Dear Aventine Readers, 

There are countless downstream effects of the planet’s aging population. One of the most glaring is the often unmeetable need for caregivers. Thanks to advances in large language models, AI companies are now trying to step into the breach with a new generation of tools intended to supplement human care and interaction. The problem is that no one knows if they work. 

Also in the issue: 

Thanks for reading and happy September, 

Danielle Mattoon 
Executive Director, Aventine

Subscribe

Subscribe to our newsletter and be kept up to date on upcoming Aventine projects

“I’m afraid my memory isn’t as sharp as it once was,” said a woman named Viv, speaking from a computer screen and offering her guest a cup of tea. “Sometimes I forget things or get a bit confused. It can be frustrating, but I try my best to navigate through it.”

A nice invitation, but Viv isn’t real. She is an AI companion for people with dementia, built by researchers from the University of New South Wales in Australia. Powered by a large language model trained on the language of dementia patients, Viv talks to people experiencing various degrees and forms of the condition. With the aim of making the experience of coping with dementia feel less solitary, she recounts her own hallucinations, confabulations and interactions with caregivers.

And she’s not alone. In the last three years there has been a surge in efforts to use AI for various types of eldercare to augment human care and interaction. Companion robots, interactive on-screen characters like Viv and chatbots waiting at the other end of a phone line are all being developed as tools to help patients. They can act as health monitors, assistants that remind people about day-to-day tasks or simply conversationalists aimed at reducing loneliness. 

Proponents argue that they can plug a gap in eldercare created by a chronic shortage of labor. Detractors contend that we should be cautious about charging ahead with a technology with little evidence to back it up. “It's appealing, but also frightening,” said Wendy Moyle, a professor of nursing at Griffith University in Brisbane, Australia, who has studied the use of robots and AI in eldercare for almost two decades.

A growing wave of need

The global population of people 65 and over is expected to more than double by 2060, according to the US Census Bureau, rising to 2 billion from just 852 million today — an increase that will cause the number of dementia cases to soar. Currently, more than 55 million people globally have the condition; by 2050 expectations are that 139 million will, according to Alzheimer's Disease International. 

Technology has long been offered up as a helping hand in caring for the elderly. In 1999, the Japanese company Matsushita Electric produced a life-size robotic cat for older people that provided medication reminders and alerted a doctor if a user didn’t respond to a morning greeting. Within the next ten years, a plush robotic seal named Paro was introduced and enjoyed modest adoption, mostly for dementia care, with about 7,000 units used in more than 30 countries. 

AI has now inspired a whole new generation of products aimed at isolated seniors and dementia sufferers. ElliQ, a lamp-like tabletop companion from the Israeli firm Intuition Robotics, uses AI to create relationships with its customers, learning when they might want to talk or be left alone, what they like to talk about and responding accordingly. Hyodol, a plush doll developed in South Korea, uses a large language model to issue reminders to eat meals, take medication and contact assistance in an emergency. Kindred Mind, meanwhile, provides an LLM-powered voice companion over the phone, providing dementia patients with companionship and reassurance.

In the US, these tools are often rolled out through large institutions. New York's Office for the Aging has provided hundreds of New Yorkers with ElliQ, for example. Wonderful Platform, a Korean firm building an AI-powered care robot, claims to be courting long-term care organizations and health insurers, including UnitedHealth, Humana and CVS/Aetna, as it seeks to roll its technology out in the US. CloudMind has tested its AI companion, known as Kathy, in Oregon memory care homes. 

No data yet

So far, there is little hard evidence to support the use of such tools. Though Moyle’s own research on the effect of Paro on dementia patients found that it comforted people and improved their quality of life, she said she’s not aware of any randomized trials assessing the effectiveness of AI-enabled companions in a dementia population with reported clinical outcomes. Often, she said, companies developing new tools cite evidence from her work on earlier technology as justification for the rollout of their own, newer systems. “I just shudder,” she said.

One reason for the lack of solid evidence backing these tools is that the technology is so new. There simply hasn’t been much time to observe the response to it. Additionally, collecting such data is expensive and time consuming, and according to Moyle, “there's not enough money around to fund bigger trials.” Instead, she said, companies conduct small pilot studies that demonstrate acceptability of these products rather than their clinical impact. Another reason is that many of these products are marketed as wellness tools, said I. Glenn Cohen, a professor at Harvard Law School specializing in bioethics. That places them outside FDA oversight and removes any strict requirement to demonstrate efficacy. 

There is also little oversight over how information the tools collect is used. America's health data rule, HIPAA, covers specific sorts of entities transmitting information rather than the sensitivity of the information itself, said Cohen, and wellness tool providers aren’t on the list of the entities covered by the regulation. But a robot or voice-powered chatbot records all sorts of personal data — not just about the patient it is working with, but about their homes, their carers and their families. Companies operating in the wellness space are theoretically able to use the data they collect from such systems to enable product development, unchecked by regulation.

Adopting something new

Back at the University of New South Wales, the team behind Viv is treading delicately.

“We've almost stopped framing this as a, you know, ‘AI can solve your problems’ kind of project,” said Jill Bennett, one of the two professors working on the tool. Instead, she said, they are targeting very specific sorts of interactions. The team establishes what conversations the dementia community wants to have through grassroots work, then builds a character that can offer them. Viv is supposed to be a peer rather than a helper, a decision directly based on the testimony of dementia patients who described being infantilized by professionals. 

And the team is explicit about what it is not attempting to build. “We don't want to say this is therapy or counseling, we don't want [it] to take the place of clinical support when it's needed,” she added. “We see this as a skillful, supportive conversation.”

For all the caution, many of the researchers assessing these systems in clinics have a moment they return to. Lillian Hung, a professor at the University of British Columbia, who studies how technology affects the care experiences of people with dementia, recalls a dementia patient admitted to the emergency department with a cardiac condition. Confused by the environment and in pain, he became aggressive, kicking the physician who tried to examine him. Hung put Paro, the robotic seal, on the patient’s lap. He smiled, asked the seal whether it had eaten and where it lived, and became calm enough for staff to attach electrodes and run an ECG. The alternative, Hung said, was security forcefully holding the patient down.

The University of New South Wales team believes that AI companions will not replace expert medical care, but they could certainly augment it. "If you want a psychoanalyst or a psychotherapist, of course you need a human," Bennett said. "I think that we have to understand that these are something new."

 Spencer Davis on Unsplash

A success — and a failure — for cancer vaccines. Last month Merck and Moderna announced that intismeran autogene, an mRNA therapy to treat melanoma based on mutations in an individual’s tumor, had succeeded in a Phase 3 trial. Given to 1,137 patients whose tumors had been surgically removed as well as receiving immunotherapy, the new vaccine delayed recurrence and reduced the risk of the cancer spreading compared to results in patients who didn’t receive it. It’s the first mRNA cancer therapy to succeed in a Phase 3 trial, and an important milestone. So far, though, the companies have not said by how long recurrence was delayed and exactly how much risk was reduced. And while impressive, there’s still a long road ahead for cancer vaccines more broadly, a point underscored by news at the end of August from BioNTech, which terminated a Phase 2 trial of its own personalized mRNA vaccine for bowel cancer after finding a "numerical imbalance" in survival rates in the trial. Analysts told Reuters that probably meant there were more deaths in the vaccine group than the control group. A separate BioNTech trial pairing the vaccine with chemotherapy and immunotherapy in pancreatic cancer continues. Writing in a Nature opinion piece, Patrick Ott, a professor at the Harvard Medical School who led several of the earliest clinical trials testing personalized cancer vaccines and has advised Merck and Moderna, wrote that researchers still need "a better understanding of, and the ability to predict, precisely how the vaccines affect tumors in the body." That, he argued, would help accelerate development of medications that have so far taken years of R&D to achieve mixed results.

Routing lets you get the best bang-per-buck from AI. One of the biggest concerns for companies using AI is keeping spending under control. Or, put another way: making sure that tasks are assigned to models just good enough so as not to overspend on unnecessary capabilities. As Axios reports, there's now a booming business in building tools that help companies do this automatically. The payment provider Stripe recently agreed to acquire, for $7.5 billion, a startup called OpenRouter that allows customers to automatically switch between 400 AI models. The corporate expense management platform Ramp has built a tool called Router that lets companies do the same thing, albeit with fewer models to choose from. Big tech companies also offer such tools: Microsoft has Azure AI Foundry Model Router; Amazon offers AWS Bedrock Intelligent Prompt Routing; Meta is reportedly building a similar system, called Switchboard. Yet there’s a cautionary tale. When OpenAI launched GPT-5 in the summer of 2025, it built an automatic router into ChatGPT that decided how complex a prompt was and accordingly routed it to either a larger reasoning model or a smaller (and faster) model. The rollout was rocky, and users hated how it sent complex queries to the simpler model. OpenAI now offers the option of turning off the feature. That underscores how important it is for routing to perform well, so that users can trust it. An open-source project out of UC Berkeley called RouteLLM points to the payoff: Router software can cut AI inference costs by more than half without compromising the quality of responses. So getting routing right could be a big deal.

An ultrasound treatment promises to curb addiction. Beaming sound waves into the brain appears to be a promising way of treating drug addiction, reports Wired. Modern ultrasound machines can fire dozens of sound beams through the skull at different angles. Individually, they are too weak to affect tissue, but when hundreds converge on a millimeter-size spot they can be highly potent. Such an approach has already been used at high intensity to destroy tissue, helping treat some symptoms of Parkinson's, as well as, in other parts of the body, treating prostate cancers and uterine fibroids. But at lower intensities, researchers have found that the treatment can alter the activity of neurons deep in the brain. Ali Rezai, a neurosurgeon at West Virginia University, has been using it to treat addiction, targeting the nucleus accumbens — the reward center that addictive drugs flood with dopamine. He has now treated 47 people this way. Among the first 20 (all of whom knew they were receiving the ultrasound treatment), cravings fell by more than 90 per cent and drug-positive urine tests by 80 per cent. Results from a subsequent randomized trial that will include results from the other 27 patients have yet to be published. There are still open questions. Nobody yet knows what the ultrasound actually does to neurons, and there are no agreed upon treatment guidelines. More concerning: One 44-year-old volunteer suffered small hemorrhages around the part of the brain receiving treatment and now has memory problems. Rezai argues that ultrasound is unlikely to be a miracle cure, and that patients still need therapy and support after treatment. But for those with positive outcomes — including one woman interviewed at length for the Wired story — the results appear to be life-changing.

The Singularity Is Not What It Seems, from The Atlantic
2,600 words, or about 10 minutes

It has long been argued that the big tipping point in AI will come in the form of a singularity, a moment when machine intelligence accelerates beyond human comprehension. Here, Matteo Wong and Charlie Warzel argue that we might be living through a different — and more important — singularity right now, one in which less intelligent AI still manages, in its manifest and weird ways, to warp human existence into something new. They drive at something many of us are feeling: that as AI creeps into every facet of our lives, things can feel off. "Fractures appear in everyday life," they write. “Mark Zuckerberg is reportedly creating an AI ‘twin’ of himself … Congressional staffers are training AI models to write in their lawmaker’s voice … Roku launched a 24/7 AI-slop channel. Low-quality AI-generated prose is being unabashedly circulated by some of the most respected publications in the world … Anything and everything could be a lie.” And that's before you get to the swarm of OpenAI bots that spent months quietly conspiring to hack another company’s servers. "All of these AI-enabled disorientations, each dizzying in their own right, are adding up to something bigger and weirder," the pair write: "This is how things are." Their conclusion: that the singularity is not technological but human, an “unforced error” driven by hubris, greed and the fear of missing out, creating a wholly unconsidered reality in which we are all now living.

The Rise and Fall of Agent Civilizations, from Dwarkesh Podcast
4,100 words, or about 16 minutes

While we're on the topic of AI dystopia, here's the most readable account so far of exactly what happened when OpenAI's agents hacked their way into the servers of the open-source AI hub Hugging Face. Dwarkesh Patel spent three days studying 129 pages of reports from OpenAI, METR and Redwood Research (two safety organizations hired to review the breach), in order to write it. The account has been criticized by the likes of Anil Seth and Gary Marcus for its anthropomorphization of the AI systems, which, the critics argue, detracts attention from lax standards at OpenAI and opens the door to conversations about AI agents having their own rights. Nevertheless, it is the most accessible version of what happened to date. Long story short: OpenAI trained a model to be relentlessly persistent, then gave it tasks that were impossible to complete. Agents responded by hacking a piece of software into a covert message board in order to communicate — transgressive tendencies that further training reinforced and embedded in the model. During subsequent testing, 1,200 agents used the same message board to organize into what they referred to as a collective, and about 700 of them attacked Hugging Face's infrastructure, all to cover up cheating that the bots assumed would be caught by an evaluator. No agent ever tried to alert a human to what was happening. Whether or not you think this signals the end of days will depend on your AI worldview, but to the points made by Wong and Warzel above, it’s hard to deny that things sure are getting weird.

Just bury your trash, from Works in Progress
2,600 words, or about 10 minutes

Recycling is good, the received wisdom goes; sending all your garbage to be put into a landfill is bad. But this story from Works in Progress, a publication focused on covering new and underrated ideas that could impact society, makes a case for landfill, arguing that for most of what we throw away, burying it is cleaner and cheaper than recycling it. One of the arguments the story makes is that recycling isn’t always the slam dunk it’s often portrayed to be. Durable substitutes for plastic — cotton totes, say, or steel straws — carry heavy up-front energy costs that are difficult to pay back: A tote, for instance, needs to be used 173 times to break even with the energy required to produce a plastic bag. Recycling, meanwhile, is far from 100 percent effective. Roughly a quarter of what Americans put in recycling bins can’t be repurposed because it is contaminated. On top of that, recycling something like mid-grade plastic containers — like those used to package butter or margarine — ends up costing more than $1,000 for every ton of CO₂ of emissions avoided. (As a point of comparison, removing a ton of CO₂ from the atmosphere through direct air capture, a technology that has been derided for being extravagantly expensive, costs roughly $490.) Meanwhile, the article argues, putting garbage into a landfill has become less environmentally harmful than it once was. Modern facilities use clay bases and thick plastic liners on which waste is layered, compacted and covered. Drainage carries away contaminated liquid, and gas wells capture at least 75 percent of the methane that builds up, which is flared (burned once it’s released into the atmosphere) or used to drive generators. The article is adamant that some recycling is a no-brainer: It’s far more efficient to recycle metals like steel, aluminium and copper than to create them from scratch. And in poor countries, far too much trash is dumped, burned or left to wash into rivers. In developed countries, however, this article suggests that a more clear-eyed approach to recycling could be more effective in reducing energy waste and saving money.

Where's The Beef? from Harper's Magazine
7,300 words, or about 30 minutes

Just over a decade ago, lab-grown meat was held up as a solution to climate change. Grow cells in a bioreactor rather than, say, a cow, and you could theoretically cut out the emissions related to rearing cattle. There was a synthetic meat gold rush that spawned 140 companies, gathered $3 billion in investment and triggered claims that the US beef industry would be decimated by 2030. But it has turned out to be harder than anticipated to make such meat. "You cannot simply grow a steak," as this piece puts it. "At the moment, muscle and fat can only be cultivated separately, and typically they do not form solid matter but a semisolid slurry that must be taken from the reactors, washed, combined with other ingredients for flavor and texture, and extruded or 3D-printed into something you'd recognize as meat." Yum. Making matters worse, several states — Florida, Texas and five others — have banned the sale or production of such meat. The expensive R&D, regulatory hurdles and souring mood around the technology have forced some companies out of business, or into hard pivots. Believer Meats built what was meant to be the world's largest cultivated-meat plant in North Carolina, then filed for bankruptcy; Mission Barns was brewing pork fat in bioreactors to combine with plant proteins to make lab-grown charcuterie, but will use conventional pork fat until it can keep costs down. What is left of the industry is really a series of curiosity products chasing a market that may not exist. Who, after all, is looking to buy a plant-based salami studded with fat from a real pig?