Meta adds project workspaces to its AI agent

Welcome to The Signal from Superpower Daily with Maya and Theo

Today we are looking at Meta turning its Muse AI into a persistent project manager that updates itself while you are away

Yeah

We have a really packed daily briefing today

We are going to unpack this huge shift going from disposable chat to actual living workspaces

Right

and we will also examine a pretty massive reality check for AI in the laboratory

Plus a new milestone in mathematical reasoning

And we will look at some honestly extreme solutions for data center power constraints

Yeah

So let us jump right in

Sounds good

So for our lead story today

Meta has added dedicated project workspaces called Spaces to its Muse AI agent

Right

These Spaces allow connected apps to keep planners and trackers updated in the background

and they do this without needing new prompts

Which is a huge shift because

I mean we have spent the last few years treating AI like a really capable search engine

Yeah

exactly

You

encounter a problem

you open a chat window

you type a prompt

the machine gives you an answer

Right

And then you close the window

Exactly

That entire interaction is just transient

The machine forgets about you the exact moment you navigate away

Well

Meta is essentially trying to kill that transient model

They are shifting to a persistent architecture

Which is a pretty crucial evolution for software builders and for operators

It really is

So Meta launched this new feature as a sidebar of topic focused project

areas These live directly within the Muse

interface Okay So you are no longer just getting a block

of text You are getting

a customized

stateful environment

Stateful environment meaning it remembers what

happened

yesterday Right It tracks ongoing changes

over time So let us walk through how this actually feels for

the user Say you go to

the agent You describe a specific

ongoing need Like planning a complex family vacation across multiple

time

zones Right Or maybe you are managing a massive do

it yourself home renovation You feed

Muse that initial context And

Muse takes that description It builds a bespoke

tool on the fly

Like a dedicated tracker Exactly Or a calendar planner Or even

a

Kanban style project board Okay I mean

the initial creation is impressive But that is

not

really

the breakthrough here

No it is not The real breakthrough

is the background automation Right The

system connects to third party

applications It fetches new information automatically And it does this

while

you

are not even looking Wow Yeah So if

a band changes their

concert dates your entertainment space updates If

flight prices drop for your vacation the

travel tracker reflects that new price If your bank account registers

progress toward a savings goal the

financial

board just updates automatically Exactly

It is all happening silently You do not have to write a new

prompt asking for an update

You just open the space And the new data is already

there

just waiting for you Right And this matters immensely beyond the headline because it changes the

entire paradigm of digital assistance We are moving from a

reactive model to a proactive model Precisely Right now

most AI requires your constant intervention You have to actively

pull the information out of it

Let me try an analogy here The current AI model is sort of like hiring a

temporary

worker for a single afternoon Okay You

give them a very specific task they do

it they hand you

the

result and then they leave Right They have zero context about what your business

actually does on a day

to

day

basis That is a perfect comparison Yeah But

what it is building feels entirely different It is more

like

hiring a dedicated chief of staff Right

Someone

who sits in on your meetings Exactly They

have the keys to your filing cabinets You give them a

project in January and then in March they tap you on the shoulder

and

say the vendor changed their pricing Yeah And you never even asked them to check They just knew it

was their job to monitor the situation

Right So how does it actually monitor that Well that proactive

monitoring

relies on application programming interfaces or APIs Right The AI is essentially

establishing permanent communication channels

with

your other software It

listens for webhooks Okay Let us define that A webhook is just a signal from

one app

to

another that an event occurred Exactly So the flight price drops the airline app sends a webhook

the muse space catches it And then the AI interprets the data

and redraws your

project board You got it I have to say I have some serious skepticism

about handing

over those keys Oh absolutely I

mean I understand the convenience but if I link my apps does this system just start rearranging my

digital life without my supervision

That is the big question Because there is a massive difference between

an AI

tracking a flight price and an AI deciding to reallocate my budget

because of price changed Yeah You are hitting on the core tension

of autonomous agents here It is

the balance between

helpfulness

and control An agent with too little autonomy is just a fancy

notification system but an agent with too much autonomy becomes a

massive liability Have they said how they are

handling permissions We do not know exactly how MEDA will handle that yet The exact friction points

are unclear Like will it ask for approval before updating a shared

family

calendar Right We just do not have those

answers yet The examples we do have are

very specific Mark Zuckerberg cited a few personally used cases during

the rollout He mentioned building a

barbecue

game He talked about creating an event

tracker He even mentioned using Spaces as a home control

dashboard Yeah but we have to place a giant asterisk next to those examples

Why is that

Because those are anecdotes from the chief executive officer They are not rigorous independent tests of the

platform Oh

right Of

course A CEO testing a product built by his own engineers is always going to have a smooth

experience The engineers ensure the integrations work flawlessly for

him That makes total

sense We have not really seen widespread third party

validation yet No we

have not We also need to talk about

adoption numbers MEDA is heavily emphasizing their

existing

scale Right Zuckerberg noted on their October earnings call that the Muse agents had

already reached 10 million

users That is across the United States and

Canada And that milestone happened just two months after the initial

debut which is

huge Ten million users is a formidable number for a new

product but we really must be precise about what that metric actually

represents

That figure is for the Muse agent as a whole

It is an installed base

It is basically people who have tried the chat interface

So it is not a metric for the new Spaces features specifically

Exactly

It

tells us they have a massive funnel of potential users but

it does not tell

us if anyone is actually trusting the AI to manage their persistent projects Which brings me

to

what I think is the biggest vulnerability of this entire concept Let us hear it

The practical value relies

entirely on the reliability of those third party app connections Oh absolutely That

is the

fragile foundation of the modern internet Right Because APIs

break constantly A company updates their security protocols or a data provider changes their rate limits And suddenly the

connection

just fails Exactly If those background

updates

break your space becomes worse

than useless It actually becomes misleading

Yeah You might look

at

your travel tracker you see an old

price and you make a

bad financial decision based on that And the AI might

not even know the connection is broken Software engineers call that

silent failure It is the most

dangerous kind of failure Because the system looks

like it is working but it is feeding

you stale data So Meta has to

prove their architecture is resilient against these inevitable API breakages They have to design

a system that gracefully alerts the user when a data feed goes dark Exactly So what should

builders

and investors watch next I would say watch

the reliability of the ecosystem Observe whether these background

updates actually survive the chaos of the real internet Do these workspaces remain

useful between visits or do they just

degrade into broken dashboards full of error messages Right Because if Meta

solves the integration fragility they will basically

own the user's digital workflow But if they

fail spaces will just

be another forgotten tab in the browser Well said So

we are

going to explicitly close out the lead story here Moving from personal

projects to the laboratory OpenAI and biotech

company LIGO have introduced a 100 task benchmark This tests

AI on raw messy experimental

evidence GPT 6

Astra scored 66 7 of available credit

It showed real progress

but it still made critical errors that could

completely ruin experiments

This is such a fascinating collision

You have pristine digital intelligence hitting sloppy physical reality

It really is

We are so used to testing AI with clean data

Current

benchmarks often use polished curated

materials They

feed the models final figures from published scientific papers But

those published figures are heavily edited They

are cropped They are color corrected Yeah

They

are

explicitly designed by humans to communicate a single clear finding to other humans Exactly And

real

science does not look like that Real science

is completely ambiguous It is

incredibly messy Yeah And

that

is the exact gap the Biobasics Bench tries to measure It is

a

new evaluation framework It tests AI models using real everyday wet lab data

A wet lab

is where actual biology happens It involves liquids

It involves chemicals

and actual biological organisms Right So the

AI

has to look at raw gel images It has to

analyze plates covered in bacterial colonies It has to

process complex microscopy files I read it

even has to watch videos of scientists physically moving fluids around round on

a lab bench It does

And the

scale of this test was massive The researchers

generated 18 500 distinct results Yeah

They

evaluated seven different AI models and

they pushed the models across various reasoning effort levels So

the standout performer was OpenAI's GPT 6 Astra

Right

It earned a score of 66 7 at the extreme high reasoning

setting

They call that the X Phi setting

Exactly

It forces the model to spend more computational time thinking before it actually answers

And Astra comfortably defeated the older model that was GPT 5 6

Sol Sol only managed to score

43 8 So let us talk about what Astra actually

did well because it showed some surprising

deductive

skills Yeah It proved it can spot bad

experimental controls Which is a critical function of a

lead scientist You have to know when your baseline is

fundamentally flawed Right So in one

specific cloning task the AI was asked to evaluate a plate of bacteria And Astra correctly rejected the plate as

a reliable control How did

it actually know It read the physical handwriting on the side of

the

plastic dish Yeah It noticed that the label named the

wrong antibiotic That

is

incredible

Yeah And the older model Sol just completely missed that

detail Exactly Sol just looked at the center of the

plate It saw no bacterial colonies

growing So it concluded the negative control was

perfect It completely failed to realize the entire chemical environment was

wrong

Right Astra also demonstrated an ability to recognize when an image simply lacked enough evidence to make a

claim This happened in a cell imaging

task

right

Yes The researchers were looking at cells treated with specific fluorescent

dyes

Okay The older model saw bright

spots It immediately assumed those spots indicated cell

death But Astra analyzed the same image and

paused It realized a second background dye was interfering with the primary

signal

Right The visual data was

contaminated And Astra did not just

guess It actively called for better controls It told the researchers they needed a

different setup to distinguish the competing signals Recognizing insufficient

evidence is a real hallmark of scientific maturity It shows

the AI is moving away from the need to please the user with a definitive confident answer

It is

learning

to say it does not know Which sounds great

on paper but I

have to push back hard on this scoring system Go ahead

The 66

7 score is based on a rubric It awards partial credit Right

The

model gets points for showing its work It gets points for correctly identifying a visual feature It

gets points for logical intermediate steps Okay

But

what if it gets all that partial credit and then makes one fatal logical leap at the very end You

are describing the exact danger of rubric based evaluations

Because if the model tells a scientist to synthesize the completely wrong DNA sequence

the entire experiment is ruined

Right

It does not matter if the AI correctly identified the antibiotic label earlier

A 66 score implies it is two thirds of the way to being a reliable lab

partner But in

biology a single wrong variable means a 100 failure

rate

Exactly So does this scoring system create a dangerous false sense of

security That is the most important caveat of this entire

study We must explicitly state what that number actually

means

The 66 7 score does not mean Astra successfully completed 66 7 of

the experiments It means it accrued partial credit across

expert written rubrics Is this sort of like giving a

chef partial credit How

do you mean Well the chef buys

the perfect ingredients they chop

the vegetables flawlessly they season

the

meat perfectly and then they put the entire meal into a blender instead of

the

oven Right The meal is

completely

ruined Exactly The process was

mostly correct but the outcome is a

total disaster And Astra did exactly that in

some cases It made fatal

mission ending

errors Like what In

another cloning task the AI needed to identify the

correct DNA product It correctly saw the visual groupings

on

the jaw It extracted the visual

data perfectly but then it totally reversed

the logic It assumed a larger DNA product

indicated a successful

synthesis Exactly It overrode the logical comparison supported by the control

samples It extracted the right visual

facts and then it hallucinated the wrong biological

rule The downstream effects of that error are

massive If

a human researcher blindly followed that AI advice they

would purify the wrong plasmid And

a plasmid is just a small circular piece of DNA used to carry genetic information Right

If

you grab the wrong one you're

working with the wrong molecule entirely Weeks

of subsequent lab work would just be wasted You

would be trying to cure a disease with the genetic equivalent of a blank piece of paper That

is terrifying And

the models also struggled with basic visual tasks didn't

they They

did AstroMens

counted bacterial colonies by about 12 It missed clearly visible colonies It

failed to separate colonies that were growing close together And

most concerning of all It

hallucinated colonies Right

It

marked completely empty agar gel as containing biological material So

it is seeing things that are not physically there Exactly

It

is not just misinterpreting data It

is literally inventing data Correct

The

models are making undeniable progress in reading experimental evidence Sure

But

a correct visual observation does not guarantee a correct scientific interpretation The

gap between seeing a spot and understanding what that spot means chemically is still vast So

what is the main takeaway for the scientific community here What

do we watch next Watch

whether these reasoning models can cross that specific cognitive gap They

need to move

from making a correct visual observation to forming a reliable correct scientific interpretation And

they

must do

this without hallucinating the outcome Exactly Until they bridge

that

gap benchmark gains do not

equal dependable lab autonomy A high score on

a rubric means absolutely nothing if the resulting plasmid is wrong Next up in theoretical

computing Physical Superintelligence claims its

AI produced a formal mathematical proof for the 18 year old Aronson Ambonis quantum computing conjecture The

proof was accepted by the verification software Lean and

it was published for public review We

are shifting from the muddy ambiguity of biology to the absolute precision of mathematics here We

really are This

story represents a potential leap in artificial reasoning Right

So

Physical Superintelligence or PSI they

are an AI research company They

ran an unreleased anthropic model on their own private infrastructure And

they let this model run continuously for days It was not a quick prompt and response situation Right

On the third day of computation

the AI produced a full formal mathematical proof

And this is not a trivial math problem either

It addresses the fundamental limit on how classical computers simulate quantum computers

To understand why this matters beyond the headline

we have to look at the Aronson Ambonis conjecture

itself

Right Scott Aronson and Andres Ambonis proposed this problem back in

2008 It is kind of a holy grail in computational complexity

theory It deals with query

cost

Yes Let us break that down for the listener

What is a query in this specific context

Okay

Imagine a computer is trying to solve a puzzle

The puzzle is hidden inside a massive

black box of data Okay

The

computer has to ask the box questions to figure out the answer Every

time it reads a piece of data from the box that

is one query Got

it Now a classical

computer like the laptop on your

desk has to ask those questions

sequentially Right It checks one piece of

data then it checks the

next But a quantum computer behaves

differently Very

differently Because of quantum

superposition it can theoretically query many different pieces of data

simultaneously It can explore multiple paths at

once Okay So the conjecture asks a very specific question

about that

It does For certain types

of random

unstructured data just how much faster is the

quantum computer It is trying to find the mathematical limit of that

quantum

advantage Yes The 2008 conjecture predicted that a classical algorithm can always approximate the quantum algorithm's answer without needing exponentially

more queries Exponential is the key

word there We are talking about the difference between a hard problem and an

impossible

problem Exactly If the gap

is polynomial the classical computer

is slower but it

is manageable Right But if the

gap is exponential the classical computer would need more time than the lifespan of the universe

to catch up And PSI claims their new theorem provides an explicit polynomial bound on

those query gaps So it definitively rules out exponential quantum advantages for certain

unstructured random inputs

Yes it does I want to be very

careful here though Does this mean quantum computing is

a

dead end No not at all That is a vital

distinction to

make This proof only applies to unstructured random

data Like looking for a needle in a completely randomized

haystack

Exactly But quantum computers excel at finding hidden

patterns

right

Precisely Algorithms that exploit hidden mathematical structures still hold their exponential

advantage The most famous example is Shor's

algorithm

Right It factors massive prime

numbers It relies on finding hidden periodicity in the

math

And Shor's algorithm remains exponentially faster than any known classical approach

It does

This new proof does not touch that at all

It only places a

speed limit on random inputs

OK

Let us talk about how this AI proof was actually verified Because mathematics has a real hallucination problem with AI

It really does

Large language models are notorious for writing proofs that look convincing but contain massive logical flaws

Right So PSI tackled that problem by using

Lean

Yes Lean is an interactive theorem prover It is a highly rigorous software

assistant You do not just write a proof in

English You write it in a specialized programming

language And Lean checks every single logical step against the fundamental axioms of

mathematics It ensures there are no missing

steps It guarantees the logic compiles perfectly

Exactly

And PSI published this full formal Lean code on GitHub

They are not just asking us to take their word for it

Right

They provided the artifact

They did

They claim the AI even improved the proof twice after being fed related open problems during the process

But this process raises a massive red flag for me

OK

What is the red flag

The AI wrote a proof

A software program called Lean verified the proof

Right

How do we know the software did not just rubber stamp a hallucinated logic loop

It feels like two robots agreeing they are right while the humans stand outside the room

That is a very necessary skepticism

Lean is incredibly strict

If Lean says the code compiles

the logical steps from A to B are valid

OK

But software cannot evaluate the philosophical meaning of the code

It checks the grammar

not the truth of the premise

Exactly

The formal steps might be technically valid within Lean's specific constraints

But the broader logic could contain a subtle flaw

Right

The initial setup of the problem might be wrong

The

AI might have proved a theorem that is slightly different from the actual Aaronson M Bainis

conjecture Which brings us to the most important caveat of

the story This work has not been peer reviewed

by human mathematicians As of November 8th it is entirely self

reported by the company It is basically a corporate account

of an internal experiment Right It has not survived

contact with academic scrutiny Human experts have to translate that dense Lean code back

into high level mathematical concepts They have to ensure the AI

actually solved the right problem So what is the

next step for this breakthrough The next

step is independent mathematical review We

must watch the academic response Top human mathematicians are going to

download that code from GitHub They are

going to break it apart They

will look for edge cases They

will search for definitional errors We need to see if the AI's logic holds up when brilliant

humans

try to destroy it

Exactly If the proof survives AI has officially entered the business

of generating novel world class mathematics Transitioning from theoretical

computing limits to physical computing constraints Elon Musk promised to reveal details