The AlphaDev Milestone: A New Model that is Able to Discover and Improve Algorithms
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 299: We expand on the tool-augmented LLM discussion by discussing specific categories of methods. The research section covers Google’s Tool Augmented Language Model paper. Finally, we discuss Microsoft’s DeepSpeed-Chat that enables the fine tuning of LLMs.
Edge 300: Our 300 edition! We deep dive into Falcon-LLM, the new foundation model that is making a lot of noise in the open source community. Go subscribe!
📝 Editorial: The AlphaDev Milestone: A New Model that is Able to Discover and Improve Algorithms
With all the hype around LLMs and foundation models, sometimes we ignore other areas of machine learning. Last week, DeepMind unveiled a major breakthrough in AI that could signal an important element of the road to AGI.
What is a good test for Artificial General Intelligence (AGI)? This question has been central to the evolution of AI for decades. While the Turing Test is a generic answer, the real answer becomes more complicated when delving into the specifics. Is there a single test that can indicate the emergence of AGI? Obviously, there are countless variations of the Turing Test that can be modeled. However, in my opinion, there is a category that stands out as a leading indicator of AGI-like foundations—the discovery of new science and, more specifically, the discovery of new algorithms.
Algorithm modeling requires various cognitive skills, such as multi-step reasoning, planning, and empirical evaluation, among others. Many branches of mathematics, computer science, or physics rely on a set of core foundational algorithms. Despite the progress in computer science, breakthroughs in foundational algorithms have significantly slowed down due to the high bar they must meet. Foundational problems like sorting, searching, matrix multiplication, or combinatorics have had established solutions for decades. Can the new generation of foundation models help improve some of the core algorithms that were used to create them?
A few months ago, DeepMind unveiled AlphaTensor, a new model that helps discover a more efficient matrix multiplication algorithm that hasn't seen improvement in 50 years. Building on that work, DeepMind has now unveiled AlphaDev, a reinforcement learning (RL) model that discovered faster sorting algorithms and improved existing ones. The model is based on AlphaZero, an RL model that achieved superhuman performance in various games like Go, chess, and shogi. Not surprisingly, the AlphaDev environment was modeled as a single-player game in which the model observes an algorithm and experiments with different instructions to improve it.
AlphaDev was not only able to discover faster algorithms using existing methods but also some based on completely novel approaches. If you are a classically trained computer scientist, the idea of discovering new sorting algorithms may seem unfathomable.
The core discussion about the path to AGI has been centered around foundation models. Yet, DeepMind is using RL to discover new algorithms. First matrix multiplication, and now sorting. Do you know what those two techniques are foundational to? That's right: AI.
🔎 ML Research
AlphaDev
DeepMind published a paper detailing AlphaDev, a new reinforcement learning method able to discover new algorithms. The model was based on AlphaZero and trained on a single-player assembly game based on potential instructions of the algorithm —> Read more.
AVFormer
Google Research published a paper outlining AVFormer, a method for augmenting large scale audio models with visual representations. The core principle is based on injecting visual embeddings into frozen ASR models to improve their robustness —> Read more.
Visual Captions
Google Research published a paper discussing a technique that generates visuals based on real time video conference streams. The model is fine tuned using a dataset of visuals that are appropriate for video conference conversations —> Read more.
3D Understanding
Salesforce Research published papers detailing ULIP and ULIP-2, two techniques used to understand 3D objects. Both technique are based on multimodal methods that can process image, language and 3D cloud data —> Read more.
ReLM
Researchers from Carnegie Mellon University published a paper introducing ReLM, a model that can query LLMs using regular expressions. ReLM is relevant for tasks such as validating aspects such as memorization, bias or toxicity in LLMs —> Read more.
📍 Live Tutorial: Working with LLMs at Scale
This free event on June 15th explores LLMs and the two main problems they face when it comes to production: high cost and lack of domain knowledge. Discover how vector databases can be a solution by facilitating data injection and caching through the use of vector embeddings. The virtual session ends with a hands-on tutorial where you can build an LLM application using LlamaIndex and Milvus.
🤖 Cool AI Tech Releases
PaLM in Vertex
Google announced support for its PaLM and PaLM2 modes in the Vertex platform —> Read more.
Chat Notebooks
Stephen Wolfram published an incredible blog post about a new paradigm that combines LLMs and notebooks —> Read more.
CodeTF
Salesforce Research open source CodeTF, a Python library for code LLMs —> Read more.
📡AI Radar
OpenAI competitors Cohere announced a $270 million series C.
Meta announced its plans to add generative AI across all its platforms.
Contextual AI raised $20 million for its enterprise LLM platform.
Automattic, unveiled an AI assistant for Wordpress.
Enterprise AI platform Instabase announced a new $45 million funding round.
Microsoft unveiled a version of the Azure OpenAI Service for governments.
LinkedIn released new generative AI tools for copy suggestions.
Weights and Biases released new tools for building and monitoring ML apps.
Salesforce continue its investment in generative AI with the release of Commerce GPT and Marketing GPT.