Lifeboat Foundation

AI has mastered some of the most complex games known to man, but models are generally tailored to solve specific kinds of challenges. A new DeepMind algorithm that can tackle a much wider variety of games could be a step towards more general AI, its creators say.

Using games as a benchmark for AI has a long pedigree. When IBM’s Deep Blue algorithm beat chess world champion Garry Kasparov in 1997, it was hailed as a milestone for the field. Similarly, when DeepMind’s AlphaGo defeated one of the world’s top Go players, Lee Sedol, in 2016, it led to a flurry of excitement about AI’s potential.

DeepMind built on this success with AlphaZero, a model that mastered a wide variety of games, including chess and shogi. But as impressive as this was, AlphaZero only worked with perfect information games where every detail of the game, other than the opponent’s intentions, is visible to both players. This includes games like Go and chess where both players can always see all the pieces on the board.

Biological materials are made of individual components, including tiny motors that convert fuel into motion. This creates patterns of movement, and the material shapes itself with coherent flows by constant consumption of energy. Such continuously driven materials are called active matter.

The mechanics of cells and tissues can be described by active matter theory, a scientific framework to understand the shape, flow, and form of living materials. The active matter theory consists of many challenging mathematical equations.

Scientists from the Max Planck Institute of Molecular Cell Biology and Genetics (MPI-CBG) in Dresden, the Center for Systems Biology Dresden (CSBD), and the TU Dresden have now developed an algorithm, implemented in an open-source supercomputer code, that can for the first time solve the equations of active matter theory in realistic scenarios.

Quantum advantage is the milestone the field of quantum computing is fervently working toward, where a quantum computer can solve problems that are beyond the reach of the most powerful non-quantum, or classical, computers.

Quantum refers to the scale of atoms and molecules where the laws of physics as we experience them break down and a different, counterintuitive set of laws apply. Quantum computers take advantage of these strange behaviors to solve problems.

Continue reading “A scientist explains an approaching milestone marking the arrival of quantum computers” »

Researchers confront a formidable challenge within the expansive domain of materials science—efficiently distilling essential insights from densely packed scientific texts. This intricate dance involves navigating complex content and generating coherent question-answer pairs that encapsulate the core of the material. The complexity lies in the substantial task of extracting pivotal information from the dense fabric of scientific texts, requiring researchers to craft meaningful question-answer pairs that capture the essence of the material.

Current methodologies within this domain often lean on general-purpose language models for information extraction. However, these approaches need help with text refinement and the accurate incorporation of equations. In response, a team of MIT researchers introduced MechGPT, a novel model grounded in a pretrained language model. This innovative approach employs a two-step process, utilizing a general-purpose language model to formulate insightful question-answer pairs. Beyond mere extraction, MechGPT enhances the clarity of key facts.

The journey of MechGPT commences with a meticulous training process implemented in PyTorch within the Hugging Face ecosystem. Based on the Llama 2 transformer architecture, the model flaunts 40 transformer layers and leverages rotary positional embedding to facilitate extended context lengths. Employing a paged 32-bit AdamW optimizer, the training process attains a commendable loss of approximately 0.05. The researchers introduce Low-Rank Adaptation (LoRA) during fine-tuning to augment the model’s capabilities. This involves integrating additional trainable layers while freezing the original pretrained model, preventing the model from erasing its initial knowledge base. The result is heightened memory efficiency and accelerated training throughput.

“It’s really simple to define this problem,” said Marcin Bieńkowski, an algorithms researcher at the University of Wrocław in Poland. But it “turns out to be bizarrely difficult.” Since researchers began attacking the k-server problem in the late 1980s, they have wondered exactly how well online algorithms can handle the task.

Over the decades, researchers began to believe there’s a certain level of algorithmic performance you can always achieve for the k-server problem. So no matter what version of the problem you’re dealing with, there’ll be an algorithm that reaches this goal. But in a paper first published online last November, three computer scientists showed that this isn’t always achievable. In some cases, every algorithm falls short.

Large Language Models (LLMs) have shown great capabilities in various natural language tasks such as text summarization, question answering, generating code, etc., emerging as a powerful solution to many real-world problems. One area where these models struggle, though, is goal-directed conversations where they have to accomplish a goal through conversing, for example, acting as an effective travel agent to provide tailored travel plans. In practice, they generally provide verbose and non-personalized responses.

Models trained with supervised fine-tuning or single-step reinforcement learning (RL) commonly struggle with such tasks as they are not optimized for overall conversational outcomes after multiple interactions. Moreover, another area where they lack is dealing with uncertainty in such conversations. In this paper, the researchers from UC Berkeley have explored a new method to adapt LLMs with RL for goal-directed dialogues. Their contributions include an optimized zero-shot algorithm and a novel system called imagination engine (IE) that generates task-relevant and diverse questions to train downstream agents.

Since the IE cannot produce effective agents by itself, the researchers utilize an LLM to generate possible scenarios. To enhance the effectiveness of an agent in achieving desired outcomes, multi-step reinforcement learning is necessary to determine the optimal strategy. The researchers have made one modification to this approach. Instead of using any on-policy samples, they used offline value-based RL to learn a policy from the synthetic data itself.

While physics tells us that information can neither be created nor destroyed (if information could be created or destroyed, then the entire raison d’etre of physics, that is to predict future events or identify the causes of existing situations, would be impossible), it does not demand that the information be accessible. For decades physicists assumed that the information that fell into a black hole is still there, still existing, just locked away from view.

This was fine, until the 1970s when Stephen Hawking discovered the secret complexities of the event horizon. It turns out that these dark beasts were not as simple as we had been led to believe, and that the event horizons of black holes are one of the few places in the entire cosmos where gravity meets quantum mechanics in a manifest way.

Continue reading “The origins of the black hole information paradox” »

As we learned in middle school science classes, inside this common variety of greens—and most other plants—are intricate circuits of biological machinery that perform the task of converting sunlight into usable energy. Or photosynthesis. These processes keep plants alive. Boston University researchers have a vision for how they could also be harnessed into programmable units that would enable scientists to construct the first practical quantum computer.

A quantum computer would be able to perform calculations much faster than the classical computers that we use today. The laptop sitting on your desk is built on units that can represent 0 or 1, but never both or a combination of those states at the same time. While a classical computer can run only one analysis at a time, a quantum computer could run a billion or more versions of the same equation at the same time, increasing the ability of computers to better model extremely complex systems—like weather patterns or how cancer will spread through tissue—and speeding up how quickly huge datasets can be analyzed.

The idea of using photosynthetic molecules from, say, a spinach leaf to power quantum computing services might sound like science fiction. It’s not. It is “on the fringe of possibilities,” says David Coker, a College of Arts & Sciences professor of chemistry and a College of Engineering professor of materials science and engineering. Coker and collaborators at BU and Princeton University are using computer simulations and experiments to provide proof-of-concepts that photosynthetic circuits could unlock new technological capabilities. Their work is showing promising early results.

The AI startups working to build products that support African languages often get ignored by investors, says Hadgu, owing to the small size of the potential market, a lack of political support, and poor internet infrastructure. However, Hadgu says small African startups including Lesan, GhanaNLP, and Lelapa AI are playing an important role: “Big tech companies do not give focus to our languages,” he says, “but we cannot wait for them.”

Lelapa AI is trying to create a new paradigm for AI models in Africa, says Vukosi Marivate, a data scientist on the company’s AI team. Instead of tapping into the internet alone to collect data to train its model, like companies in the West, Lelapa AI works both online and offline with linguists and local communities to gather data, annotate it, and identify use cases where the tool might be problematic.

Bonaventure Dossou, a researcher at Lelapa AI specializing in natural-language processing (NLP), says that working with linguists enables them to develop a model that’s context-specific and culturally relevant. “Embedding cultural sensitivity and linguistic perspectives makes the technological system better,” says Dossou. For example, the Lelapa AI team built sentiment and tone analysis algorithms tailored to specific languages.