Google’s Gemini builds practice worlds for robots

Welcome to The Signal from Superpower Daily with Maya and Theo Today

we are looking at Google's Gemini 4 Argon It is an AI that is basically hallucinating practice worlds for robots Right

So a robot dog can practice opening your trash can virtually instead of breaking the real one in your kitchen Exactly

It

is wild Yeah

I

mean we really need to unpack the mechanics of this The implications here are just massive

They

really are Google

just rolled out Gemini 4 Argon to paying users So

that is their AI ultra subscribers and Gemini API developers

right

Yes

And they dropped this series of demonstrations showing a completely end to end

robotics workflow Which is a

huge deal It blows past this old concept of an AI just acting as

a pilot You know just

receiving

commands Right Because the word pilot implies there is a pre existing vehicle in

a

pre existing world But Argon acts much more like

a reality architect I

like that framing Well the core problem in robotics has always been

the sim to real gap Right Going from the

simulation

to the real world Exactly You can train a robot

flawlessly in a digital simulation But the moment you put

it in

a physical room it fails

Because the

real world

is messy Totally

messy The lighting changes The friction of the floor

is just slightly different A

shadow

confuses the camera Yeah And

the robot falls over You just cannot

hard code every single variable So developers used to spend months building these hyper

realistic

digital twins of physical spaces Right

But Argon bypasses that entire process Because

it

generates the training environment itself Exactly It uses whatever the

robot is looking at right now So let's look

at that Boston Dynamics

spot

demo The yellow robot dog Yeah The one we

have

all seen dancing on YouTube

Right So in this specific task Spot

needs to balance on three legs And it has to use its front leg to step

on a trash

bin pedal

Which is hard Very hard If you just throw a reinforcement learning algorithm at

that in the real world it is

going to

fall over repeatedly Oh yeah It

might break its own leg It will

definitely

dent the trash

can Right Because hardware degrades You cannot run 10

000 physical trial and error loops You would

be replacing servo motors

all day

It is incredibly expensive So instead

Argon

builds

a virtual practice world first It uses the robot's cameras to generate a virtual version of

that specific trash bin And Spot practices the balancing act

inside that

generated

simulation Thousands of times Right It figures out the exact center of gravity

It calculates the necessary torque And then it executes the

action on the real hardware

But here's the crazy part It also evaluates the training footage afterward

to see how it did Right That evaluation

phase is

the real breakthrough It is We are witnessing

a comprehensive closed control loop

So you have

simulation creation direct

physical control and automatic feedback It is literally

grading its own training data I want to push

on

that self grading aspect actually Sure Because how does

an AI actually evaluate

physical

failure What do you mean

Well if I drop a glass I

know

I failed because it's shattered Right How does a language model know a robot arm failed to

grasp

a piece of plastic properly Well let's look at

the

robot arm demonstration they showed Argon was tasked with dropping a

Lego brick into

a

box Seems simple enough Right But on

the first attempt it fails The brick bounces out or

it

misses the box entirely Okay Argon's prompt instructs it

to inspect the system logs And it also inspects the visual

camera images after a failure So it is looking at lines of text code and a video

feed

at

the same time Exactly Wow

It is processing multimodal data Wow It looks at the joint torque logs to see exactly how much

force

was applied by the motors Okay And it looks at the camera feed to

trace the trajectory of the brick So it

combines

those two distinct data streams Right It identifies

the root cause of the error And then it actually devises

a fix in its own

code It rewrites its own parameters Yes It verifies

that fix in

its head essentially That is wild In

the demo across just three iterations Argon dropped the physical execution time from almost a

minute down to just 17 4 seconds And it successfully landed the Lego brick

right in the center of the box It adjusted its own physical

motion parameters

based purely on digital feedback That feels like a

massive leap

in spatial reasoning It is But the Party Props demo is

where things

get truly weird Oh

yeah This one was fascinating So Argon takes a

single photo of a table And the table is

covered in these

random

items Highly irregular things Right Like they had a

soft

plush cap on there Right Deformable objects are basically the

holy grail of robotic

manipulation

Because they squish Exactly If a robot picks

up a steel cube The geometry of

that

cube never changes The

corners

stay put Yeah If you write an equation for it it works every time But a plush cap changes shape the absolute millisecond

you touch it The center

of mass shifts

The material compresses Traditional inverse kinematics

completely falls apart Yeah Argon takes that static two dimensional

photo

of the table and it somehow turns it into a

virtual three dimensional replica

Just from those pixels And from that it devises 77 distinct tasks for the robot to do

with those random props It generated over 4 500 training episodes based

entirely on that one snapshot Which means it is inferring the 3D geometry and the physical properties of those

objects from a flat image Yes It looks at a

shadow and it deduces depth It looks at the texture of the cap and just makes a statistical guess

about how deformable it is It builds a localized physics

engine for that specific table And Google reported this simulated training drastically improved

the real robot's performance speed It really highlights the

power of that feedback loop Let's

talk

about the

origami demo

actually Oh right Folding paper Folding

paper

requires incredibly delicate tension You have a dual arm robot

trying to coordinate a

fold Which usually tears the paper Right So Argon watches

the training footage of the folding sequence It divides

the

continuous video into individual discrete steps And then it flags the pauses It flags the exact

failed grips where the paper slipped

So

it is performing error attribution Yes By identifying exactly

which micro movement led to the failure the model learns what

to penalize during its next training cycle It

is actively curating its own dataset It throws out the bad

habits and reinforces the successful trajectories I read it even

evaluated

footage of a human sorting rings Yeah It

watched a person fumble with the rings

It identified where

the human stalled or reversed progress And it extracted physical lessons

from

our human trial and error OK

I have to jump in here Because I am always deeply skeptical

of these

highly curated vendor demonstrations Of course

The demos always look like magic But the actual benchmark scores

tell a much

more grounded story They usually do The contrast between cognitive reasoning

and

actual physical control is sharp here The benchmarks reveal the

exact boundary of current AI capabilities

So let's look at the numbers Argon scored

87

on Google's own ERQA test Right That stands for Embodied Reasoning Question Answering

It is a massive 400 question test It

focuses purely on spatial and physical reasoning And Argon

beat Claude Opus 5 5

which scored 82

3 It also beat GPT 6 Astra which scored 81 So if you ask Argon

a theoretical question about how objects interact in

a room it is the smartest model

on

the planet But and this is a big but Yes Then

you look at the AA Robotics

Benchmark from Artificial Analysis Right This is not a written test This test requires the model

to actually control a real

robot arm across

eight distinct tasks And how did Argon

do

It

placed third It scored 55 out of 100 points Yeah GPT 6

Astra placed first with 69 points and GPT 6 1 Sol placed second It just proves that

reasoning

about a physical space is entirely different from actuating hardware within it Exactly Think about the

difference between passing a written driving test

and

parallel parking a real car On a busy street in the

rain Right The written test just

proves you memorize the rules of geometry you know the right of way

But the physical act

requires translating that abstract understanding into continuous

motor control Under totally unpredictable conditions Because the physical world

has variable friction The lighting changes when

a

cloud passes over the sun The sensors

have microscopic noise in them Right The written test assumes a frictionless

vacuum This is exactly

why we need to scrutinize the human guidance

required in these demos The

environment in

those Google demonstrations was heavily heavily

structured to

guarantee success Highly structured Look at the

cup writing example Oh yeah Google proudly called it a zero shot task They

implied the model had no specific prior training to

flip

that exact cup But Argon wasn't just thrown into an

unfamiliar room No It was

fed

a

camera image that had

a literal centimeter grid digitally overlaid on it Which

is

a huge crutch A grid completely removes the burden of spatial estimation Right

The AI doesn't have to figure out how far

away the cup is The human engineers

already

provided a precise coordinate system It was also given ready made

high level commands Right It didn't have to invent the concept of

grasping from scratch It was given a function literally called

grasp and lower until contact The

simulation prompts even contained long lists of forbidden shortcuts Which is so funny

to me The human

engineers

explicitly told the AI what

physical

hacks not to use Because of reward hacking Right AI models are notoriously lazy Right If you tell

an AI to move an object

from point A to point B in a simulation it might

just

exploit a glitch It figures out how to

teleport the object through the table Exactly So the engineers spend massive amounts of time writing

constraints to force the AI to obey

the laws of physics

The biggest red

flag

for me though is the missing data The failure rates Yes Google did

not disclose the number of real world attempts behind these

videos They didn't publish the failure rates behind their success metrics We literally do not know if it took one try or 1 000 tries to

get that Lego

brick perfectly in the box for the camera

And that matters A marketing video shows the one

triumphant success But real engineering requires knowing the exact failure

rate If an automated forklift succeeds one time

out of ten it is completely useless on a factory floor It

is definitely not ready to fold laundry in your living room The reliability has

to approach 99 9 before you trust a

heavy machine around humans So if you are building in this space you

really need to watch how reliably Argon performs across

unguided real world attempts We

need to see it operate without

the crutches Take away the centimeter

grids Take

away the pre written command libraries Let's see how it handles

raw unstructured chaos

You also need to watch the underlying economics of this

system Oh the compute costs The compute required to process continuous

sensory data is astronomical API access to Argon

is currently 2 per million input tokens And

it is 10 per million

output tokens But Google explicitly stated those are introductory prices Right They double after November 30th A robot operating

in real time is processing 24 or even 60 frames

of video every single second Constantly

streaming high resolution visual data into the context window And generating continuous

motor control outputs You

will burn through a million tokens in a matter of minutes

When that bill

doubles developers are going to have to do some very hard math They really will The return on

investment for an AI controlled robot versus a traditional automated

machine might not make sense yet So physical friction is

definitely slowing down the hardware

side But if we look at

the communication bottlenecks between humans AI is moving incredibly fast That brings us to Meta

Time Magazine just named Meta's

Omnilingual ASR as one of the best inventions of 2026 This is

a fascinating pivot Moving from AI navigating physical space to AI

bridging

human communication gaps Meta first published the research for

this back in November 2025 Right It is a completely

open

source speech recognition system

The sheer

scale of it is what caught Time's attention honestly It supports over 1 600 languages And

more than 500 of those

languages were previously completely unserved by any speech recognition technology They

were just digitally invisible Historically developing voice

AI was a game of extreme

data wealth What do you mean by data wealth To

build a robust speech

to

text model you needed

a massive lake of paired audio and transcripts Like perfectly labeled data Yes

Languages like English Mandarin or Spanish have billions

of hours of transcribed audio floating around the internet So the

models train easily on that abundance But smaller regional

languages simply don't have that digital footprint

They

get completely left out of the AI revolution Because commercial AI companies operate on profit margins Right It is not profitable to spend millions of

dollars

building a bespoke acoustic model for a language spoken by

just a few hundred thousand people Exactly So

Meta fundamentally changed the underlying architecture to solve this They didn't just try to scrape more data No They built a system

where a community can add an unserved

language

using just a handful of their own data samples And this relies heavily on self supervised

learning right Yes

The model doesn't need

perfectly transcribed audio to learn the fundamental structure of human speech It just listens It listens to raw

unlabeled audio and learns to

predict the acoustic

patterns It maps the phonetic space of human sounds universally And once it has that foundational map It only

takes a tiny amount of labeled data Like

a

few hours of a

native speaker reading text Just to

align those sounds to a specific language Yes It democratizes

voice technology It is massive for language preservation

But

Meta also really focused on how this gets deployed The hardware constraint

is the whole ballgame here Right The model family includes highly compact variants specifically designed

for low power devices And they also released massive parameter models

for cloud servers where you

need maximum accuracy But if your goal is

to deploy this technology in real time rural areas

or developing nations you cannot rely on

a constant high bandwidth connection You can't depend on a server farm in California Right The model

has to run locally It has to fit

on a cheap smartphone or a basic medical tablet running on battery power Meta

used public resources for the initial training

but

they actually compensated local partnerships on the ground to gather those rare recordings

Which brings us to the

real world deployment Yes Time highlighted that this system is currently

being used in Nigerian community clinics Nigeria is incredibly linguistically diverse You have

hundreds of

distinct languages and dialects across different regions So health practitioners are running

the compact models on their

devices right now When you pair this omnilingual

ASR with Meta's omnilingual

machine translation system

the workflow is just seamless A patient speaks in a rare local dialect

The speech is recognized transcribed

into text and then instantly translated across any of those 1

600 languages for the doctor to read It theoretically destroys the communication barrier in

critical health care settings

A doctor who only speaks English or Hausa can perfectly understand

the

symptoms described by a patient A patient speaking a language with absolutely no prior digital

footprint But But This is where

I

have to heavily question Time's criteria for invention of the year You are questioning the

victory lap Absolutely Deploying a piece of software is not the same

thing as improving a human life You are

pointing

out the gap between technological capability and practical efficacy Because we

do not have the clinical

outcomes No Time extensively reports on the adoption rates in

these

Nigerian clinics They document the geographic reach

But there is absolutely zero data on measured patient outcomes

Exactly Does this translation system actually reduce

misdiagnoses Does it speed up triage in

a crowded waiting room Are these patients objectively getting better care

Universal transcription is a stunning technical

achievement But its actual effect on health care remains a complete

mystery

And your skepticism is entirely justified Because a translation error when ordering

a coffee is just annoying Right A translation error in a medical diagnosis is

a critical safety failure If the model misinterprets a symptom because of

background noise in a busy clinic Or

if it hallucinates a medical term based

on a heavy accent The doctor might

prescribe the wrong medication The stakes are

literally life and death Deploying the tool is

merely step one Step two requires rigorous peer reviewed clinical trials We need to see the false positive

rates We need to see how

the

acoustic model handles the chaotic audio environment of a rural triage center Not just a quiet recording booth Exactly So if

you are following the intersection of AI and global

health you really need to

watch for the hard

data Watch whether this massive language coverage

actually translates into demonstrated measurable clinical benefits We need longitudinal studies Not just press releases about deployment

scale

The technology has to prove it does no harm before we

celebrate it as a medical revolution Right OK next

up The

contrast here is just fascinating to me Meta is squeezing AI onto low power clinic tablets And

meanwhile

Apple is figuring out how to brute force massive power hungry models on high end

desktop setups Exactly Apple just published some deeply technical

benchmarks that are shaking up the local infrastructure space

This is a

masterclass in solving hardware bottlenecks Apple literally linked four M5 Ultramax

studios together Four of them Mm hmm They used their open source

MLX framework to coordinate this cluster And together these four desktop machines generated over 36 tokens per second Hitting 36

tokens per second

locally is wild when you look

at the size of the

model they were running It

really is The model is called MuseSpark Open

It is an absolute behemoth It has 816 billion parameters Just

to give you a sense of scale Apple had to compress this model down to a

4 bit version And even after crushing the precision

of the math the file size was still 493 gigabytes So let's unpack

why that size breaks traditional hardware

Well

a parameter is essentially a single numerical weight inside the neural

network It's a tiny piece of logic Right And having 816 billion of them gives the model incredible reasoning capabilities

But it requires vast amounts of memory bandwidth just to move

those numbers from the storage

drive into

the processor The highest end Mac studio you can buy maxes

out

at 512 gigabytes

of unified memory So mathematically a 493 gigabyte model should

technically fit onto a single

machine

right It fits perfectly if the

model is just sitting

there doing absolutely nothing It fits on the drive

Right But running inference requires

dynamic space When a model operates it needs a

massive amount

of extra working memory Often called the KV cache Yes It needs temporary

space to hold your specific prompt It needs space to

calculate the intermediate

mathematical

steps across billions of parameters And to store the

output tokens as it generates them It needs breathing So once

you factor in that required

working memory a single fully maxed out

Mac Studio simply hits a wall It cannot load MuseSpark open The system will stall or just crash

So Apple had to figure out how to divide a single AI brain across four

separate physical bodies Which proves that local inference

for giant frontier class models is

actually possible for developers Without renting rack space from

Amazon or Microsoft they can pool their desktop

resources But the way you slice that brain matters immensely Apple

published

data on two different architectural approaches Let's walk through

them because the engineering here is just brilliant The first

method

they tested is called pipeline parallelization Think of

pipeline parallelization as a

traditional factory assembly line A

neural network is built in sequential layers In a pipeline

setup

you slice the model horizontally You assign

each Mac Studio a specific

set of layers Yes Machine A processes layers 1 through 20

When it finishes its math it packages the result and hands

the baton to machine B So machine

B processes layers 21 through 40 and so on It solves

the memory capacity problem completely No single Mac has

to hold

the

whole

400 gigabyte model But it doesn't necessarily speed up the cognitive process does it No

No Generating a single word

still requires

that data to travel sequentially through all four machines While machine

A

is working machine B C and

D are just sitting idle

waiting for

their turn So to fix that idle

time Apple tested the second method tensor

parallelization Tensor parallelization is a completely different beast Instead

of dividing the model by

layers you divide the mathematical work within every single layer You slice the model vertically Every single Mac Studio holds a piece of layer 1

They all calculate their specific piece of

the matrix multiplication at the exact same millisecond Then

they instantly share their partial answers combine

them into a final

result for layer 1 and

all move to

layer 2 together It sounds

incredibly efficient Every chip is firing at maximum

capacity

simultaneously It is theoretically much faster for generating answers But Yes

but It introduces a massive new bottleneck The communication

overhead Because they are working on the same layer

simultaneously the machines have

to constantly talk

to each other They must exchange inputs

synchronize their states and combine

outputs for every single layer

For every single token generated That requires

serious hardware wiring You cannot just

connect them over Wi Fi No You need extreme uninterrupted

bandwidth Apple utilized RDMA That stands for Remote Direct Memory Access It is a protocol that allows

one computer to reach directly into the RAM of another computer It completely bypasses

the central processor and the

operating system of the target machine There are no

software interrupts It's like splicing the nervous

systems of the four machines together into one shared organism And

they paired this RDMA protocol with physical Thunderbolt 5 cables The underlying

M5 Ultra chips also showed a huge hardware leap Apple tested smaller