Google’s Somewhat “Moat-less “ AI Week
Sundays, The Sequence Scope brings a summary of the most important research papers, technology releases and VC funding deals in the artificial intelligence space.
Next Week in The Sequence:
Edge 291: Our series about new techniques in foundation models continues with an overview of the technique that made ChatGPT possible: reinforcement learning with human feedback(RLHF). We review the RLHF paper and the Transformer RLHF Library!
Edge 292: We discuss OpenChatKit, an open source alternative to ChatGPT.
📝 Editorial: Google’s Somewhat “Moat-less “ AI Week
Google I/O is one of the most anticipated tech events of the year, and for the past few years, AI has been at the forefront of it. Unsurprisingly, the 2023 edition of I/O was brimming with news about generative AI.
Much attention was given to the new PaLM 2 model, which boasts updated conversational capabilities and the ability to write and debug code. Since PaLM 2 serves as the foundation for Bard, the latter also showcased notable improvements. Bard aims to support 40 languages and can generate images in responses through integration with Adobe's Firefly. Additionally, Google has started experimenting with conversational features to enhance its search experience.
Generative music emerged as another intriguing aspect of Google's announcements, with the introduction of MusicLM, a tool capable of transforming text into music. Codey is Google's anticipated response to GitHub's Copilot, enabling users to ask coding questions. Another noteworthy announcement was Sidekick, a tool designed to elevate the writing experience by offering contextual suggestions within Google Docs.
Google Cloud received significant updates in the field of generative AI as well. Vertex welcomed the Imagen text-to-image model, serving as a counterpart to models like DALL-E or Stable Diffusion. Another notable addition was the A3 supercomputer, optimized for LLM workloads.
Numerous other announcements related to generative AI spanned across mobile and consumer apps. However, amidst the impressive array of unveilings, one couldn't help but notice the absence of Google's characteristic innovation. Many of the new products seemed to be mere responses to competitors like Microsoft or OpenAI, lacking the striking impact one would expect. Google's traditional strengths of speed, size (in search), and open-source (Android) were notably absent. Generative AI appears to be a technology trend where Google is struggling to find its competitive edge.
📍 The Future of Vector Search: Insights from HNSW Author Yury Malkov
The Hierarchical Navigable Small World (HNSW) is one of the most popular graph-based indices for nearest neighbor search. Join us for a fireside chat with Yury Malkov on May 16, as we discuss the proliferation and increasing importance of LLMs and thoughts on the future of vector search. —>Learn more
🔎 ML Research
GPT Interpretability
OpenAI published a paper outlining an interpretability technique that uses GPT-4 to derive explanations about the behavior of neurons in other LLMs. The method generates and scores explanations for neuron behaviors and applies them to other neurons in the model —> Read more.
ImageBind
Meta AI published a paper detailing ImageBind, the first AI model that works across six different modalities. The multimodal architecture supports text, image/video, audio, 3D, thermal and inertial measurements units —> Read more.
Differential Privacy for Foundation Models
Amazon Science published two papers proposing differential privacy techniques for large neural networks. Automatic gradient clipping and bias-term-only fine-tuning can improve the differential privacy during the training phase —> Read more.
LLMs and Mobile UIs
Google Research published a paper exploring the viability of using LLMs to improve conversational interactions with mobile UIs. The core idea is that LLMs will understand the intricacies of mobile UIs to accomplish specific tasks —> Read more.
🤖 Cool AI Tech Releases
Google I/O Announcements
Google unveiled several AI technologies at Google I/O including PALM-2 and MusicLM —> Read more.
TensorFlow and Keras Enhancements
At Google I/O, the TensorFlow team unveiled a series of enhancements to its core stack and the Keras framework designed to better support the new generation of generative AI models —> Read more.
Claude 100K
Anthropic announced that it has expanded capacity of its Claude model to 100K tokens —> Read more.
MPT-7B
MosaicML released MPT-7B, an open source transformer trained in 1T text and code tokens —> Read more.
🛠 Real World ML
Entity Resolution at Walmart
Walmart Global Tech provides an overview of their use of ML for entity resolution —> Read more.
📡AI Radar
Microsoft made a strategic investment in Builder.ai to streamline collaboration on AI applications.
Allen Institute for AI raised $30 million for its incubator platform.
Ashton Kutcher’s Sound Ventures has already deployed half of its $240 million VC fund into OpenAI, Anthropic and Stability AI.
IBM unveiled Watsonx, a new platform for generative AI models for businesses.
Seattle´s VC firm Ascend announced a new $25 million pre-seed fund focused on AI.
Amazon acquired AI startup Snackable to improve its podcast capabilities.
H2O.AI released H2OGPT and the LLM studio to foster LLM development.
AI-powered VC firm Vela Partners launched with a new $25 million fund.
Meta announced a new set of generative AI tools for advertisers.
The Allen Institute for Artificial Intelligence announced their initial work on an LLM for scientific research.