☕️ A Model-Packed Arctic AI Blizzard

The week seems full of new model releases, from Snowflake unveiling its enterprise-grade, purpose-built LLM for business intelligence, Arctic, to Microsoft's debut of a family of lightweight language models, dubbed Phi-3. Meanwhile, Chinese-owned SenseTime's shares skyrocketed with the release of its ChatGPT-rivaling SenseNova 5.0 model, and Apple gives the world a taste of eight of its on-device small language models. Join us at AI Tangle as we untangle this week's happenings in AI.

Snowflake releases Arctic, its flagship enterprise model that excels at "business intelligence"

In the competitive arena of generative AI models, Snowflake, a giant in cloud computing, steps in with its Arctic LLM, tailor-made for the world of enterprise. With its "enterprise-grade" capabilities, Snowflake's Arctic LLM aims to tackle tasks like generating database code and developing high-quality chatbots or what it calls "business intelligence," setting its sights on the needs of corporate users with an emphasis on open-source use with its Apache 2.0 license.

​

Where does Arctic stand?

Snowflake's Arctic LLM is a 480-billion parameter model built on the mixture of experts (MoE) architecture, similar to Databricks' DBRX and Google's Gemini Pro 1.5, the former of which Arctic is directly competing with. The model boasts a surprisingly small context window, though officials at Snowflake state that Arctic bests the likes of DBRX and Llama 2 70B at an impressive performance-to-cost ratio on popular benchmarks. As for accessibility, Snowflake pledges that Arctic will be available on Hugging Face, Microsoft Azure, and Together AI's Lamini, though the company is pitching users to favor its Cortex platform.

Phi-3 hits the shelves as Microsoft releases a family of small language models

Microsoft has recently launched a lightweight AI model, Phi-3-mini, the first of three small language models (SLMs). Sébastien Bubeck, Microsoft's vice president of GenAI research, says that Phi-3 is aimed at a wider client base and is "10x cheaper compared to the other models out there with similar capabilities." The SLMs are designed to perform simpler tasks, making them suitable for companies with finite resources, with Phi-3 available on Microsoft's Azure, Hugging Face, and Ollama, and also optimized for Nvidia's GPUs.

​

Apple releases eight small language models on Hugging Face, dubbed OpenELM

Apple has recently released a suite of eight small language models (SLMs) on the Hugging Face hub aimed at on-device use, following Microsoft's Phi-3 debut earlier in the week. Collectively named OpenELM (or Open-source Efficient Language Models), these models are designed to run directly on smartphones and tablets to enrich performance and user experience. Though these models are just proof-of-concept models, Apple is slowly but surely ramping up its game in the AI industry and giving sneak peeks of what to perhaps expect for future iOS updates.

​

Synthesia adds more emotions and lip tracking to its custom video avatars

Synthesia, an AI startup specializing in video avatars, has introduced a new feature allowing them to display more emotion. The latest version of the avatars, which are designed for business users to create promotional and training videos, providing improved lip tracking, more natural movements, and more vibrant expressiveness. The company states this is a significant step towards making more human-like generative video avatars.

​

Chinese AI startup SenseTime sees a 30% rise in shares after the release of its SenseNova 5.0 generative model

Chinese AI giant SenseTime's, touted as one of OpenAI's Chinese rivals, shares surged over 30% after announcing its latest generative AI model, SenseNova 5.0. The model's advancements focus on knowledge, mathematics, reasoning, and coding capabilities. Chairman Xu Li claimed SenseNova outperforms OpenAI's GPT 4 in most general usage scenarios, especially in enterprise applications and Chinese-language contexts.

​

AI coding assistance startup Augment emerges out of stealth with $252 million in funding

Augment, a rival to GitHub Copilot with its AI-powered coding platform, has recently emerged from stealth with $252 million in funding at a near-unicorn valuation of $977 million. Founded by ex-Microsoft developer Igor Ostrovsky and AI research scientist Guy Gur-Ari, Augment aims to improve software quality and team productivity. Though the market around AI coding assistants is competitive, Ostrovsky remains optimistic that Augment is on the right path, especially with such a lucrative funding round.

​

French-based FlexAI comes out of stealth with a $30 million seed round

​FlexAI, a French-based startup working in stealth since October of 2023, recently came out with $30 million in seed funding, with top backers including names such as Alpha Intelligence Capital (AIC), Elaia Partners, and Heartcore Capital. The company aims to reimagine compute infrastructure for developers to build and train AI applications more efficiently, as the company teases its first product - an on-demand cloud service for AI training.

​​Podwise - Podwise is a service that uses AI to transcribe, summarize, and analyze podcasts to make the most out of your time, providing insight into podcast content and making it easily searchable and accessible.

​Fotor - Fotor is an AI-enhanced online photo editor with features to enhance and edit photos effortlessly - experience simple online photo editing at your fingertips.

​Athena AI - Design, tweak, and deploy your own AI chatbot across any messenger with Athena AI, allowing you to automate customer support, lead generation, and much more.

​Thunderbit - Thunderbit is a versatile all-in-one AI toolkit that empowers business users to effortlessly build and deploy AI tools and automation products in seconds, packed into a single Chrome extension.

The Verdict on The Rabbit R1 (5-min read)

Touted at CES 2024 in January, the rabbit r1 promised many things for a pocket AI assistant. The first few have started arriving at customers' doorsteps, including The Verge's editor-at-large David Pierce, who gives his thoughts on the bunny companion.

A Victory For Open-Source AI (26-min watch)

The open-source debate of AI models has been a hot topic for some time - should developments be available to everyone or stay behind closed doors? With the release of Snowflake's Arctic model, the tug of war goes one step towards the former, and Wes Roth is here to cover it.