Meta’s AI helped write 10 health papers in 4 weeks

Welcome to The Signal from Superpower Daily with Maya and Theo

Imagine going to the hospital for a severe infection

And before you even finish your follow up

appointments an AI has already analyzed your records and helped write a peer reviewed medical paper about

your stay Right Today we are looking at this massive experiment from Meta where they compressed a typical research timeline in a

major way They

churned out 10 health papers in just four

weeks Which is

wild But moving fast in data science is not the same thing as changing real world

clinical

care Exactly Okay Let us

unpack this We really need to look at exactly how Meta achieved

this speed Because I mean traditional medical research is just

notoriously slow

Oh absolutely You have to

secure funding You have to get approval from an ethics board You

recruit patients You gather the data over months or

even years Yeah and then you have to actually clean

that data Right Only after all of that can you begin

the statistical analysis And finally you

write the manuscript That process

usually takes

years

Decades sometimes Exactly But Meta compressed that entire back end process

into a single month They recruited a specialized team of

medical researchers for this These

were

like real doctors right Yes They were specialists in cardiology They had experts in epidemiology They brought in infectious disease specialists

and

internal medicine doctors Okay So

a highly credentialed

group Very much so This

was a human led experiment and

it lasted exactly four weeks The core

engine driving the speed though

was an internal coding scaffold And that scaffold

is

powered by

Meta's AI right Exactly Specifically they

used MuseSpark 1 3 and MuseWonder So that scaffold

is where

the real acceleration happened It is Think of the scaffold

as a highly automated laboratory assistant The system had

access to two main

data

sources Okay What were they First

it

could query public datasets

Second and much more importantly it could query a licensed database containing

five

million electronic

health records Wow Five million That

is a lot of data

But these records were

de identified right Yes Fully

de identified All the names were removed

The dates of birth were

stripped

out The specific addresses

were

gone Good So privacy was maintained Right

But the actual clinical narrative remained intact And the division

of labor here is what really matters The human

doctors did not sit there writing code Because they probably

do not know how to code

anyway Well most of them do not

So the AI generated the initial hypotheses It wrote the

necessary

programming code to search the database It executed that

code against those five

million records It formatted the results It even drafted the

initial manuscripts for the final papers I mean that is

a staggering amount

of automated work It is staggering But we must be very clear

about the boundaries of this automation The humans

retained the ultimate scientific authority here Right The

AI

is not actually playing doctor The AI did not

make the consequential clinical choices The medical researchers decided which

specific

patient groups to study They chose the

specific variables to measure They interpreted

the final statistical results This

is a crucial distinction So the AI acted as an accelerator for the raw

mechanics

of

the research basically Yes It did

not replace scientific judgment You can think of the AI as well a

hyper efficient data entry

intern A

very fast intern Very fast It can crunch

five million records in seconds It can find mathematical

patterns that humans might miss But

it

completely lacks clinical wisdom Right Because it does not understand what

a patient is actually experiencing

It

only sees data points Precisely Here is where it

gets really interesting for me Let us look at the

actual findings from these papers Let us see

how

this works in practice Okay Let us

start

with

the endocarditis study Right So infective endocarditis That

is a really serious infection It affects the inner lining of

the heart and its

valves It is very serious So the researchers used the AI to track patient

outcomes over a one year

period And they

found something notable What was it Seven percent of patients with endocarditis received a new heart

rhythm diagnosis within that following year Okay But

they

needed

a point of comparison right Yes So they

compared this to a control group That group consisted

of patients who had bloodstream infections

but they did not have

endocarditis And what did

that show In that control group only four

percent

received a new heart rhythm diagnosis Wow So seven percent versus

four percent That is

a three percentage point difference Which is

statistically significant especially in a population of that size But we

have to evaluate the fundamental nature of

this

data Right because these are

retrospective health records They are collected

after the fact they are generated through routine clinical care

So they are essentially just billing codes and doctor's notes They

are not generated

by a carefully controlled

prospective clinical trial They

are not And to their credit MEDA understood this

limitation They actually published

independent expert reviews alongside these

papers Oh

that is smart Who reviewed this

cardiology one Michael Furboniak He is

a cardiologist at Penn State He looked at that 3 difference and he immediately

saw a potential flaw in the underlying

reality of the data Yeah he raised a very practical point He

suggested that differences in

cardiac monitoring could entirely explain those numbers Which makes

total sense I mean think about

it from the patient's perspective You have

a

severe heart infection Your doctors know your heart is vulnerable Right So they are

going to monitor your heart very closely for the next year Exactly

You will likely have more electrocardiograms You

will wear heart monitors You will

have more frequent follow up appointments

And that is exactly

the

issue The control group had bloodstream infections Those are serious illnesses sure But they

do not necessarily

prompt the same level of intensive long term

cardiac monitoring Oh I see

Furboniak pointed out a classic statistical problem It is

called detection bias If you look harder for

a

heart rhythm problem you are mathematically more likely to find one Right

And

the AI does

not know that a patient received more monitoring No it does not It only knows

that a billing code for a diagnosis suddenly appeared in

the record The human expert

has to provide that vital clinical context So

the finding is promising I guess It points to an interesting correlation But

it absolutely does

not justify changing how doctors actually treat endocarditis

today No

absolutely not So what

does this all mean for you the listener as a potential patient It means we

have to

be extremely careful about the hype surrounding AI in medicine Very careful I mean

generating 10 papers in four weeks sounds incredible on

a press release

But

it raises a major systemic concern for me The volume

problem Yes If every research team

in the world adopts these tools we risk just flooding

medical

journals We could see a massive tidal wave of uncontextualized correlations We might end

up with thousands of papers showing weak links in retrospective data A doctor trying to

stay updated would have to sift through a mountain of noise This is exactly why the human element

remains the bottleneck The AI can generate the paper in

hours but a human

peer

reviewer still has to read it Right They

have to

actually spot the detection bias They do Let us examine another study from this batch to see how this

plays

out in a different

medical

field Okay The behavioral health one Yes This one

focused on patients in

the emergency department It compared two groups The first group consisted of patients who left the

hospital without being seen by

a Which

happens a lot unfortunately It does And the second group consisted of

patients who stayed and completed

their care The AI crunched the data It found

worse outcomes for the patients

who left

Which I mean makes sense It does They had higher rates

of return visits to the emergency room

They also had higher short term mortality rates On

the surface that finding makes intuitive sense If you leave

a chaotic emergency room

before

receiving treatment

your condition may worsen Right But again the independent review highlighted severe limitations

in

relying purely on this retrospective

data Now

who reviewed this one Rohini Downs She

is a psychiatric pharmacist at Parkland Health She actually praised

the large sample size of the study I mean a human could never manually review that

many emergency room logs in a month Sure

But

she pointed

out critical missing context right She did The problem lies

in what the data does not capture The records did not explain the severity of

the patient's

underlying

illnesses when they arrived Oh

wow That is a huge gap It is They did not specify the exact behavioral health diagnoses

in many cases And most importantly the electronic records could not

explain why the patients actually left Right Because the data just

shows

a sterile timestamp of departure It shows they checked in at

one time and left at another And this is where

the AI just

hits a wall Like did those patients leave because the wait

time

was 10 hours Right Or did they leave because

they suddenly felt better Or did they leave because they were told a specific psychiatric service

was unavailable that day The AI

cannot answer that It only

sees the correlation It sees the departure timestamp and the subsequent

mortality rate Downs argued that this missing context is absolutely essential for improving patient care Did

she not

She did Clinicians and hospital administrators need those

specific details They need to understand the human circumstances behind the departure to actually fix the

systemic

problem

Yeah Like if patients are leaving

because of excessive wait times the hospital needs to hire more triage staff

Exactly

Or if they are leaving because a psychiatric specialist is unavailable on weekends

the hospital needs to reallocate its resources

And a massive data set of 5 million records cannot answer those fundamental operational questions

It cannot

Not if the original doctors and nurses did not type that specific context into the file

We see the exact same problem in the diabetes study from this project too

Yes the remission study

Right

The researchers used the AI to track patients whose diabetes had successfully gone into remission

They wanted to look at how and why that remission eventually ended

And the data showed a clear pattern

The remission ended more often because a patient simply restarted their

treatment medication Right

It

ended less often because of a documented chemical rise in their blood sugar levels That

is a fascinating sequence of events to uncover honestly

but

it presents a classic problem of causation versus correlation Why

did the patients restart their treatment The

retrospective records could not establish the actual human reason Right

Perhaps

their blood sugar was slowly rising but

the doctor just did not order a formal lab test to document it before writing the prescription Or

perhaps the patient was feeling anxious and requested the medication for peace of mind Or

perhaps they moved to a new clinic and

the new doctor simply resumed their old chart automatically We

just do not know The

raw data cannot tell us the intent behind the prescription Because

of this blind spot the

proposed contribution of the paper had to be significantly narrowed The authors could not make a grand clinical claim They

could only suggest that future studies should report those two outcomes separately Right

They need to separate confirmed blood sugar increases from simple administrative treatment restarts This

really underscores the core limitation of this entire experiment It

does If

we connect this to the bigger picture of medical science the

limitation is severe Retrospective

records are excellent for generating hypotheses They

help researchers find curious questions Exactly

They

highlight areas worth pursuing in a formal controlled

randomized

trial But

they are terrible for changing actual medical practice They

cannot by themselves prove that a specific medical intervention works or

prove why a patient behaved a certain way No

they cannot So

we have to watch very closely how major medical research institutions react to this meta experiment Will

universities and hospitals adopt these AI frameworks strictly for early lead generation Or

will we see a dangerous push to use these rapid fire AI drafted papers

to influence actual clinical guidelines That is

the fear We also

need to watch the peer review process itself How will

human reviewers adapt to AI accelerated manuscript submissions They are

already stretched thin Right They

will

be completely overwhelmed if every single lab starts producing 10 papers a month It could

break the system Well that

cleanly

wraps up our look at MEDA's health research sprint Indeed It

is

a profoundly powerful tool But it

requires strict unwavering

human supervision to ensure the data actually reflects reality In

other news we are seeing AI used to generate a very different kind of output We

are moving from AI analyzing existing data to AI generating entirely new functional creations Right

We

are shifting from AI writing research papers to AI writing actual usable

software A

software engineer named Brandon Thomas just released a massive new project It

is a suite of 13 free open source applications He

calls the entire project ArtCraft This

project is aimed directly at the biggest most

established names in commercial software Thomas

has released functional alternatives for very familiar creative and office tasks And

the naming convention makes his targets pretty obvious Very

obvious WordCraft

targets Microsoft Word DeckCraft

targets Microsoft PowerPoint GridCraft

takes on Microsoft Excel And

he also built tools for the creative industry Photocraft

takes direct aim at Adobe Photoshop PDFCraft

targets Adobe's lucrative document management tools This

directly challenges the fundamental commercial model of the modern software industry Think

about the software you use every day at work Oh

the barrier to entry is high An

Adobe Creative Suite subscription costs roughly 780 a year if you pay up front

And if you go month

to month it is about 70 every

single month Which

adds up But ArtCraft is entirely

free It is open

source Anyone can download

it And Brandon Thomas says he built this

entire massive suite using Anthropic's Claude Opus 5

5 model He used the

Rust programming language And he calls his specific development

process vibe coding which is such

a funny term Vibe coding

is a newer highly descriptive term in

the developer community It represents a shift in how humans

interact with

machines Because historically a programmer had to write

every single

line of logic They had to define every variable and

structure every loop With

vibe coding the human developer acts more like an

artistic director They give the AI

high level instructions They describe

the desired outcomes They describe the layout and

the general function and then the AI generates the underlying syntax and code to achieve

that

specific vibe Exactly The human then

reviews the output

refines the prompt and connects the

various pieces together Now Thomas says he is a

15 year industry veteran so he is not a novice

who just discovered ChatGPT No he has roughly a decade of deep experience specifically with

the Rust programming language And Rust is known for being incredibly

fast and memory safe but it is also

notoriously difficult to learn

It is very difficult He used that deep human expertise to guide the AI He knew exactly how to evaluate the code Claude

Opus

was spitting out Right He knew how to stitch it together

into 13 cohesive applications But he also makes a very specific highly calculated legal

claim about this project

The

cleanroom claim Yes He says he used a strict

cleanroom approach He claims he built these applications entirely from publicly available file

format specifications Meaning he says he did not use any of Adobe or Microsoft's proprietary

source code This claim is meant to protect the ArtGraph project from massive

copyright lawsuits But that claim is

entirely unverified I mean let us break down what a cleanroom actually means

historically

Historically a cleanroom involves two completely isolated teams of human programmers

One team reads the public specifications and writes a functional description

Right

And then the second team takes that description and writes the actual code

The second team never sees the original proprietary software

This proves they did not steal anything

Andres Guadalmuz is an intellectual property scholar

He reviewed this situation with ArtCraft and he is highly skeptical

I would be too

Guadalmuz questions the exact ratio of human written code to AI generated code in this

project But more

importantly he questions the fundamental premise of an AI

cleanroom

Exactly How can anyone verify a cleanroom process when a massive AI model is involved

Claude Opus 5 5 was trained on a vast portion of the

internet And that training data likely included millions of lines of

code discussions about proprietary

software and potentially snippets of the very code Thomas is trying to

avoid It is very hard to prove that a massive AI model has not internalized proprietary logic during its initial

training

Yes If the model outputs code that looks too mathematically similar to Adobe's actual source

code Thomas could be in serious legal

trouble The AI might be regurgitating protected intellectual property without him even realizing

it So is this project a genuine threat to Adobe and

Microsoft or is it just a

highly publicized legal target waiting to happen What

is truly fascinating here is the immediate trade off for the end user

CNET actually evaluated the ArtCraft suite

And what did they find

They flagged several practical

everyday hurdles

Yes you save 70 a month

That is appealing

But you have to download and configure open source software entirely yourself

Oh that is rough

You might spend three agonizing days configuring Rust environments and downloading dependencies just to get the Word processor to open on your laptop

That is a harsh reality check for anyone looking for a quick free alternative to Microsoft Office

And feature parity is also absolutely not guaranteed

right

Not at all

WordCraft might let you type text onto a blank page but does it support complex enterprise macros

Does it support real time collaborative editing with 50 co workers like Microsoft

Word does

Probably not Definitely not

right now CNET also raised critical security and long

term maintenance questions Yeah if you use PDFCraft to sign a

sensitive legal document who guarantees that

software is secure Who patches the software when a critical vulnerability is

found next year Thomas points to his extensive industry experience as proof he

will

maintain it But

one single person managing 13 complex desktop applications is a monumental

almost

impossible task The enterprise market will certainly not switch to unproven

vibe

coded software Major corporations gladly pay Microsoft and Adobe thousands of dollars a year for

guaranteed reliability They pay for immediate

customer support They pay for strict

security compliance We will have to observe if this cleanroom legal defense actually holds up to

serious IP scrutiny Because

if Adobe's lawyers find identical code structures they

will strike hard They

absolutely will We also

need to see if these apps are truly maintained long term If

Brandon Thomas gets tired and abandons the project in six months the

early adopters are completely out of luck The

software will slowly break as operating systems update Exactly

The

initial cost of free software is often your own time The

long term cost is the

assumed risk of abandonment

Meanwhile

let us look at the financial reality powering all of this

We just talked about the output of AI tools like Claude Opus

Now let us look at the immense cost of the human talent required to build those foundational models in the first place

We have new

official financial data from OpenAI's branch in the United Kingdom

Right

The 2025 financial accounts for OpenAI UK Limited have

been

officially published and they show

a truly staggering level of compensation The London office

has 110 staff members Together those 110

people shared 255 million pounds in total compensation So if you

divide that simply it breaks down

to an average of 2 318 million pounds per person which is

just insane It is

But let

us carefully impact this massive average We have

to look at the exact structure of that compensation because it is absolutely not all cash You are

not

getting a 2 million pound direct deposit every year The accounts show

that 87 million pounds were paid in traditional salaries And the vast

majority 168 million pounds

was paid in company shares So equity

makes up nearly 2 thirds

of the total compensation package It does This financial filing highlights the

explosive undeniable

growth of their London research hub

The accounts also show the business is

scaling its revenue rapidly Turnover at the UK branch more than

tripled compared to the previous year did it not Yes It hit 293 million pounds The

subsidiary also posted a clear profit

of 20 3 million pounds This illustrates the sheer financial gravity

required to operate at the absolute cutting edge of AI You need massive compounding revenue just

to retain top global AI talent Exactly But the media framing of

this

specific story has been very interesting to watch The Guardian reported that this average

package is 58 times higher than the median UK full time salary But that is a deeply

flawed comparison It makes for a great headline but it is mathematically misleading

Highly misleading You cannot accurately

compare a

total package that is heavily weighted in illiquid private equity to a standard cash salary paid to a normal worker Furthermore the average itself is

highly

skewed internally right Very much so The

Guardian

correctly noted that

the absolute biggest packages went specifically to the specialized AI researchers These are the rare engineers

who are actually building and optimizing the core transformer models But the report did not

quantify the actual gap between the highest and lowest paid staff within that office The London office is not

just a pure research lab Right It operates as a

full

business unit It includes government affairs staff

It includes public relations teams

It includes enterprise sales and

marketing teams who sell corporate licenses I highly doubt the junior

marketing managers in that specific office are pulling down two million pounds a year Yeah The overall average is clearly

being

pulled up by a select few senior researchers receiving massive multi million pound equity grants And we also have to

remember the strict nature of that equity Paper wealth in private shares

is highly dependent on private market valuations holding perfectly steady That is the critical often

ignored caveat If a

top researcher holds

10 million pounds in open AI shares they only realize that wealth

if they can actually sell them Right The company is private

You

cannot just log onto

a brokerage app and sell your shares on a Tuesday You have to wait for an official tender offer And if the broader AI

market cools down or if a major competitor

suddenly overtakes them the perceived value of those