{"text":[[{"start":6.5,"text":"Andreas Hoepner co-leads the GreenWatch team at University College Dublin. He also heads the Data Science Hub of the EU Platform on Sustainable Finance"}],[{"start":16.85,"text":"While most commentators see the current surge in AI use as an obstacle to achieving the goals of the Paris agreement due to the increased energy consumption, my view is that the opportunities probably outweigh the risks."}],[{"start":29.950000000000003,"text":"I do not base my view, however, on large language models (LLMs), whose gazillions of instant autocomplete decisions lack any humanlike understanding of knowledge, but on the potential for AI to substantially speed up battery technology development, which is the prime requirement for the green energy transition."}],[{"start":47.650000000000006,"text":"The Chinese philosopher Confucius differentiated between three ways of learning wisdom: learning from imitation (the easiest way), from experience (often the bitterest) and from reflection (the noblest). This third approach requires an in-depth humanlike understanding of what knowledge is — a so-called epistemology."}],[{"start":69.60000000000001,"text":"The machine “learning” underlying the current LLM-driven AI rally is nearly exclusively based on neural networks — binary computing systems whose nomenclature is inspired by the human brain. These networks are trained to classify information into a binary, probabilistic world view at lightning pace. That is very useful for imitating behaviour (“let’s do it” vs “let’s not do it”) or to try enormous possibilities in a short space of time such as in protein folding or materials discovery for batteries."}],[{"start":102.20000000000002,"text":"But learning by reflection, and having an epistemology, requires something more complex than what these binary results provide. For example, a human learner can reason about appropriate behaviour based on moral principles or heuristics, whereas neural networks can only imitate this behaviour once observed from a large number of humans."}],[{"start":121.95000000000002,"text":"In fact, when asking half a dozen major LLMs if they possess an epistemology, all decline except Claude and ChatGPT. One LLM, Meta AI, is particularly candid, stating that it does not experience doubt, justification or truth in the way humans do."}],[{"start":139.50000000000003,"text":"This may change with the introduction of quantum computing. Until then, the computational power that is required to layer more complex outcome functions over each other in high frequencies is virtually impossible to access."}],[{"start":153.00000000000003,"text":"The neural networks that are used in most LLMs are incredibly fast in imitating language, usually scraped from (nearly) the entire internet. They can determine the most likely letter succeeding and preceding a set of text."}],[{"start":166.80000000000004,"text":"An LLM completes trillions of text auto-completes in seconds to provide its user with the most probable plausible response to a prompt. Essentially, LLMs are hugely powered auto-complete machines."}],[{"start":179.75000000000003,"text":"This is fantastic when asking a general question on a subject where the LLM has millions of letters of text to imitate, such as corporate net zero targets. But for a more specific question on a subject where it has a much smaller corpus of letters available for imitation, actual CO₂ emissions for example, the results will be limited."}],[{"start":198.80000000000004,"text":"For instance, when asking six LLMs if a British energy company has a net zero target, they all provide the correct answer. Asking the same LLMs for the 2025 Scope 1 greenhouse gas emissions value of the same company, less than one in five manages to return the correct value and two-thirds return an incorrect value. One humbly declines."}],[{"start":219.60000000000005,"text":"Consequently, while the current neural network-based LLMs are the fastest ghostwriters and have impressive artistic skills, they are in themselves unlikely to revolutionise the acquisition or processing of emissions data. In fact, the two-thirds of the LLMs which provided the incorrect value spent a lot more energy than a human would for no benefit. Given this substantially increased energy use, some assume that the rise of AI is a net negative for the green transition."}],[{"start":247.90000000000006,"text":"Several AI hyperscaling companies aim to meet that energy demand predominantly with renewable energy but the majority appears to use the average energy mix. Hence, when LLMs are unsuccessful for data retrieval, they may waste energy in principle."}],[{"start":262.6500000000001,"text":"That being said, the quantities of carbon emissions from a company’s purchased energy generation, known as Scope 2, are still tiny."}],[{"start":272.0000000000001,"text":"Most crucially perhaps, the ability of AI to make a large number of binary screening decisions in a short period of time is hugely promising for developing battery technology, probably the biggest bottleneck in terms of the electrification needed to achieve the goals of the Paris agreement."}],[{"start":288.0500000000001,"text":"Tens of millions of materials are potential candidates for batteries, and many are yet to be explored. In the past three years, teams at hyperscaling companies have partnered with laboratories to conduct material discovery a hundred times faster than before."}],[{"start":303.90000000000015,"text":"Google DeepMind scientists John Jumper and Sir Demis Hassabis won half of the chemistry Nobel Prize in 2024 for developing an AI model to predict the complex structures of proteins."}],[{"start":316.3000000000001,"text":"Maybe DeepMind or a team at another hyperscaler can revolutionise battery technology at a similar pace ahead of the Paris agreement’s 2050 target year. Time is tight but I am optimistic."}],[{"start":335.70000000000016,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1781783991_4422.mp3"}