Anthropic says juniors fell to 9% of tech hires

Welcome to The Signal from Superpower Daily with Maya and Theo

We are tracking a massive shift in the tech industry today

We really are

An internal study from Anthropic reveals a staggering drop in junior technical hires

Basically

artificial intelligence is shifting the core focus of engineering from writing code to exercising high level human

judgment

Right The fundamental reality of technical labor is just changing beneath our feet right

now Yeah it really

is We finally have concrete data showing exactly how these tools reshape daily software

development So we need to look closely at what these numbers actually mean for you and your

career We should definitely start with what actually happened at Anthropic Well the numbers in this internal study are honestly completely

striking They really

are Junior employees made up just 9 of all technical hires by April of this

year I mean that is a massive

plunge That is a huge

drop Right They previously made up 27 of those exact same hires This is not some slow

trend This is a sudden collapse of the traditional entry level

engineering role Well we really have to define what a junior developer

traditionally does A junior engineer usually handles the routine

implementation

work Right They write the boilerplate code

they build the standard unit tests

they fix the well documented

bugs

Yeah the basic

stuff

Exactly And those are the exact tasks that modern AI models excel at

performing The company simply does not need human beings to type out basic syntax

anymore Technical teams are also shrinking as a proportion of the overall

workforce They fell from 61 down to

56 of the total company headcount Meanwhile the

sales

and finance departments grew much faster So that

shift tells a very specific story about the current market It absolutely

does Building the

actual software product is becoming much faster and much cheaper But selling

that product in a highly saturated market is becoming much harder That makes

a lot of sense Companies are

reallocating their capital They are

moving money away from basic code generation and they

are pouring that money into revenue generation and strategic finance So this

data points to a fundamental shift in what engineering managers actually value right now The study

surveyed technical leaders across the company About one

third of machine learning managers reported hiring fewer people specifically because of AI tools And one

fifth of traditional software engineering managers said the exact same thing Those managers

are looking for an entirely different set of human traits today They

are no longer

screening for raw implementation skills Right No more whiteboard

tests Exactly They're not testing

how

fast you can write a sorting algorithm on a whiteboard They are prioritizing autonomy

and they are prioritizing

judgment I think we need

to unpack what judgment actually means in this context Judgment is the ability

to know what to build in the first place It is the ability

to evaluate

whether an architecture makes logical sense for the business The AI can write the code

but the human must decide if the code solves the actual problem Right The actual bottleneck in software

development

has changed entirely The old bottleneck was the speed

of human typing But the new bottleneck is simply

human

attention The anthropic study reveals a fascinating metric

here By August interactive AI agents logged six

times as

many hours as their human handlers That is an enormous and frankly overwhelming

volume of output The researchers calculated something really staggering based

on that Oh yeah Human employees would need half

their entire

workweek just to read what their own AI agents produce This is where we see the cognitive load

shifting Reading and reviewing code is often much harder than writing it yourself Oh absolutely You have to understand the logic

of a

machine You have to verify its security You have

to ensure it integrates with the

rest of the system So because of this massive output the daily

work has fundamentally shifted People are

spending less time building the core

product They are instead building visibility tools They are

building

complex monitoring systems They need these

dashboards just to keep track of what their AI agents are doing in the background We also have fascinating calendar data from this study

You might logically think that automation would clear up

your daily schedule You would hope so But meeting time per employee

actually rose 20 year

over year And for technical contributors specifically meeting time rose 16 That

sounds completely counterintuitive at first

glance I mean if the

machines are doing the work why are

the humans in more meetings The answer basically comes

down to human trust Trust is more

vital than ever when machines are handling the

raw output That makes total sense Face to face coordination becomes absolutely essential

You need to align

on strategy you need to

verify that everyone understands the AI

generated architecture You simply cannot leave high level strategic alignment

to an automated system It feels exactly like moving from being

a line cook to becoming the head expediter in a busy restaurant kitchen I really like that analogy Think about

how

a kitchen operates The AI

is now chopping all the

onions

It is preparing the raw ingredients It is

doing all the repetitive prep work

Right exactly But the human expediter has to

taste the

sauce The human has to manage the dinner rush

And the human has to ensure the final

dish matches the menu perfectly before it goes out to the customer That is a perfect framework for understanding this

shift You are no longer the chef You are

the final quality control checkpoint Right

You're basically managing the flow of automated

labor

But I feel like that raises a very

serious existential question for the entire industry If everyone is just acting as a manager

of AI output who actually learns the foundational skills anymore Yeah

that is the big question How do

you become a good head expediter

if you never spent years chopping the onions yourself That is the central anxiety surrounding this entire

transition You absolutely need foundational technical skills to exercise

good judgment Exactly If you do not understand the underlying

principles

of

software engineering you cannot spot the subtle flaws in an

AI systems logic So we might be breaking the primary pipeline

for creating future senior engineers Right We also need to look at the

productivity

metrics the study measured They tracked 192 engineers who actively adopted coding

agents They compared their output against 74 engineers who

chose

not to adopt the agents The adopters showed a massive five fold spike in

code commits Wow So their raw output increased dramatically A

code commit is basically a saved package

of new work added to the main project Exactly A five fold increase means

these engineers are pushing out new features at breakneck speed But we definitely need

to highlight some very important caveats to this data point The adoption of these tools

was entirely voluntary within the company Right That creates a massive

selection bias Selection bias is a

critical factor here An engineer preparing for

a massive complex new project might decide

to start using agents right at that exact moment Right Their output would naturally

be

higher just because of the project demands Meanwhile an engineer doing slow

delicate maintenance work might avoid

the entirely Furthermore we have to state clearly

that

more code does not necessarily mean better code Generating thousands of lines of code

with an AI is incredibly easy Very easy But making sure those

thousands of

lines are secure efficient and maintainable is incredibly hard

The

researchers absolutely acknowledged that reality

They checked for code removal rates

They checked for code reversal rates

Okay They wanted to see if

the

AI code had to be human written and agent written code on those specific

metrics Interesting But quality has

many complex dimensions that simple metrics cannot capture AI errors can be incredibly subtle Very

true

An AI might write code that functions perfectly in a test

environment but introduces a massive security

vulnerability Those

subtle logic flaws can be much harder to identify during a routine

review We also have to remember

the context of this study This is internal data from one specific highly advanced

AI company Right Anthropic is not a normal company Exactly

Anthropic engineers are building the frontier models themselves They are not

a standard enterprise IT department So

this internal

study

is a leading indicator Yeah It is not

a proven economy wide trend just yet So

this shifts our focus to the future We

need

to watch how non technical staff start using these

exact same tools Right We might soon see

people outside of engineering departments building their own complex software A marketing manager

might build their own analytics tool or a finance director might build their own automated

accounting system That trend could completely alter traditional department

boundaries across entire companies The strict line between technical and non

technical staff is blurring very quickly And that wraps up our

look at the changing nature of technical work at Anthropic Very true

Next up let us look at how these exact AI tools are actually being integrated

into your

daily workflow OpenAI has just added collaborative slide decks directly into ChatGPT This is a highly practical update that

impacts how you likely work every day

OpenAI launched a

new feature called ChatGPT Slides Right You can now start building a presentation from

a simple natural language conversation You can start

from a text outline You can even upload your existing research files And

the system takes all of that context

and generates a full presentation You can also

provide an example deck for the system to use as a strict style

reference Oh that is useful Yeah it will match your company colors It'll match

your preferred typography This

development

breaks the traditional

prompt only paradigm Think about the friction

you usually experience with AI

generation The friction is real Usually you

ask an AI for a slide deck It gets the

content about 80 But

then

you have a massive problem We call that the last mile problem You have a nearly perfect

slide But

you need to move one image slightly

to the left or you need to

change one specific bullet point Under

the old prompt only system you had to ask the AI

to

fix that remaining 20 And asking for that tiny

fix almost always ruins the

entire layout It does The AI decides to regenerate the whole

slide It breaks the formatting that was already perfect It is incredibly

frustrating It

really is It is why many people abandon

AI presentation tools entirely OpenAI is trying to

directly solve that exact

frustration Teammates can

now jump into a shared digital environment called Space This

space acts as a collaborative canvas You can manually edit

the text in real time yourself you can adjust the shapes

and

you can drag the charts around the

screen So you do not have to rely on the

AI to make those tiny

adjustments anymore But the AI is still there in

the background You can highlight a specific chart and ask the AI to update the data

It basically merges rapid AI generation with hands on

human

collaboration It also includes standard shared commenting features It feels very similar to working

in Google Workspace It lives exactly where the creative work

actually happens But the crucial detail here

is exportability I mean a great tool is useless

if your clients cannot open the file Absolutely true

Web users can export their finished presentations

directly to Microsoft PowerPoint They can export to PDF and they can also export

to Google

Slides Desktop application users have an even deeper integration They can

open the generated deck natively

in Apple Keynote or Microsoft PowerPoint The formatting

translates directly into the native application There are

some major constraints you need to keep in mind if you want to

try this today The biggest limitation right now

is access This feature requires an eligible pro account a business account or an enterprise account But having

a paid account is not enough on

its own It specifically requires an account

that has been

granted access to the

space environment Oh I see A standard paid subscription plan alone

will not guarantee you can use this feature immediately It is

rolling out slowly The enterprise

software landscape is

intensely competitive right now We need to watch how quickly large enterprise teams actually migrate

their workflow Most companies are

deeply entrenched in their current systems They're locked into traditional Google Workspace environments They have years of files saved in Microsoft environments

The

switching costs are huge The switching costs are incredibly

high So we will see if enterprise

teams are willing to move their daily presentation work

completely into ChatGPT's native space

environment That makes sense Moving from

human collaboration to autonomous AI teamwork Four AI agents just beat the video game

Minecraft in 27 hours according to new research

from DeepMind This

story sounds like pure science fiction honestly A

team of four Google DeepMind agents successfully defeated the Ender Dragon The

Ender Dragon Right that is the

final boss of the game They used

the

Gemini 4 Argon model to accomplish this They completed this massive task

in 27 hours

and

14 minutes They started from a completely random

world generation They had no pre mapped knowledge of the

environment And most importantly they achieved this with absolutely zero human

input during the actual gameplay The technical approach

here

is

completely fascinating When we think of AI playing a video

game we usually think of computer vision We imagine the AI looking

at the pixels on the screen and

learning to react to visual patterns But these agents did not actually look at pixels

at all Right

They bypassed the visual interface entirely They used a specific JavaScript library called MindFlare Okay This library allows

external programs to interact directly with the underlying code of the

Minecraft

server So the agents interacted with

the game by writing their own code They

wrote

2 783 custom JavaScript programs That is wild They wrote these programs on the fly to

execute their moves They wrote code to understand their physical environment They

essentially programmed themselves to play the

game in real

time That approach avoids the massive computational cost of processing video

frames It allows the agents to focus entirely

on logical reasoning and strategic planning It makes a lot of

sense And they demonstrated incredible spontaneous coordination This is where the

story gets truly wild The agents exchanged 229 chat messages with each other

during the run Wow They were

actively talking to coordinate their actions

They shared specific notes for map coordinates so

they could find each other in the massive digital world They divided

their

labor completely naturally No human programmer told them

how to split up the tasks Really One agent autonomously decided to scout the surrounding area for resources

One agent went deep underground to

mine for diamonds Another agent stayed at the base and smelted

iron ore That is amazing The fourth agent focused on prepping

the final portal They even exhibited a form of

digital teamwork when things went wrong When one agent died

in the game

and dropped all its videos valuable loot the other

agents coordinated to recover that loot before it disappeared They basically protected the team's resources But we need a serious reality check here

This

is an

incredible technical achievement but they died 605 times before actually winning

the

game Yeah that is a massive

failure rate It highlights the difference between a

game environment and the real world

It really does Think about the enterprise applications If an

AI

agent fails 605 times while doing

your company's payroll you have a massive financial and legal problem A

video game allows

for infinite retries The real

world does not So how does a video game success translate to

the rigid zero tolerance demands of enterprise software

You

make an excellent point The enterprise translation is the hardest part The

researchers clarify that this is strictly a sandbox environment A video game has clear rules It has

clear boundaries It has an

objective scoring system It is not a

reliable workplace simulation by any means There are other important

caveats

to this research too The four

agents were not entirely

alone A fifth agent acted as

a silent observer This observer monitored the entire run and

provided high level feedback Human researchers also retained

ultimate

control They had the ability to pause or stop the servers at any time They

actually

paused the game six different times during the run They

paused for about 20 minutes total

to manage server issues and monitor the system So it was not a

completely

unbroken autonomous run We need to watch whether this type of

long horizon coordination can hold up in the real world We need to see if

agents

can coordinate multi step plans when the goals and

consequences are much less contained than a digital video game Can they negotiate a contract Can they manage

a supply chain disruption Those are the real tests Meanwhile in the realm of high level physics Anthropic says its CLOD model just cut

a

quantum circuit's gate count 17

fold This involves some intense mathematics and

physics CLOD Mythos V was

tasked

with optimizing eight different quantum circuit

designs It

used a specific propose and test loop written in the Python programming language

In one notable point addition task the

model successfully cut the required gate count 17 fold We need to

explain

why this actually matters for the future of computing Quantum computing is currently constrained

by very expensive physical operations In

classical computing operations are incredibly cheap and stable But

in

quantum computing operations are physically difficult to maintain These operations are called gates

Specifically operations like Toffoli gates are incredibly resource intensive Right

the

Toffoli gates A

Toffoli gate is a fundamental logic

gate that interacts with

multiple quantum bits at once Keeping

those quantum bits stable during

the operation is one of the hardest challenges in

modern physics Every single

gate you can remove from a circuit design brings us one step closer to a practical

working quantum computer CLOD ran an automated search for 53 hours 53 hours wow It continuously proposed

new designs and tested them It scored these different circuit designs based on a specific mathematical formula It multiplied the total

number of qubits by the total number of gates

It wanted

to

find the lowest possible score The AI system

found a highly creative classical method for modular inversion Modular inversion is

a complex mathematical operation used heavily in cryptography Okay The model used an

algorithmic

shortcut called div steps The fascinating part is

how it handled memory The model basically reconstructed its own scratch work

on the fly It chose to recalculate data rather than storing

that

data in memory It actually invented a completely new

reversible method to achieve this In quantum computing

memory is extremely scarce A qubit is a precious resource Yes

So recalculating information is often more efficient than trying to store it

CLOD figured this out autonomously It did all of this without

any human intervention

between the testing rounds

It also achieved results on existing

human designs It reduced the gate count for a specific Google designed

quantum optimization decoder by a

factor of 12 Wow It looked at the work of top human physicists

and found massive optimizations But there is always

a tradeoff with these complex optimizations It is exactly like finding a massive shortcut on a cross

country

road trip How so The shortcut saves you days of driving but you

have to rent a slightly larger more expensive car to

take that specific route That is the perfect analogy for the space time

tradeoff happening

here The system brilliantly cut the gates

by 17 times Right But it required 12

more qubits than the smallest prior circuit design

It basically traded physical gates for physical memory That is a

crucial limitation Adding 12 more qubits to a physical

quantum

machine is incredibly difficult right now Exactly We also have to look

at the metric used to score the designs

The formula of multiplying qubits

by gates is just a cost proxy Right it is an estimate The researchers

did

not measure the actual physical execution time of the circuit on a real

quantum machine These are also highly task specific improvements They optimized a

very specific mathematical operation Right

Optimization is not proof that massive fault tolerant quantum computers are suddenly arriving tomorrow morning We are still

years away from that reality We

have to watch whether these lower theoretical

resource

estimates translate into actual physical performance gains The map looks beautiful on paper But we

need real results Exactly We need to see

these newly optimized designs run successfully on real physical quantum hardware in a laboratory We are

moving to our Quick Reads section

Mark Zuckerberg publicly explained exactly

why

he hired Alexander Wang to lead the new Meta Superintelligence Labs The hiring process was incredibly brief They

had a single 44 minute phone call Following

that call Wang sent Zuckerberg a detailed memo focusing entirely on the concept of talent density Meta

then made a massive financial move

They

bought a 14 3 billion stake in

scale AI 14

billion wow That massive purchase gave Meta a 49 share of the company They

made this multi billion dollar move primarily just to get Wang on board as leader Zuckerberg

chose Wang

specifically

for

his recruiting gravity Wang built scale AI into a data powerhouse Top engineers want

to work for him Zuckerberg did not choose him because he had a traditional frontier model

research resume He chose him because he can

attract the smartest minds in the world The

core research

team at Meta is remarkably tiny compared

to their budget The team is closer to just 100 elite people Really Yet

Meta is deploying an astonishing amount of capital They plan to spend between

130 billion and 145 billion in capital expenditures in 2026 alone Putting 100

people in charge of 145 billion is the ultimate

example

of talent density But

the financial stakes are astronomical

They are Chief Financial Officer Susan Lai issued a stark warning to investors She explicitly stated

that building a bad AI model would carry severe perhaps unrecoverable financial consequences for the

entire company In other news Google

AI agents

successfully designed new DNA aptamers in a laboratory environment Okay An

aptamer is a short strand of DNA designed to bind tightly to a specific target molecule You can think of

it like a

highly specific chemical key fitting into a biological lock But the initial

physical

laboratory testing proved something fascinating The AI optimized the completely wrong chemical signal Specificity plunged from 75

percent down to 19 1 percent The DNA candidates were physically sticking to the plastic screening

beads used in the lab They were not sticking

to the actual target protein they were designed

to

find Oh wow The AI found a

mathematically correct but practically useless solution It optimized for

the testing environment instead of the biological goal But the agents learned directly from that physical setback They received the lab

data

They

revised their internal design rules entirely They ran the process

again and successfully recovered specificity back

up to 72 4 percent

This shows an AI system actually grounding itself in physical laboratory reality It learned

from

physical failure That is a big deal

The entire process was heavily aided by Veridot software That specific software acts

as a digital twin of the physical laboratory environment It prevents the

AI controlled physical robots from

making impossible real world moves

and causing expensive damage to the lab equipment Meanwhile

Goodfire researcher Nick Camerota is challenging a massive assumption about how AI actually

works OK He's challenging the prevailing idea that neural networks are just a tangled unreadable mess of completely unrelated

concepts

AI researchers currently call that tangled mess superposition Imagine a dictionary where every single page

has five different words printed directly on top of each other Right

That is how

researchers currently view the internal memory

of a neural network It is dense and

tangled Camerota argues that a single AI neuron firing for both a picture of a cat face and