ChatGPT and Codex will share one usage budget

Welcome to the signal from superpower daily with Maya and Theo

Hello

Imagine waking up tomorrow and finding out your unlimited cell phone plan suddenly charges you by the text message

Well That is exactly what is about to happen to your AI

Yeah Starting in

December open AI is emerging chat GPT and codex usage limits for

plus pro and business subscribers Into a single

token

based allowance right I mean it is basically the end

of an era

Irelia's we are watching the AI economy Fundamentally shift away from that all you can eat buffet model and this transition into

a

strictly metered utility Well it is going to change how every single one of us

interacts

with these tools totally So let us look at

what

actually happened here Sure this applies directly

to the

paid

tiers those plus pro and business subscribers are losing

their separate allowances instead They are going to draw from

one

shared token budget Yeah though instant chat is explicitly excluded

from

this shared limit right But everything else goes

right against the meter exactly and the reason why this matters beyond the headline is just

tied to operational realities I mean open AI is operationally merging their

standard chat interface

with actual autonomous work the agentic stuff Yes They are bringing agentic capabilities

directly into the main chat GPT

experience And when you shift a system from just answering questions

to actively executing multi step

tasks The computing math changes entirely I do want to push

back

on

this a little bit though Okay because if you look

at this from a user perspective it feels a lot like a stealth

limit

Disguised as a feature integration

Oh I see what you mean It brings me right back

to the early days of smartphones You know we all

had those

beautiful unlimited data

plans

Oh yeah I missed those right then the

carriers transitioned us to metered gigabytes Suddenly everyone was

watching their usage like a hawk You were turning

off cellular data for specific

apps

You were hunting for Wi Fi Yeah is

this not just the exact same playbook Well that cell

phone data analogy is

actually incredibly precise And think about the infrastructure like a

text

message costs the cellular carrier virtually nothing sure It just piggybacks on the

constant pinging between the phone and

the tower But streaming a

high

definition movie that cost them massive bandwidth right So the AI equivalent of a

simple text message is a

standard chat prompt you ask

for a recipe it gives you a recipe The AI equivalent of

streaming

that 4k video is an agentic workflow Okay so for the listener who

is just starting to use these

advanced

features

let us really unpack that workflow Yeah what exactly makes an agent so much more incredibly expensive

to run than just a standard query It basically all

comes down to loops and context

windows

a standard chat is entirely linear right

You type a question

into the box

the model reads it

calculates the probabilities and generates

an answer

It is one clean transaction makes sense but an

agent operates on a completely different architecture

You do not just ask a

question You give it a

high

level goal like write me an app Exactly and then

it breaks that goal down into actionable steps So let's say

you

ask it to build a web

scraper First it might

write some Python code then in the

background

it runs that code in a sandbox Okay

if the code fails and

throws an error the agent reads that error it feeds

the error back into its

own context window It writes

new corrected code

and it runs it again Wait wait wait So it is essentially just

sitting

there talking to itself in the background Yes Literally every single time it does that it is

burning through tokens

on both ends of the conversation Exactly

both ends and remember how these models process

data a token is just a

unit of information Sometimes it is a

whole

word Sometimes it's a chunk of word right The model takes

input tokens in and it produces

output tokens out So under this

new metering

system both directions count against

your limit Oh wow Yeah the prompt you

wrote counts the background

context it holds in

memory counts

the generated output counts Adds up fast

It does every single iteration of

that agentic loop

You know every

time it fails reads the

error and tries again it as to the tally you could easily burn through Thousands of tokens

while you're just sitting there watching a loading spinner and obviously using their most advanced frontier models

is going

to drain this new Allowance exponentially faster Oh absolutely If I am asking a

highly complex coding question to a top

tier model It is

going to absolutely

chew through this budget It has to just based

on the physics of the data center advanced models

require significantly more compute power

per

token They use much larger

neural networks

right with billions of

more parameters They require require massive memory bandwidth The

physical hardware the GPUs and the server racks

they work incredibly hard and draw immense power Right So OpenAI simply

cannot offer unlimited agent loops on frontier models No The economics of running a massive data center simply do not work if users

are allowed to spin up infinite background processes But here is

the most important caveat to this entire announcement and it is

honestly the part

that

is driving developers crazy The missing numbers Yes The

exact numbers are completely missing from the announcement OpenAI has not

published the token allotments for each specific plan Which is just

an absolute

nightmare if you were a business Oh yeah I mean if you were trying to

forecast your software costs for the upcoming quarter we know

the fundamental mechanism of the platform is changing We just do not

know the actual budget we are being given And

they also haven't provided a specific reset

schedule

Like will this allowance

reset

daily hourly weekly We just don't know And they haven't even

given an exact launch date in December They are just vaguely stating that they quote unquote expect most

users will not face more frequent wait times Well an

expectation is a forecast based on

internal

modeling It is absolutely not a Right They are looking at their historical usage curves They are predicting how people will likely use the

new

agentic features based on beta testing Sure But human

behavior with new technology is notoriously unpredictable Once regular users realize

what these autonomous agents can actually accomplish usage could

easily spike way beyond their internal models So what actually happens to

you when you hit that invisible wall Say you are in

the middle of a massive coding project on a Tuesday afternoon and the meter

runs out What are your options Well your options are going to

be

very limited and very transactional You can stop working and wait

for the allowance to reset You can upgrade your base subscription plan to

a

higher more expensive tier Or you can buy additional a la carte credits for eligible features

to tide you over Which brings us right back to buying a data top up on a prepaid cell phone plan just so you can use

Google Maps to get home It is

the exact same psychological shift You go from feeling completely unrestricted and creatively free

to

feeling rationed Users are going to start doing mental math before they hit enter on

a prompt They will ask themselves if a particular task is actually worth

the tokens And that fundamentally changes your relationship with

the tool It really makes you hesitant

to

experiment If I know a complex open ended coding task might drain my entire daily budget and leave me locked out for

the afternoon I might just not try it

at all That is the calculated risk OpenAI is

taking

here Friction absolutely reduces user adoption and

experimentation But they are being forced to balance that friction against their own soaring infrastructure

costs

Right

Compute is expensive

It is arguably the most constrained and expensive resource in the global economy right now

They have to allocate it efficiently

even if it annoys their core user base

So what you need to watch next is the release of those official plan level budgets

The moment those exact numbers drop

businesses and solo developers are going to have to do some very serious software cost forecasting

Enterprise clients in particular

They are going to need to build token management directly into their departmental budgets

They

will need dashboards to track which teams are burning through allowances efficiently and which are wasting compute This

represents the real maturation of the AI industry Right

We

are officially moving out of the endless hype phase and moving firmly into the accounting phase They

are definitely in the accounting phase now And

the honeymoon period of free unlimited

supercomputing is over Keep

a

very close eye on those settings menus when December rolls around We will

see how accurate OpenAI's expectation of smooth sailing really turns

out to be In other news

MIT has formally

abandoned a plan to use AI for campus camera monitoring The university ended

a highly controversial pilot program after significant pushback from faculty and students This is such

a fascinating collision of digital technology and physical public space It really is

The pilot in

question was run in partnership with a surveillance company called Ambient ai Let us

look at what actually happened here They tested

real time safety alerts using staged scenarios across the campus This

included testing the AI's ability to detect things like brandished firearms or unattended packages left in hallways The

overarching goal was to move toward proactive

monitoring Currently

police

dispatchers on university campuses mostly use camera networks reactively They

pull the tape after the fact Exactly

They pull up the footage after a crime or an incident is reported to see what happened Layering

AI on top would have completely changed that dynamic The

automated

system would actively watch all the live feeds simultaneously It would

computationally flag specific events and

alert human

dispatchers in real time But MIT confirmed

on October 27th that this evaluation is officially over They stated it

will not be revisited Furthermore all the

pilot

data that was sitting in the ambient ai portal was completely purged And the

resistance that forced this shutdown is what makes this story matter far beyond the headline Yeah The

pushback from the MIT community was not just a technical debate about AI accuracy or the annoyance of false positives Right

It

went

much deeper into the philosophy of how a campus should operate Faculty and students heavily protested over privacy concerns Yeah They

protested over issues of access control and data governance And they were highly concerned about the total lack of democratic community input before the pilot was quietly initiated There

was a pervasive sense that this massive intelligent infrastructure was being silently layered onto the campus environment

without consent Which

is huge It

is always a massive trigger for academic communities especially

at a place like MIT They

inherently value transparency open

debate and

technical scrutiny The

protest organized by Professor Catherine D Ignazio was absolutely brilliant

Oh it was so

good She didn't vandalize the equipment or cut

wires She decorated the

cameras She went around campus using rhinestones

feathers and giant googly eyes to dress up the surveillance

tech It was a deeply symbolic and highly effective act of protest I mean the entire design philosophy of modern security cameras is to make them blend

in

They're

supposed to be invisible They're meant

to

be ignored They become part

of the architectural background By putting giant

googly eyes on them she made the

invisible infrastructure highly visible again She effectively forced

a democratic dialogue You really cannot

ignore a surveillance camera that is literally staring back at you with craft supplies Right It makes

the

invisible observer suddenly very present It perfectly highlights

the core psychological issue of modern surveillance It is a

one way mirror Yeah The

authorities

can see you track your movements and log your associations but you

have no idea when they are actually looking at you The AI

implementation would have taken that uncertainty and made the monitoring constant absolute and

fully

automated But here's

the most important limitation to this whole victory for the students MIT is

dropping the AI software layer They are

absolutely not dropping the cameras That is

the crucial distinction that often gets lost in these tech policy victories They are

fixing the symptom which is

the automated algorithmic tracking They are

not

fixing the disease The underlying physical

architecture of mass monitoring remains fully intact and operational across the campus And the physical

scale of this network is massive The campus network

currently has approximately 4 600 active cameras They just

added 575 of those over the past year alone That is

a staggering density of surveillance for an open university

campus They cover critical infrastructure

labs dorm entrances and

vast

exterior public

areas Now the footage is

stored on site for 30 days There is currently no audio recording enabled and there is

no facial recognition being used Right But student

union representatives continue to argue that the sheer overwhelming scale of the network is enough

to chill political expression all on its own even without the AI component The chilling

effect they're describing is a well documented sociological phenomenon You

fundamentally behave differently when you know you are being recorded Even

without

an AI watching in real time a network of 4 600 cameras creates a classic panopticon effect You

just have to assume someone could pull the tape at any moment If

you decide to attend a political protest on the campus quad

your physical presence who

you walked with

and how long you've

stayed is being logged in a secure server somewhere Then that highly sensitive

movement data just sits there for a month The community is being

asked to blindly trust the specific people holding the administrative keys to that server Right They have to

trust

that the governance rules won't be abused MIT's official stance is

that footage is generally only used for incident review or specific criminal investigations But administrative rules can

change overnight Policies can be quietly

rewritten

by a new administration and the physical cameras are

already bolted to the brick walls Which means the true capability

for total surveillance is fully deployed Activating AI monitoring in the

future wouldn't require closing the campus for construction Right It would just be

a

silent software update pushed out over the network The financial and physical friction

of installing the hardware is already gone So what you need to

watch next is how universities across the country govern these existing non AI surveillance systems They have to actively

earn community trust Having restrictive rules written

on

a piece of paper in an administrative office is one thing Enforcing them transparently

and proving they aren't being abused is another entirely Academic institutions have to prove they

are

responsible ethical stewards of this immense data collection

Otherwise the community will always view the

cameras

with deep suspicion Right The googly eyes might eventually come

down

but the underlying distrust of the institution

will remain Next up Anthropic is rolling out text

watermarks for

several cloud models but two older versions contain a built

in bypass This is a highly complex technical

story but it carries incredibly heavy regulatory implications for the entire AI industry It really does It revolves around

a new implementation

of a technology called Synth ID text Let us get into what

actually happened Starting November 3rd Anthropic is

officially expanding this watermarking technology It will cover Cloud

Mythos 5

Opus 4 7 Opus 4 6

and Sonnet

4 6

And the core

goal here is to embed a statistical mathematical pattern directly into the generated text itself We need to

be clear about what this is not Okay This is

not

a visible label or a stamp at the bottom of a generated essay It does not

use hidden Unicode characters embedded in the spaces It is entirely

invisible to the human eye but on the

older models specifically Opus 4

6 and Sonnet 4 6 there is a massive glaring loophole Yeah

If a user simply changes the generation settings in the API away from their default values

the watermark is disabled entirely

Specifically if a developer changes the temperature

the top P

or the top K settings

the system just

turns the watermark off

Wow

And this is not some clever hacker exploiting a bug

It is a fully documented behavior of the implementation on those specific older models

So why does this matter beyond the headline

Because Anthropic isn't just doing this for fun They are implementing this specifically to meet the European Union AI acts looming December deadline

Right

They are trying to provide their enterprise customers with a robust compliance tool

European regulators desperately want a reliable way to identify AI generated content in the

wild They want transparency for

consumers A technical watermark seems like an incredibly elegant solution on

paper which is why it ended up in the

legislation I think we need to unpack how this statistical watermark actually works in

practice because it sounds like

magic It

does How does an algorithm change the text without changing the actual meaning or the grammar of the

sentence To understand that you really have to understand how these models actually choose

words Large language models don't think in

concepts They predict the next word based on massive probability

tables

Right So sometimes multiple words are equally good choices for a given

sentence For example a sentence about the weather could perfectly well end with the word

gray Or it could end with the word

overcast Both words make perfect grammatical sense

Exactly

They convey the exact same atmospheric idea to the

reader

Exactly It adds no latency to the

generation adds no financial

cost it uses no extra

tokens But that API bypass is a huge structural

problem

Yeah It feels exactly like a manufacturer installing a mandatory compliant seat belt in

a car but the seat belt automatically unbuckles itself if you drive over

the speed limit That is a brilliant way to frame the

absurdity of it The bypass exists precisely because of how those specific generation settings interact with

the probability tables Let us take

the temperature

setting for instance Temperature generally controls the creativity

of

the output right If I want a

highly creative poem I turn

the

temperature up Technically it controls the randomness of

the probability distribution A very low temperature makes the model highly

conservative

and deterministic Okay It basically forces the model to always pick the absolute most

likely next word stripping

away the options High temperature flattens

the probability curve making it

much more adventurous It starts picking mathematically

less likely words Wait so if temperature controls randomness and top P and top K do similar things by restricting the pool

of

available words Yes Top K strictly limits the model's choices to a hard number of the

most dynamic cumulative probability threshold They all actively

mess with the

underlying math of word selection So if you change those parameters

you completely disrupt the carefully controlled

coin flips that the watermark absolutely relies on to embed its pattern Exactly The delicate statistical signal just

gets

completely lost in the noise of the altered probabilities Precisely And for whatever technical reason

on

Opus 4 6 and Sonnet

4 6 Anthropic apparently could not

make the synth ID text implementation robust enough to mathematically survive those parameter changes Wow So instead

of

outputting broken text the system just

quietly shuts the watermark off This is a

major caveat for developers who are blindly relying on this feature for

strict

EU compliance If a user tweaks a standard

highly

common

API

setting just to get a

slightly more creative response their mandatory compliance feature vanishes into thin air

Furthermore the detection API the tool you actually

use to check if

the watermark is there it is still locked in a private preview This means

the vast majority of Anthropic's customers cannot

even

verify the watermark is functioning themselves They just have to blindly trust Anthropic that it is

working in the background And even when it works detection only estimates a statistical likelihood

It cannot definitively prove human authorship Plus it is notoriously easily broken by short text samples heavy

code

snippets

or just running the text through a heavy manual

editing pass If you take a beautifully watermarked paragraph and run it through a

heavy editing pass changing a few adjectives restructuring a sentence the continuous statistical pattern is

instantly destroyed It is gone The

underlying reality is

that statistical text watermarking is incredibly fragile

You really need

to watch how EU regulators react to this rollout Can a

mandatory compliance feature truly be considered valid under the law if it

can be toggled off with a simple standard API tweak Regulators generally do not take kindly to massive technical loopholes but they're also being

forced to grapple with the rigid technical reality of how

these transformer

models actually function Statistical watermarking is not a magic regulatory wand It is just math and math can easily be

scrambled

Meanwhile OpenAI CEO Sam Altman is arguing that

scarce computer memory should

live in cloud data centers rather

than

in personal computers This takes us away from specific software features and straight into a deeply philosophical and economic divide in the tech

industry Definitely It is a fundamental debate about the future architecture of

global computing What actually happened is that Altman did

an

interview on the Cheeky Pint podcast He was speaking with John Collison

and he used the opportunity to aggressively push back against the open source movement's

massive drive for local AI inference

The open source community and a growing segment of power users they desperately want powerful AI models running locally directly

on your laptop or your desktop hardware They want decentralized AI

that operates independently of the major tech corporations

Altman

explicitly

argued that this decentralized approach is terribly

inefficient for the planet He stated that the world currently has a

heavily limited supply of DRAM dynamic random access memory Yes He believes it

is a tragic waste of global resources for that premium memory to sit

idle in personal computers

He

specifically characterized lightly used personal machines as utilizing their installed DRAM perhaps 5 of the time over their lifespan I

have to say that is the ultimate billionaire perspective on global resource allocation Yeah He basically views your privately owned

personal property as nothing more than an inefficiently utilized node in a global hardware

network

that he thinks he could manage better Right My car sits idle 95 of the time in my driveway but that doesn't mean I want it absorbed into

a

centralized corporate fleet Well from a purely utilitarian hardware efficiency standpoint

his raw math is not entirely wrong Think about a server rack in

a

massive open AI data center Okay That

hardware is running at absolute maximum thermal capacity

24 hours a day seven days a week It is constantly crunching

numbers

and generating tokens for thousands of users simultaneously Whereas a high end gaming PC sitting in a

suburban bedroom is completely turned off while the owner is at work or

at school or asleep If high bandwidth silicon memory is arguably the most precious and constrained resource on earth right now seeing it hoarded in

millions

of idle desktop machines looks deeply inefficient to a hyperscale cloud provider who's desperate for more

compute But this highlights a massive glaring caveat to his entire argument Altman's 5 utilization figure is a rhetorical characterization designed to

make a point It is absolutely not a hard peer reviewed

statistic across all global PCs And more importantly consumer RAM serves many vital

purposes

completely outside of running AI models

Exactly The PC gaming community was