• Posts
  • Meta AI develops compact language model for mobile devices

Meta AI develops compact language model for mobile devices

Also: Microsoft drops ‘MInference’ demo, challenges status quo of AI processing

Good morning,

In today’s newsletter, we delve into the launch of Thrive AI Health, a new venture by OpenAI Startup Fund and Thrive Global that leverages generative AI for personalized health coaching. We’ll also discuss the enduring dominance of Google Search over ChatGPT, Meta AI’s innovative compact language model for mobile devices, and Microsoft’s breakthrough in AI processing speed with MInference. Additionally, we explore how AI tools are revolutionizing vacation planning, and highlight YouTube’s experimental AI-generated custom radio stations. Finally, we’ll cover the latest venture capital updates, including Hebbia’s significant Series B funding round led by a16z and the substantial growth of AI’s influence in the VC market.


  • 📱 Meta AI develops compact language model for mobile devices

  • 💵 OpenAI Startup Fund, Thrive Global launch new company Thrive AI Health

  • 🔎 ChatGPT might rule the AI chatbots — but it can't beat Google Search

  • 🤖 Microsoft drops ‘MInference’ demo, challenges status quo of AI processing

  • 🏝️ How to use AI to plan your next vacation

Meta AI has introduced MobileLLM, a new approach to creating efficient language models optimized for smartphones and other resource-constrained devices. Developed by researchers from Meta Reality Labs, PyTorch, and Meta AI Research, MobileLLM prioritizes model depth over width and uses innovative techniques such as embedding sharing, grouped-query attention, and block-wise weight sharing. These optimizations allow MobileLLM, with fewer than 1 billion parameters, to outperform previous models of similar size by up to 4.3% on benchmark tasks. The 350 million parameter version of MobileLLM even demonstrated comparable accuracy to the larger 7 billion parameter LLaMA-2 model on certain tasks. This development reflects a shift towards more compact and efficient AI models, potentially enabling advanced AI features on mobile devices. Meta has open-sourced the pre-training code, inviting further research and development in this area.

OpenAI Startup Fund and Thrive Global have announced the creation of Thrive AI Health, a new company that will offer a generative AI-enabled personalized health coach. This innovative coach aims to facilitate behavioral change in five key areas: connection, sleep, fitness, stress management, and food, by providing disease prevention and treatment optimization. Leveraging OpenAI and Thrive Global’s behavioral change methodologies, the AI coach will use peer-reviewed science, personal preferences, and medical data to generate personalized insights and recommendations. Leading this new venture as CEO is DeCarlos Love, a former product leader at Google with extensive experience in AI and health tech. The initiative is backed by strategic investments from the Alice L. Walton Foundation and has established research partnerships with notable institutions such as Stanford Medicine and the Rockefeller Neuroscience Institute. Thrive AI Health aims to revolutionize health outcomes and reduce healthcare costs by addressing chronic diseases through personalized, proactive, and data-driven coaching.

Since its debut in 2022, ChatGPT has significantly impacted how people seek information, yet it struggles to dethrone Google Search, which has seen increased dominance. Google’s market share grew to 91.1% in June, bolstered by rising ad revenue and enhanced AI integration through its Gemini AI. Meanwhile, ChatGPT’s web traffic declined by 12%. Despite early fears that AI chatbots might undercut traditional search engines, Google’s strategic use of AI to enhance its search capabilities has proven advantageous, increasing both user engagement and monetization. However, Google’s growing market dominance is under scrutiny due to ongoing antitrust investigations by the U.S. Department of Justice, with rulings expected later this year.

Microsoft has introduced MInference, a breakthrough in processing speed for large language models. Demonstrated on the AI platform Hugging Face, MInference significantly accelerates the “pre-filling” stage of language model processing, reducing processing time by up to 90% for inputs of one million tokens while maintaining accuracy. This advance addresses the computational challenges of handling lengthy text inputs, a critical bottleneck in AI development. MInference showcases up to 10x latency reduction and is expected to impact AI energy consumption positively by reducing computational resources. This innovation positions Microsoft at the forefront of efficient AI processing, potentially influencing future research and the competitive landscape in the AI industry.

AI tools are transforming the way people plan vacations by offering a range of functionalities that streamline the process from initial research to on-the-ground activities. Large language models like ChatGPT, Gemini, and Copilot can quickly generate lists of potential destinations based on user preferences, such as climate, family-friendliness, and nightlife. Once a destination is chosen, these models can help create detailed itineraries and suggest activities, enhancing the travel experience. AI can also assist with language translation through apps like Google Translate, which uses camera integration for real-time translations of menus and signs. For booking logistics, tools like Google’s Gemini integrate with Google Flights and Hotels to provide real-time information, making it easier to compare prices and schedules. Additionally, AI-powered tools can generate content for online reviews and social media, helping travelers document their experiences efficiently. While these tools offer significant convenience, users are advised to double-check AI-generated information for accuracy due to the propensity of AI to produce errors.

🛠️ AI tools updates

YouTube has introduced a new AI tool that creates custom radio stations tailored to users’ specific music preferences. Currently in an experimental phase and available to a limited number of users, this feature allows users to request music in various ways, either by typing or speaking. For instance, asking for “queer hip hop beats” can generate a station featuring artists like Janelle Monáe, Lil Nas X, and RuPaul. The tool also allows real-time adjustments to the stations through a chat interface. Similar to Spotify’s “AI DJ” and the PlaylistAI app, this feature aims to enhance the music discovery experience by leveraging AI to offer more personalized and convenient music recommendations. It is yet to be determined whether this feature will become widely available or if it will be exclusive to paid subscribers.

💵 Venture Capital updates

Hebbia, an AI startup that helps businesses analyze complex data to answer multi-step questions, has secured $130 million in a Series B funding round led by Andreessen Horowitz. The round also saw participation from Index Ventures, Google Ventures, and Peter Thiel, valuing the company at approximately $700 million. Founded in 2020, Hebbia enables companies to process both structured and unstructured data, such as regulatory filings and PDFs, to address intricate business queries. The startup has experienced rapid growth, increasing revenue 15-fold and quintupling its headcount over the past 18 months. Hebbia aims to revolutionize how businesses interact with AI by creating products that leverage advanced AI capabilities to drive technological revolutions.

AI is significantly influencing the venture capital market, accounting for over 40% of new U.S. unicorns and more than 60% of the increase in total venture-backed valuation in the first half of this year. Between January 1 and June 25, 2024, 13 new AI unicorns emerged, adding $116 billion in aggregate value, with Elon Musk’s xAI contributing $24 billion alone. Overall, U.S. unicorn value grew by $162 billion during this period. AI’s share of venture investment in North America and Europe has risen to about 20% in 2023 and continues to grow. Despite the challenges faced by some AI companies, venture capital interest in AI remains strong, driven by the transformative potential of AI technologies and their applications across various sectors.

🫡 Meme of the day

⭐️ Generative AI image of the day