Welcome back!

Google is making waves once again with the launch of Gemma 2, an open-source model which offers Llama 3 70B performance at half the size, and significant upgrades to Gemini 1.5 Pro. These advancements demonstrate Google's commitment to top-tier AI performance and efficiency, further consolidating its position as a leader. Let’s dive in…

Today’s Summary:

• Google releases Gemma 2 models

• Meta tests creators chatbots on Instagram

• OpenAI introduces CriticGPT

• Meta LLM compiler for code

• New multimodal vision benchmark

• ChatGPT Voice Mode demo

• Toys”R”Us AI-powered commercial

• AI drops hats on New Yorkers

• 2 new tools

TOP STORY

Google Boosts AI Lineup: Gemma 2 Launch and Gemini 1.5 Pro Upgrades

The Summary: Google has unveiled major AI advancements by releasing the new Gemma 2 open-source models and several upgrades to Gemini 1.5 Pro.

Gemma 2 offers top-tier performance in 9B and 27B sizes, with 27B surpassing Llama-3-70B, while Gemini 1.5 Pro has now a huge 2-million token context window and new code execution capabilities.

Key details:

• Gemma 2 available in 9B and 27B parameter sizes

• Knowledge Distillation used to optimize model size

• 9B ranks as the best < 15B local open-source model

• 27B beats Llama-3 70B in LMSys Elo benchmarks

• Gemini 1.5 Pro upgraded to 2-million token context window

• Added code execution capabilities in Gemini API

• Access Gemma 2 and Gemini upgrades via Google AI Studio

• Test Gemma 2 27B on Hugging Chat

Why it matters: Gemma 2 raises again the bar for high-performance open source models, while reducing deployment costs. The open nature of Gemma 2 allows for customization and integration into community projects. These updates confirm that Google keeps pushing AI forward and remains a top AI innovator and a key player to watch.

META

Meta Tests Creator AI Chatbots on Instagram

The Summary: Meta is testing AI chatbots on Instagram, enabling creators to make AI versions of themselves. These chatbots, clearly marked as AI, will interact in messaging. The test starts in the US with about 50 creators and a small user group, with plans for a full launch by August. CEO Mark Zuckerberg views this as a way for creators and businesses to enhance community engagement.

Key details:

• Initial test limited to the US

• Chatbots clearly marked as AI

• Test includes creators like Wasted and Don Allen Stevenson III

• Gradual roll out to more users over next few months

• Meta AI Studio used for chatbot creation

• Zuckerberg describes it as "early days and first beta version"

Why it matters: This move by Meta opens new ways for creator-fan interaction, potentially transforming influencer engagement. However, it also raises questions about authenticity in social media and the increasing role of AI in content creation.

OPENAI

CriticGPT: The AI That Checks Other AIs

The Summary: OpenAI has developed CriticGPT, an AI system designed to identify mistakes in AI-generated code. CriticGPT outperformed human contractors in catching bugs and was often preferred over human-written critiques.

CriticGPT uses reinforcement learning from human feedback (RLHF) and a novel search technique to improve AI safety as models become more advanced and complex.

Key details:

• CriticGPT identified 85% of bugs vs 25% by human contractors

• Model critiques preferred over human critiques 80% of the time

• Uses a "tampering" process to insert subtle bugs into AI-generated code

• Trained with RLHF and a new "Force Sampling Beam Search" technique

• Generalized beyond code to detect errors in 24% of previously "flawless" AI responses

Why it matters: As AI systems advance, evaluating their outputs becomes increasingly difficult. This research shows a promising approach to scalable oversight, where AI systems help in assessing and improving other AI systems. By catching errors, CriticGPT could lead to safer and more reliable AI assistants.

QUICK NEWS

Quick news

• Meta LLM Compiler for code optimization and compiler reasoning

• New multimodal benchmark for vision language models

• Yet another ChatGPT Voice Mode demo

• Toys”R”Us created a brand commercial with Sora

• AI drops virtual hats onto New Yorkers

TOOLS

🥇 New tools

• Figma Slides - Create any presentations and slides

• Character Calls - Phone calls to AI characters