This Week in AI

Rabbit R1 first impressions

Perhaps the first new device of the AI era, the Rabbit R1 has finally shipped to first adopters. The impressive little orange box is built on Perplexity’s language model so it can understand your questions and deliver high quality answers. But it’s the much-vaunted action model that most people are excited about, making the R1 a useful assistant to help you get things done and handle tasks for you. At launch, it’s unfinished and only connects with four services—Uber, DoorDash, Midjourney, and Spotify—but many more integrations are planned. Reviews from The Verge and Mashable.

Groq’s breakthrough AI chip achieves blistering performance

Using an LLM like ChatGPT requires…..patience. Generating the tokens needed to create an image, respond to an inquiry, or write code takes time. In our time-pressed world, a few seconds can feel like an age, and destroys any sense of immediacy, something we expect with typical human-to-human conversations. Groq’s new tensor streaming processor just achieved 800 tokens per second (about 500 words per second) on Meta’s LLaMA 3 model. That’s really fast, and puts Groq in a great position in the AI inferencing marketplace (their chips aren’t useful for AI training).

Adobe’s impressive AI upscaling makes blurry videos look HD

Time to dust off all the video you took in the ‘90s and 2000s but never look at because it’s so low resolution it looks terrible on your giant HDTV. Adobe’s latest AI innovation, VideoGigaGAN, published in a recent paper, upscales video by up to eight times. Demos are quite compelling.

Video: How to build an AI data center

Nvidia shared an animation at their March GTC event showing how their new Blackwell AI chip fits into a huge AI data center, something Nvidia calls an “AI factory.” The company just released that video. Even if you don’t get what all the parts are, or read the geeky specs, it’s cool to see how it comes together (2 min, 28 sec)

AI Innovation

Generative AI arrives in the gene editing world of CRISPR

A hot startup has published a research paper outlining a new technology to generate blueprints for microscopic biological mechanisms to edit your DNA, which may lead to a future where scientists can battle diseases with great precision and speed, using hyper-personalized gene and cell therapy.

Microsoft introduces Phi-3, a model that redefines what’s possible with SLMs

Often you don’t need a large language model to get the job done; sometimes a smaller model (SLM) is the most cost-effective and efficient solution for tasks including language, reasoning, coding, and math. Microsoft is making new Phi-3 SLM models open and available to developers. Context windows range from 4K to 128K tokens, model sizes range from 3.8B (mini) to 7B (small), and 14B parameters, and they are tuned for instruction following. Not to be left out, Apple also released eight new even smaller models, named OpenELM, ranging from 270M to 3B parameters, which hint at how AI might come to the iPhone.

Intel builds world’s largest neuromorphic system for more sustainable AI

The current trajectory of AI is unsustainable. Today’s large language models use a brute force approach. They consume vast computing resources and gobble up energy. As models scale, the number of global users expands, and AI becomes integrated into every business, we are on an unsustainable path. We probably need a different approach. Recognizing the challenge, Intel has built the world’s largest neuromorphic computer in partnership with Sandia National Laboratories. Today’s AI systems use GPUs to simulate simple neural networks that are based on the rudimentary understanding of the brain we had in the 1960s. By contrast, neuromorphic computers take inspiration from the latest scientific understanding of how brains work and use artificial neurons. The human brain consumes about 20W of energy (similar to a single LED light bulb), a proof point that we can do much better. Here’s Intel’s press release with more details.

Meta challenges transformer architecture with Megalodon

Transformers, invented by Google in 2017, underpin all of the generative AI capabilities delivered in the last few years, from ChatGPT to Microsoft Copilot, to Sora. Transformers are amazing, but have limitations and don’t scale well. The new Megalodon model, proposed as a successor to the transformer, can have a context window with millions of tokens without requiring huge amounts of memory. When you double the size of the input to a transformer, the memory and computation time required to process that input quadruples. Other transformer alternatives include Mamba and liquid neural networks.

AI Insights

Why the AI industry’s thirst for new data centers can’t be satisfied

Two weeks ago, ARM’s CEO warned AI could devour 25% of U.S. electricity by 2030, and computer makers are always scrambling to get AI chips from companies like Nvidia and Intel. Now, the Wall Street Journal reports that the frenzy to build AI data centers is leading to a shortage of property and other computer parts including cooling systems, transceivers, and backup generators. A new data center currently pops up somewhere in the world every 3 days and Microsoft spent $30 billion building data centers in 2023 alone. In a related story, Reuters reports that AI is expected to fuel a natural gas boom.

Visualizing AI patents by country

Guess which country filed the most AI-related patents since 2012? Here’s a clue: it’s the same country that filed the most AI patents in 2022 (the most recent year data is available for) and it’s NOT the United States. Not by a long shot.

When memory meets privacy, an LLM becomes a life-long companion

Psychology Today explores how private LLMs are set to revolutionize personal assistants with hyper personalized interactions that help users reach their full potential, support personal growth, and optimize routines. The key to success will be privacy, security, empowering user control, and giving AI a memory to enable continuity and personalization.

Toolkit for the Future

Get smarter on AI.

What’s the secret to staying ahead of the curve in the world of AI? Information. Luckily, you can join early adopters reading The Rundown– the free newsletter that makes you smarter on AI with just a 5-minute read per day.

Sign up with one click

Capacity: Deflect tickets, emails and calls

Relieve pressure on your support team and serve customers better and faster with a support automation platform. AI answers 90% of queries and escalates the rest to your team. It automatically updates its knowledge database so it’ll never escalate that same issue again. Supports SMS, social, email, and phone communication.

Maika AI: AI companion for content creators

A versatile AI assistant built for content creators. Free to try and helps you to write text, edit content, change writing tone, create images, generate audio, easily translate text into other languages, quickly summarize hours-long YouTube videos, generate memes, and more.

Firstbase: The easy way to start a new company

Incorporate your company, access one-click growth tools, stay compliant, and manage everything your business needs — all online, from anywhere. Launch your U.S. business in minutes with no paperwork or legal headaches.

Browse AI: Extract and monitor data from any website

Extract data from websites and place it in a spreadsheet that fills itself and notifies you of any changes. Pre-built robots make it a breeze to get started. Extract whatever you need: job listings, property details, TikTok videos, hotel info, or competitor pricing. Free to try. No coding required.

Reclaim.AI: AI assistant for your calendar

Maintain healthy calendar habits, improve your productivity, optimize cross-team meetings, boost collaboration, and improve work-life balance. Reclaim schedules 1:1s and defends your calendar so you can focus. Syncs multiple calendars and integrates with Google Calendar, Zoom, Slack, and HubSpot.