Google's Gemma 2: Primed for Practical Deployments
Plus: OpenAI’s CriticGPT finds GPT-4’s mistakes with GPT-4, Google's partnerships to help AI with real-world facts.
Hello Engineering Leaders and AI Enthusiasts!
Welcome to the 307th edition of The AI Edge newsletter. This edition features Google’s open Gemma 2 models.
And a huge shoutout to our amazing readers. We appreciate you😊
In today’s edition:
🚀 Google releases Gemma 2, a set of 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
📚 Knowledge Nugget: Enlightenment AI by
Let’s go!
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.
Why does it matter?
The models sound like they are built for practical deployments. They come in practical sizes so that they can be easily deployed while being amazing in quality due to best-in-class performances.
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.
Why does it matter?
With more advances in reasoning and model behavior, AI models’ mistakes can become more subtle for AI trainers to spot. CriticGPT is a step towards addressing this fundamental limitation of RLHF.
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.
Why does it matter?
Earlier, Google announced efforts to ground Vertex AI results using web data and a plan to allow companies to ground AI systems in their own internal data.
Now, it is grounding these systems in known factual data from third parties, which could significantly lessen hallucinations and make AI more trustworthy for enterprise customers.
Enjoying the daily updates?
Refer your pals to subscribe to our daily 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.
Knowledge Nugget: Enlightenment AI
In this interesting post, author
writes about a curious, recurrent pattern he’s observed in long, winding conversations with AI chatbots: its propensity for “enlightenment thinking.”He draws on a dialogue with Claude and narrates how it seemed to reach a state of ecstatic enlightenment. The explorations raise questions about its inner world model and the nature of the fluid landscape of meaning humans and AI traverse together, the semioscape.
Why does it matter?
While not conclusive, the article contributes to ongoing debates about AI consciousness and whether current models can be said to understand or experience in any meaningful sense.
What Else Is Happening❗
🤝TIME and OpenAI announced a multi-year content deal
OpenAI will gain access to current and historic content from TIME's extensive archives from the last 101 years to enhance its products. It will also enable TIME to gain access to OpenAI's technology to develop new products for its audiences. (Link)
🌍Google is using AI to add 110 new languages to Google Translate
It is Google’s largest expansion ever, thanks to its PaLM 2 LLM. It includes languages like Cantonese, NKo, and Tamazight, representing more than 614 million speakers and opening up translations for around 8% of the world’s population. (Link)
🎼YouTube is in talks with major record labels for an AI music deal
It is offering to pay Universal Music Group (UMG), Sony Music Entertainment, and Warner Records “lump sums of cash” in exchange for legally licensing their songs to train new AI music tools. These will likely be one-off payments, not royalty-based arrangements. (Link)
🤖Meta to start testing user-created AI chatbots on Instagram
CEO Mark Zuckerberg announced yesterday that Meta will begin to surface AI characters made by creators through Meta AI studio on Instagram, starting in the U.S. These will primarily show up in messaging for now and will be clearly labeled as AI. (Link)
📞Character.AI now allows users to talk with AI avatars over calls
Users can initiate calls with a user-generated AI character directly with a button tap. Users can also switch between calling and texting seamlessly and stop the AI from talking through a “Tap to interrupt” option. The feature currently supports only a few languages. (Link)
New to the newsletter?
The AI Edge keeps engineering leaders & AI enthusiasts like you on the cutting edge of AI. From machine learning to ChatGPT to generative AI and large language models, we break down the latest AI developments and how you can apply them in your work.
Thanks for reading, and see you tomorrow. 😊