Dear Aventine Readers, 

As 2025 begins, we’re looking ahead at how science and technology are primed to change the way we live in the coming year and have come up with ten important trends to watch for. 

As you may expect, AI will play a big role. Commercialized systems will increase AI’s presence in the background of our lives — organizing our days (if we ask it to), assuming more back-of-office and customer service tasks, even helping doctors cut down on paperwork. AI-enhanced robotics will become more dexterous and intuitive, maybe paving the way for a new wave of automation that operates beyond the confines of predictable factory and warehouse environments. 

Beyond AI, we might see life-altering cancer vaccines, meaningful advances in quantum computing and expanded development of both geothermal and nuclear energy. Then there’s crypto, which might finally become legit if a change in regulations gets through the Senate. 

It’s going to be a big year! We’re looking forward to sharing it with you. 

All best, 

Danielle Mattoon 
Executive Director, Aventine

Autonomous AI systems that accomplish tasks without having to be asked to do so by humans are poised to make our lives more efficient, particularly at work.

AI agents, as these systems are known, are already being made available by some of the big tech companies and don’t rely on prompts the way ChatGPT does. Instead, once you sign up, the agent will sense when something needs to be done and do it, using multiple forms of AI in the process. For now, that might look like telling an agent to scour your inbox and send a weekly roundup of everything you need to know about a certain topic. In the future the options could be far more nuanced and useful: A travel bot, for instance, might notice changes to your meeting schedule, take the initiative and book and rebook your flights and hotels to ensure that your itinerary is always up to date.

In theory, so-called agentic AI might not just save us from tedious tasks but also move artificial intelligence into a world where we don’t need to keep track of which tasks need doing and which ones can be automated. “You used to create the automation; it can now self-create,” said Humayun Sheikh, CEO of Fetch.ai, a marketplace for AI agents.

There’s a growing army of these agents, which can be used by anyone but are predominantly focused on automating business tasks. Salesforce has AI agents designed to take on sales, marketing, commerce and customer service tasks. Microsoft released a suite of bots capable of similar tasks that are being tested by companies including Clifford Chance, McKinsey & Company and Thomson Reuters. Open AI has unveiled its own agents in Swarm. And a slew of startups are building specialized agents for businesses, to automate activities like recruitment, medical billing and debt collection, according to senior leaders at Y Combinator, who think agentic AI could become as big a market as Software as a Service (SaaS). Deloitte predicts that 25 percent of companies that use generative AI will carry out trials with AI agents in 2025.

Many of the concerns around LLMs apply to agents as well. Reliability, accuracy, privacy and security are all potential stumbling blocks. That’s why Sheikh predicts adoption will be seen first in “low-risk areas” such as customer service and gig economy apps. But he thinks 2025 will be a big year: “There's going to be a lot of good examples of agentic applications coming up,” he said.

The tidal wave of progress in AI is spilling over into the physical world, helping robots undertake tasks with unprecedented competence.

Historically, robots have excelled at performing repetitive tasks in predictable environments and failed when it comes to general tasks in chaotic surroundings. But over the past 12 to 18 months that has begun to change, with robots increasingly accomplishing tasks such as frying shrimp, turning the pages of books or roughly folding laundry — actions that require generalized dexterity and were until recently impossible to imagine a robot doing competently.

Improvements in two long-established approaches to training robots are fueling these advances, said Pulkit Agrawal, an associate professor at MIT focused on AI and robotics. The first is reinforcement learning performed in software simulation, in which a robot model is given a goal and then uses trial and error to achieve it, getting rewarded when it does well and penalized when it does poorly — feedback that informs subsequent efforts. The second is behavior cloning, in which humans use controllers to physically manipulate a robot so that an AI system can learn from the movements, a technique that has been supercharged by generative AI. Both approaches hope to overcome the most significant obstacle to using AI to advance robotics: a paucity of high-quality data on which to train models. 

Improvements in large language models and image recognition have also helped, said Agrawal, enhancing task planning and perception of surroundings respectively. Finally, an emerging area of research is taking on the data problem directly, using existing video of human activity to train robots, which — if achievable — would reduce the expensive and time-consuming need to create training data from scratch. 

The excitement around the sudden ability to make robots more agile and dexterous is palpable in the robotics community, at least among investors. In late 2024, the same year it was founded, robotics startup Physical Intelligence (PI) raised $400 million in financing from investors including Jeff Bezos and OpenAI to pursue its R&D in imitation learning. Also last year, the so-called godmother of AI, Fei-Fei Li of Stanford University, raised $230 million for her startup, World Labs, which is focused on spatial understanding and reasoning. These ventures join a large and growing collection of other startups, university labs and large private companies all focused on applying cutting-edge AI to the development of robotics.

A string of advances in quantum computing through 2024 has paved the way for the technology, which has long faced fundamental challenges, to finally become useful.

Over the past decade there has been enormous progress in our ability to combine multiple quantum bits, or qubits, to create ever larger quantum computers. But there has been a persistent problem: These qubits, defined by quantum particles such as electrons and photons, are easily disrupted, a process that introduces errors into quantum calculations. “When we try to implement the algorithms that we would like to use quantum computers for, we get unreliable results,” said Roberto Bondesan, senior lecturer in quantum computing at Imperial College London.

In recent years, there’s been an effort to build error-free qubits called logical qubits — collections of qubits that work together to reduce errors. In 2024, a series of research results from Google, Microsoft, AWS and Yale not only confirmed that logical qubits do reduce errors but, more important, demonstrated that they are scalable and that it’s possible to use them to perform calculations. In other words, when more logical qubits are combined, the rate of errors decreases exponentially and software can run on them, which means that building larger quantum computers should make them work far, far better. In December Google unveiled the physical embodiment of these advances in its new Willow quantum chip.

These achievements are regarded in the quantum computing community as a significant inflection point, though some engineering challenges remain. Next on the to-do list is to use quantum chips like Willow to perform commercially useful applications, such as simulating drug performance or new chemical reactions, which can’t be performed on classical computers. Another research push will attempt to make logical qubits more efficient in order to make more efficient devices. Time, investment and patience are still required to commercialize quantum computing, but it increasingly seems like the technology has passed the “when-not-if” milestone.

After decades of opposition spurred in large part by accidents at the Three Mile Island and Chernobyl and Fukushima nuclear power plants, the tides are turning for nuclear power.

Global electricity demand is growing rapidly, driven in particular by data centers powering artificial intelligence. At the same time, most nations are seeking to reduce carbon emissions and are in search of always-on power sources for when solar and wind can’t deliver. Nuclear energy provides an obvious solution. 

In the U.S., energy companies are trying to restart shuttered nuclear power plants. The Palisades nuclear power plant on Lake Michigan is set to reopen late 2025, and Microsoft is supporting the reopening of one of the Three Mile Island reactors in Londonderry Township, Pennsylvania, which is expected back online in 2028. Other large tech companies are also getting involved: Google and Amazon made investments in small modular reactors — as yet unproven facilities that generate about a third of what a large nuclear reactor produces but are theoretically easier and cheaper to manufacture and install.

Such enthusiasm is a global trend. Some 31 countries have signed a declaration of intent to triple global nuclear generation capacity by 2050. Around the world, there are 65 nuclear reactors under construction, 87 more that have been approved for construction and another 344 that have been proposed, according to the World Nuclear Association. Those would augment the 440 reactors that are currently operational. Against this backdrop a wave of innovation is taking place, particularly in advanced modular reactors, which will use new approaches to increase efficiency, and in microreactors, which are expected to be small enough to be moved by road or rail.

The appetite for building new large-scale reactors during the Trump administration remains to be seen, but those in the industry remain hopeful about the support for nuclear energy in general. “I’m cautiously optimistic nuclear will continue to have momentum,” wrote Patrick O’Brien, director of government affairs and communications at Holtec, the company currently restarting Palisades, in an email to Aventine. “Seems the incoming administration has continued to signal their support for nuclear as a baseload clean energy provider and the Congressional support remains bipartisan.”

Artificial intelligence is accelerating our attempts to understand the world, and may soon make its own discoveries.

AI is pervading science. Many of the advances we cover in this newsletter are now made possible by AI, whether it’s predicting the effects of climate change or constructing gigantic atlases of the brain. “Applications of AI are found across all STEM fields,” according to a 2024 report by the Royal Society, the world's oldest continuous scientific society, helping scientists obtain insights from data, synthesize information and write code, among many other things. And many of the advances made possible by AI are beginning to show results. The Economist reports, for instance, that 65 AI-inspired molecules are new drug candidates in human trials of some sort, and at least four or five are expected to progress to the final phase of testing this year. AI that is able to predict protein structures — a breakthrough that was awarded a Nobel Prize this year — can now accurately determine the structure and interactions of all life’s molecules, helping scientists to better understand all biological processes. 

But AI is also gaining scientific autonomy. An AI “scientist” developed earlier this year can form hypotheses, check the originality of such hypotheses, test them and write papers about the results. Its research isn’t groundbreaking, but it has delivered incremental advances, such as tweaks to machine-learning algorithms to make them more effective or efficient. Two AI models developed by DeepMind, meanwhile, are able to perform mathematical proofs, and as these sorts of models are trained on larger datasets of mathematical knowledge it is anticipated that they could begin to prove theorems that humans have so far been unable to.

The prospect of AI-based scientific discovery that exceeds what humans are capable of may seem far-fetched. But “it could happen at any moment,” said Professor Yang-Hui He, a mathematical physicist who works with AI tools and is a fellow at the London Institute of Mathematical Sciences. “I don't think [AI will solve] something like the Riemann hypothesis” as its first independent discovery, he added, referring to what is arguably the best-known unsolved math problem in the world. “But it could!”

Riemann hypothesis or not, in 2025 you should absolutely expect to see AI making ever larger contributions to science, freeing researchers to make greater advances and playing an increasingly large role in the discovery process itself.

AI promises to save physicians hours of work every day by automating the mundane task of note-taking.

The idea behind the medical AI scribe is straightforward: A software program listens to and summarizes patient-physician consultations in real time using natural language processing and large language models, and creates the notes that a doctor would usually write during and after an appointment. Audio recordings are typically not stored and AI-generated notes are encrypted to help satisfy medical data protection requirements around the world, such as the U.S. Health Insurance Portability and Accountability Act (HIPAA), and doctors must check the AI-generated notes to ensure they are error-free. 

It is already a crowded market: Earlier this year, Stat reported that more than 50 companies are already offering automated medical documentation software for healthcare providers. A trial by the Permanente Medical Group found that, in a trial that included 3,442 physicians using these tools in 303,266 patient encounters, it saved them on average one hour every day.

Adoption to date is modest but picking up quickly. Zaahir Moloo, a physician and co-founder of a medical scribe startup called ScribeBerry that helps doctors write notes and referral letters, said that some doctors are concerned about the products’ potential to hallucinate or introduce bias into text, and there are lingering worries about data privacy issues. But use of these tools is growing: A survey of Australian medics revealed that the percentage of general practitioners who use AI scribes regularly in their consultations has risen from just under 3 percent to just over 8 percent in the space of four months, which suggests the technology will break into the mainstream in 2025.

Advances in personalized medicine and mRNA vaccines are leading the way to new approaches that encourage a patient’s immune system to fight cancer. 

The current momentum is driven by a vaccine called mRNA-4157 that has been developed to fight melanoma by Moderna and Merck. The vaccine is performing well in company-sponsored trials, and while its launch is scheduled for 2027, Moderna’s CEO has said that its release could be accelerated to 2025 in some countries that choose to expedite its approval. The ultimate goal is for such vaccines to be preventive, but they are currently being tested on and developed for people with existing cancers. 

The mRNA-4157 vaccine and others like it — including those for colorectal and pancreatic cancers — are tailored to each individual. First, biopsies from a tumor are taken and studied to understand how a patient’s immune system will most effectively fight off cancerous cells. Then instructions are encoded into messenger RNA (mRNA), which prompts cells in the body to produce proteins that trigger an immune response against those cells. The process typically takes one to two months from biopsy to treatment. 

“The great advantage of preventive cancer vaccines lies in the potential to harness a less-compromised immune system in vaccine recipients before their immune responses become affected by the advanced status of the disease itself or by aggressive treatments such as chemotherapy,” wrote Michele Graciotti and Lana Kandalaft of the Ludwig Institute for Cancer Research in Lausanne, Switzerland, in a Nature Reviews Drug Discovery article published in late 2024.

In addition to mRNA-4157 and its associated trials, the National Health Service in the U.K. has established a Cancer Vaccine Launch Pad to accelerate clinical trials with the technique, and in 2024 it enrolled its first patients into a colorectal cancer trial. Many different vaccines — for lung, ovarian and prostate cancers among others — are entering trials around the world. And companies such as BioNTech, a pioneer in mRNA vaccines that was heavily involved with the Covid-19 shots, is trying to develop “off-the-shelf” cancer vaccines that would not require the expensive personalization process. 

For now, the vaccines remain costly, so the approval of mRNA-4157 this year would help only a small number of individuals. But the achievement would mark a huge shift in cancer treatment, enabling doctors to more easily head off tumors before they become problematic.

Geothermal energy is becoming an ever more compelling source of clean energy thanks to a combination of technological advances and a friendly political environment. 

Regular readers of Aventine will know that geothermal is having a moment after being overlooked and underfunded for years. Drilling techniques borrowed from the oil and gas industry, along with recently developed technologies, are enabling “next generation” geothermal plants to drill into hotter rock found deeper in the earth than traditional geothermal, creating vast quantities of power. At the same time, startups like Dandelion and Bedrock are taking regular geothermal approaches into urban settings to provide heating for homes, offices and other buildings. 

“The word is out that geothermal is something of a ‘one-stop-shop’ energy solution, offering baseload power, building heating and cooling, direct heat, and thermal energy storage,” wrote Lauren Boyd, director of the geothermal technologies office at the Department of Energy, in an email to Aventine. “It’s hard to find something comparable in the energy space.”

The technology has a number of virtues, including the fact that — once the infrastructure becomes standardized and less costly — advanced geothermal can produce always-on power almost anywhere in the world. It also benefits from the skills and resources of the oil and gas industry, making it one of the least politically controversial sources of renewable energy. In November, the House passed two bills supported by both Democrats and Republicans that will benefit the industry by removing federal permitting regulations around drilling. 

Meanwhile, oil and gas companies including Devon Energy, Chesapeake Energy, BP and Chevron are backing the advanced geothermal startups Fervo Energy, Sage Geosystems and Eavor. 

“This coming year, and the next four years, look very good,” said Joseph Moore, a principal investigator at Utah FORGE, a geothermal laboratory backed by the Department of Energy. “We’ll see plants come online, we’ll see [power] generation, we’ll see a flurry of new startup activities.”

An incoming administration and anticipated legislation that are both highly crypto-friendly could make 2025 a boom year for the industry.

Two years ago, after Sam Bankman-Fried’s FTX imploded, the crypto industry was on the ropes. Gary Gensler, chair of the Securities and Exchange Commission (SEC), called it rife with fraud and vowed that his agency would be “a cop on the beat.” Senator Elizabeth Warren of Massachusetts recruited an “anti-crypto army” in Washington. 

That army has retreated. 

No industry gained more power more quickly than crypto, thanks to November’s elections. Donald Trump embraced the industry while campaigning, promising to make the U.S. the “crypto capital of the planet.” And now that he is headed for the White House with control of both houses of Congress, Washington is closer than ever to enshrining new laws tailored to digital assets.

The most relevant piece of legislation is the Financial Innovation and Technology for the 21st Century Act (FIT21), which was passed by the House of Representatives in May, but stalled in the Senate. The legislation, which will likely be reintroduced this year, intends to settle an important and long-running dispute over whether cryptocurrencies should comply with the same regulations governing the stock and bond markets. 

FIT21 creates distinct categories: digital commodities regulated by the Commodity Futures Trading Commission (CFTC) and “restricted digital assets” regulated by the SEC. To distinguish between them, the bill proposes a formal definition of “decentralization.” Tokens meeting the criteria for decentralization, which many would, would be commodities — not securities — and thus subject to lighter-touch regulation. 

Crypto advocates are divided over some of the bill’s details, but few would disagree that its passage would not only remove legal uncertainty that has discouraged certain crypto-related business models, but that it would also be a monumental statement of legitimacy for the crypto industry. FIT21 stalled in the Democratic-controlled Senate, but it can be reintroduced in 2025, and Republican leaders have said crypto rules will be a “top priority” once they take control. If it is reintroduced and passed, 2025 could be the year that crypto finally goes mainstream.

Approaches from the world of blockchain could allow for a future in which individuals control how their data is used to build artificial intelligence. 

AI models being developed today are owned and operated by a few massive, powerful companies with the resources needed to train AI models. But such a top-heavy industry risks  replicating the negative side effects of today’s internet: bias, disinformation, monopolistic business practices and large-scale data leaks. 

A growing cadre of AI researchers and entrepreneurs hope to create an alternative path forward for artificial intelligence. The vision: decentralized AI. The idea is that AI models can be built on distributed computing networks whose nodes work in concert to provide computing power and host training data despite not knowing or trusting one another — in other words, blockchains. 

If this sounds like more crypto hype, don’t tell the founders and researchers working in this space, who have seen funding in the technology grow to $436 million in 2024 — nearly triple what it was in 2023, according to Pitchbook. 

Anna Kazlauskas, CEO of the startup Vana, is among them. She points out that users who run the Ethereum blockchain together account for 50 times the computing power used to train OpenAI’s GPT 4. Vana boasts early success in corralling that power: It built a system in which users can download their Reddit data and use it to train a foundation model. Data remains in the users’ control, and users get a vote on how the model is used.

It’s just one example of a rising trend. Both Berkeley and MIT have kicked off efforts to foster the community of developers and researchers interested in taking a decentralized approach to building AI. The data protection offered by the technology could help avoid damaging leaks that plague sectors that trade in sensitive personal information such as healthcare and finance. And the ability for users to decide how and when to share data could give an advantage to smaller companies — especially if the firms can demonstrate they’ll treat people’s data with care. 

To listen to current discussions about the technology is in some sense to relive the heady days of the early internet: Participants talk of building models and agents that put people rather than huge corporations first. In 2025, as conversations around the power of the likes of OpenAI, Microsoft and Anthropic grow ever louder, decentralized AI looks like a pragmatic solution to many moral and ethical concerns surrounding the technology. 

It would be premature to suggest that the tech idealists behind decentralized AI will outcompete the giants. But they have something that previous generations lacked: They grew up with the internet as we know it, and they want to do things differently.