Tonight’s news
is hard to overlook because of one number:
5 trillion.
A new AI company called:
Accelerated Understanding
announced that
their new model can handle:
5 trillion data units
in a single run during testing.
Reuters even compares this scale to top language models today,
estimating it to be about
millions of times larger.
That sounds incredible.
But if you think this news is simply about
"Another AI with a huge context window,"
you’ve missed the real point.
Because what makes it truly special is:
it’s not designed to read text at all.
This AI isn’t about predicting “the next word”
Most familiar large language models like ChatGPT, Claude, and Gemini
focus on language,
tokens,
and sequences.
Simply put,
they predict
what is most likely to come next.
This approach has proven powerful,
enabling models to
write articles,
code,
analyze data,
use tools,
and act as agents.
But the physical world poses a problem:
It is not a linear sequence like a text.
Air isn’t a token-by-token string
When a typhoon moves,
it’s not a sequence of:
first word,
second word,
third word.
It’s the entire atmosphere
changing simultaneously:
temperature,
pressure,
humidity,
wind speed,
oceans,
and terrain
all interacting,
and every point potentially changing
at the same time.
Chip cooling is similar.
Heat doesn’t flow like sentences;
it diffuses continuously through materials.
Accelerated Understanding aims to solve this problem
The founders,
Anima Anandkumar
and
Benedikt Jenik,
today officially unveiled their physics-based AI.
Anandkumar is a professor of Computing and Mathematical Sciences at Caltech,
with past research at Amazon
and leadership of AI research teams at Nvidia.
One of her key research focuses is:
Neural Operators.
What is a Neural Operator?
Without math,
think of typical AI as learning
"one input leads to one output."
Neural Operators learn
"how an entire system changes from one state to another."
For example, given the current state of the atmosphere with
temperature,
pressure,
humidity,
and wind fields,
what could it look like in a few hours?
It learns not a single number,
but
the whole functional transformation of the physical field.
Why it differs so much from language models
Accelerated Understanding’s founders emphasize
their system does not use
Transformers,
one of the key architectures behind most large language models today.
Instead, it follows the
Neural Operator route.
Anandkumar sums it up neatly:
Language models are human-centered,
while theirs is nature-centered.
It’s not about what humans wrote,
but about how
nature itself changes.
Why a different structure is necessary for AI
Different problems have fundamentally different data structures.
Text is a string of
words.
Code is
symbols.
Scientific problems involve whole 3D spaces,
time series,
fluids,
materials,
temperature fields,
pressure fields,
and electromagnetic fields.
Forcing everything into tokenized text is
not always the most efficient approach.
What does 5 trillion actually mean?
The company told Reuters
their AI can process
5 trillion data units in a single run.
Reuters compares this to what top Anthropic or Google flagship models handle,
estimating the difference around
5 million times.
But be cautious.
This is not a benchmark directly comparable to a context window.
Why can’t we just say it has 5 million times Claude’s context?
Because they process different data types.
Claude or Gemini’s context is defined in
tokens,
representing text.
This physics AI processes
physical data — numeric fields across space and time.
So “5 trillion” mainly tells us this AI tackles
massive scientific problems.
It doesn’t mean
“It reads 5 million times more info than Claude, so it’s 5 million times better.”
The capabilities are not comparable.
It’s like comparing a telescope to a printer on “who sees farther”
They serve completely different functions.
The real takeaway today is not rankings,
but
AI architecture diversification.
Before, many assumed
Transformers’ success meant
they just needed to scale Transformers for all AI problems.
Now more researchers say
that’s not necessarily true.
First application: chip design
This connects nicely to earlier topics on companies shifting to different AI value chain segments.
This company says physics AI can be used for
chip design.
Chips involve more than circuits —
including temperature, materials, heat transfer, current, and structure.
The faster an AI chip, the harder cooling becomes
Packing more transistors, compute power, and memory in one system makes heat a major challenge.
Engineers currently simulate, prototype, test,
adjust materials,
and retest.
Each cycle is expensive in time and money.
Accelerated Understanding’s idea: teach AI the physical system first
For example, if material changes to A,
how does the temperature distribution change?
If cooling structure switches to B,
where might overheating occur?
If package size shifts to C,
how do performance and temperature interact?
If AI can turn many experiments into quick simulations,
lab experiments needed can be dramatically reduced.
This is one of AI for Science’s big values.
Second application: weather
Anandkumar has worked on this before,
during her Nvidia days, helping develop weather AI models using Neural Operators.
Caltech research shows Neural Operators are used for fluids, materials, and climate simulations.
Her team has built high-resolution AI weather models using similar approaches.
Why is AI weather forecast especially suited for this?
Traditional numerical weather models
solve many physical equations by
dividing Earth into grids
where temperature, pressure, wind, and humidity
are computed over time.
They’re accurate but very computation-heavy.
Neural Operator methods hope to
learn how these fields evolve from vast historical data,
enabling much faster future predictions.
But this doesn’t mean physics models can retire
This aligns with previous SasaDaily coverage on AI typhoon forecasts.
AI can be quick,
and sometimes very accurate in path prediction,
but extreme events are challenging
due to the rarity of training data.
Accurate path prediction doesn’t guarantee accurate intensity prediction.
Mature methods will combine
AI, traditional physics models, observational data,
and weather experts.
Third application: robotics
Robots entering
factories,
stores,
or homes
need deep understanding of physics,
such as:
how heavy something is,
how hard to grip a glass,
whether the floor is slippery,
where objects will fall if hit,
and what happens next when the robotic arm moves.
These can’t be solved by language understanding alone.
LLMs can tell a robot “pick up the glass”
But physics control
depends on real-world properties like
weight,
slipperiness,
temperature,
water inside,
placement,
and people nearby.
The challenge of Physical AI is to
turn “knowing what to do” into “knowing how the real world will respond.”
That’s why
World Models, Physics Models, and Embodied AI
are increasingly important.
Fourth application: energy and geology
Accelerated Understanding says
their model could handle underground geology data,
helping energy companies
understand subsurface structures,
predict physical changes,
and reduce costly cycles of
simulation,
exploration,
and experiments.
This is a major economic difference from general chatbots.
A chatbot might save you 20 minutes writing emails, which is great,
but a scientific AI that avoids a $5 million experiment
represents a vastly higher commercial value.
So this company isn’t starting as a consumer chatbot
The founders clearly aim initially at
enterprise clients.
That makes sense,
since most consumers won’t ask
“Help me simulate heat transfer in this material.”
But chip manufacturers, energy companies, robotics firms, meteorological agencies, and engineering companies
deal with these questions daily.
There’s an interesting backstory: Jeff Bezos tried to recruit them for Project Prometheus
Documents obtained by Reuters reveal that in late 2024,
Vik Bajaj, co-founder of Project Prometheus with Jeff Bezos,
discussed collaboration with the founders,
even issuing a proposal with terms.
The terms were notable
Documents show Anandkumar could
become the company’s external face,
join the board,
and lead the scientific direction.
She and Jenik together could receive
35% equity.
Their combined salary would start at $1 million,
rising to
$2 million after three months.
The documents also mention
over $2 billion in committed funding.
But they ultimately didn’t accept
Choosing instead to stay with their own company.
Project Prometheus went on to complete a
$12 billion Series B in June 2026.
Prometheus also works on AI + complex physical systems manufacturing.
This highlights a key trend:
Investors now seriously back Physical AI and Science AI,
not just chatbots.
Once the biggest question was “Who will build the next ChatGPT?”
Now there’s a new question:
Who will develop foundational AI models that understand the real world?
For example:
Fei-Fei Li pushes World Labs,
Yann LeCun advocates World Models,
Nvidia builds Cosmos,
many robotics companies pursue Physical AI,
and Accelerated Understanding chooses
physics equations and Neural Operators.
These paths aren’t identical
World Models often learn from
videos,
images,
actions,
and spatial data
to understand the world.
For instance, an AI might view
many videos of vehicles moving
and slowly learn how objects move.
Neural Operators are closer to
directly learning continuous physical systems, like
fluid dynamics,
heat transfer,
atmospheric flows,
and materials behavior.
They may eventually cooperate
rather than one winning outright.
In the future, one robot may need three AI types
First layer:
Language model
understands what people want.
Second layer:
World Model
predicts how the environment changes.
Third layer:
Physics Model
precisely understands forces, materials, fluids, and heat.
Viewed this way, AI won’t be a single supermodel covering everything,
but rather
a combination of complementary intelligences.
This ties into this morning’s theme of vertical integration
Today we covered Thomson Reuters developing its own specialized models,
Xiaomi pursuing AI chips,
and Nvidia investing in AI applications.
Tonight, we see
the question of why all AI must use the same model architecture
emerging.
Industry is moving from
everyone competing on the same leaderboard
to
different AI for different problems.
The “strongest model” question is losing relevance
If you ask,
“Which car is best?”
the answer depends on the purpose.
Racecar?
Truck?
Off-roader?
Family car?
AI will be the same.
The best for coding might be one model,
legal research another,
real-time voice a third,
robotics a fourth,
and physical simulation something totally different.
Transformers won’t suddenly become obsolete
This is important.
Tonight’s news is not
"The death of Transformers."
Far from it.
Transformers have proven strong for
language,
code,
multimodal tasks,
and agents.
Accelerated Understanding just suggests
some physical problems are better served by other architectures.
Technological progress is rarely about
new tech completely replacing old,
but
adding new tools to the toolbox.
Neural Operators weren’t invented today
Anandkumar’s team has researched them for many years.
Caltech studies show they can accelerate simulations of
partial differential equations,
fluids,
materials,
and climate systems.
So the novelty today isn’t first hearing of Neural Operators,
but
someone building an enterprise-grade AI company based on this line of research.
This is the key step from research to product
Papers can say,
“The model is fast.”
“Benchmarks are great.”
But real businesses ask:
Can my data be used?
Is the accuracy sufficient?
Can errors be quantified?
Are results reproducible?
Can it integrate with existing simulation software?
Is it cheaper than current methods?
The true test is after going to market.
The 5 trillion number needs a big asterisk
Because it’s currently
only company-reported test results.
Reuters interviewed and reported the company’s claims,
but we can’t yet say there is a full independent third-party benchmark
showing Accelerated Understanding’s model
outranks existing science AI overall.
Especially, how “5 trillion data units” translates into real task
performance, accuracy, and cost
still requires more validation.
Scientific AI’s key KPIs aren’t just “how much data it consumes”
Really important metrics include:
Accuracy
How close are predictions to reality?
Speed
How much faster than existing simulations?
Cost
How much GPU time, wall clock time, and experiments are saved?
Generalization
Does it remain reliable on unseen materials, weather, or geometries?
Verifiability
Can outputs be confirmed by physical experiments?
These matter far more than a single huge number.
Because scientific AI’s biggest risk isn’t “a wrong answer”
If ChatGPT writes
a poor email,
you can rewrite it.
But if physics AI says
a material can withstand certain conditions,
a chip won’t overheat,
or a structure is safe,
and it’s wrong,
the consequences can be severe.
As AI moves closer to the physical world,
verification becomes ever more vital.
This is why scientific AI can’t rely only on AI to verify AI
A model predicting a new material is viable,
and another model concurring,
doesn’t guarantee it’s truly viable.
Experiments, measurements, and real-world testing remain essential.
AI’s greatest impact might be
reducing 1,000 physical experiments to just 50
rather than eliminating all of them.
If achieved, the commercial value is huge
Human science and engineering advances
aren’t often limited by ideas,
but by costly verification.
New drugs, materials, chips, energy systems, aircraft, and climate solutions
all need expensive trials.
If AI can filter out
90% of unpromising ideas,
labs can focus on the best 10%,
radically changing R&D.
This might be the biggest revolution in AI for Science
Not
just “AI reading more scientific papers.”
That’s valuable too.
The bigger possible advance is
AI dramatically shortening the “idea → simulation → experiment → verification” cycle.
Chatbots help us process information faster;
Physics AI aims to help us understand the world faster.
These are two very different goals.
The real takeaway tonight isn’t 5 trillion
5 trillion is eye-catching.
But a year from now, the question won’t be
whether it’s 10 trillion or more,
but whether it really helped
reduce chip experiment cycles,
improve weather modeling speed and accuracy,
make robots safer by understanding physics better,
or accelerate underground structure analysis for energy companies.
Success on these fronts means a truly
new class of AI.
One final thought for tonight
Over the past three years,
we’ve come to associate
AI with large language models.
Accelerated Understanding reminds us these are
not synonyms.
Language is just one part of the world.
Reality also includes
time,
space,
forces,
temperature,
fluids,
materials,
weather,
and geology.
If AI’s next phase really leaves the chat window behind
to enter chips,
science,
robotics, and the physical world,
the next major breakthrough
might not be a bigger Transformer,
but
an AI fundamentally not designed for text.
That’s the truly noteworthy aspect of tonight’s news.
Today, let’s improve together with AI.
Learn one AI skill every day.
Save a little time daily.
Enhance your abilities bit by bit.
SasaDaily, growing with you.
