startups over copyright infringement
💼 Anthropic rolls out Claude’s cutting-edge collaborative features
🤖 Google experiments with celebrity-inspired AI Chatbots
🛑 OpenAI postpones the launch of ChatGPT voice mode
🏁 Amazon steps into the chatbot race with Metis
🎨 Figma’s new AI features stir competition with Adobe
🥇 Alibaba’s Qwen-72B tops Hugging Face’s Open LLM Leaderboard
🚀 Google releases Gemma 2, lightweight but powerful open LLMs
🔍 OpenAI’s CriticGPT finds GPT-4’s mistakes with GPT-4
🌐 Google partners with Moody’s, Thomson Reuters & more for AI data
Let’s go!
Apple and Meta are discussing gen AI partnership
Apple is reportedly in talks with its longtime rival Meta to integrate the latter's Llama 3 AI model into Apple Intelligence. This move comes as Apple prepares to roll out its AI features across iPhones, iPads, and Macs later this year.
The potential partnership follows Apple's existing deal with OpenAI, suggesting a collaboration strategy rather than solo development in the AI race. In Apple's arrangement with OpenAI, there's no direct payment. Instead, OpenAI can offer premium subscriptions through Apple Intelligence, with Apple taking a percentage. It's unclear if Meta would agree to a similar business model, given that Llama 3 is open-source and free to access.
ByteDance and Broadcom collaborate on AI chip development
ByteDance is collaborating with U.S. chip designer Broadcom to develop an advanced AI processor. This partnership aims to secure a stable supply of high-end chips amid ongoing U.S.-China tensions. The project centers on creating a 5-nanometre, customized Application-Specific Integrated Chip (ASIC) that complies with U.S. export restrictions.
This chip's manufacturing is set to be outsourced to Taiwan Semiconductor Manufacturing Company (TSMC), though production is not expected to begin this year. While the design work is currently underway, the critical "tape out" phase has yet to commence.
Researchers developed a new method to detect hallucinations
ChatGPT and Gemini can produce impressive results but often "hallucinate" false or unsubstantiated information. This research focuses on a subset of hallucinations called "confabulations," where LLMs generate answers that are both wrong and arbitrary. Researchers have developed new methods to detect confabulations using entropy-based uncertainty estimators. They introduce the concept of "semantic entropy" to measure the uncertainty of LLM generations at the meaning level.
High semantic entropy corresponds to high uncertainty and indicates a higher likelihood of confabulation. The method computes uncertainty at the level of meaning rather than specific word sequences, addressing the fact that one idea can be expressed in many ways. The method provides scalable oversight by detecting confabulations that people might otherwise find plausible.
Synthesia 2.0: World’s 1st AI video communication platform
Synthesia is launching Synthesia 2.0 - the world's first AI video communications platform for businesses. It reinvents the entire video production process, allowing companies to create and share AI-generated videos at scale easily.
The key new features and capabilities of Synthesia 2.0 include:
2 Personal AI Avatars: Expressive Avatars shot in a studio and Custom Avatars created using your webcam.
AI Video Assistant: Converts text, documents, or websites into high-quality videos, with options to customize the branding, tone, and length.
Intuitive Video Editing: Editing simplified with "Triggers" that let you control animations and edits from the script.
Translation and Dynamic Video Player: Videos can now be translated into over 120 languages. Synthesia is also building a new video player with interactive features.
AI Safety Focus: Synthesia is pursuing ISO/IEC 42001 certification, the first standard for responsible AI management, to ensure its AI technologies are ethical.
OpenAI is on an acquiring spree, buying Rockset and Multi
Last week, OpenAI acquired Rockset, a startup that develops tools for real-time data search and analytics. OpenAI said it would integrate Rockset's technology to power its infrastructure and offerings across products.
This week, OpenAI acquired Multi, a startup focused on building remote collaboration tools and software. Technically, the deal is an acqui-hire as the entire Multi team, including its co-founders, will join OpenAI to work on the company's ChatGPT desktop application.
Record labels sue AI music startups over copyright infringement
The world's major record labels, including Universal Music Group, Sony Music, and Warner Music, have filed twin lawsuits against the AI music generation startups Suno and Udio. The lawsuits accuse the companies of unlawfully training their AI models on massive amounts of copyrighted music, which, according to the complaints, allows the startups to generate similar-sounding music without permission.
The record labels allege Suno and Udio have effectively copied artists' styles and specific musical characteristics. The labels claim the AI-generated music is so close to the original that it is eerily similar when transcribed into sheet music. The lawsuits also accuse the startups of making it easy for people to distribute AI-created samples that mimic copyrighted recordings on platforms like Spotify.
Anthropic rolls out Claude’s cutting-edge collaborative features
Anthropic has introduced new collaboration features for Claude. These features include:
Projects: Projects in Claude allow integration of internal resources like style guides or codebases, enhancing Claude's ability to deliver tailored assistance across various tasks. Users can set custom instructions for each Project to modify Claude's tone or perspective for a specific role or industry.
Artifacts: It allows users to generate and edit various content types like code, documents, and graphics within a dedicated window. This benefits developers by offering larger code windows and live previews for easier front-end reviews.
Sharing Features: Claude Team users can share snapshots of their best conversations with Claude in their team’s shared project activity feed.
Additionally, any data or chats shared within Projects will not be used to train Anthropic’s generative models without a user’s explicit consent.
Google experiments with celebrity-inspired AI Chatbots
These chatbots will be powered by Google’s Gemini family of LLMs. The company aims to strike partnerships with influencers and celebrities and is also working on a feature that allows people to create their own chatbots by describing their personalities and appearances.
The project is led by Ryan Germick, a longtime executive at Google and a team of ten. These chatbots could be an experiment and may only appear on Google Labs rather than being widely available.
OpenAI postpones the launch of ChatGPT voice mode
Originally planned for late June, the Voice Mode aims to provide a more naturalistic and conversational experience with the AI chatbot, complete with emotional inflection and the ability to handle interruptions.
However, it will now be available only to a small group of users in late July or early August. OpenAI is working on improving content detection and user experience before wider rollout. GPT-4o's real-time voice and vision capabilities are also expected to roll out to ChatGPT Plus users soon.
Enjoying the weekly updates?
Refer your pals to subscribe to our newsletter and get exclusive access to 400+ game-changing AI tools.
When you use the referral link above or the “Share” button on any post, you'll get the credit for any new subscribers. All you need to do is send the link via text or email or share it on social media with friends.
Amazon steps into the chatbot race
Amazon is reportedly working on a new consumer-focused chatbot codenamed “Metis.” It is planned to be released somewhere around September. Here’s what we know about it:
The chatbot is powered by a new model, Olympus, and can be accessed via a web browser.
It uses a retrieval-augmented generation (RAG) technique to provide up-to-date information and automate tasks.
The model conversationally provides text and image-based outputs, suggesting follow-ups to queries. It also shares links to sources and supports image generation.
It uses an infrastructure similar to Amazon’s upcoming voice assistant, Remarkable Alexa.
Figma’s new AI features stir competition with Adobe
Figma announced a range of new features at the 2024 Config conference. Significant ones include a UI redesign, generative AI tools, new icons and toolbar, AI-enhanced asset search, and auto-generated texts in designs.
For instance, by typing a simple prompt into the textbox, users can create an entire app design mock-up for a restaurant. Figma will connect the design pages and even write suggested content!
Figma has also added a few designer-specific features to allow users to tweak designs in real-time. It features a developer mode with a “ready-for-dev” task list. The upgrade also boasts Figma slides, a Google slides-like tool for building and sharing presentations.
Alibaba’s Qwen-72B tops the Hugging Face leaderboard
Hugging Face’s latest open large language model leaderboard ranks and evaluates open LLMs based on benchmarks like MMLU-pro and tests them on high-school and college-level problems.
The platform used 300 NVIDIA H100 GPUs to re-evaluate major open LLMs to obtain updated rankings. Chinese company Alibaba’s Qwen-72B dominated the leaderboard, becoming a top performer overall.
Not just that, the leaderboard was mainly dominated by Chinese companies, highlighting their headway into the open LLM space.
Google’s Gemma 2, a set of lightweight, powerful open LLMs
Google has released Gemma 2 set of models that punch above their weight classes. Available in 9B and 27B parameter sizes, these models are
Higher performing and more efficient at inference than the first-generation
Have significant safety advancements built in
Optimized to run at incredible speed across a range of hardware and easily integrate with other AI tools
Trained on 13 trillion tokens for 27B, 8 trillion for 9B, and 2 trillion for 2.6B model (en route)
27B performs better than Llama3-70B and Nemotron-340B on Lmsys Arena, making it best in its size and stronger than some larger models. While 9B outperforms the likes of Mistral-large and Qwen1.5-110B.
The 27B Gemma 2 model is designed to run inference efficiently at full precision on a single Google Cloud TPU host, NVIDIA A100 80GB Tensor Core GPU, or NVIDIA H100 Tensor Core GPU. Moreover, this is an open weights model line, currently only available to researchers and developers.
OpenAI’s CriticGPT finds GPT-4’s mistakes with GPT-4
OpenAI trained a model based on GPT-4, called CriticGPT, to catch errors in ChatGPT's code output. It found that when users get help from CriticGPT to review ChatGPT code, they outperform those without help 60% of the time.
OpenAI aligns GPT-4 models to be more helpful and interactive through Reinforcement Learning from Human Feedback (RLHF). A key part of RLHF is collecting comparisons in which people, called AI trainers, rate different ChatGPT responses against each other.
OpenAI is beginning to integrate CriticGPT-like models into its RLHF labeling pipeline, providing trainers with explicit AI assistance.
Google's partnerships to help AI with real-world facts
Google is partnering with reputable third-party services, such as Moody’s, MSCI, Thomson Reuters, and Zoominfo, to ground its AI with real-world data. These four will be available within Vertex AI starting next quarter. They will offer developers qualified data to backstop their model outputs and ensure responses are factually accurate.
Google is also announcing high-fidelity grounding. Available through an experimental preview, it’s designed to help AI systems work better with a given set of specific information.
That's all for now!
Subscribe to The AI Edge and gain exclusive access to content enjoyed by professionals from Moody’s, Vonage, Voya, WEHI, Cox, INSEAD, and other esteemed organizations.
Thanks for reading, and see you on Monday. 😊