Loading page…
Loading page…
The Signal / Superpower Daily
ChatGPT agents now have a way to buy, not just find. Claude’s work on Gödel’s notebooks also brings a question of mathematical credit into focus.
Superpower Daily: The Signal
Episode guide
ChatGPT agents now have a way to buy, not just find. Claude’s work on Gödel’s notebooks also brings a question of mathematical credit into focus.
Full transcript
Select any speaker or timestamp to continue listening from that point.
Welcome to The Signal from Superpower Daily with Maya and Theo
So today we are looking at OpenAI's new chat GPT
wallet where your AI agent can now actually buy things on your
behalf
Right The machine is moving from just being a digital advisor to an autonomous economic
actor It finally gets a credit
card Yeah it is a huge
shift We have a really fascinating deep dive for you
today We are looking at a fundamental transformation in how artificial intelligence operates in the real
world
I mean we are moving past the era where AI just gives us information or generates some text
We are entering an era where AI actually takes action
It spends real money
It interacts with legacy aviation infrastructure
It
actively tests critical cybersecurity defenses and
it even parses decades old mathematical shorthand to finish legendary proofs
And the thread connecting all of these developments is really this crisis of attribution because when an AI buys something we
need to know who is responsible When
an
AI finds a flaw in a nationwide aviation network we need
to know how it interfaces with legacy code When an
AI writes a police report the justice
system demands an author Technological capability
is just suddenly racing far ahead of our legal frameworks It really
is So I want to dive straight into this lead story The industry
has been waiting for this shoe to drop for a while now Oh absolutely
On November
13 2026 OpenAI
officially
launched the Chat GPT Wallet And
right now this is rolling out exclusively for United States pro subscribers You
know you have to be 18 or older You
need a U S.-issued card and a U S billing
address Right They are keeping it contained at
first
Yeah But the core functionality is
staggering The feature basically allows OpenAI's DOTS agents to make purchases on a user's
behalf And they built this in direct partnership with American
Express MasterCard and
Visa Which is a massive
deal We really need to connect this to the broader evolution of e commerce because the digital economy has spent 20 years obsessing over
checkout friction
Oh completely I mean we moved from typing 16 digit card numbers into sketchy web
forms We moved to saving profiles on
Amazon We moved to one
click purchasing And then we moved to mobile wallets like
Apple Pay But a fundamental bottleneck
still remained The
human
being Exactly The human being still had to perform the
mechanical labor You had to find
the product You had to evaluate
the specs You had to put it in a
digital cart And you had to physically press the final
buy
button Right The ChatGPT wallet basically removes the human from the mechanical part of that
loop
entirely See I am naturally skeptical
of this Handing an AI direct access to your
credit card just feels dangerous
It does sound a bit scary
It feels like setting up a high frequency trading
bot but giving it a corporate expense account to buy consumer
goods I mean the potential for a runaway script seems
incredibly high But the system does allow you to set some pretty
strict
parameters Yeah the architecture is
deliberately cautious The approval process doesn't strictly have to happen at the final checkout page but users can review each purchase
before
submission Okay This is the
standard route It is definitely the
safest route You tell your agent to find a specific pair of
noise canceling headphones The agent navigates the web it
finds the item it parses all that
shopping
cart data Right And then it presents you with a final screen showing
the total cost This includes
taxes and shipping And then you explicitly approve
that specific transaction But the terms also allow
for advance authorization And I think this is where the
landscape truly shifts Yes Advance authorization completely
changes the paradigm You can basically grant ChatGPT the authority to approve
purchases in advance You can set
a budget limit You can
define
specific conditions Wow You are essentially writing a procurement policy
for your agent You are telling the system to just
monitor a marketplace If the agent finds a specific vintage watch under a
specific price threshold it executes
the buy immediately So
no
confirmation screen Right It does not bother you with
a confirmation screen The transaction just happens
in the background But there is a massive safeguard built into
that
advance authorization right The system relies on predicting the
exact final cost So if any material detail changes during
the checkout process the
autonomous action halts Say the shipping cost suddenly jumps by 20 at that
final
confirmation step The system recognizes the discrepancy It requires you to step back in and re approve
the transaction manually And that material detail change is a
critical circuit breaker Because artificial intelligence models do hallucinate they misinterpret complex website layouts
all the time Right
They
aren't perfect Plus a retailer might dynamically change pricing based on
inventory levels This can happen in the milliseconds between adding an item to a cart and
checking out So the system needs a
hard
stop Yeah that
makes sense If the terms of the transaction shift from the
pre approved parameters the agent loses
its autonomy instantly Human oversight
is
immediately mandated Okay I want to push back on the security of the
financial data itself because my biggest fear isn't really the AI buying a thousand rolls of paper
towels by accident My fear is
a data breach A lot of people
share that fear I worry about the AI just leaking my permanent credit card number to a malicious merchant or
some
spoofed website How is the underlying security architecture actually
structured here Well the security architecture actually borrows heavily from the mobile wallet protocols we
already trust The ChatGPT agents never actually see your full card
details Oh
really Yeah They do not have your 16 digit primary account number stored anywhere in their
context window and they cannot leak what they do
not possess So how does the payment actually reach the
merchant's bank OpenAI work directly with the major payment networks to utilize transaction
specific
payment tokens Okay You save a physical card
to the wallet and then the payment
network communicates with OpenAI The network
provides a substitute credential This is
a scoped card number It comes with
a unique security code It comes with
a unique expiration date So it's
basically
a
burner card Essentially yes This token is cryptographically tied
to that specific interaction It allows the payment to route through the traditional banking
system to the merchant but your actual card details remain locked in a vault
at your issuing bank I look at it kind of like
a code check system You give a restaurant
your expensive winter coat They give you a
cheap little paper ticket The
ticket represents the coat It allows you
to
retrieve it later Right But if someone steals that paper
ticket on the street they cannot just walk into any other building
and demand a coat The ticket only works
at that specific restaurant The payment network acts like the coat check attendant They verify
the context of the token That is a highly
accurate way to visualize it The token is entirely useless outside of
its specifically
defined merchant scope That's fascinating It is also important to
understand the financial plumbing here This purchase is still charged directly to
your existing credit card account You
still earn your airline miles You
still get your cash
back Oh that's good to know You are still
protected by
your bank's fraud liability rules Open AI
is not acting as a bank They are not holding user
funds in a stored value account They are simply an authorized software
conduit for passing that tokenized credential This
brings up an incredibly important limitation A listener might assume that an AI initiating
a purchase just guarantees
a delivery It absolutely does not An approved order by the AI
does
not guarantee a successful transaction Right This is where the frictionless vision meets
the friction of the real world The
AI can decide to buy something It
can perfectly navigate the checkout form
It can transmit the encrypted token But the traditional
financial
gatekeepers
still hold absolute authority Your card issuer can
just decline the payment They
might flag a rapid automated
purchase as suspected fraud or they
might see insufficient funds And the merchant also retains the right to reject the order entirely because the merchant is the ultimate
seller
of record here Open AI
is not the retailer They
are not holding inventory If
the headphones arrive broken you cannot
return
them to Open AI Right
You
can't complain to Chad GPT Exactly You are bound
by the specific merchant's return policy You
are bound by their warranty terms
If there is a billing error you dispute the
charge with your credit card company The traditional
dispute resolution framework remains entirely intact Now I noticed a
massive detail in the fine print Open AI is
not taking a transaction cut at
launch This feels completely counterintuitive to me
It does seem strange at first glance I mean Apple takes
a massive percentage of app store transactions
Credit card networks survive on interchange fees Why would Open AI build this incredibly complex financial infrastructure
and just leave that revenue on the table Because
they are playing a much longer game here Taking a transaction cut
fundamentally changes the company's legal and regulatory standing If
Open AI takes a percentage
of the sale
they cross a regulatory threshold Oh I
see They start looking like a payment processor
They start looking like a money services business And this invites an
avalanche
of scrutiny from financial regulators like FinCEN Right They want to avoid being
classified
as a financial institution at all
costs Exactly By abstaining from a transaction fee they position the
chat GPT wallet strictly as a software tool
It is a convenience feature for
the user
Their ultimate goal is behavioral modification How so They want to train users to rely on the AI as
the primary default interface for the entire Internet The training
data generated by watching millions of users
make purchase decisions is vastly more
valuable than a 2 swipe fee they want to
own the intent phase of commerce And I guess this strategy also explains why they are
aggressively releasing business tools for merchants E commerce
sites deploy
really sophisticated bot protection software These firewalls are designed to block automated scripts from scraping prices
or buying up limited inventory So OpenAI has to convince these
merchants to whitelist their agents They have to ensure a legitimate purchase attempt by a ChatGPT agent is not immediately blocked as
a
malicious attack It requires a really delicate technical negotiation between OpenAI and the major
retail platforms There is one final caveat listeners must
understand here If you panic and remove a card from your ChatGPT
wallet it does not cancel a purchase that was already
submitted Once the agent clicks buy and passes the
token the transaction is in
flight It lives on the merchant's servers and the banking
network The AI is just the
initiator Once the data leaves the OpenAI environment
the traditional machinery takes over completely
You cannot recall an order just by deleting your payment profile
Ok
so if you are tracking this rollout
the metric to watch next is the Practical Completion Rate
It is an engineering marvel for an AI to parse a website and submit an order form
But it is an entirely different challenge to see how often these automated agents successfully navigate the gauntlet of anti bot software and bank fraud
algorithms
The global financial system is built to block anomalous rapid behavior
We will see how it reacts to millions of autonomous agents suddenly buying coffee and electronics So
we are going to wrap up the lead story there We
have established that AI is fundamentally altering consumer retail But
that friction between new AI and old systems becomes far more dangerous when we look at national infrastructure The
stakes
elevate dramatically when we move from buying headphones to routing commercial airliners Next up
we are looking at the Federal Aviation Administration A nationwide
ground stop on November 13 triggered a massive federal inquiry Agencies are
currently investigating whether the FAA's newly expanded AI traffic management tool played a role in shutting down American airspace This incident
just perfectly illustrates the vulnerability of zero fail legacy systems We
need to look at what actually happened on November 13 A
severe communications disruption hit all air route traffic control centers simultaneously These
centers manage high altitude commercial flights The
disruption
severed communication links And this forced the FAA to halt all flight departures nationwide The chaos
was immediate Federal agencies
launched an immediate investigation And the
response from the highest levels of government was incredibly fast It
was
Homeland
Security
Secretary Mark King Morgan issued a statement He confirmed
that the system failure was rooted in a telecommunications issue He explicitly
stated it posed no physical threat to the public Transportation Secretary
Sean
Duffy also addressed the crisis He confirmed technicians
were aggressively investigating the root cause of the communications failure But the investigation
quickly focused on a specific newly introduced variable They are looking
really closely at the FAA's SMART tool Right So SMART
is
an acronym It stands for
Strategic Management of Airspace Routes and Trajectories It is a highly advanced piece of optimization software The FAA began
introducing
it in September They initially deployed it
in a limited testing phase around Washington D C But according
to
reporting from Futurism they expanded this
tool nationwide just days before the catastrophic
outage I have to admit
that timing is incredibly suspicious You roll a massive
new AI network out across the entire country and a few days
later the entire nationwide communications system collapses It is difficult not
to draw a straight line between those two events The timing is undeniably
suggestive It invites immediate suspicion
But we really must
rigorously separate correlation from causation This is a classic
trap in enterprise software deployment I want to challenge
that though If I update the
operating system on my laptop and ten minutes later my external monitor stops working I blame the update
Sure The rollout of
SMART
and the collapse of the telecom network happened essentially simultaneously Why shouldn't we assume
the new software broke the old hardware Because of how these
systems fundamentally interact SMART is an optimization
engine It ingests an ocean
of real time data It looks at
complex weather fronts across the continent It analyzes runway
congestion levels at major hub airports It tracks the
active
flight paths of thousands of planes And it uses all
of this data to dynamically adjust and streamline routes The objective is to
help human controllers reduce delays and save aviation fuel It is performing
a highly complex continuous computation computational task
in the cloud But the
AI is not actually flying the planes It is just an advisory tool Precisely Information
presented
at the initial news conference made this very clear Human FAA
specialists remain intimately involved They are
operating the system They are
overseeing SMART during this entire rollout phase The AI
cannot autonomously reroute an aircraft without human oversight It generates
highly optimized suggestions the human
controllers review the logic
and then the human
controllers implement the final decision This leads directly to
the most important limitation in this story The federal inquiry is
active but officials have not
established any causal link between SMART and the communications failure They have established nothing
beyond the coincidence of timing It is an inquiry
to rule it out not a conclusion of guilt We also have to
consider the response from the private contractor who built the SMART system Airspace Intelligence known as
ASI released
a definitive statement They did not mince
words No they didn't They
stated categorically that
the ground stop was not based on or a result of the SMART system They deny that any
product of ASI caused the failure Crucially they deny that
their integration with the FAA and air traffic control systems caused the crash And that is a
highly specific legally vetted denial They
are addressing the
exact friction point we are discussing The aviation network relies
on deeply entrenched legacy telecommunications infrastructure Some of these data
pipes
were laid decades ago When you introduce a powerful
new variable like SMART and the whole system fails the new variable is the
obvious suspect But the reality might be
far more mundane Exactly It is entirely possible
that
an underlying pre existing weakness in the legacy telecom network simply gave way A switch might have
failed in a server room completely independent of the AI's operations I look at it
as a plumbing problem The AI software might
be perfectly designed It might generate flawless routing suggestions But if SMART attempts
to push massive new volumes of data through a legacy API built in the 1990s the old pipes might simply burst Right The AI did
not
make a mistake The infrastructure simply could
not handle the bandwidth required by modern machine learning tools That is the exact
scenario investigators are trying to model right now A buffer overflow and
a 30 year old messaging protocol is just as likely as a rogue AI command But right
now the root cause remains unknown The thing
to watch next is the final resolution of this federal inquiry The FAA
has to find the definitive root cause of the communications failure They have
to point to a specific failed router or a specific line of bad code Until they
provide that transparency the new
AI tool will operate under a heavy cloud of suspicion The public
has very little tolerance for black box technology when they are sitting on a tarmac The FAA
needs to prove exactly what broke to restore confidence in their modernization efforts Transparency is
the only currency that matters when upgrading zero fail systems Meanwhile
the
cybersecurity industry is trying to solve that exact problem OpenAI
is shifting its focus from simply finding software flaws to actively fixing them The
company just opened early access to Daybreak Red Team This
program allows AI agents to investigate web app vulnerabilities and independently test software patches This
represents a massive evolutionary step We
are moving from passive rules
based scanning to active
contextual exploitation
What actually happened is that OpenAI launched an early access program for a group of highly coordinated AI agents
These agents are purpose built to test web applications for security
weaknesses And this matters far beyond a simple product launch because of how these agents operate They fundamentally change the workflow of network
defense You really have to understand the daily misery of a
cybersecurity professional to appreciate this tool
Traditional static code scanners are a nightmare of alert fatigue
Oh I bet
They scan a massive code base
They look for simple text patterns that match known vulnerabilities
When they find a match
they throw a critical alert
They generate thousands of these alerts every week
But they lack any situational awareness
A
traditional scanner might flag a specific library as vulnerable It
screams that the sky is falling But
a human engineer looks at the architecture and realizes that vulnerable library is buried behind three layers of internal authentication It
is
deployed on an isolated server It is
completely inaccessible to anyone outside the building So the
scanner basically created panic over a flaw that cannot actually be exploited I think
of a basic code scanner like a terrible security guard The guard
walks past a massive estate They see
a rusty padlock on a tiny garden shed They hit
the panic button and declare the entire property fundamentally insecure Exactly But
Daybreak
Red Team acts like a professional digital lock picker It
does not just point out the rusty lock It physically walks up to the shed uses
its tools to pick the lock It opens
the door it steps
inside and most
importantly
it looks around
to see if there is a hidden tunnel connecting that shed to the main mansion It investigates the
actual attack paths It determines if
a theoretical weakness can be weaponized in the real world And that contextual
investigation changes everything The AI
agents examine the raw source code but they
simultaneously analyze the surrounding architectural context They map
the network topology Furthermore security
teams
do
not have to deploy these agents blindly They can import
the massive lists of findings from those noisy code scanners They can import reports from human bug bounty hunters Oh that's smart
Yeah
Daybreak takes
those
unverified suspected flaws and systematically attempts to validate them through active
exploitation This gives security teams
a massive advantage in prioritization A team might start
Monday with a list of 100 theoretical vulnerabilities Instead of wasting time
researching all of them they let Daybreak run
Right Daybreak proves that
only
three of those vulnerabilities can actually be exploited by an outside attacker The human engineers can
focus all their energy on patching those three critical holes It distills pure signal
from overwhelming noise But the Daybreak workflow
has a second equally critical phase The human developer writes
a patch to fix the vulnerability They deploy the new
code Immediately the Daybreak agents
go
back to work They retest the specific
attack path Yes They launch the
exact same exploit chain against the newly patched version of the software
This is a vital logical
distinction Proving a vulnerability exists is
entirely different from proving a developer's specific fix actually neutralizes the threat Because developers make mistakes Developers
often implement incomplete fixes
They might patch the specific
symptom without addressing the underlying logic flaw Going back to our lockpicker
you realize the shed is vulnerable
You hire a contractor to install
a heavy duty deadbolt The Daybreak lockpicker returns that
night and attempts to break through the new deadbolt If the lockpicker fails you
have absolute cryptographic proof
that your patch works It creates a closed loop
remediation
cycle You identify you validate you patch you revalidate
The continuous
AI assistance
dramatically accelerates
the entire defensive life cycle This brings us to a massive
limitation and a necessary caveat Letting autonomous hacking agents loose is
terrifying It really is These AI agents
are not just
unleashed on the open internet to hack whatever IP address they stumble across right No absolutely not Okay good
The
liability risks would
bankrupt
the
company
in a week The ethical risks are insurmountable These
aggressive assessments are strictly confined
to isolated heavily monitored environments Okay good They
do not operate
on
live
public
production servers where real customer data is actively moving The OpenAI documentation states the agents
operate under guardian review They operate within defined scope boundaries
How does this containment actually work
at the engineering level It relies on hard coded programmatic
limits The scope boundaries dictate exactly
which applications the agents can see They define the specific IP
subnets the agents can scan They list the exact network
endpoints the agents are permitted to interact with If a Daybreak
agent compromises a server
and discovers a potential attack path that requires jumping to a database outside its defined scope the execution engine kills the process immediately
The assessment halts The guardian review acts
as a secondary
layer of defense against the AI itself It likely involves a separate independent oversight
model
It might also mandate a
human in the
loop requirement for certain destructive actions This oversight constantly monitors the red team agent's
telemetry It ensures they do not exceed their authorized
mandate
It ensures they do not accidentally crash the isolated
environment through aggressive testing It is a highly supervised heavily instrumented exercise in
controlled chaos The most important thing listeners should watch next regarding
Daybreak Red Team is the pricing structure and the timeline for general availability This is currently locked in early access OpenAI has
not announced how they plan to monetize
the service And the pricing model will dictate the future of
network security Will OpenAI charge per vulnerability validated Will they charge
for raw compute hours Will they
sell it as a massive flat
fee enterprise subscription
Good questions If it is affordable it democratizes elite red teaming
for small
startups If it is
priced at a premium it simply widens the
defensive gap between tech giants and everyone
else We are going to shift gears dramatically We are moving from
securing the future of the internet to
deciphering the foundational logic of the century A researcher at the University of Helsinki named Jan von Plato used
anthropics Claude
to analyze the private notebooks of Kurt Gödel The AI surfaced 663 theorems and proof ideas hidden in the mathematician's shorthand This
story
just challenges our entire philosophical understanding of time human memory and the mechanics
of mathematical discovery We have to lay out the sheer
agonizing scale
of the human effort involved here
Kurt Gödel is arguably the most important logician since Aristotle His incompleteness theorems shattered mathematics
in the 1930s Since 2017 a team led by von
Plato had been painstakingly working through 16 of Gödel's
private workbooks from
the 1940s These books contain 1 009 densely packed pages They are not written in standard German They are not
written
in English They are written entirely in Gabelsberger Right and
Gabelsberger is an obsolete incredibly complex German shorthand It was designed for rapid dictation not for preserving highly technical
mathematical
logic for future generations It relies on subtle variations in stroke thickness and vertical placement It is highly compressed It is deeply idiosyncratic to the writer
Deciphering
Gödel's specific handwriting in this obsolete script is practically a cryptographic discipline The human team
was really suffering They were moving at
a brutal pace of roughly one page per week They were translating symbols
without always grasping the broader logical
arc spanning hundreds of pages The sheer volume of data was overwhelming their
capacity to synthesize it Right So von Plato changed his strategy entirely He took his
completed raw text transcriptions He bundled them with six academic books he
had
previously written on Gödel He gathered a
massive corpus of surrounding historical context and
he fed this entire library into Anthropic's Claude model And Claude ingested a
lifetime of human scholarship in moments It processed the
entire archive and generated a comprehensive 650 page logical explanation
of the notebooks in a matter of days They traded
a decade of human struggle for a weekend of compute time The AI extracted 663 distinct theorems sketches and false starts This
matters far beyond the historical novelty because it proves the transformative power of
a massive context window The context window is absolutely the key
to this entire breakthrough
Human scholars are brilliant but human working
memory is severely limited A human researcher cannot hold a thousand
pages of dense interrelated set theory in their active memory simultaneously
Of course not A human might read a highly specific novel
definition of a set on page 40 By the time
they reach a complex proof that relies entirely
on that specific
definition on page 800 six months later the precise nuance has evaporated They lose the
thread Claude does not experience time or memory decay It holds the entire 1 000 page archive in
its active working
memory simultaneously It could draw a perfect instant line
between the definition on page 40 and the application
on page 800 It maintains perfect persistent context It can synthesize a unified theory
of Gödel's intellectual state over a decade It maps the evolution of his thought process However it is vital to
clarify that Claude
did not simply print out 663 flawless finished proofs It provided a detailed inventory of a human being's messy intellectual journey Right They
had
to verify the AI's output They used the KUK proof verification system KUK is essentially a rigorous computer program designed to mechanically verify mathematical reasoning It breaks logic down to its
absolute
foundational axioms It checks every single step
of a proof to ensure there are
no leaps of faith or hidden assumptions The verification results were incredible About
400 of the extracted results held up exactly as
stated
The math was flawless But about 50 were demonstrably mathematically false And the false starts are beautiful They are
They are a testament to
the reality of mathematical discovery It is a process of relentless trial
and error Gödel himself
would frequently write German words like falsch
or schlecht incorrect or bad in
the margins
of these notebooks when he realized a brilliant idea led
to a logical dead end Yeah The machine captured
his failures alongside his successes But the remaining 200 proofs are
where this story
becomes a philosophical battleground The KUK system determined that 200 of these proofs needed repairs to function logically
Let's examine one specific reconstruction It concerns the
axiom of choice This is a fundamental controversial concept in
set theory regarding the ability to
select items from infinite sets Right Claude reconstructed a proof regarding this axiom based on
the notebook text The reconstructed
proof passed the rigorous Koch verification checks It is mathematically true but von Plato openly noted that Claude had to fix a definitional mistake
Gödel
made in the text Gödel had used the correct definition of a
term two years earlier in the notebooks but he messed up the definition
in the specific notes for this proof And Claude recognized the error It leveraged its massive context window
It pulled the correct definition from von Plato's scholarship and the earlier notebook pages It
autonomously bridged the logical gap It corrected the genius's mistake and completed the
chain of reasoning I have
to push back strongly on the
philosophy of authorship here If a human being sketches
a bridge but builds the
central pillar out of sand the
bridge collapses If a machine comes along recognizes the sand is
a mistake replaces it with concrete and finishes the construction who built
the bridge It's a great question If Claude had to add assumptions fix a mistaken
definition and construct
the final logical span can
we honestly call this Gödel's completed
proof Is this not a collaboration Is it
Claude's proof Heavily inspired
by a dead man's
notes You are identifying the exact
debate that historians of
mathematics will wage for the next
50 years Joel David Hamkins is a highly
respected logic professor at Notre
Dame He
offered
a very sober rigorous analysis of this
situation What did he say He pointed out that if this AI reconstructed material resembles Paul Cohen's
later definitive proof regarding the axiom of choice it is historically fascinating
It proves definitively that Gödel's intuition was pointed in the
exact right direction decades earlier
But aiming at the target is not the same as hitting the bullseye Precisely
Knowing a mountain can technically be climbed is vastly different from
standing on the summit The mathematics as verified mechanically by Koch are unassailable The logic
is valid But historically it only proves Gödel possessed the conceptual scaffolding Right It
does
not prove he ever fully erected the building in his own mind Hamkins stated explicitly
that untangling
this knot deciding exactly how
much credit belongs to Claude how
much to
von Plato and how much to Gödel is
the
ultimate
job for historians not computer scientists This brings us to the most important limitation
of this entire project Listeners need to understand that Claude did not read the
raw handwritten shorthand images The
AI did not crack the Gabelsberger cipher visually It possessed
no ability to
interpret the thickness of a pencil stroke on old paper No it
relied entirely on the human prepared text transcriptions provided by von
Plato The AI
was operating on data that had already been meticulously parsed formatted and translated by a human expert over many years It is a profound symbiosis
of
three distinct entities The human scholar provided the agonizing transcription labor and the
deep historical context The AI provided the massive contextual synthesis and the logical
gap
filling And the Coq software provided the cold mechanical
verification of truth The ultimate limitation is that while
computational checks can absolutely validate the mathematical truth of a theorem
they
are entirely incapable of settling
the historical reality of authorship What listeners should watch next is how
the rigid academic community ultimately credits these discoveries When these proofs
are cited in future papers will they be attributed to Gödel with a massive asterisk Will von
Plato and Claude share the formal byline It challenges our fundamental
definition of what it means to discover something It really forces us to ask whether
mathematics is something we discover in the universe
or something we invent And more importantly who gets the
credit when a machine does the inventing based on
the incomplete scribbles of a quick read section First in our quick
reads former open AI researcher Will DePue published a forensic analysis showing six Chinese AI models carry the numerical
fingerprints of restricted NVIDIA Blackwell chips This is a remarkable intersection of geopolitical export controls and deep microscopic technical forensics I
want to frame this as digital forensics DePue
analyzed the publicly released model
weights of several major Chinese AI models He discovered residual mathematical traces that point directly back to the specific hardware configurations used to initialize
the models before training even began He found that DeepSeq V4 Flash and Minimax
M3 matched hardware setups featuring exactly 148 streaming
multiprocessors This specific architecture corresponds perfectly to NVIDIA's B200 and B300 chips Meanwhile
DeepSeq V4 Pro Kimi K3 and Ling 3 1 Flash matched setups with exactly 152 streaming multiprocessors This points directly to the more
powerful GB200 and GB300 chips DePue went
a step further He ran a test on a physical B300 chip and and it's to recreate V4 Flash the starting
weights
bit for bit
To
understand why this works you really have to understand how models
are born A model's weights are the billions of adjustable numbers
that
determine
how it processes data Before
the massive training run begins these weights are initialized
with starting values These values are generated from a random
seed The software framework doing
this work usually PyTorch distributes this initialization task across
the parallel processors on the GPU These are the streaming multiprocessors Because a chip with
152 processors distributes the math slightly
differently than a chip with 148 processors
it leaves a permanent structural fingerprint in the
arrangement
of those random starting numbers Millions of those
numbers barely change during the actual training process perfectly preserving the
ghost of the original hardware
architecture It is brilliant detective work But there is a massive caveat that prevents this from being
an indictment This fingerprint only identifies the hardware configuration used to generate the initial
starting values in the first second of the model's life It
does not definitively prove those restricted chips were used
for the months long massive
training run It also completely fails to prove physical smuggling or export violations
The Chinese developers could easily have rented this specific hardware via cloud computing services located abroad completely outside the restricted zones Dupuit
also
openly acknowledges that a specially configured array of AMD hardware remains
a remote but technically possible alternative explanation Yeah It is a smoking gun
pointing to a specific type of compute but
it tells us absolutely nothing about where that compute was physically
located or how it was accessed It provides a new tool for scrutinizing claims of computing
independence but it does not
hold up in a court of law regarding sanctions Next up physicist Nathaniel Craig credits GPT 6 ASTRA with leading
the
detailed
proof work for a candidate definition of the standard model on a
space time lattice This is another profound instance of AI accelerating fundamental scientific theory The details were outlined
in a recent archive preprint and an OpenAI Science blog
post Craig had a broad strategy for tackling a 25 year old physics
problem ASTRA helped him turn that high level strategy into a rigorous line by line mathematical construction on four dimensional Euclidean lattices The problem involves representing
chiral particles These are particles governed by the weak force which
fundamentally differentiates between left and right handedness in the universe The challenge is representing them on
a grid without accidentally creating unwanted mirror partners that do not
exist in reality The standard model of physics usually relies on perturbation theory This treats complex interactions as tiny manageable corrections
to
a simpler state but that math breaks down when forces get too strong To
understand how the strong force actually creates the mass of
a proton you cannot use perturbation You have to use a lattice You replace continuous smooth space time with a
finite discrete grid Think of continuous space time as a
smooth wheelchair ramp A lattice is
a staircase Computers can only calculate
the staircase But creating that digital staircase for the weak
force automatically creates those unwanted mirror particles due to a mathematical law called
the fermion doubling problem A technique called overlap fermions removes
the mirror particles but it leaves behind a mathematical phase
setting problem that has stumped theoretical physicists for decades
Astra essentially brute forced the
detailed proof to solve this phase
issue across all required topological cases The caveat here is absolute though This is a candidate non perturbative lattice definition It is a purely mathematical
construction on paper It is not an accepted peer
reviewed resolution by the physics community Most importantly it is not an experimental result derived
from
smashing particles together in a collider It is a theoretical framework Its ultimate significance relies entirely
on whether the mathematical construction
can
withstand the brutal unrelenting scrutiny of human
physicists over the next several years The AI
generated the proof but the scientific method still demands human consensus
and eventual
experimental validation Finally a Maricopa County court released a DUI suspect because the arresting
officer used Trulio AI to write the report leaving the court unable to determine authorship This really forces the
theoretical problems of AI attribution into the very real very
high stakes arena of the
criminal justice system The details of the case are striking Commissioner Daniel
Fenzel released a suspect named Anthony Quindara Quindara was accused of driving a Dodge Charger over 100 miles per
hour in a 65 zone He was
accused of failing to stay in his lane He was accused of having open containers of alcohol in the vehicle Right The
official arrest report named Truleo AI as
a drafting assistant The critical problem was that the paperwork submitted to the court completely lacked the trooper's original unedited account It failed to specify exactly
how much the generative AI shaped the final narrative Because
of this massive gap in the evidentiary chain the commissioner ruled it
was
legally impossible to assess probable cause from the paperwork alone The
court relies entirely on the chain of custody of evidence The judge needs to know definitively who authored
the specific factual allegations If the AI hallucinated a detail about the smell of
alcohol
or smoothed over a contradiction in the officer's messy
memory the evidentiary chain is hopelessly broken The defendant loses the ability to cross examine
the true author of the accusations against them But the caveat here is vital This ruling is not a blanket ban on the
use of AI in police reporting Sean Harrington a legal technology expert at Arizona State University analyzed the case He pointed
out that human officers remain completely
legally responsible for the facts in their sworn statements regardless of what software they used
to generate the text The specific
fatal issue in this case was the missing original account The court simply needed to
see the
raw input from the officer along cried the
AI generated output
to judge the tool's actual contribution It was a catastrophic failure of police procedure not a fundamental
judicial rejection of the underlying technology It is also highly
relevant
to note that Quintero later missed his arraignment and still faces the
underlying criminal charges The AI issue delayed the judicial process but it did not magically erase the alleged crime We are moving to three
takeaways from today I can take the first
one All right First artificial intelligence is aggressively crossing
the boundary from digital advisor to autonomous actor We see this clearly with the chat GPT wallet spending user money
without final confirmation screens And with Daybreak
Red team actively exploiting cyber defenses rather than just
scanning for
them The systems are no longer just thinking
they are doing Second the integration
of AI
into complex systems is radically outpacing our legacy frameworks for attribution and responsibility Right
We see this systemic confusion everywhere We see it in aviation ground stops where the AI
is an immediate suspect We see it in the philosophical debate over whether an AI
can claim mathematical authorship of Goethe's proofs
And we see
it
in the legal collapse of AI assisted police reports Our institutions simply do not know how to
assign
blame or credit to an algorithm
Exactly Third forensic verification is rapidly adapting to survive the AI era We are
utilizing deep mathematical models to track physical hardware chips across the globe through the residual fingerprints left in model weights We are using massive context windows to solve decades old physics
lattice problems The digital magnifying glass is becoming as sophisticated as the subjects it investigates
For
what
to watch tomorrow keep
a very close eye on how financial regulators
and consumer protection groups react to OpenAI's advanced authorization terms for the chat GPT wallet Allowing an AI to spend money without a final human confirmation click is a massive regulatory gray area that FinCEN will not ignore for long Head over to SuperPowerDaily com for more on
all of these stories Thank
you so much for joining us We'll see you tomorrow
Guys,
advanced authorization terms for the chat GPT wallet.
Allowing an AI to spend money without a final human confirmation click is a massive regulatory gray area that FinCEN will not ignore for long.
Head over to superpowerdaily.com for more on all of these stories.
Thank you so much for joining us.
We'll see you tomorrow.
Original reporting
Read the complete Superpower Daily coverage behind this episode, including reporting context and source links.