CM3Leon and Llama 2, Generative TV & AI Showrunner Agents, Custom instructions for ChatGPT, WormGPT, Bing Chat Enterprise and more
Greetings and welcome to this week's AI Brews for a concise roundup of the week's major developments in AI.
In today’s issue (Issue #24):
AI Pulse: Weekly News & Insights at a Glance
AI Toolbox: Product Picks of the Week
AI Skillset: Learn & Build
🗞️🗞️ AI Pulse: Weekly News & Insights at a Glance
🔥 News
Meta released Llama 2, the next generation of Meta’s open source Large Language Model, available for research & commercial use. Compared to Llama v1, it was trained on more data (~2 trillion tokens) and supports context windows up to 4k tokens. Llama 2 outperforms other open source language models on many external benchmarks, including reasoning, coding, proficiency, and knowledge tests. Microsoft is Meta’s preferred partner for Llama 2, which will be optimized to run locally on Windows [Details ].
Llama 2 70B Chat model is available free on HuggingChat.
San Francisco startup Fable presents SHOW-1, a Showrunner AI tech that can create personalized TV episodes, from a prompt, with the user as the star . The AI Showrunner Agents, outlined in Fable's research paper, have the ability to write, produce, direct, cast, edit, voice, and animate TV episodes [Details | Paper].
Meta has developed CM3Leon, a new multi-modal language model that excels in text-to-image generation and image captioning. Unlike most image generators that rely on diffusion, CM3Leon is a transformer model. It is more efficient, requiring five times less compute and a smaller training dataset than previous transformer-based methods [Details | Paper].
OpenAI is rolling out custom instructions for ChatGPT, that will persist from conversation to conversation. By setting preferences, like a teacher specifying they're teaching 3rd-grade science or a developer wanting non-Python efficient code, ChatGPT will consider them in all future interactions. This feature isn't currently available in the UK and EU [Details].
Google Deepmind presents CoDoC (Complementarity-driven Deferral-to-Clinical Workflow), an AI system that learns to decide when to rely on the opinions of predictive AI tools or defer to a clinician for the most accurate interpretation of medical images. The code is open-source [Details].
Stability AI launch new developer platform site, with integrated sandbox environment merging the product and code surface areas [Details |Developer platform].
Researchers present TokenFlow - a framework for text-driven video editing. It creates high-quality videos from a source video and a text-prompt, maintaining the input video's spatial layout and dynamics, without needing training or fine-tuning [Details].
MosaicML released MPT-7B-8K, a 7B parameter open-source LLM with 8k context length. It can be fine-tuned on domain-specific data on the MosaicML platform [Details].
AssemblyAI announced Conformer-2, their latest AI model for automatic speech recognition trained on 1.1M hours of English audio data with improvements on proper nouns, alphanumerics, and robustness to noise [Details].
LangChain launches LangSmith, a unified developer platform for debugging, testing, evaluating, and monitoring LLM applications [Details].
Microsoft announced, at its annual Inspire conference, new AI features to Azure, including the public preview of Vector search in Azure Cognitive Search and Document Generative AI solution to chat with documents [Details].
Microsoft is rolling out Bing Chat Enterprise for businesses - Chat data is not saved, no one at Microsoft can view it or use it to train the models [Details].
OpenAI is raising the ChatGPT Plus message limit for GPT-4 customers to 50 every 3 hours, to be rolled out in the coming week [Details].
Qualcomm and Meta will enable Llama 2, to run on Qualcomm chips on phones and PCs starting in 2024 [Details].
Wix’s new generative AI tool can create entire websites from prompts [Details].
Apple has been working on its own AI chatbot ‘Apple GPT’ and framework, codenamed ‘Ajax’, to create large language models [Details].
FTC investigates OpenAI over data leak and ChatGPT’s inaccuracy [Details].
SAP invests in generative AI startups Anthropic, Cohere and Aleph Alpha [Details].
🔦 Weekly Spotlight
WormGPT – The Generative AI tool cybercriminals are using to launch business email compromise attacks [Link].
A Twitter thread on using Bard's new features, such as extracting a text summary from an invoice image, and converting an image of a mathematical equation into Latex etc. [Link].
Study claims ChatGPT is losing capability, but some experts aren’t convinced [Link].
🔍 🛠️ AI Toolbox: Product Picks of the Week
💎 Air: Air can perform up to 40 minute long sales & customer service calls over the phone that sound like a human. It can also perform actions autonomously across 5,000 unique applications. In closed beta.
💎 Simplescraper AI: Pull insights from any Website using AI. Summarize, Analyze, and extract understanding from any data on the web
💎 InstaVerse: Powered by Blockade Labs, InstaVerse is an AI-powered 3D asset generator and visualizer, that creates explorable worlds directly from text input.
💎 Superhuman AI: Generative AI features launched in the popular email client, Superhuman. Superhuman AI matches the voice and tone in the emails you've already sent, applying that to everything it creates.
📕 📚 AI Skillset: Learn & Build
Generative AI in Financial Services with Moody's at Microsoft Inspire 2023 [YouTube link].
How to use AI to do stuff: An opinionated guide [Link].
Generate images with “Hidden” text using Stable Diffusion and ControlNet [Link].
Building LLM applications for production [Link].
If you find value in AI Brews, consider sharing it with a friend. Thanks for reading and have a nice weekend! 🎉 Mariam.
Just a heads up, MS Edge tags all of your links as "not private."