AI helps scientists screen 70 million cells

Welcome to The Signal from Superpower Daily with Maya and Theo Today scientists are using AI to find a single needle in a haystack of 70 million cells and then actually pulling that exact needle out to see what it is made of Right It is an incredible shift We are seeing AI truly escape the digital world I mean it is starting to physically alter our reality It really is So let's jump right into this I want to start with a study called SPARX It was just published in the journal Cell Yes the SPARX paper This paper completely stopped me in my tracks It is wild Researchers

paired AI cell screening with physical isolation in a 70 million cell study That scale is just unprecedented Right They basically linked complex microscope images to underlying genetic changes And then they physically recovered those specific cells for deep protein analysis This is a massive breakthrough in a field called spatial biology The method itself is called spatially resolved CRISPR screening That is where the SPARX acronym comes from I see It was developed by a very impressive consortium You have researchers from LMU's Gene Center the Max Planck Institute and Helmholtz Munich You really have to picture the physical scale of this experiment to grasp it Definitely We are

talking about 70 million individual human cells Right They are sitting in laboratory culture plates They are alive They are constantly interacting It is a living environment Exactly And scientists need to find the handful of cells that are doing something highly unusual It is an incredible logistical challenge Finding that needle requires a very specific process So where do they even begin Well they use something called genome wide CRISPR knockouts That is the crucial first step Let's explain what a knockout actually is Sure I think people hear CRISPR and they immediately think of you know gene editing designing perfect DNA Right But in this context they are

using CRISPR as a set of genetic scissors Okay They intentionally disable specific genes across this massive population of cells Just breaking them on purpose Basically yes They make a targeted cut They break the gene Every single cell in that massive culture gets a different gene turned off So they essentially create a massive living library of genetic errors Exactly A library of errors Because they want to see what those specific errors actually do to the physical structure of the cell And that is exactly where the machine learning comes in Right Traditional genetic screening cannot easily process complex visual features It struggles with nuance I always thought

of traditional screening as like a very basic spreadsheet That is a good way to put it You have a list of genes on one side You track simple binary traits on the other Did the cell survive the drug or did it turn a fluorescent green color Yes It's just flat data Exactly But biology is not binary It is structural It is spatial I mean it is incredibly messy Right Complex patterns in microscope images are notoriously hard to link to underlying genes A cell might subtly change its internal shape Or move things around inside Exactly It might move its organelles to a different corner A spreadsheet

cannot capture that nuance But the AI can Yes The machine learning model is trained to look for those subtle visual shifts Okay It identifies the exact cells that look structurally different And it flags them based on their genetic knockout The AI flags the target But here is the truly incredible part The physical part Yeah They do not just take a digital picture and move on to the next cell No They use automated laser micro dissection This is the part that feels like science fiction Totally They use a microscopic laser They physically cut those exact flagged cells out of the culture Right Right there in place

Wow The laser isolates the individual cell from the plate And it leaves the surrounding millions of cells completely untouched It is a level of automated physical precision that was just you know previously unimaginable at this scale They are letting the AI guide a physical surgical tool And SPARKS solves a huge problem in biology doing this It adds a direct molecular readout to the visual data Meaning they can see what is inside Exactly It bridges the gap between what a cell looks like and what it is actually made of Right Right Because the cell is recovered intact by the laser the scientist can actually analyze its

physical contents Yes Using a technique called mass spectrometry So how does that work Well mass spectrometry is essentially a highly precise scale You take that single isolated cell you break it open you feed its proteins into the machine And the machine weighs them Exactly The machine uses magnetic fields to weigh every single protein piece Because different proteins have different exact physical masses Correct The machine reads those masses It gives you a complete exhaustive inventory of the cell's internal machinery That is wild And they actually tested this entire pipeline on two real world biological processes Right The first experiment looked at cellular recycling Biologists call this

process autophagy Yeah Autophagy is basically how a cell cleans up its own garbage It has to happen A healthy cell needs to degrade damaged internal components It breaks them down It recycles the raw materials It is crucial for preventing disease So in this first experiment the SPARC system successfully found the specific genes already known to govern this recycling process Yes Which proved the laser and the AI actually work together It validated the known science perfectly But then it pushed the boundary further What did it find The system also identified entirely new genetic contributors to autophagy It found genes we did not know were involved in

the cleanup process Oh wow That is a big deal for aging research Massive deal But the second experiment is even more fascinating The STING experiment Yes It focused on something called the STING immune sensor Right STING is a very crucial part of our innate immune system How does it work You can think of it as a cellular burglar alarm It detects foreign threats Specifically it looks for stray viral DNA floating inside the cell And the researchers wanted to understand exactly how this alarm activates Yes They used the SPARKS pipeline to track its physical movement So they knocked out genes They let the AI watch the

STING proteins Then they used the laser to pull the interesting cells Exactly And they proved that internal cellular acidity matters deeply to this alarm system Specifically they looked at the acidity in the Golgi apparatus Yes The Golgi is essentially the cell's central shipping and receiving center They found that a specific protein influences this local acidity The protein is called GPHR Okay The mass spectrometry data confirmed this connection The acidity level directly affects STING's early movement It basically controls how quickly the alarm triggers So they successfully connected a subtle visual change in a cell to a specific immune activation pathway Yes That is a massive triumph

for the technology It is a remarkable demonstration But we need to establish a very clear boundary here This is the crucial limitation of the entire study The computational modeling aspect Let's talk about that The researchers used advanced computational modeling to nominate further STING regulators Meaning they guessed other genes Well they generated a long list of other genes that might be involved in the alarm system based on the data But those remain candidates They are not confirmed biological findings Okay So the AI is giving us a suspect list It is not handing us a conviction That is exactly right Got it A suspect list Every single

proposed regulator on that list still requires independent experimental validation Right SPARKS can point the way It can narrow the search But it cannot bypass the fundamental scientific method So the biological proof must still be done in a physical lab Yes The AI cannot do the final experiment for you So what should people watch next with this Look for the independent validation of those STING candidates The global scientific community is going to start testing them And researchers can actually access the public mass spectrometry data right now Yes they can It is hosted online in Proteome Exchange The data set is PXD 082653 That public data access

is crucial for moving the field forward Absolutely That wraps up our lead story on molecular isolation We are going to pause here Next up let's look at a very different kind of physical automation over on the streets of London Right Pony ai and Uber are making a major strategic move They are Pony ai's Gen 7 Robotaxis will begin testing on London streets in the coming weeks Very soon And the companies plan to add real passenger rides soon This directly extends a partnership that began in May 2025 And London is just the next target for a massive you know planned deployment How big They are currently

planning a 2 000 Robotaxi European rollout Wow But you know there are significant missing details in this announcement right now First of all they have not set an actual launch date for the passenger service No I mean testing is one thing Putting paying customers in the back seat is another entirely Exactly And more crucially they have not named who will own and operate the London fleet This raises an incredibly obvious question Why wouldn't Uber just own the cars themselves You would think they might They certainly have the capital They have the global network They have the user base Well because owning and maintaining physical vehicles

is a terrible capital intensive business model Ah Uber learned this lesson years ago They burned billions trying to build their own autonomous hardware Now they want a strict capital lite template We can actually see this exact template working right now in Zagreb Croatia Yes Zagreb is the blueprint for this European expansion Let's break down how that specific partnership functions It is a tripartite model There are three distinct specialized players involved Okay First you have Pony AI They supply the digital driving system They are the artificial brain and they also supply the physical cars Right These are ARC Fox Alpha T5 vehicles They developed them jointly

with the Chinese automaker BAIC Correct So Pony AI brings the software brain and the metal shell but they deliberately do not run the taxi service themselves The second player is a startup called Vern Vern Yes Vern was founded by Matej Rimac You might know him from Rimac Automobili Oh right Vern actually owns and operates the physical fleet in Zagreb So they handle all the messy physical reality of the business All of it They handle the charging infrastructure They handle the daily cleaning They deal with the local municipal regulations They basically take on all the operational friction Yes And then there is the third player That

is Uber Just the app Right Uber simply connects the riders to the cars via its app They provide the immediate consumer demand They take a percentage cut of the fare and they completely avoid the massive capital expenditure of buying and maintaining thousands of robo taxis It is a very elegant division of corporate labor It really is But there is a major caveat here for the new London announcement Which is The 2 000 vehicle figure is for all of Europe It is not a dedicated commitment for the London fleet alone Ah Good to know Furthermore the Zagreb tripartite model has not been officially confirmed for London

yet So we do not know who the Vern of London will be Exactly Someone has to step up to physically operate those vehicles in the UK Right What should you watch next Watch for the announcement of that local London fleet operator That will signal that the commercial service is actually imminent Also you need to zoom out and look at Uber's broader strategy Uber is pushing hard globally They want to offer autonomous trips in up to 15 cities worldwide by the end of 2026 Ambitious Very We will see who takes on the heavy lifting in London Let's pause for a moment Meanwhile let's talk about how

AI navigates abstract logic We have a fascinating new study led by Yale researchers This one is deep It explores the hidden architecture of artificial thought The researchers found evidence of actual symbolic structure inside the numerical representations of seven different large language models Yes A new preprint shows something incredible Replacing an LLM's internal representation process with structured approximations left its behavior largely unchanged This gets to the heart of a massive debate in computer science Let's get into it The research was led by Tom McCoy The team also included Paul Soulos from Microsoft and Paul Smolenski from Microsoft Research OK They wanted to know how these models

actually hold information So they substituted the model's usual vector lists They used something highly specific called tensor product representations Right To understand why this substitution matters we have to talk about the role filler concept Oh this is a linguistics thing Yes It's a very fundamental idea in both linguistics and classical logic Imagine a simple sentence Cats chase dogs Cats chase dogs That's a basic action The words cats and dogs are the fillers They are the actual content Right But they sit inside specific structural roles Cats is the subject Dogs is the object This reminds me of playing structural Mad Libs Yes When you play Mad

Libs you have a blank space on the paper that says noun That blank space is the role And the silly word you choose to write in is the filler That is a very accurate analogy The breakthrough here is profound Because of the debate Right For a long time critics argued that large language models are just stochastic parrots Meaning They argued the models just guess the next isolated word based on statistics The critics claim the models have no true internal structure They claim the models don't actually understand relationships Yes But this study suggests LLMs might actually be preserving distinct structural relationships deep inside their math They

ran a specific targeted edit experiment to prove this This is where the evidence gets undeniable What did they do They looked at a test phrase The phrase was the clever doctor helped a lawyer The clever doctor helped a lawyer Got it Inside the model they located the exact mathematical encoding for the word crever They mapped its numerical coordinates And then they physically moved that encoding They shifted it from the subject role to the object role Yes They literally dragged the mathematical concept across the artificial neural network They just moved the math And the model immediately shifted its behavior Wow It started acting as if it

had read the doctor helped a clever lawyer So it preserved the logical structure perfectly It just cleanly swapped the adjective's target Yes It did not break the sentence It did not output gibberish That implies a highly organized internal structure That is incredible It implies the model actually maps concepts to roles rather than just memorizing text patterns It does But we need to outline the limitation here Yes The authors are very careful about their claims This new evidence only supports one particular role filler structure Okay It is not a generalized explanation for how all neural networks encode all information So the black box is cracking open

slightly but it is not fully open yet We do not have a universal theory of artificial cognition Exactly What should you watch next Watch how far this specific role filler structure actually extends Right The researchers tested tasks across language arithmetic logic and even coding We need to see if this structural mapping holds true for much more complex reasoning tasks Yes we do It is a massive step toward understanding how these systems actually think Let's take a breath here In other news the economic reality of the AI boom is hitting the ground hard in San Francisco Yes Yeah This is a tough one New reporting from

The Guardian and data from the San Francisco Rent Board paints a very stark picture It is intense San Francisco one bedroom rents have risen more than 25 recently The average is now 4 400 Simultaneously eviction notices have spiked 44 The situation on the ground has reached a breaking point Mayor Daniel Lurie has actually declared a formal rent emergency in the city The reporting details landlords aggressively seeking higher returns They want to match the rapidly rising AI salaries in the city This financial pressure is even hitting the city's rent controlled units San Francisco has about 160 000 of these pre 1979 buildings Let's clarify the math

here We need to be very precise about what the data actually proves Right We have to push back a little on the direct causality Can we definitively blame the AI industry for every single rent hike in the city Mathematically no We cannot Okay The figures show a severe localized housing squeeze This squeeze is happening right alongside massive AI wealth generation But the data cannot mathematically isolate AI's exact percentage contribution to the rent spike So it is a strong correlation It is a highly visible one but it is not a proven isolated cause Other economic factors are always at play Still landlords have a huge financial

incentive to clear out below market tenants Yes And the reporting highlights the specific legal and physical tactics used Like what Well some property owners use the Ellis Act This is a state law It allows owners to clear a building if they are officially going out of the rental business Ah And others are reportedly using disruptive renovations Yes They start massive construction projects Or they simply neglect critical repairs until the tenant finally leaves out of frustration Exactly Once the unit is legally vacant the rent resets to the new astronomical market rate We have to look at the human cost here The demand on local tenant services

is overwhelming The Housing Rights Committee gets 200 desperate calls weekly Their seven person clinic faces more than 150 voicemails every single Monday morning The geography of the Bay Area workforce is changing drastically because of this It really is Essential hospitality workers are driving in from Tracy or Modesto Those are massive punishing daily commutes Just brutal They are doing this just to find rent they can actually afford The reporting details people delaying 10 medical co pays They are skipping basic health care just to keep their housing That is terrible But there is an important caveat to the eviction data we should mention What is that The

44 spike represents eviction notices It does not represent completed physical evictions Right A notice is simply the start of a long legal process It does not always end in removal But it certainly signals a massive surge in displacement pressure The threat alone causes immense instability Absolutely So what should you watch next Watch Mayor Lurie's proposed countermeasures He wants to fund eviction legal aid He wants mandatory displacement payments Most significantly he wants strict caps on newly vacant rent controlled units That would be huge The major question is whether these proposals can actually survive the inevitable political and legal challenges Landlord groups will certainly fight them We

will see if local policy can actually protect tenants from macroeconomic forces Let's pause We are moving into our quick read section Let's pick up the pace a bit First quick read Astrophysicist Bryce Menard used cloud science to create the first complete map of the ultraviolet sky Right NASA's GALAX mission originally missed a third of the sky It had to deliberately avoid pointing its sensitive sensors at bright stars It left massive data gaps Yes Claude used a technique called in painting to fix this In painting is fascinating The AI learned the complex relationships between visible light infrared and radio data Okay It used those learned patterns

to estimate the missing UV gaps It came within 10 accuracy on tested hidden areas But there is a major caveat here Automation still requires intense human inspection Why Menard had to manually catch 38 000 faint atmospheric glow circles Oh wow Yeah These were visual artifacts Yes They survived two entire rounds of AI review Claude did eventually correct them in a few hours once heavily prompted but it missed them initially Watch how AI agents handle multi step scientific data work going forward Human oversight remains completely non negotiable for real science Absolutely Next quick read The University of Washington's Institute for Protein Design received a massive hardware

grant They secured nearly 7 million GPU hours The grant comes from the Gensun and Lori Huang Foundation It is delivered via the cloud provider CoreWeave This compute power targets precision cancer therapeutics durable vaccines and plastic degrading enzymes It will cut AI model learning time from months down to mere days It accelerates the pace of discovery exponentially But note the caveat This is projected research infrastructure It is not a finished working model Exactly More compute does not automatically equal validated biology Watch for the initial open source model releases They are expected late 2026 into early 2027 This timeline is entirely conditional on experimental validation in a

physical lab Makes sense Final quick read Neiman Lab found ChatGPT appending real cartoonist signatures to AI generated images The AI generated cartoon images it clearly did not create But it appended recognizable names like Brendan Loper and Emily Flake It creates totally false synthetic authorship Condé Nast actually signed a 2024 content deal with OpenAI But that specific deal did not authorize cartoon training Interesting The core caveat here is technical AI image verification tools only check for digital OpenAI signals They cannot verify true human authorship Watch how OpenAI addresses this weird hallucination They added a custom warning about New Yorker style cartoons But the warning did not

stop all the falsely signed outputs And they have not explained how the physical signatures actually appeared in the training data We are moving to three takeaways from today First AI is rapidly moving from digital prediction to physical molecular isolation and design in the life sciences We clearly see this with the SPARC cell screening We see it with the massive UW protein design compute grant The technology is permanently entering the physical lab Second the black box of AI logic is cracking slightly The Yale study shows that neural networks likely preserve distinct relationships between concepts They are not just generating statistical text patterns They are maintaining actual

symbolic structures Third the secondary blast radius of the AI boom is getting highly personal It is displacing workers from housing markets in San Francisco It is attaching real artists identities to synthetic work The collateral damage of this transition is incredibly tangible I keep thinking about how these stories intersect If AI can map the structural logic of language physical labs can automate molecular design how long until an AI generates a completely synthetic cell from scratch It is a profound question There is one key development to watch tomorrow Look out for how the city of San Francisco attempts to codify those proposed rent caps on vacant units

The displacement pressure is mounting daily The policy response will be critical To dive deeper into any of these stories visit superpowerdaily com Thank you for listening We'll see you tomorrow