For years, AI researchers have anticipated the moment when AI systems will be able to improve themselves better than humans could. With investors pouring money into a new generation of research-driven AI labs, there are more resources than ever available to pursue the goal. Now, one of those neolabs has taken a major step towards making it real.

On Wednesday, Adaption introduced a new product called AutoScientist that helps models learn specific capabilities quickly by using an automated approach to conventional fine-tuning. The techniques are applicable to a wide range of fields, but the Adaptation team is particularly focused on the potential for speeding up and easing the process of training and fine-tuning a frontier-level AI model.

According to co-founder and CEO Sara Hooker, who previously worked as VP of AI research at Cohere, AutoScientist represents a new way to approach the AI training process. “What’s super exciting about it is that it co-optimizes both the data and the model, and learns the best way to basically learn any capability,” Hooker told TechCrunch. “It suggests we can finally allow for successful frontier AI trainings outside of these labs”

AutoScientist builds on the company’s existing data offering, Adaptive Data, which aims to make it easier to build high-quality datasets over time. AutoScientist, meanwhile, is designed to turn those continuously improving datasets into continuously improving AI models. “Our view at Adaption is that the whole stack should be completely adaptable, and should basically optimize on the fly to whatever task you have,” Hooker says.

Of course, that approach will only be as good as the results. In its launch materials, Adaption boasts that AutoScientist has more than doubled win-rates across different models — impressive numbers, but difficult to put into context. Since the system is built to adapt models to specific tasks, conventional benchmarks like SWE-Bench or ARC-AGI aren’t applicable.

Still, Adaption is confident that users will see the difference once they try AutoScientist out — so confident that the lab is making the tool free to use for the first 30 days after its release.

“The same way that code generation unlocked a lot of tasks, this is going to unlock a lot of innovation at the frontier of different fields,” Hooker says.

Neil Batlivala has spent seven years building a healthcare company that most of the tech industry has never heard of and that serves a patient population most of Silicon Valley ignores. But last month, that work put him at the center of something much bigger.

His company, Pair Team, announced on April 30 it had been accepted into ACCESS, a Medicare program — as one of 150 participants chosen by the Centers for Medicare & Medicaid Services to test what AI-driven medical care could look like at federal scale. The program goes live July 5.

“The government is creating swim lanes for AI innovation in traditionally regulated industries,” he told me over a Zoom call a few days later. “The best solution wins, which, in regulated industries like healthcare — that’s not been the case.”

ACCESS — Advancing Chronic Care with Effective, Scalable Solutions — is a 10-year CMS program testing a payment model that rewards health outcomes rather than required activities (like a certain number of check-ins). Participating organizations like Pair Team receive predictable payments for managing qualifying conditions and earn the full amount only when patients meet measurable health goals, like lower blood pressure or reduced pain. It covers diabetes, hypertension, chronic kidney disease, obesity, depression, and anxiety.

That payment structure is the real news.

Traditional Medicare reimburses based on time spent with a clinician. There’s no mechanism to pay for an AI agent that monitors a patient between visits, calls to check in, coordinates a housing referral, or makes sure someone picks up their medication. ACCESS creates that mechanism for the first time.

“It’s a payment model transformation,” Batlivala said. “You just couldn’t do this before.”

The first cohort spans a wide range of participants — AI doctor startups, virtual nutrition therapy providers, connected device companies, and wearable makers like Whoop. Batlivala is skeptical of some of them.

“I’m a big fan of wearables, but for a senior who’s struggling with food insecurity, I don’t know how much Whoop is going to be able to do,” he said, adding of his own company, “We’ve been building toward this for five-plus years now.”

Pair Team launched in 2019 with a specific kind of patient in mind: people managing chronic conditions who were also dealing with unstable housing, too little food, or lack of transportation. About a third of Americans fall somewhere in that category.

The company’s premise was that you can’t improve health outcomes without addressing the full context of someone’s life. It now employs roughly 850 clinical professionals, runs what it describes as the largest community health workforce in California, and, per Batlivala, generates revenue above nine figures. It has raised about $30 million, backed by Kleiner Perkins, Kraft Ventures, and Next Ventures.

The model has peer-reviewed evidence behind it. A study, co-authored by Pair Team researchers and peer-reviewed by the Journal of General Internal Medicine, evaluated Pair Team’s community-integrated model, which blends medical, behavioral, and social care for Medicaid members with high rates of homelessness, serious mental illness, and chronic disease and it showed strong patient engagement and significant reductions in avoidable emergency and inpatient utilization. Batlivala says one in four hospital visits and one in two ER visits don’t happen when a patient is in his company’s care.

But for years, delivering that level of care required human teams, which limited how fast and cheaply it could scale. Then, about nine months ago, Pair Team deployed a voice AI agent called Flora as its primary patient-facing interface. Flora is available 24 hours a day, handles intake, coordinates referrals, and does the check-ins that keep patients engaged between clinical visits.

The first call that shifted his thinking was with a 67-year-old woman living out of her car, managing PTSD and congestive heart failure. She spoke with Flora for over an hour. “It was both incredible and depressing,” Batlivala told me. “Flora was probably the only ‘person’ she’d talked to in weeks about her situation.” Now, hour-long conversations with Flora are routine. “That’s the companionship piece,” he said. “And it turns out that is truly an intervention.”

The architects of ACCESS are themselves former startup operators. The program was designed by Abe Sutton, Director of the CMS Innovation Center, and Jacob Shiff, Chief AI and Technology Officer of the CMS Innovation Center. Sutton was previously a venture capitalist at a healthcare fund called Rubicon Founders. Shiff is a former healthcare founder. Both joined CMS under the Trump administration and their startup backgrounds are reflected in the program’s design: outcome-based payments, direct-to-consumer enrollment, and a deliberate push for competition.

There are real risks. Participants are feeding extraordinarily sensitive patient data — intimate conversations about housing and diseases and mental illness — into a federal infrastructure with a documented history of breaches, including exposed Social Security numbers. For the vulnerable populations ACCESS is designed to serve, that’s not an impractical concern.

There are financial risks, too. The track record of CMS innovation programs is mixed. A 2023 Congressional Budget Office analysis found that the CMS Innovation Center increased federal spending by $5.4 billion during its first decade rather than producing the projected savings. CMS is also paying less per patient per month than many participants anticipated, which means the math only works for organizations that have fully automated most of their patient interactions.

Batlivala’s answer to the reimbursement concern is that it’s a feature, not a bug. “If you want to build a model that truly incentivizes the use of AI, the reimbursement rates have to be low,” he told me. “The economics only work if you’re running a lean, AI-first operation.”

Pair Team says it right now has partnerships in place that give it access to roughly 500,000 potential patients, and that it wants to reach a million within three years.

Healthcare investors have been watching this closely. Digital health funding hit its highest Q1 total since the pandemic this year, with AI companies capturing the bulk of it. But ACCESS has barely registered outside health tech trade press.