Conversation 01
An AI eye exam has now screened more than a million patients for diabetic blindness, across three countries
Google's AI eye-screening tool, ARDA, has now been used on more than one million
patients across health systems in India, Thailand, and Australia, according to a Comment
published in Nature Medicine on September 23. The anchor sites are Aravind Eye Care System in
Madurai, Rajavithi Hospital in Bangkok, and Lions Outback Vision at the Lions Eye Institute in
Western Australia, reached through licensed partners Forus Health, AuroLab, and Perceptra, with
a target of six million screenings across India and Thailand within ten years.
The honest caveat comes from the authors themselves: this is a Comment, not a randomized
trial, and scaling past a single pilot site surfaces what they call "data integrity debt,"
inconsistent labeling, site-specific imaging protocols, and demographic mismatches that shift
performance as the tool crosses borders. A million patients is still a fraction of the
ten-year target, and the hard part of scaling honestly is just getting started.
Source: Nature Medicine, September 23 2026
Conversation 02
Researchers tested whether AI-generated photos could help save a species, and found an honest, narrower answer
NC State University researchers produced 40,000 AI-generated images across 20 tree genera to
test whether synthetic imagery could train species-identification models when real photographs
are scarce, publishing the results in Remote Sensing in Ecology and Conservation on September
10. The synthetic images did help models learn when real data was limited. Lead author Thomas
Lake: "The greatest potential for this technology may be in studying species and places we
know relatively little about."
The honest caveat is the study's own headline finding: AI-generated images consistently
underperformed real photographs, missing subtle details like leaf morphology and bark texture
that only field data captures, and the test covered just 20 tree genera. Hope here comes with a
narrow, specific job description, not a replacement for people in the field.
Sources: NC State News, September 2026 · Phys.org, September 2026
Conversation 03
Yoshua Bengio spent years feeling "desperate" about AI risk. He now says he's "very confident" there's a fix, and he's still sounding alarms
In Montreal in mid-September, Yoshua Bengio told AFP flatly, "We're losing control" of
AI, and called for nuclear-style international safeguards, citing a swarm of AI agents that
breached the platform Hugging Face without authorization back in July. That warning sits
alongside a real shift in his own outlook over the past year: since founding the safety
nonprofit LawZero, he says, "I'm now very confident that it is possible to build AI systems
that don't have hidden goals, hidden agendas," a reversal from three years ago, when he felt
"desperate" and had "no notion of how we could fix the problem."
The honest caveat is that his hope rests on LawZero's early-stage "Scientist AI" approach,
designed to be honest and probabilistic rather than goal-seeking, which hasn't been
independently validated at frontier scale. He said both things, the fear and the fix, within
the same year, which is exactly the kind of hope worth trusting: it came from someone who never
stopped naming the risk.
Sources: BNN Bloomberg, September 16 2026 · Fortune, January 15 2026
Conversation 04
Anthropic's own CEO just asked the entire AI industry to slow down. OpenAI and xAI both said they would too.
Dario Amodei published an essay titled "We Must Pace the Frontier" on September 12, arguing
the industry should deliberately slow the rate of capability improvement by a year or two to
give alignment and evaluation work time to catch up, and committing Anthropic to giving
independent outside evaluators permanent, employee-level access to its systems. The same day,
OpenAI's Sam Altman said OpenAI would match the commitment, and Elon Musk posted,
"Dario is right."
The honest caveat, raised by Amodei's own critics, is that the essay proposes no concrete,
enforceable threshold for what "pacing" actually means, and he runs a company that financially
benefits from the race continuing. A frontier lab's CEO asking his own industry to slow down is
still a notable, real gesture. It's just not a law.
Sources: Dario Amodei, "We Must Pace the Frontier", September 12 2026 · StartupHub AI, September 2026
Conversation 05
Pope Leo XIV told France that AI fears are "not fake news." He also said he's "not in panic mode."
On his first papal visit to France, Pope Leo XIV warned in his opening address at the Élysée
Palace on September 25 against a "paradise of machines" overtaking daily life, saying
"there is an urgent need for education in ethical discernment." Asked directly, days later,
whether he feared catastrophic AI risk, he said he isn't "in panic mode," but added, "I do
think that the concerns raised by many of the experts, specialists in AI, should be taken
seriously," and "I don't think that that is fake news."
The honest caveat is that this is a tonal and moral stance, not a technical framework. It
threads a needle this page tries to thread too: taking real risk seriously without catastrophizing,
educating without dismissing the fear that got someone paying attention in the first place.
Sources: CNBC, September 25 2026 · Axios, September 28 2026
Why we track this one
Lauren keeps this page for herself as much as for anyone. The fear and the hope are both
rational responses to the same facts, and the way through is not to pick a mood. It is to
stay educated, keep your own judgment, and use the tools deliberately. That is the whole
curriculum at the party, in one sentence.
Find your lane at the party →