Loading page…
Loading page…
The Signal / Superpower Daily
Meta’s pocket AI device has a reported price, not a confirmed one. Anthropic’s strong stock-market opening, meanwhile, gave way to an intraday retreat.
Superpower Daily: The Signal
Episode guide
Meta’s pocket AI device has a reported price, not a confirmed one. Anthropic’s strong stock-market opening, meanwhile, gave way to an intraday retreat.
Full transcript
Select any speaker or timestamp to continue listening from that point.
Imagine you are the mayor of a major coastal city
right
You're sitting in an emergency operations center just staring at a screen
Which is pretty much the most stressful place you could possibly be And on
that
screen is the output of a brand new multi billion dollar artificial intelligence model I mean
this supercomputer has systematically beaten every human meteorologist on Earth for the last four years Right Statistically
it's practically flawless And
right
now it is telling you that the incoming hurricane is going to peak as a
manageable Category 2 Okay So you make
the
call you hold off on
the mandatory evacuations you tell the public to just shelter in place But what if I told you
that by relying on that mathematically perfect AI you just sentence your city to
a Category 5 disaster It is a terrifying scenario But
the crazy part is it's
no longer just a hypothetical one No it's not I mean we've
spent years thinking
about artificial intelligence as this well basically a software experiment Right A
novelty that writes your
emails
generates some images maybe occasionally hallucinates a weird answer
Yeah or a really smart chatbot
Exactly But the parameters of that experiment
have
fundamentally changed The digital world is violently colliding with
physical reality right now Which is exactly why we're here today
Welcome to today's Deep Dive We are
looking at a landscape that
is just shifting way faster than the daily headlines can track And we're bringing you along for it We really are We're synthesizing a whole stack
of recent intelligence
from superpower daily to uncover the actual mechanics of what is happening under the hood of the tech industry Right Because we are moving past the era
of chatbots We are looking at physical hardware We're looking
at autonomous agents actively taking control
of our screens and frankly the massive geopolitical engines that are
funding all
of it The narrative arc across these sources is incredibly
clear if you actually look at the underlying data Let's do it Yeah Where do we start
We have to
start
at the macro
level We're going to trace the unprecedented flow
of billions of dollars through the stock market and political spheres Then we're going to zoom into the
software layer right examining how AI is shifting from
a passive tool to an active agent An agent that does things for you
Yes And from there we'll look at
how companies are trying to force these agents into physical hardware devices that you carry in your pocket And then finally we'll test all of
this technology against one of the most chaotic forces on Earth which is atmospheric weather So if the
physical build out of
this
technology is becoming impossibly expensive which it is these AI
companies desperately need
guaranteed massive revenue
streams right and bottomless pools of capital just to
survive
They absolutely do We can't talk about what
an AI agent
is doing on your laptop without first understanding the financial war chest that built it And that brings us to Anthropic's historic debut
on the NASDAQ on November 19th 2026 Right Anthropic the company behind the Claude family
of
models
they executed what can really only be described
as a tectonic event in the public market It was massive It was Let's look at
the mechanics of
the initial public offering the IPO They went into the book building
process planning
to sell 825 million shares at a price target of somewhere between 70 and 78 dollars a share which is already a huge offering But institutional
demand So we're talking pension funds
massive
sovereign wealth vehicles
mutual funds It was so aggressive that they expanded
the offering
They didn't just bump the price right They expanded
the actual float to 980 million shares
pricing it at 83 dollars And to
put that in
perspective for you listening
that
resulted in
an 81 3 billion dollar offering That completely
eclipses SpaceX to become the largest U
S IPO in history It's staggering And when the bell rang on
the
NASDAQ the stock immediately gapped up It opened at 111 dollars and 50
cents That's a 34 percent jump right
out of the gate which briefly implied a
company valuation above two trillion dollars Two trillion Though it
did retreat
below 100 dollars in intraday trading which is pretty
standard for an IPO of this magnitude as early
institutional allocators take their immediate profits Right They
cash out a bit Yeah But
the
daily price action is almost a distraction from the underlying structural reality here This offering
combined with
their previous private funding rounds and debt structuring it gives
Anthropic a 220 billion dollar cash reserve Let
me just repeat that 220
billion in cash Yes
And
that is strictly the cash on the balance sheet That
sits alongside these massive long term computing infrastructure commitments that they've locked in
Oh right
The cloud providers
Exactly
110 billion dollars in compute from Amazon and another
121 3 billion from Google The
scale of capital here is just no longer comparable to traditional
software startups They
are operating with the balance sheets of nation states It
really is nation state level
And the CEO
Dario Amadei
came out and immediately addressed this
He claimed that having a financial fortress of this magnitude is actually going to reinforce their safety commitments
Reinforce them yeah
The argument he made is basically if they have 220 billion in the bank
they don't have to bend to the typical quarterly pressures of Wall Street
They can pace their training
delay deployments if a model is unsafe
and keep the systems strictly under human control without worrying about going bankrupt
Right
That is the narrative of the benevolent monopoly
The theory is that financial independence breeds responsible development
But I look at that premise and I just have to push back
I mean does having 220 billion in index fund and mutual fund money really
insulate you from pressure
That's a great question
Because it feels like it makes you entirely beholden to the most ruthless return on investment engines on the planet
When you are valued at 2 trillion
your shareholders aren't going to accept a delayed product launch because of some theoretical safety concern
Oh absolutely not
The fiduciary duty to maintain that valuation seems like it would crush any internal safety culture the second those two things conflict
You are identifying the core friction of corporate governance in the AI space right now
Amodei is trying to argue that the capital is a moat protecting their research
But the market reality
as the superpower daily editorial analysis points out
is that the capital is an anchor
An anchor
Right
Yeah Because when you take public money at a 2 trillion implied valuation you
are promising astronomical future cash flows You
cannot generate those cash flows if you're constantly hitting the brakes on deployment You
have to ship products Exactly
The
real test of Amodei's pledge isn't going to happen on a day when the stock is up 34 It's going
to happen on a day when a competitor releases a faster model Anthropic stock
plummets 15 and the board demands an immediate potentially unsafe
deployment just to stop the bleeding And speaking
of the market bleeding this is
where the macro story gets genuinely bizarre So Anthropic
has this record shattering standing ovation IPO But
if you zoom out and look at the broader U S market on that exact same day
it tumbled
It really did
The S P 500 fell 2 03 It was the fifth straight
session drop And it wasn't just retail or you know
legacy sectors pulling back The most fascinating data point from the market reaction was the performance of
the AI infrastructure
suppliers
The hardware guys Right Nvidia dropped for
a sixth consecutive session completely ignoring their
own upbeat
revenue
forecasts
Micron Broadcom Corweave Nebious The entire physical supply chain of artificial intelligence sold off while the
software
builder went parabolic
OK let's unpack this It's like watching a modern
day gold rush right Anthropic is standing on a stage announcing they found
a massive vein of gold The crowd goes wild They're
valued at two
trillion
dollars Best day ever Yeah But the actual companies selling the pickaxes
the wheelbarrows the mining equipment they're
suddenly going out of business Why is there such a violent disconnect between the software
hype and the hardware reality Because the AI rally is violently hitting the physical thermodynamic
constraints of
the real world Thermodynamic constraints Yes You have portfolio managers explicitly stating that their expectations for data
center capital spending are easing And they're easing because building a data center is no longer a matter of
just you know ordering more server racks Let's break down those physical constraints because I think people hear data center and they picture like a warehouse
with
some really heavy duty air
conditioning Right That was the old model But the current generation of AI models requires computing density
that
makes
traditional air conditioning completely obsolete When you rack thousands of
the latest AI accelerators together they generate so much heat You have to
use direct
to chip liquid cooling Oh wow So you're
literally piping liquid over the electronics You are literally piping
specialized coolant directly over the silicone You cannot easily retrofit
a five year old data center for that You have
to build custom infrastructure from the ground
up And
that requires massive amounts of
copper
specialized piping and entirely
new supply chains Exactly And even more importantly it requires electricity at a
scale that local power grids simply cannot provide We are seeing data center projects stalled for three to four years simply waiting in the interconnection queue for
a utility company to build a new
substation Three to four years
just for
power Just for the
power hookup The software like the Anthropic model can be updated
and scaled globally in a matter of weeks But the physical hardware required to
run it
is bogged down in
zoning laws transformer shortages and grid capacity limits And then you
add the macroeconomic pressure on top of that The markets are pricing in a 94 chance of
a quarter point Federal Reserve rate increase in
December Which is the
killing
blow for speculative infrastructure Because it's all debt financed
Exactly When you are building a 5
billion data center you
are financing that construction through debt When interest
rates
rise the cost of servicing that debt just explodes Right So the market
is looking at Anthropic and cheering for the software margins but they're looking at the hardware builders and realizing that the physical build
out is getting squeezed by interest rates and supply chain
realities The return on invested capital for physical infrastructure is collapsing Okay so
if the commercial market is groaning under the weight of these interest rates and physical bottlenecks who has
the capital
and the
authority to keep this engine
running The government The answer unequivocally is the federal government Which brings us
to the political and financial disclosures highlighted in the reports And I want to be extremely precise here because we are
looking at this
strictly through the lens of SEC
filings government contract databases and financial disclosures reviewed by CNBC
We are examining the mechanics of capital and
state power
Yeah We are not taking political sides
here Absolutely not It's just impartial reporting on the data So the context here is a September disclosure
covering 751 trades made by former President Trump's investment accounts And CNBC's review basically mapped these stock purchases
directly against major federal government actions involving those exact companies
And we need to state the explicit
caveats CNBC included
These
accounts
are run by independent managers Very important detail Right There
is zero evidence that Trump directed these specific trades nor is there evidence that the independent
managers had non public advanced knowledge of the government contracts And the filings
use broad dollar ranges like 1 million to 5 million
So
calculating the exact profit or volume is literally impossible Yeah And the importance
of the data isn't about alleging some specific
backroom deal
The importance of the data is the timeline How so It perfectly illustrates how deeply entrenched the
valuations of
these tech
companies are with the federal government Take Palantir for example The disclosures show the accounts
purchased
shares of Palantir on September 2nd and September 9th
Okay And then weeks later the federal contracts
start dropping The Department of Housing and Urban Development HUD
awards Palantir 10 4 million to oversee grants and detect fraud Right Then later in the month on September
29th the Department of Homeland Security awards them another 32 1 million contract And on that exact same day Palantir CEO Alex Karp is at
the
White House meeting with Trump
and other executives to discuss AI infrastructure Exactly But looking at this data
I have to be a little
skeptical about the
direct cause and effect I mean Palantir is a massive multibillion dollar enterprise A
10 4 million HUD contract is literally a rounding error for their
quarterly revenue Is the broader stock market really moving
on a 10
million grant Or is there a different signal being broadcast
here You nailed it The market doesn't care about the 10 million in isolation
The market cares
about
regulatory capture and the expansion of the defense industrial base The broader trend When a company like Palantir
which has traditionally built tools for the CIA and the Department of
Defense starts winning contracts from civilian agencies like
HUD to detect
grant fraud it signals a massive total addressable market expansion Oh I see The government is
signaling that AI is no longer just for drone targeting or signals intelligence It
is going to be the foundational operating system for the entire
federal bureaucracy It's the normalization of the technology across the
whole
state apparatus And Palantir isn't the only example in these
filings right No The disclosure also tracks purchases of SpaceX debt and equity Those purchases bracketed a September 18 announcement where NASA
expanded its contract with
SpaceX
by 946 million for three additional crew missions Almost a billion dollars Yeah Now NASA had actually signaled this expansion back
in May in their public procurement documents So the
information was technically public though
the
final execution happened right after the trades And then you have NVIDIA Right The accounts showed four specific September purchases of NVIDIA and CNBC
noted
at
least 108
NVIDIA
transactions since Trump returned to office in 2025 Wow Yeah 108 transactions And this high volume
trading occurred while NVIDIA executives maintained incredibly
close access to the administration including state dinners and
dedicated infrastructure meetings When you
pull all of these macro threads together Anthropic's 220 billion war chest the stalling physical infrastructure due to interest
rates and the tens of millions in
federal contracts flowing to AI firms
alongside massive political
investments it fundamentally redefines what AI is
It really does We aren't talking about
consumer software anymore We are talking about critical state infrastructure The survival of these companies is entirely dependent on securing favorable regulatory
environments
and bottomless government contracts to subsidize that physical build out Exactly The sheer gravity of the
money involved means these systems are being hardwired into
the economy which forces us to
ask
the next logical question What exactly requires
this much capital Right What are
they actually doing with it What are they building that demands
100 billion in compute
from Amazon and a 2 trillion valuation They are building
agents We're
moving away from the paradigm of a chatbot For the last few years we've
treated
AI like an
incredibly smart encyclopedia right
A text generator
Yeah Ask a question It generates text You read it But the human is the one executing the
physical action based on the AI's advice The companies we just talked about are using their billions
to shift the architecture from generative tools to autonomous
agents that take control of
your hardware This is the agentic
shift And it is no longer theoretical It is actively
rolling out to consumer devices right now On November 19th Google updated
the Gemini Mac app in
the U S to include background task automation And this
fundamentally changes the user relationship with the computer Gemini isn't just giving you instructions on how to
resize a photo or find a
document anymore No it's doing it
for you
It is actually handling the clicking the
navigating and the typing
across your local files
your third party desktop apps and
your web browser To understand how massive
this is look at the specific workflows Google
demonstrated
Let's say you run a small logistics business okay You have a local folder on your Mac desktop
filled with dozens of
invoices and shipping manifests
A nightmare of paperwork Right
In the old
paradigm
you would
open each PDF copy
the weights dimensions and addresses
open a web browser go to FedEx or UPS and
paste the
data into
a multi box shipping quote
form Just mind numbing repetitive friction With the new
Gemini
update you instruct the agent to handle the shipping quotes The AI uses accessibility
APIs within macOS
to visually process the local folder Visually process
it Wow It opens
the PDFs extracts the unstructured data launches the web browser navigates the DOM the document object model of the shipping website and actively fills in
the They also showed it pulling live metrics and screenshots from multiple regional sales dashboards and automatically formatting a Google slide
presentation Yeah that was
impressive Or taking a local folder of clothing
photos independently researching the market value of similar items
online and prepping draft listings for a resale
site It is simulating human labor at the operating system
level
However there is a massive security and user experience constraint designed into this initial
rollout I was going to say this sounds a little
dangerous It could
be Which is why Gemini does not take over your actual mouse
cursor It doesn't hijack your screen and pervade you from
working It executes all of these actions inside an isolated background window that provides a live chat preview of what it is
doing Ah
okay A
digital
sandbox Precisely You
the user have to explicitly grant the agent permission to access
specific apps You can watch its progress in the
background window You can pause its execution at
any moment
And critically you must manually approve key actions like finalizing a purchase or sending an email before they are
actually executed Okay let's stop right there because this constraint
fascinates me Google is hyping this as the future
of
automation
right Right But if I have to sit there staring at a little background window to babysit my
AI intern constantly verifying that it didn't pull the wrong address from the PDF or accidentally approve a 5
000 freight charge hasn't Google completely defeated the purpose of automation The cognitive load of supervising an unreliable agent
is often higher than just doing the task
yourself You are hitting on the exact trust barrier that is terrifying every major tech company right
now They desperately want to sell the enterprise market on the dream of zero labor
automation But they are paralyzed by the legal and reputational liability of an AI going rogue on a local
machine Because if it goes
wrong Think about the risk
surface If an AI hallucinates a command and accidentally attaches your unencrypted tax returns to a mass
email the damage is
catastrophic Oh
absolutely The isolated background window is a
compromise They are giving the model the mechanical capability to navigate an
operating system system but they're keeping the human locked in as the final fail safe to absorb
the liability It's the illusion
of
automation really But it isn't just Google trying to crack
this open OpenAI is approaching the agentic shift from the
audio layer They just released GPT Live 1 1 to developers
via their API Yeah
the voice model This is designed for applications that can listen
and speak simultaneously what they call a
full duplex conversation A full duplex architecture solves the latency and interruption problems that made older voice bots
so incredibly frustrating If the GPT Live agent is speaking to you and you interrupt it
to
say wait go back to the
previous point the model physically stops its
audio generation It
just stops
It stops ingests your interruption in real time and adjusts
its response OpenAI claims version 1 1 features significantly more accurate transcriptions and steadier accents
across various languages Although we have to note they haven't published quantified peer reviewed benchmarks for
those specific claims yet Still
the capability is impressive but the economics of how they are charging developers for this API reveals exactly how
these systems are built
behind the scenes That's the
most
interesting part Right OpenAI charges developers 5 cents per active minute when the voice
model is engaged but they introduced
a new rate 2
cents per minute And this lower rate only applies to qualifying WebRTC sessions while the model's inference
is paused It isn't just for when the user is
quietly thinking It is a specific
technical state This is where we need to dive into the mechanics of modern
AI
architecture That pricing structure exposes something called delegation
architecture Let's unpack delegation architecture Because I think most people assume that when they talk to an advanced
AI they are talking to one unified omnipotent neural network that is doing the
listening the thinking and the speaking all at
once That is exactly the illusion the companies want to
maintain But
technically it would be prohibitively expensive
and agonizingly
slow Instead developers separate the interface from the reasoning
engine
OK how does that
work The GPT Live model is essentially just the ears and the
mouth Its entire job is to process the incoming audio via
WebRTC which is a protocol that streams data in ultra low
latency chunks and to synthesize natural sounding speech
in return So
it's fast but it's
computationally
shallow Exactly It handles
the conversation But let's say the user asks the voice agent to search a massive company database for a specific client record or execute a complex multi
step math problem The GPT Live model doesn't have the architecture to
do that efficiently
So what happens The developer programs the system to pause the
voice model's inference its
actual token generation The voice model delegates that hard thinking to
a completely separate
heavy duty back
end model Like an older more powerful text model
running on a different server Right
The back end
model crunches the data finds the answer and passes a
text string back to the voice model The voice model then
unpauses and speaks the answer out loud The two cent standby rate is what OpenAI charges the developer to keep the WebRTC audio connection open and listening while the
back
end model
is grinding away at the problem And crucially the developer has
to pay for the back end compute entirely separately I
use the analogy of a corporate office for this The OpenAI GPT Live setup is like hiring a highly charismatic
receptionist who
can
handle
the phone lines
flawlessly right I love this analogy But the moment a customer asks a highly technical accounting question the receptionist doesn't know the answer They have to put the customer on hold slide a note under a locked door to an invisible accountant in the back room wait for the accountant to slide a note back
and
then read it to the customer Yep And you the developer you have to pay the accountant a separate hourly
wage on top of
the receptionist's salary The analogy
holds perfectly
but
it highlights
a massive glaring vulnerability Which is The receptionist is fantastic at reading the note but the receptionist has absolutely no
idea if the accountant is
lying hallucinating or providing dangerous information The voice interface is completely agnostic to
the truth of the back end data Which brings
us
to the safety testing blind spot Exactly When OpenAI ran their system
card safety tests on GPT Live 1 1
testing it for
things like generating illicit
behavior offering self harm advice or executing prompt injections they evaluated
the voice
model with the back end delegation completely disabled Oh wow So
they tested the
receptionist without the accountant in the room They tested the interface in total isolation They proved that the voice
model itself won't spontaneously generate unsafe audio
But in a real world fully deployed application the
voice
model is constantly delegating to the back end Right The safety of the final answer provided to the user depends entirely
on the safeguards built into that invisible back end model OpenAI's safety tests for the Voice API do not predict how often unsafe or hallucinated answers might occur in a live
environment because they aren't testing the actual delegated workflow That is a staggering loophole We are unleashing these agents
into live voice calls acting on behalf of companies and the flagship safety
tests
only validate the
outermost layer of the software Yeah it's pretty concerning And OpenAI even
admitted
there are strict
technical constraints on this model too Yes there
are The 1 1 API supports streaming and function calling
but it explicitly does not support structured outputs or fine tuning Explain why a
developer would care about structured outputs or fine tuning in this context Fine tuning allows
a developer to bake specialized
knowledge directly into the weights of the model which you
can't do
here You have to rely entirely on
context windows and delegation Got it
But structured outputs are even
more critical When an AI generates a response it usually outputs raw text A structured
output forces the AI
to format its answer in a strict data schema like a JSON
file
so that other
traditional software systems can easily read and execute it Without structured outputs it becomes incredibly difficult to guarantee that the
voice model can reliably trigger other APIs without breaking the code syntax So we have software agents
that are
highly capable but structurally isolated and heavily constrained by safety liability Google
is keeping Gemini trapped in a background window on your Mac
OpenAI
is acting as a highly constrained receptionist for your apps But why keep these agents
trapped inside a laptop or hidden behind an API call
If the agent is meant to
navigate the physical world shouldn't it have a physical presence Which
brings us to the hardware layer of the deep dive Meta
believes
the next paradigm isn't on a screen at
all According to reports from TechCrunch and The Information Meta is preparing to launch a pocket sized AI device
called the MuseCharm with a reported price tag
of 300 Meta hasn't officially confirmed the final
retail price but the supply chain
leaks indicate a premium positioning A
300 dedicated
piece of hardware The
reports describe it as Tamagotchi like It is a physical carryable object whose entire purpose is to
provide a dedicated microphone
and speaker to interact with Muse Meta's real time voice
agent But this is the detail that people
need
to grasp The Charm
is not the Muse AI
itself It's a crucial distinction Muse is already fully deployed You can access it
right now for free via the
Meta app on WhatsApp on the web and on Mac The Charm is strictly a dedicated
hardware conduit And when you look
at the broader market
300 is a very aggressive price point
It really is It sits above a competitor called Friend which is a 250 AI persona wearable that focuses on text
and voice companionship
And it
is priced drastically above Amazon's new
device the Bee Frontier Let's look at the Amazon Bee The regular price is 75
though they recently had it on sale for around 52 Its
hardware focus is on transcribing daily conversations and tracking tasks but Amazon requires a 99 annual
subscription to connect the hardware to their broader app
ecosystem It is two entirely different theories of hardware economics Amazon is employing the classic razor and blades model They are likely taking a loss or breaking even on the 52 physical
hardware
just to trap the user inside
a
99 recurring subscription ecosystem Right Get them in the door cheaply Exactly Meta conversely appears to
be seeking a massive upfront hardware margin with that 300 price tag To access an ecosystem they are otherwise
giving away
for
free to keep you engaged in their social network See I look at this
as a consumer and the value proposition just baffles me
I have a
1
000 smartphone in my pocket right
now Most
people do Right
It has a state of
the art microphone array incredible speakers a high resolution screen and it runs WhatsApp natively
I can talk to the Muse AI perfectly fine on the device I already own Why in the world would I pay
300 to carry around a secondary plastic walkie talkie to talk to the exact same AI Because you
are underestimating the psychological weight of user friction and the industry's current obsession with ambient computing Ambient computing being the idea that technology should
fade
into the background of our environment rather than requiring us to
stare
at a screen
Exactly Let's break down the physical
friction of using your smartphone's
AI Okay Let's do it First you reach into
your pocket You pull out the phone
You look at it to bypass the
biometric lock You swipe to
find the specific app
You wait for the
app to load You tap
a microphone icon wait for the
chime and then you speak It takes a few
seconds In the context of human
computer interaction that is an eternity of latency and cognitive
load
Not to mention the distraction factor The moment
I unlock my phone to
ask the AI a question it's
gone I see a push notification for my email a text
from my spouse and an alert from a
news app The context switching destroys
whatever train of thought I originally had The smartphone is basically an engine
of distraction A physical
device like the
Charm something you can
blindly grasp in your pocket or attach to
your bag creates a zero latency zero distraction connection Just press
and talk You press a tactile
button and you speak The Tamagotchi comparison is apt not just because of the size but because of the psychological bond
A Tamagotchi wasn't just
a game running on a PC It was an object taking up physical space in your life Oh that makes sense Meta is betting that by giving their
AI a physical embodiment you will interact with it tenfold more often than if it were buried in an app folder It is a fascinating
reversal
of the last
15 years of consumer technology We spent
a decade unbundling
everything and merging
it into the smartphone
Right the great convergence We used to carry a digital camera an MP3 player a Garmin
GPS and a cell phone Apple came along and consolidated
all of that physical hardware into a single
pane of
glass Now the tech industry is trying to unbundle the smartphone We
are paying 300 to carry a separate single purpose brain Because the brain is
no longer just a feature of the phone
The brain is attempting to become
the primary
interface for your entire digital life But as we discussed with the software agents these models are still deeply flawed Very flawed And
while an AI hallucinating a
shipping quote on your Mac or giving a bad answer on a 300 pocket
charm is frustrating the stakes are relatively low Which
brings us to the final and
most critical layer of this deep dive What happens when we take
the training wheels off This is where it gets real What happens when we take these AI
architectures out of the sterile predictable rules based environment of a
computer operating
system
and we ask them to model the physical world We are looking at the ultimate stress test for artificial intelligence chaos physics Specifically we are
looking at an independent peer reviewed study
out of the University
at Albany that evaluated Google's WeatherNext Cyclone model against official human led meteorological forecasts This
wasn't a
small sample size either The researchers Jacob Will Noor
El Golad and Professor Tang evaluated the AI's
performance across
385 tropical cyclones worldwide 385 storms Over a four year period from June 2022 to May 2026 Right And they were testing a very specific metric
storm
intensity They wanted to know if the AI could accurately predict the maximum wind speeds of a
hurricane compared to what the storm actually produced in reality And
they benchmarked those predictions against the official forecasts from NOAA and the Joint Typhoon Warning Center And the headline result
is exactly what
Google wants to see WeatherNext Cyclone beat NOAA's state of the art physics based hurricane
model at every single lead time from one to five
days out Every single one It was vastly superior at looking five days into the
future and plotting the storm It even produced fewer false alarms
for rapid intensification events which is when a storm's winds jump 35 miles
per
hour in a 24 hour window On paper across a
massive data set beating the combined
output of
NOAA's supercomputers and human
meteorologists is a monumental
scientific achievement Definitely But as you dig into the methodology of the study
a deeply concerning structural flaw
emerges The
catch
There's always a catch There's always a catch The researchers found that while the AI
was
technically
more accurate on average it consistently underestimated storm intensity Underestimated it Yes And
that underestimation grew significantly
worse the farther out in time
it tried to predict Furthermore the AI's uncertainty ranges the statistical bounds of what it thought was possible were far too narrow To put that
in practical terms the AI
would predict that a storm's winds would peak
between
80 and 90 miles per hour It was mathematically highly confident in that range But in the physical
world the storm would hit 115 miles per hour The actual
real world devastation frequently landed completely outside the
bounds of what the AI calculated
was even possible And the performance also varied wildly
depending on the ocean basin The
AI was strongest in the North Atlantic where historical data is incredibly rich but
it was weakest in the eastern North
Pacific Let me use an analogy here It sounds like a student who figures out how to game a standardized test
Oh I like where this is going The student gets an A plus on
all the multiple choice questions by playing the odds and constantly guessing the
statistical average But when the test throws a curveball essay question about an extreme unprecedented event the
student completely
fails because they only know how to play it safe That is an excellent way to describe a phenomenon
known in machine learning as the smoothing effect
smoothing effect Yes And we need to
unpack how a neural network actually learns during
its training phase to understand this
How
does the smoothing
effect actually work mathematically When Google trains an AI weather model
they feed it decades of historical atmospheric data But the AI
doesn't understand
physics Right It's just code It
doesn't know what a cloud is
or
how thermodynamics work It is purely looking for mathematical
correlations
in the data grid During training the algorithm uses a loss function usually optimizing for mean squared error Okay so it's trying to keep its error
score low Its sole objective
is
to adjust
its internal weights to minimize the average difference between
its predictions
and the historical reality across
billions of data points
And
most weather is boring Broad normal everyday weather makes
up 99 of the data set Truly catastrophic rapidly intensifying Category 5
hurricanes are massive statistical outliers Because the AI is mathematically punished for
being wrong it learns that predicting
extreme outliers is dangerous to its overall score Predicting an extreme event that
doesn't happen ruins its average error rate So it just avoids predicting them So the
neural network learns to smooth its predictions toward the historical mean It literally strips away the chaotic high variance extremes to find the
mathematically safe middle ground But if you are the
mayor in
our opening scenario trying
to decide whether to issue an evacuation order
for a million people an AI that structurally underestimates
extreme events is catastrophic It's deadly Over
a 385 storm data set the AI might have a better average score But during the
one storm
that hits your
city
that smoothing
effect will get people killed This is exactly why Professor Tang
emphasizes that AI models must be integrated as additional
tools
not replacements for human forecasters
Tools not replacements We are comparing two fundamentally different
ways
of understanding the world NOAA uses physics based models that
calculate fluid dynamics thermodynamics and the Navier Stokes equations They are simulating the actual physics of the atmosphere Yes Whereas the AI is just doing
hyper advanced pattern recognition The AI
is incredible
at looking at the initial state of the atmosphere and
instantly calculating the where and the when
of a storm track But it fundamentally lacks operational judgment Operational judgment A
human meteorologist can look at the data and
say the models are predicting a Category 2 but I
see a pocket of incredibly warm
water and low wind shear The
physics allow for an explosion in intensity We have to
override the average and warn the public And the AI can't do that The AI cannot make that leap
It is trapped by its
own optimization for the average And it is critical
to note that the Albany researchers were evaluating
an earlier version of the Google model
not the newest 2026 release which perfectly highlights the terrifying lag in this industry The
science just can't keep up with the deployment Rigorous peer reviewed science takes years
to conduct By the time scientists publish a paper proving a model has a deadly smoothing effect the
tech companies have
already deployed version 2 0 and integrated it
into public infrastructure The gap between technical capability and reliable safety is the defining tension of this entire ecosystem right now
It really is Whether it is an AI predicting a hurricane an API delegating a task to
an untested back end model or an agent silently navigating the operating system on your
Mac we're deploying statistical prediction engines into chaotic physical environments and just hoping
the math holds up So let's pull all of these threads together because we
have covered massive ground today We started by looking at the sheer financial gravity of the AI
movement Anthropic pulling in an 81
billion
IPO
and building a 220 billion war chest Well the broader stock market and the physical hardware infrastructure crack under the realities of interest rates power grids and liquid cooling retrofits Right We saw
how that physical bottleneck is driving these companies deeper
into the federal government with tens of millions in defense
and civilian
contracts flowing to AI firms right alongside massive political investments Which
proves AI is no longer a software
product It is a foundational layer of state and economic security From there we watched the software evolve from text generators into
autonomous agents Google Gemini using accessibility APIs to take over the mouse and keyboard on our Macs contained only by a
fragile digital
sandbox born out of liability terror And we
saw OpenAI splitting the charismatic voice of their AI from the invisible reasoning engine
in the back room creating massive untested loopholes in
how we deploy safe models We watched Meta attempt to pull those agents out of
our digital screens and force them into physical 300 pocket charms hoping to overcome user
friction and establish a zero latency psychological bond through ambient computing And finally we watch the
AI collide with the ultimate chaos of
atmospheric weather It proved that while it can
predict the historical average better than
any human physics model it structurally smooths out the catastrophic extremes leaving us entirely
reliant on human operational judgment when the stakes are
highest
It is a comprehensive picture of a technology
that possesses unprecedented capital and capability but is fundamentally
struggling to adapt to the messy unpredictable and highly regulated nature of the real world It's an incredible time
to be watching this space
Which leaves you with one final provocative thought to ponder as you
look at your
own devices this week Throughout this entire deep dive we
have been focused on how humans interact with these AI agents Humans
and machines We are the ones pausing Gemini in the background window We are
the ones pressing the physical button
on the meta muse charm We are the mayors reading the weather next forecasts
and
making the final call The human is always the failsafe But the human bottleneck
is exactly what these companies are trying
to engineer away Exactly What
happens next year when these autonomous
agents start
interacting directly with each other Oh that's where the liability gets messy Think about
the delegation architecture and the lack of structured outputs we
discussed Imagine you ask your 300 meta muse agent to
dispute
a massive freight charge The muse agent calls the logistics company's customer service line which happens to be powered by OpenAI's
GPT Live agent And they are arguing over a multi box shipping quote that was originally generated in
the background by your Google Gemini agent You have three different corporate AI
agents all running on different architectures all delegating to invisible
back end models negotiating financial terms in a fraction
of a second with zero human oversight If one
of those back end
models hallucinates a tariff
code or if their interaction creates an unpredicted
feedback loop that executes thousands of dollars
in erroneous transactions who is legally liable
It's a legal nightmare Is it the agent who spoke the words The back end model that did the reasoning The hardware
manufacturer of the
pocket charm As artificial intelligence moves from passive software that we use to active agents that act on
our behalf in the
physical world the legal economic and physical shockwaves are
going to make
Enthropic's historic 81 billion IPO look
Original reporting
Read the complete Superpower Daily coverage behind this episode, including reporting context and source links.