{{ 'now' | timezone: 'America/New_York' | date: '%b %d, %Y' }}
|
|
|
|
Welcome to The Hidden Layer. I’m Ian Krietzberg.
Today, news and notes on
Uber’s latest attempt to automate its fleet, Microsoft’s support of Anthropic’s burgeoning legal battle against the Pentagon, and an update on a specially designed A.I. therapist that proves that… well, the machines may be better than the actual headshrinkers after all. Sort of.
Also discussed in this issue: Ross Harper, Tobias Hauser, Zachary Ginder, Arthur Evans, Waymo, Zoox, Uber, Andy
Macdonald, Elon Musk, Donald Trump, Microsoft, and more…
|
Two Things You Should
Know…
|
- Uber’s
Zoox deal: Zoox, Amazon’s boxy-looking robotaxi project, announced a partnership with Uber yesterday. It’s another “multiyear agreement” that will make a select fleet of Zoox robotaxis available through the Uber app in Las Vegas (later this year) and Los Angeles (next year). Zoox, which has accelerated its expansion across Vegas and San Francisco over the past year or so, is a rare
competitor to Waymo in the robotaxi race. But it’s also far smaller: Its vehicles have covered more than a million autonomous miles, while Waymo has done over 100 million.
This partnership marks the latest example of Uber’s slow push to return to robotaxis following the death of a pedestrian in 2018 that forced the company to shelve its self-driving ambitions. In July, it unveiled a three-way partnership with Lucid, an electric vehicle maker, and Nuro, a self-driving tech company, to build and deploy some 20,000 robotaxis across its fleet. For what it’s worth, though, a recent report from Consumer Watchdog found that Nuro covered only about
150,000 miles in California in 2025. Further, human drivers had to disengage the system 244 times, meaning the Nuro could cover only about 700 miles between disengagements—instances where humans have to take over for whatever reason. Waymo, meanwhile, covered 3.3 million miles in California in 2025, racking up 174 disengagements (one every 19,000 miles), and Zoox covered 1.2 million miles with only 20 disengagements (one every 60,000 miles). Nuro didn’t respond to a request for comment. - Microsoft support: The docket for Anthropic v. U.S. Department of War is quickly filling up. Among the documents posted over the past few days are a number of amicus briefs filed in support of Anthropic’s position. One came from a cadre of former military officials, another from
OpenAI and Google employees, and yet another from a coalition of advocacy organizations. There’s also a
declaration of support from Microsoft.
Microsoft, a partner-investor in Anthropic, asked the court to temporarily prevent the implementation of the D.O.W.’s “supply chain risk” designation, at least until the case has been decided. “This is not the time to put at risk the very A.I. ecosystem that the Administration has helped to
champion,” the brief concludes.
|
|
|
|
A MESSAGE FROM OUR SPONSOR
|
How can healthcare organizations capture the next wave of AI value?
Many healthcare systems and hospitals have deployed AI to address specific clinical and operational challenges. But without enterprise-wide integration, these tools can create fragmentation and limit impact. Real transformation will require a more connected approach. In The coming evolution of healthcare AI toward a modular architecture, McKinsey
explores how shifting to a modular, interoperable AI architecture can unlock scale, improve data governance, and enable holistic innovation. The article outlines the strategic moves leaders must make now to build a more resilient, AI-enabled healthcare system. Read the full article
here
|
|
|
|
Quote of the Week:
Software Necromancy
|
“It’s entirely possible that I don’t need SaaS-based platforms anymore, if I can use the L.L.M.s.
But I’ll be spending [money] in a different way—I don’t see this in any way as a cost-saver going forward. At least in legal, there are some good applications here, but I think there’s a human element that has to be involved. A.I. is a Formula One car, but it doesn’t mean that everyone’s a race car driver.” —Andy Macdonald, the C.E.O. of legal tech and advisory firm Consilio, to me the other day regarding the efficacy of his recent A.I. experiments.
And now for the
main event…
|
|
|
|
Limbic, a British company backed by Khosla Ventures, is working toward a future where people
use A.I. to supplement their shrinks rather than rely on the tech altogether—one of the darker trends of the artificial intelligence era so far.
|
|
|
|
A few months back, OpenAI, the Musk antagonist and Trump administration
ally, released some data that contained stirring and deeply unsettling insights into the fragility of the increasingly artificial human condition. In response to a wrongful-death lawsuit filed against the company, OpenAI announced in October that 0.15 percent of its weekly users have had “conversations that include explicit indicators of potential
suicidal planning or intent.” Some 0.07 percent of users, meanwhile, showed signs of “serious mental health concerns, such as psychosis and mania.” Those numbers add up: That’s about 1.2 million and 560,000 people, respectively, each week, based on ChatGPT’s contemporaneous base of 800 million weekly active users.
Like it or not, A.I. has become inextricable from mental healthcare. According to reports
published in 2024 and 2025 by Harvard Business Review, people use generative A.I. for therapy and companionship more than anything else. (A more recent HBR report on Gen Z, interestingly enough, found companionship to
be a significantly less popular use case.) Nevertheless, there are enough users to form a subreddit of 23,000 people called TherapyGPT—a community full of stories of users going so far as to abandon their human therapists entirely in favor of artificial intervention.
The appeal isn’t hard to divine. An L.L.M. is always there:
cheap, readily available, nonjudgmental, and it doesn’t require insurance. And yet, obviously, therapy bots are neither trained nor licensed; they remain unregulated and, consequently, lacking in the kind of rigorous evidence that would suggest they’re either effective or safe for therapy. Last year, the American Psychological Association called the use of generic chatbots for mental health support a “dangerous trend.” APA C.E.O. Arthur Evans said at the time that “APA envisions a future where A.I. tools play a meaningful role in addressing the nation’s mental health crisis, but these tools must be grounded in psychological science.”
It’s an effort that Limbic, a British company operating at the frontier of mental health A.I., has been undertaking since its foundation. And in a paper
published Thursday morning in Nature Medicine, Limbic provided the results of a double-blind, randomized controlled trial that found, among other things, that A.I. actually outperformed human therapists at therapy.
|
Caveats, of course, abound. The study was specific to text-based “psychotherapy interactions”; the
outperformance was across “key clinical competencies required for delivering high-quality cognitive behavioral therapy,” rather than therapy sessions and formats writ large; the study focused on single sessions rather than multiple sessions; and participants were told that they were evaluating therapy agents—rather than receiving a full-on mental healthcare intervention—“and were asked to construct a realistic scenario to discuss.” Still, a consortium of 22 clinicians blindly reviewed
transcripts resulting from both A.I. and human sessions (the study featured only six human therapists as a comparator), and the result was that “74.3 percent of all cognitive layer sessions surpassed the top 10 percent of scores given to the human therapists.”
|
|
|
|
A MESSAGE FROM OUR SPONSOR
|
How can healthcare organizations capture the next wave of AI value?
Many healthcare systems and hospitals have deployed AI to address specific clinical and operational challenges. But without enterprise-wide integration, these tools can create fragmentation and limit impact. Real transformation will require a more connected approach. In The coming evolution of healthcare AI toward a modular architecture, McKinsey
explores how shifting to a modular, interoperable AI architecture can unlock scale, improve data governance, and enable holistic innovation. The article outlines the strategic moves leaders must make now to build a more resilient, AI-enabled healthcare system. Read the full article
here
|
|
|
|
Limbic was intent on clinically validating one of its two core products: a clinical reasoning harness
designed to make standard L.L.M.s safe and effective for therapeutic interventions. This model-agnostic architecture leverages custom-built machine learning models trained on clinical mental health information to closely control, guide, and review a given L.L.M.’s responses. Today, Limbic works directly with clinicians; one of its offerings involves a (similarly
validated) A.I. system to aid in patient intake and assessment. The other, called Limbic Care, is an A.I. system designed to integrate into existing care plans.
That product, which has
already been deployed, was the subject of the study, which also found that the A.I. agents—when imbued with Limbic’s specialized clinical layer—“significantly outperformed stand-alone L.L.M.s.” The clinicians preferred those specialized agents over the stand-alone L.L.M.s 82 percent of the time. The findings, according to the paper, “provide early empirical evidence that L.L.M.s enhanced by specialist A.I. systems show capabilities required for interpersonal psychotherapeutic interactions.” In a
real-world observational analysis of 9,000 users across the U.S. and U.K., the study found that users with higher exposure to the clinical layer reported a recovery rate 20 percentage points higher than those with lower exposure.
Professor Tobias Hauser, a computational psychiatrist and senior author of the study, offered various caveats, but noted that the big takeaway was that “standard models are not made for mental health support and are much more likely to fail. To
provide good mental well-being support, one needs to divide the excellent language abilities of the L.L.M. from the reasoning and clinical expertise that is done by other, complementary models. Simply put, someone who can talk well and fluently might not be the best at providing therapy.”
Dr. Ross Harper, Limbic’s C.E.O., told me that the company decided from the beginning not to attempt to solve the language modeling problem. Instead, he was focused on addressing the
“thorny question of how you can unlock this technology within clinical settings in a way that is not only clinically validated and performant, but also explainable and auditable.” The results, he said, give Limbic and the field a “clear path to how we can use this new technology within a mental healthcare setting.” A.I., he added, will be adopted in healthcare “at the speed of trust.”
Limbic remains focused on working with clinicians, but Harper didn’t rule out pursuing a
direct-to-consumer offering eventually. Even then, though, he doesn’t expect therapists to go away. “There is an intractable supply-demand imbalance in psychological therapies. We have graduated existing human clinicians into more senior roles, where they will spend every minute of their valuable time working on the things that only they can do, like handling complexity,” he told me. “It’s about deploying our scarce resources in a way that actually scales the industry. That’s how I see things
playing out in practice. I think it’s sensational to talk about substitution or replacement.” He added that the company, which was backed by Khosla Ventures in a $14 million round in 2024, experienced an “inflection point” in growth last year, though
he declined to elaborate.
Perhaps none of this is perfect or all that shocking: Humans require support, technology has always been a means through which people express their id, and therapy is one of many fields where A.I.’s great potential is still mostly unrealized, and far from straightforward. “I don’t think this is our golden ticket to resolving our mental health crisis or the shortage of therapists out there,” Dr. Zachary Ginder, a California-licensed psychologist
and a practicing therapist, told me. “But that said, it’s promising.” Dr. Sahib Khalsa, a psychiatrist and neuroscientist at UCLA, agreed, telling me that the study marks “a step in the right direction. I’m optimistic about the possibilities, but I’m also skeptical until there’s a body of evidence.”
Ginder added that he could see a lot of potential value in something like Limbic—say, where it acts as a bridge between sessions, an option while people are stuck on
a waitlist, or even a decent, frictionless option for “subclinical” people. “As far as I’m concerned, if we’re looking at it as support for therapists, then it’s not like one has to be better than the other.” But as a stand-alone option, he said, the risks are significantly higher.
As machines become both more normalized and more complementary to human specialists, in this field as in others, Ginder noted that some of the most impactful aspects of therapy remain rooted in a kind of
humanity that the machines of today, at least, can’t really automate. “We have a thing called co-regulation in therapy: voice pacing, a person’s presence, just attunement with another person,” Ginder said. “And we are biologically hardwired for interaction and relationship. I often find that healing happens in relationship—so that’s missing from this process. I think that leads to the conclusion that this won’t be for everyone, but for a certain population I think it could be very
beneficial.”
|
As Congress remains stalled out on frameworks and attempts to actually regulate A.I., China continues to do
the opposite. And increasingly, the Chinese government has turned its legislative attention to A.I. companions. [Carnegie Endowment]
This piece offers a good look at the realities of A.I. delusion and the many different forms it takes.
[404 Media]
Somewhat relatedly, in a recent study undertaken by CNN and the Center for Countering Digital Hate, chatbots helped teens (hypothetically) plan shootings. [CNN]
Grammarly
has already been served with a class-action lawsuit over its recent rollout (since rolled back) of an A.I. tool that claimed to offer writing insights from real-life authors, journalists, and academics. In keeping with the tradition of the A.I. industry, they didn’t ask for consent first. [Wired]
|
That’s all for today. I’ll see you next week. Ian
|
|
|
|
Join Emmy Award-winning journalist Peter Hamby, along with the team of expert journalists at Puck, as they let you in on the
conversations insiders are having across the four corners of power in America: Wall Street, Washington, Silicon Valley, and Hollywood. Presented in partnership with Audacy, new episodes publish daily, Monday through Friday.
|
|
|
|
Puck’s daily political newsletter from Washington on what’s really happening in this town, from the White House to the Pentagon
to Capitol Hill, K Street, and the campaign trail.
|
|
|
|
Need help? Review our
FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.
You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.
|
Puck is published by Heat Media LLC. 107 Greenwich St., New York, NY 10006
|
|
|
|
|