#24 | Regenerating Heart Tissue

Quote of the Week 💬 

'Space Hairdryer' Regrows Heart Tissue In Patient Study

⌛ The Seven Second Summary: Researchers at Austria's Medical University of Innsbruck designed a handheld device to stimulate the growth of new blood vessels around areas damaged by or scarred after heart attacks.

🔬 How It Was Done:

• 63 patients who required open-heart bypass surgery were randomly divided into two groups. One group received a standard bypass operation, and the other group underwent the same procedure except they were also treated with the researcher’s device during surgery.

• The team applied their device to each patient's heart during surgery to deliver mild shockwaves to damaged areas.

• The shockwaves bounced tiny bubble-like structures off the surfaces of heart cells, which activated an immune response from the organ and led to the growth of new blood vessels and heart muscle cells.

🧮 Key Results:

• A year after the procedure, the shockwave group’s hearts (11.3%) were able to produce nearly 2x the amount amount of oxygenated blood compared to the control group (6.3%).

• Furthermore, a year after the procedure, the shockwave group was able to walk about 100 meters (~110 yards) farther than the control group during a 6 minute test, and they reported a better overall quality of life.

🔎 Elements To Consider: It took the researchers less than 15 minutes to administer their device during surgery, and none of the patients who received the shockwave procedure developed any device-related adverse events. It will be interesting to see if these benefits persist in larger, follow-up studies.

📚 Learn More: BBC. European Heart Journal.

Stat of the Week 📊 

Anthropic Introduces Claude 3.5 Sonnet

3.5

⌛ The Seven Second Summary: Anthropic released their latest AI model, Claude 3.5 Sonnet, and it outperforms competitors like Chat GPT-4o, Gemini 1.5 Pro, and Anthropic’s best predecessor, Claude 3 Opus, in various benchmark evaluations.

🔬 How It Was Done: No technical details were mentioned to explain how Anthropic made these breakthroughs in their model, but we have mentioned recent examples of model innovation, architectural design and high-quality data as distinguishing factors that consistently improve AI performance. It is also possible hardware advancements have also contributed to these developments, but not enough information is known just yet.

🧮 Key Results:

• The benchmarks Anthropic shared indicates Claude 3.5 Sonnet has modestly improved in writing and reasoning capabilities, though the system’s most significant advancements appear to be in its speed, costs and agentic coding capabilities.

  • Specifically, Claude 3.5 Sonnet solved 64% of Anthropic’s internal agentic coding problems, compared to just 38% from their 2nd best model.

  • Furthermore, Claude 3.5 Sonnet operates at 2x the speed of Anthropic’s 2nd best model, and is 80% cheaper than Claude 3 Opus.

🔎 Elements To Consider: Claude 3.5 Sonnet’s speed and cost advancements are proportionate to assessments from Microsoft’s CTO, Kevin Scott, about the exponential curves this technology is on. He notes “it’s 12x cheaper and 6x faster“ to make calls to GPT-4 since its initial launch ~18 months ago, and both metrics are critical to creating useful tools.  

📚 Learn More: Anthropic. Claude.ai.

AI x Science 🤖

A Novel Approach To Improve AI Reasoning

Last week, we shared a number of recent studies demonstrating how to improve AI reasoning capabilities. The common framework amongst the half dozen examples we shared was the following: one subsystem searches for and generates ideas to solve a problem, and other subsystems tests, evaluates, and provides feedback for the ideas it receives. Today, we’re going to cover one other approach we found, since it also appears promising and may be a path to models that can achieve superhuman intelligence after receiving just a handful of historical examples.

This model framework is from the team that holds the current world record at solving the ARC challenge, which is a new benchmark with a $1,000,000 cash prize to evaluate the intelligence of AI systems by presenting a variety of visual reasoning tasks for a system to complete. Each task consists of a set of grids with different colors and numbers, and is designed in a way to minimize the impact of AI systems from utilizing their training data to solve new ARC problems.

The current record holders were able to score 34% on a recent exam, and accomplished their feat by using a technique similar to active inference. This is when models make assumptions and predictions about its environment after receiving a limited amount of information to work with, then refines its knowledge about its environment once it receives feedback and new information. The team’s approach is similar insofar as they designed their model to infer patterns within an ARC question, generate many synthetic examples based the context of the specific question it received, and train itself on the self-generated examples it produced until the system is proficient at solving the visual reasoning task.

This is akin to how humans use imagination to think of new scenarios in mathematics, linguistics and a range of other subjects to challenge ourselves and reinforce new information about a topic we want to learn. The team’s 34% performance may appear modest at face value, but when state-of-the-art models attempt these ARC problems without any prior fine-tuning, they consistently score less than 2%.

As the research team has pointed out, the key to their system’s success is heavily dependent on the model’s ability to infer and create reliable examples for itself when exposed to a new problem. This is noteworthy because there is an ongoing debate about whether or not AI capabilities will plateau once they have been trained on most of the world’s human generated data, and its training becomes more reliant on data generated from other AI models to improve performance. We are probably a few years away from reaching this hypothetical plateau, although, if AI research continues to push the frontiers of inference capabilities and synthetic data generation, the plateau may never arise because the systems will never run out of worthwhile training data. This is why the data generation news and research we have previously highlighted from NVIDIA and Google are meaningful and why we initially shared it. We will stay on the lookout for other important updates so you don’t miss a thing. ARC Challenge Leaderboard.

Our Full AI Index

• Bidirectional BCI Functionality: Researchers at Carnegie Mellon University achieved a breakthrough in noninvasive brain-computer interfaces (BCIs), allowing people who are wearing their devices to control technology with their minds and also receive feedback from the technology. In a study with 25 people, their methodology allowed individuals to spell out phrases like "Carnegie Mellon", and the device’s feedback mechanism helped the patients maximize their focus and concentration. Previous BCI technology stories we have highlighted did not include bidirectional feedback capabilities, so it will be interesting to see how this innovation improves the lives of people with motor or communication disabilities. Carnegie Mellon University. Nature.

• Analyzing Plant Roots With AI: Scientists at Lawrence Berkeley National Laboratory developed RhizoNet, an AI-powered tool to automate their plant root analysis with exceptional accuracy. RhizoNet uses a deep learning model to segment plant roots and track their growth and biomass. This allows the researchers to study the behavior of different plant roots under various environmental conditions, which can allow them improve agricultural yields and develop climate-resilient crops. Berkeley Lab. Nature.

• Olympic-Level AI Benchmark Tests: Researchers from Shanghai Jiao Tong University developed an Olympic Arena benchmark to assess how well AI models can adrress advanced questions and capabilities across a broad spectrum of Olympic-level challenges. Their benchmark features 11,163 problems from 62 distinct competitions, and focuses on logical and visual reasoning abilities. Their leaderboard highlights physics, chemistry, astronomy and a variety of other academic disciplines, and while we already know fine-tuned models are capable of solving Olympic-level geometry questions about as well as a gold medalist, it will be interesting to see the field’s progress as it builds a general system with gold medal problem-solving capabilities across all Olympic subject matters. arXiv. Github. Leaderboard.

Other Observations 📰

Astronomers Observe A Black Hole As It Awakens

Astronomers from the European Southern Observatory (ESO) believe they are observing a supermassive black hole at the center of a distant galaxy 300 million light-years away. The galaxy has been quiet and inactive for years, but began to emit ultraviolet, optical, and infrared light in December 2019. The brightness of the galaxy continued to increase over the next few years, with X-ray emissions detected in February 2024. To understand the phenomenon, the team combed through archival data and made new space observations through cutting-edge instruments from ESO and NASA.

After their analysis, the team concluded the galaxy's black hole has a mass of ~1 million solar masses, which is > 100,000x the mass of our Sun. They also concluded the supermassive black hole is likely the source of the galaxy’s increased brightness. Furthermore, the team ruled out other possible explanations, such as supernova explosions or tidal disruption events as the cause of the galaxy’s light activity over the past 4 years, in part, because these events typically last for less than a year at a time. In any case, these observations are unprecedented and are providing a valuable source of insight into how black holes possibly grow and evolve over time to shape our universe. ESO. Astronomy & Astrophysics.

Our Full Science Index

• The Dawn Of The Solar Age: The Economist published a fantastic write up about the evolution of the solar industry. It covers everything from the world’s growing energy demands since the 20^th century, to the technological advancements of solar panels, to the ongoing renewable energy transformation the world is undertaking and some of the unit economics driving the shift. The Economist Archive.

• Device Reduces Epileptic Seizures: A 13-year-old boy with severe epilepsy became the first patient in the United Kingdom to participate in a clinical trial using deep brain stimulation. A cross-disciplinary team developed a rechargeable device to mount on top of a patient’s skull, and stimulate the surrounding region of the patient’s brain thalamus to block seizure activity. Thus far, the treatment has reduced the teenager’s daytime seizures by 80%. After the device showcased these remarkable results, the team is now recruiting three additional patients to expand their study, and they are planning to conduct a trial with 22 patients in the months ahead. GOSH.

• Chimps Use Plant Medicines: A multinational team of researchers observed wild chimpanzees in Uganda's Budongo Forest for 4 months, recording their behavior and plant consumption. By the end of their studies, they found 51 sick chimps who selectively ate 13 different plant species with antibacterial and anti-inflammatory properties. The team tested these plants for medicinal properties and discovered 11 of them are already used by traditional medicine practitioners in the local area. This suggests both chimps and humans have independently identified the same medicinal plants to alleviate a variety of common ailments. These findings build upon existing evidence that animals such as dolphins are capable of self-medication. Science.

Media of the Week 📸 

Watch A Robot Bug Fly & Crawl Around A Room

Roboticists from Shanghai Jong Tong University developed a robotic bug called JT-fly, that can crawl, fly, hover and land. Their robot has four wings and six legs, has a wingspan of 33 centimeters and weighs 35 grams. It can fly up to 5 meters per second and crawl at 0.3 meters per second. We have previously mentioned how recent robotics and AI developments are appearing more anthropomorphic than ever before. While we still believe this to be the case, we find it just as fascinating to see the same technological breakthroughs applied to make entomomorphic machines. IEEE Xplore.

Mapping The Biology Of Spinal Cord Injuries In Great Detail

EPFL researchers created a comprehensive biological atlas of mice with spinal cord injuries by mapping the cellular and molecular dynamics of their paralysis in unprecedented detail. The team accomplished this feat by integrating advanced molecular mapping technologies with AI. Specifically, they used two powerful tools to gather data: one to analyze individual cells, and another to show how cells are organized in relation to each other. Their AI system then analyzed this data to identify patterns and connections between millions of spinal cord cells, while preserving the spatial context and relationships between different cell types. The team's efforts identified specific neurons and genes that play key roles in enabling young mice to heal from spinal injuries, and develop a better understanding for why older mice lose similar capabilities as their bodies age. This insight led the researchers to develop new therapy treatments targeting these neurons and genes, in hopes of creating a treatment to help humans recover from similar injuries. EPFL. Nature. Github.

JWST Captures A First-Of-Its-Kind Photo

NASA's James Webb Space Telescope captured a stunning new image of the Serpens Nebula. The gas cluster is ~1,300 light-years from Earth, and is the latest demonstration of the instrument’s unparalleled spatial resolution and sensitivity to capture a spectrum of near-infrared wavelengths. Rainbows are beautiful enough, yet there’s something about a cosmic rainbow we find more enchanting. NASA. YouTube.

Atomic Structure Of An Enzyme Neurons Use To Communicate

Scientists at SickKids used advanced imaging technology to capture high-resolution images of the enzyme V-ATPase. These photos reveal the enzyme’s atomic structure, and by developing novel biochemical and computational methods to analyze the images, the scientists were able to better understand its role in neuronal communication. The team also created 3D models of the enzyme to simulate its interactions with synapses and other proteins. This allowed them to predict how targeted therapy treatments may improve the conditions of patients with neurodegenerative diseases with dysfunctional V-ATPase enzymes. SickKids. Science.