📐 AI Solves Olympiad Geometry | #3

Quote of the Week 💬 

AlphaGeometry Can Solve Olympiad-Level Geometry Problems

⌛ The Seven Second Summary: Google Deepmind introduced AlphaGeometry, an AI system capable of solving complex geometry problems at a level approaching a human Olympiad gold-medalist.

🔬 How It Was Done:

• The AI system uses a dual approach: first, there is a neural language model that is designed to understand and predict patterns in language. This model is integrated with a symbolic deduction engine, which operates on formal logic and clear rules to make deliberate and rational decisions.

• When faced with geometry problems, the language model predicts which additional constructs, such as points, lines, or circles, might be useful to solve problems. These predictions then guide the symbolic engine in its deductive process to find the best possible solution.

• AlphaGeometry was also trained without human supervision. To train itself, the AI system self-generated 100 million unique diagrams of geometric objects and exhaustively derived all the relationships between the points and lines in each diagram. As the system found all the proofs contained in each diagram, it was able to work backwards and understand if any additional geometric constructs were needed to arrive at its proofs.

🧮 Key Results:

• AlphaGeometry solved 25 out of 30 Olympiad geometry problems while under competition time limits, a significant advancement compared to previous state-of-the-art models that could only solve 10 problems under similar constraints.

• AlphaGeometry’s performance already surpasses the average score of human silver medalists (22.9) and is near the performance of an human gold medalist (25.9).

💡 Why This May Matter: Since the AI system was trained on synthetic data, it reinforces the potential of large-scale synthetic datasets to overcome data limitations and still significantly advance AI capabilities in the future.

🔎 Elements To Consider: While AlphaGeometry excels in geometry, its application is limited to just one-third of all Olympiad problems. In competitive settings, Olympiad problems involve algebra, number theory, and other mathematical topics, so there is still a significant gap between a single AI system like AlphaGeometry vs the average gold medalist across all Olympiad domains.

🧵 Thematic Thread: AlphaGeometry demonstrates AI’s growing ability to develop mathematical reasoning. If, we can infer it is possible for AI systems to develop reasoning in other domains as well. When the time comes, this will illuminate all kinds of new knowledge to advance society forward.

📚 Learn More: Deepmind. Nature. Github.

Stat of the Week 📊 

High Res Device Measures Brain Activity, Neuron By Neuron

1 neuron

⌛ The Seven Second Summary: A team of researchers have developed an instrument to record brain activity with significantly higher resolution details than standard clinical instruments.

🔬 How It Was Done:

• The devices used to record brain signals is incredibly thin, measuring less than 1/5th the width of a human hair. This property allows far more recording devices to be placed within the same surface area of a brain compared to traditional clinical devices. The result is a more precise recording of the brain.

• The length of these thin devices is still relatively long despite their thinness, extending up to 10 cm. This length allows the recording devices to access deep parts of the brain and capture a holistic view of a patient’s brain activity.

🧮 Key Results:

• These new devices can simultaneously record signals from up to 128 different points in the brain, a significant advancement compared to the 8-16 recording points found in today's clinical devices.

• In practice, these devices are so precise that they can effectively read signals originating from the individual activity of one or two neurons. Furthermore, this instrument can ‘zoom out’ and record an aggregate activity view of many neurons in a specific brain region.

💡 Why This May Matter: Isolating and measuring the activity of individual neurons can significantly enhance our understanding of fundamental neural processes. As we deepen our knowledge of the brain, we gain insights into improving cognitive and motor functions, and we can develop treatments for neurological decline and disorders.

🔎 Elements To Consider: This instrument has only been tested on two humans and a handful of rats, pigs and primates. Therefore, the instrument still needs to go through many additional sets of safety and efficacy studies before it is adopted in clinical practices.

🧵 Thematic Thread: The brain is one of biology’s most intricate and complex black boxes. However, advances in technologies like this may one day unravel and illuminate all of its mysteries to humanity.

📚 Learn More: UCSD. Nature. 

AI x Science 🤖

ML Models Teach Each Other To Identify Molecular Properties

Biomedical engineers devised a method to improve the effectiveness of machine learning (ML) algorithms to identify and characterize molecules to use in potential drug therapeutics or other materials.

This process, coined YoDeL, relies on two ML models. The initial model is trained to assess the informativeness or relevance of various data for its designated task. Subsequently, this model shares its selected data with a second ML model. The second model leverages the received data from the first model to complete its different assignments.

The YoDeL methodology demonstrated promising results. While YoDeL only outperformed a benchmark active learning ML model 17% of the time across 252 different tasks, it performed comparably to the benchmark in many cases. Furthermore, the YoDeL methodology had significantly faster processing times than the benchmark active learning ML model, where it completed tasks in minutes compared to hours or days to its benchmark.

Thus, the speed and relative performance of YoDeL may already provide practical value for teams handling large data sets and approaching deadlines. More importantly, this methodology is a demonstration of what machine-to-machine collaboration may look like in our new AI era. Duke. Science Direct. Github.

Our Full AI Index

• Research: Over the holiday season, a team of researchers announced Coscientist, an AI system capable of understanding, designing and executing chemical reactions with minimal human intervention. NSF. Nature.

• Business: Coursera’s CEO announced the company added a new user every minute, on average, for its AI courses in 2023. Were you one of them? Reuters.

• Ethics: The World Health Organization released guidance for generative AI applications across health care. The guidance outlines over 40 recommendations to address risks of inaccurate, bias or poor quality responses. WHO. 

• Text-to-Audio: Researchers at Meta unveiled MAGNeT, an open-source AI model capable of generating music, sound effects, and noises from text prompts. arXiv. Hugging Face.

• Policy: The Australian government is assembling an advisory body to consider guardrails for AI design, development and deployment. One such consideration is a requirement for companies to watermark or label content generated by AI. Australian Ministry for Industry and Science.

🧵 Thematic Thread:  We find it illuminating how this story is a reminder that good collaboration is key to drive innovation, no matter if the relationship is human-to-human or machine-to-machine.

Other Observations 📰

Fuel Cell Harvests Energy From Microbes In Soil To Power Sensors

Researchers created a fuel cell, approximately the size of a paperback book, capable of harvesting energy from soil. The device is made of conductive materials and is buried in the ground to attract electrons released by soil microbes. As these electrons gravitate toward the device, they generate an electrical current that can be harnessed to power electronic devices.

To test the fuel cell, the researchers attached an antenna to the device and monitored its ability to wirelessly transmit data about its soil environment. The experiment revealed the efficacy of the device's power generation was largely dependent on factors such as the soil’s water concentration. Nevertheless, after 161 days, not only was the fuel cell able to generate 68x more total power than it needed to self-sustain itself, it also outlasted similar baseline technologies by 120%.

In spite of these achievements, the fuel cell's current energy output is nowhere near large enough to power advanced devices like modern smartphones. However, if paired with an energy storage unit, the device may have value in its ability to power basic components used in agricultural settings or climate studies, such as sensors to log and display soil data in real-time. Northwestern. ACM.

Our Full Science Index

• Internet Accessibility: The World Bank announced a range of its projects contributed to a 115% increase in the number of Internet users in Sub-Saharan Africa, from 19% in 2016 to 36% in 2021. When considering the entire continent, The World Bank reports its projects resulted in 160 million Africans gaining broadband Internet access between 2019 - 2022. The World Bank. 

• Public Health: The WHO recently released its tobacco trends report, revealing a 39% reduction in adult consumption since the turn of the century. The report indicates approximately 1 in 5 adults worldwide consumed tobacco in 2022, compared to 1 in 3 adults in 2000. WHO.

• Energy: The U.S. Energy Information Administration released projections indicating an anticipated 75% increase in the country's solar generation capacity between 2023 and 2025. The country’s wind capacity is expected to increase 11% during the same time period. EIA.

• Space: Just last year, India joined the United States, China and the Soviet Union as the only countries to successfully land a spacecraft on the moon. Last week, Japan became the fifth country to accomplish this feat. AP News.

• FDA: In December, the FDA announced its first approval of a CRISPR-based medicine, Casgevy, to treat sickle cell disease. Last week, the FDA announced it is expanding its approval of Casgevy to also treat patients 12 and older with beta-thalassemia. Similar to sickle cell, beta thalassemia is an inherited blood disorder, and patients with this condition experience life-long anemia, requiring frequent blood transfusions. FDA.

🧵 Thematic Thread: There is energy and electricity all around us. Even when buried deep underground, it is possible to harness this latent power through innovative technologies.

Media of the Week 📸 

A Four Wheeled Robot Navigates Stairs & Obstacles

Behold as a four legged robot with wheels for feet maneuvers between, around, up, and down various obstacles in an outdoor city environment. Coming soon to a plaza near you?

Hubble Telescope Captures Image Of Galaxies Colliding

The Hubble Space Telescope captured two galaxies, roughly 570 million light-years from Earth, in the midst of a collision. The forthcoming collision may result in a single, merged galaxy, but it can take hundreds of millions of years for the aftermath of these sorts of collisions to unfold. Let’s wait and see. NASA.

Scientists Discover Four New Species Of Deep-Sea Octopus

A team of researchers recently announced they discovered at least four new octopus species during a December deep-sea expedition off the coast of Costa Rica. There were two deep-sea expeditions in this oceanic region last year, and over 300 animal specimens have been collected from these trips to research the biodiversity and underwater ecosystems in the area. Schmidt Ocean Institute.

This Week In The Cosmos 🪐

January 25: the wolf moon is the first full moon of the year

Until Next Time 💭

If you have more feedback you want to share, send us a message!