#28 | The Ocean Floor Produces Oxygen

Quote of the Week 💬

The Ocean Floor Produces Its Own Oxygen

⌛ The Seven Second Summary: An international team of researchers discovered that metallic minerals on the deep-ocean floor produce oxygen at depths of 13,000 feet, even in complete darkness.

🔬How It Was Done: 

• Researchers used a device called a benthic chamber lander to collect data on oxygen levels in the Clarion-Clipperton Zone — a mountainous submarine ridge in the Pacific Ocean.

• They deployed the device multiple times in different seafloor locations and at different times to understand a range of baseline oxygen levels in their deep-sea environment.

• The device had three separate chambers that were pushed into the ocean floor to create isolated microcosms. The researchers also used special sensors called oxygen optodes to measure the oxygen levels in these chambers.

• To confirm their findings, the researchers also conducted additional experiments on sediment cores they collected from the ocean floor. They tested these cores with different chemicals and measured their oxygen levels to see if they could replicate the results from their chamber device.

🧮Key Results:

• The team found the ocean floor produces over 2x the amount of oxygen it consumes in a day.

• The researchers checked if they made a mistake in their measurements or if certain factors they did not consider were increasing an area’s oxygen levels, but they realized their two primary explanations would not produce enough oxygen to match what they observed.

• They ultimately discovered that the oxygen production was linked to natural mineral deposits on the ocean floor, which contain metals like cobalt, nickel, and copper. These deposits had electrical potentials with enough voltage to create chemical reactions that produce oxygen.

💡Why This May Matter: It has long been understood that photosynthesis creates oxygen, and this process is responsible for producing the Earth's oxygen other organisms depend on to survive. If oxygen can be produced without photosynthesis, then our understanding of the Earth's oxygen cycle and the evolution of life on Earth may need to be revised.

🔎 Elements To Consider: These findings also highlight the uncertainty of deep-sea mining. There is still so much we do not know about the ocean, and if mineral deposits are an important source of oxygen for deep-sea life, then extracting these deposits may have unintended consequences for the health of the entire ocean ecosystem.

📚 Learn More: Northwestern. Nature.

Stat of the Week 📊

AI System Achieves Silver Medal Standard At Math Olympics

28

⌛ The Seven Second Summary: Google Deepmind introduced two new AI systems — AlphaProof and AlphaGeometry 2 — in the latest International Mathematical Olympiad (IMO) competition.

🔬How It Was Done:

• AlphaProof uses a fine-tuned Gemini model to translate around ~1 million informal math problems into a formal math language. This created a large library of formal problems for the system to reference as it searches for proofs or disproofs of various mathematical statements.

• The team embedded AlphaProof with an AlphaZero reinforcement learning algorithm to verify the steps the system takes to prove a mathematical proof. Once the system completed millions of practice problems, its reinforcement mechanisms were refined well enough to produce a model capable of solving difficult mathematical reasoning problems.

• Meanwhile, AlphaGeometry 2 is the next generation AI system we first mentioned back in January. It now combines a Gemini-based language model with a faster symbolic engine to search for and predict useful geometric constructs — such as points, lines, circles, to solve geometry problems

🧮Key Results:

• AlphaProof and AlphaGeometry 2 collectively solved 4 out of 6 Olympiad problems during the competition — 2 algebra questions, 1 number theory proof, and 1 geometry problem.

• Both systems earned a perfect score on each problem they solved correctly. This included the hardest problem in the competition, which only 5 out of 609 contestants were able to solve. 

• The system’s performance was equivalent to the top end score of an Olympiad silver-medalist. It was just 1 point shy of reaching a gold medalist threshold, which 58 contestants were able to achieve during the competition.

• Before this year’s competition, AlphaGeometry 2 could solve 83% of all historical IMO geometry problems from the past 25 years, compared to the 53% rate achieved by its predecessor, For IMO 2024, AlphaGeometry 2 solved Problem 4 within 19 seconds after receiving its formalization.

💡Why This May Matter: When we highlighted AlphaGeometry earlier this year, we mentioned the system was unable to solve algebra, number theory, and other mathematical topics. Deepmind’s plan to close this gap is to build separate AI systems with specialty knowledge and training to solve different sets of problems. Notably, AlphaGeometry 2 is far more capable than its predecessor after just 6 months. As Deepmind notes, AlphaGeometry 2 is able to solve 83% of all historical IMO geometry problems from the past 25 years, compared to just 53% from the original version.

🔎 Elements To Consider: In an official math Olympiad competition, contests submit answers over the course of 9 hours. Deepmind’s system was able to solve some problems in seconds and others within minutes, but a couple required multiple days to solve.

📚 Learn More: Google Deepmind.

AI x Science 🤖

Energy Bottlenecks & How It Impacts AI Development

Last week, we shared several methodologies Meta followed to build their Llama 3.1 models. Their 92-page paper is filled with valuable insights, and it would take multiple breakdowns to delve into each notable finding from the team’s work. One area of interest we may save for another time are the scaling laws the team observed to predict how a model will perform based on the budget they allocated for training and compute.

With that said, a notable challenge the paper mentions is hardware related. Specifically, the team mentioned how day-to-day weather patterns and temperature fluctuations impacted hardware performance by 1-2%, depending on the time of day and how significant the temperature fluctuation was. We found this notable, in part, because it is an example of how far-reaching climate change has already become — its secondary and tertiary effects impact nearly everything on the planet, even seemingly unrelated parts of society, like software development.

Meta also observed that temperature changes caused fluctuations in their data center’s energy usage. They highlighted fluctuations in power consumption “on the order of tens of megawatts (MW), stretching the limits of the power grid.“ For context, most data centers are less than 150 MW, so fluctuations in the 10s of MW are significant impacts of energy usage.

Mark Zuckerberg mentioned the energy constraints to train frontier AI models in an interview earlier this year, when Meta released the first version of their Llama 3 models back in April. Now that his company has released their Llama 3.1 paper, the sentiments he shared can be appreciated through a new lens.

Future AI models need to run on more quantities of more advanced, energy-intensive hardware in order to continuously improve for the foreseeable future. In order to meet these energy demands, massive renewable power plants need to be built, which can take years of planning and construction to complete. This means that even with near limitless amounts of capital, the AI industry may hit a wall where energy constrains its progress.

Distributed training methods may offer a solution to this bottleneck, but their feasibility and efficiency are still uncertain. This is yet another reason why the field's efficiency and cost improvements, which we regularly highlight, are so important. They are a bridge to enable the field to push forward with meaningful improvements while longer-term issues like AI governance, interpretability work, and power consumption demands are sorted out.

Our Full AI Index

• AI System Helps Doctors In England Detect More Cancer: The C the Signs AI tool is now used in 1,400 primary care physician practices in England. Since the tool’s implementation in 2021, cancer detection rates increased from 58.7% to 66.0%. This improvement is primarily driven by the tool’s ability to pull together patient data, such as someone’s medical history, test results, and personal characteristics to identify subtle signs of different cancers a physician may otherwise overlook. The report highlights that the tool has showcased a low 2.8% false negative rate since its rollout, which is a hopeful indication the country's early cancer detection screening has systemically improved. The Guardian. Journal of Clinical Oncology.

• AI System Predicts How Multiple Sclerosis Will Progress: Researchers developed machine learning models to predict the progression of multiple sclerosis (MS) by analyzing data from 15,240 patients across 40 countries. The models were trained on 2 years of patient data, forecasts disability progression over the next 2 years, and outperforms predictions made from other clinical assessments. PLOS Digital Health.

• A Hardware Device To Make AI More Energy Efficient: Researchers at the University of Minnesota developed a device to reduce energy consumption for AI applications by an order of magnitude. Their system does this by performing computations directly where information is stored, rather than using time and energy to move data in places to perform the computations. The team suspects their innovation may improve device’s energy efficiency, with estimates suggesting energy efficiency by 1000x, though these claims need to be replicated and reproduced at scale to see if the results are consistent. University of Minnesota. Nature.

Other Observations 📰

The World’s First Nuclear Reactor Passes A Loss Of Cooling Test

Researchers from China’s Shidaowan nuclear power plant shared results of their 2023 nuclear reactor cooling test with a Pebble Bed Reactor. This is a Gen 4 nuclear reactor designed to be ‘meltdown proof’, and the team's study confirms the reactor’s ability to cool down naturally, which is an industry-first milestone where commercial-scale nuclear energy has inherent fail-safe safety designs in place.

This is achieved by taking small billiard ball-sized pebbles with uranium in their core and dropping them into a central reactor chamber. When the uranium is close together, they start a chain reaction and generate heat. The pebbles are surrounded by a helium gas to capture the heat produced by the uranium’s reaction, which is then used to warm up water, produce steam and turn a turbine to generate electricity.

Since this design uses less uranium and encases the uranium in graphite pebbles, the heat generated from a nuclear reaction is spread out over a larger surface area to make the reaction easier to control. As the temperature of the uranium rises, the surrounding graphite expands, which makes it harder for the neutrons inside the uranium to collide and sustain a chain reaction. Consequently, when the temperature cools, the graphite contracts, the reaction rate increases, and the nuclear reaction resumes itself. This self-cooling mechanism prevents a reactor from overheating without relying on pumps, fans or any other cooling system to prevent a nuclear meltdown.

The paper released by the team are the results of tests where they turn off their reactor’s cooling systems to simulate a failure in a power plant’s critical safety operations. However, the Pebble Bed Reactor maintained its ability to self-regulate and control its temperature without any assistive cooling system throughout the team’s tests. This is a massive development, because cooling system failures are what caused the Fukushima disaster, as well as other catastrophic examples of nuclear fallout.

It is encouraging to see the the results of Shidaowan’s study the same week as Meta’s commentary about energy constraints to train AI systems. Nuclear energy is an opportunity to generate thousands of megawatts of renewable power to meet the world’s growing energy demands, and the chief safety concern of nuclear energy appears to be solvable through innovative system designs. Time will tell how popular nuclear energy ultimately becomes. In China, they plan to build 150 new nuclear reactors before 2035. Tsinghua University. Cell Joule.

Our Full Science Index

• 7th Patient ‘Cured’ Of HIV From Stem Cell Transplant: A new case study presented at AIDS 2024 reveals the world’s 7^th person was cured of HIV after receiving a stem cell transplant from a donor with a CCR5-delta32 mutation. CCR5-delta32 is a receptor that sits on the outside of cells, and this mutation helps to prevent HIV from entering a cell’s body. This breakthrough may lead to more scalable and innovative HIV cure strategies, which is desperately needed to end the epidemic by 2030. International AIDS Society.

• A New Way To Make Heavier Atoms: Scientists at Berkeley Lab successfully created element 116 using a titanium beam to fuse two atoms over the course of a 22 day operation. This new fusion techniques enables the lab to attempt to create element 120, which would be the heaviest atom ever created and sit on the 8^th row of the periodic table. Berkeley Lab. arXiv.

• Child Air Pollution Deaths Down Dramatically: The 2024 State of Global Air report was recently released, and while it found that air pollution was the second-largest risk factor of death for children under 5 years old in 2021, the death rate from air pollution for this group decreased by 53% since 2000. Household air pollution deaths also decreased by 36% during this time frame, largely driven by reductions in exposure in China and South Asia. 2024 State of Global Air Report.

Media of the Week 📸

Paralympian Carries Olympic Torch To Kick Off Summer Games

To kick off the summer games, French Paralympian Kevin Piette made history by carrying the Olympic torch using a robotic exoskeleton. The exoskeleton Piette used is called Atalante X and is from Wandercraft. The EU approved the machine in 2019, and has treated hundreds of patients with significant mobility issues since. The exoskeleton works by using sensors and a hand controller to detect a person’s movement intentions and support the wearer's legs to walk.

A Fully Functional Mechanical Heart Transplant

The Texas Heart Institute and BiVACOR successfully implanted the first artificial heart made entirely of titanium to replace a man’s failing heart. This artificial heart is designed to serve as a temporary bridge of support until a natural heart is available through an organ transplant. However, it also has the potential to be a long-term or permanent solution for patients with end-stage heart failure, depending on how durable and resilient the machine shows to be over the course of this early-stage clinical trial. The Texas Heart Institute.

Chandra Shares Never-Before-Seen Photos On Its Anniversary

NASA's Chandra X-ray Observatory released 25 never-before-seen images to celebrate its 25th anniversary. Over the observatory’s quarter-century lifespan, Chandra has made nearly 25,000 observations, leading to over 10,000 peer-reviewed papers. Cheers to the next 25 years! NASA. Chandra X-Ray Observatory.

This Week In The Cosmos 🪐

August 4: a new moon. The best time to stargaze!