Sponsored by

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each Thursday, we aggregate the major developments in artificial intelligence; we pass along the news, useful resources, tools and services, and highlight the top research in the field as well as exciting developments in open source. Even if you aren’t an engineer, we’ll keep you in touch with what’s going on in AI.

NEWS & OPINION

Google rebrands Bard to Gemini

-------------------------

Google has announced a rebranding and expansion of its AI chatbot and assistant, Bard, into Gemini. The rebrand marks a new era for Google's AI offerings, with advanced capabilities, a dedicated app, and a redefined subscription model to access their most powerful model for Gemini, Ultra. Here's everything you need to know about Gemini:

• Performance and availability: Last month, Google upgraded Bard with Gemini Pro, the second most capable model in the Gemini series. This improvement led to Bard recently surpassing all models besides GPT-4 Turbo on a crowdsourced LLM ranking/leaderboard. With a Gemini Advanced subscription, users can now access the Ultra 1.0 model - it’s more capable than Pro at reasoning, following instructions, coding, and has an expanded context window. It’s available (in English) starting today in more than 150 countries and territories.

• Free vs. subscription: Analogous to ChatGPT’s subscription service, users will be able to use the free tier of Gemini that leverages Gemini Pro, or pay $20/month to access the cutting edge model. The new subscription is called “Google One AI Premium” and also comes with 2TB of Google Drive storage and all the other features of the Google One subscription. Google is offering two months free with the announcement.

• Dedicated app and platforms: Gemini is accessible through a new dedicated app for Android, and its features are useable within the Google app on iOS. You can also now set Gemini as your default assistant, replacing Google Assistant. All of these AI services will be rolled into Gemini - that includes all of Google’s emerging AI features inside workspace apps like Gmail and Docs.

• Integration of Imagen: Late last week, Google integrated Imagen with Bard (now Gemini), providing users with a free-to-use AI image generator. It’s not quite state of the art, but it’s available on Gemini’s free tier and excels at embedding text into images, something many of the other top models still struggle with.

• More to come: CEO Sundar Pichai announced that next week, Google will have another major announcement aimed at developers and Cloud customers.

It’s worth noting that there is justified skepticism about some of Google’s benchmarking claims, and an early demo of Gemini drew criticism for being deceptive. We’re still excited to dig in (this announcement was just a few hours ago) and experiment with Gemini Advanced. It’s clear at this point that Google’s vision for Gemini is to be an all-in-one, multimodal AI assistant that will eventually be able to help you complete any digitally-based task. We’ll have more to report next week.

New Hampshire AG identifies Texas companies behind deepfake Joe Biden robocalls

-------------------------

The State of New Hampshire has issued subpoenas and cease-and-desist orders to two Texas companies linked to a robocall campaign that used a Joe Biden AI voice clone in an attempt to convince New Hampshire democrats not to vote in the state primary. The calls were made by a telecoms company called Lingo, which claims to have been transmitting on behalf of Life Corp - a Texas-based company that was cited by the FCC for similar activity back in 2003. New Hampshire AG John Formella described the calls as the clearest and possibly first known attempt to use AI to interfere with an election in the US. A criminal investigation is underway.

The voice clone likely originated from popular AI voice generation platform ElevenLabs, and the calls were spoofed to appear to come from the spouse of former New Hampshire Democratic Party official Kathy Sullivan.

Roblox releases a real-time AI chat translator

-------------------------

Online gaming giant Roblox has introduced a new feature to their platform enabling global users to communicate instantly in different languages. The service, built on a custom language model, currently supports 16 languages and allows users around the world to communicate effortlessly, with messages being translated into the viewer's native language in real time. The translation AI leverages linguistic similarities between languages for faster processing and is trained on open-source data, human-labeled translations, as well as common Roblox chat phrases.

Roblox has over 70 million daily users and over 2.4 billion messages are exchanged daily in 180 countries. Looking ahead, Roblox plans to extend this translation capability to other interface aspects and even introduce AI-powered voice translation for a fully localized user experience.

EU’s AI Act finally clears legislative process

-------------------------

On Friday, all 27 European Union countries approved the finalized version of the EU AI Act, the world’s first comprehensive lawbook for artificial intelligence. While the finalized copy has not yet been released by an official body, an 892 page section-by-section breakdown was leaked at the end of January by an EU policy journalist. It was not a given that the Act would be approved, as a few countries (notably France, who is developing one of the most capable models in the world) came in with major concerns about over-regulation of foundation models.

The Act categorizes AI systems into banned, high-risk, and general-purpose types based on potential impact/risk to society. Each category will have specific regulations that the new EU AI Office will help national authorities enforce.

Within the next few months, the AI Act will be published in the Official Journal of the European Union - and it will enter into force twenty days after that. From there, AI companies will have somewhere between 6 months and 2 years to comply (depending on the compliance issue/model) or face potential penalties up to 7% of global revenue.

• Confessions of an AI clickbait Kingpin

• Microsoft Copilot gets a big redesign and a new way to edit AI-generated images

• Police departments are turning to AI to sift through millions of hours of unreviewed body-cam footage

• Can AI unlock the secrets of the ancient world?

• AI’s massive cash needs: are a handful of big tech companies monopolizing the boom?

• More than 80% of company pitches now involve AI

• AI lobbying spikes 185% as calls for regulation surge

• Both OpenAI and Meta are trying to make it easy to spot AI-generated images

• Apple bought a record 32 AI startups in 2023

💥FREE AI & ChatGPT Workshop (Holiday Season Offer) 🎁

🚀Join this 3-hour power-packed workshop (worth $99) for FREE and learn hacks to 10X work output and grow your business. 

👉Click here to register (FREE for First 100) 🎁

With AI & ChatGPT, you will be able to:

✅ Make smarter decisions based on data in seconds using AI 

✅ Automate daily tasks and increase productivity & creativity

✅ Skyrocket your business growth by leveraging the power of AI

✅ Save 1000s of dollars by using ChatGPT to simplify complex problems 

👉 Click here to register (Limited seats: FREE for First 100 people only)🎁

TECHNICAL, RESEARCH & OPEN SOURCE

Hugging Face launches an AI assistant maker to answer OpenAI’s custom GPTs

-------------------------

Open source AI/ML platform Hugging Face has introduced Hugging Chat Assistants - a free space that enables users to effortlessly create tailored AI chatbots using various open-source LLMs, including Mixtral and Llama 2. The move challenges OpenAI's custom GPT Builder, providing more accessible and adaptable options for developers.

Phillip Schmid, Hugging Face's Technical Lead & LLMs Director, highlighted the simplicity of creating a new personal Hugging Face Chat Assistant "in 2 clicks!" Hugging Face has also established a central repository for sharing and accessing these customized assistants that mirrors OpenAI's GPT Store layout. As a user, you can view both the instructions for an Assistant and which model it uses.

Although some desired functionalities (like web search, RAG and image generation) aren’t available yet, the emphasis on user customization and cost efficiency positions Hugging Chat Assistants as a strong competitor to GPTs.

Apple open-sources a new AI model for instruction-based image editing

-------------------------

Apple has unveiled a groundbreaking open source AI model called "MGIE" (MultimodalLLM-Guided Image Editing), which promises to revolutionize the way we interact with image editing through natural language instructions. Utilizing multimodal large language models, MGIE can take user inputs to execute image manipulation down to the pixel, covering a broad spectrum of editing tasks. MGIE is the fruit of a collaboration with the University of California, Santa Barbara, showcased in a research paper for the International Conference on Learning Representations 2024.

MGIE operates by leveraging MLLMs to a) interpret user instructions into precise editing actions and b) to conceptualize the intended visual outcomes. This dual approach enables the model to grasp the essence of the user's request and create a visual representation that guides the actual image modification process. The model boasts a wide array of editing functionalities, including expressive instruction-based editing, Photoshop-style adjustments, global photo optimization, and detailed local editing. One of the UCSB researchers is hosting a demo of MGIE on Hugging Face.

Allen Institute for AI (AI2) takes a shot at Meta’s Llama 2 with OLMo

-------------------------

Late last week, AI2 (Microsoft co-founder Paul Allen's nonprofit research group) unveiled its large language model, OLMo (Open Language Model), a “truly open source” large language model. Following in the footsteps of fully open models like Hugging Face’s BLOOM, OLMo was published alongside its model code, weights as well as training code, data and evaluation suite - meaning users can see exactly how it was designed, trained and evaluated.

OLMo was built on Dolma, a dataset comprised of three trillion tokens that AI2 built from web content, academic publications, and books. The dataset is generally available for commercial applications and can be accessed via Hugging Face.

AI2 said OLMo will “empower academics and researchers to study the science of language models collectively.” By providing access to the full underlying training aspects, OLMo can enable less carbon to be used when fine-tuning the model as an open approach “radically reduces developmental redundancies, which is critical in the decarbonization of AI.”

MORE IN T/R/OS

• Meta AI research: efficient tool use with “Chain-of-Abstraction” reasoning

• Research on state space models (SSMs): Can Mamba learn how to learn?

• Adeus: the world’s first open-source AI wearable device

• MIT and IBM find clever AI ways around brute-force math

• Abacus AI’s ‘Smaug-72B’ just topped Hugging Face LLM leaderboard

That’s it for this week! We’ll see you next Thursday.