Sens. Richard Blumenthal (D-Conn.) and Josh Hawley (R-Mo.) are calling on the nation’s largest Medicare Advantage insurers to provide internal records and detailed information on their use of artificial intelligence to block rehabilitative care.
The lawmakers’ request — a moment of bipartisan scrutiny on a controversial federal health care program — comes just one month after a government investigation unearthed a continuing pattern of denials within Medicare Advantage.
In letters provided to STAT, Blumenthal and Hawley told executives at UnitedHealth Group, Humana, and CVS Health that recent findings by the Office of the Inspector General for the Health and Human Services Department undercut their companies’ claims to have reduced barriers to crucial medical services.
Prescription medicines purchased in the U.S. under a controversial government discount program amounted to $100 billion in 2025, a 22.8% increase from the previous year, according to the Health Resources and Services Administration, which oversees the program.
Expensive medicines represented an increasing proportion of spending in the 340B Drug Discount Program, accounting for $61.9 billion, or nearly 62% of all prescription drugs purchased through the program. Nearly $8.9 billion was spent on Merck’s Keytruda immunotherapy treatment, followed by more than $4.47 billion on Biktarvy, an HIV medicine sold by Gilead Sciences.
The data mark a steady rise in sales under the 340B program, which requires drugmakers to offer discounts that are typically estimated to be 25% to 50% — but could be higher — off all outpatient drugs to hospitals and clinics that primarily serve lower-income patients.
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology.
What Anthropic’s latest AI discovery does—and doesn’t—show
—James O’Donnell
When Anthropic announced last week that it had found a new window into its models’ “internal thoughts” as they reason through answers, there was one colleague I had to talk to: senior editor Will Douglas Heaven.
Aside from having a PhD in computer science, Will has spent a lot of time digging into what we can say about how AI models work. I spoke with him about what we should take from Anthropic’s new (and typically quirky) research. Here’s what he had to say.
This article is from The Algorithm, our weekly AI newsletter. Sign up to receive it in your inbox every Monday.
How will AI understand the real world?
Today’s AI systems can generate text, images, and code with impressive skill, but they still struggle with the complexities of the physical world. To bridge this gap, many researchers believe you need something called a world model.
At a LinkedIn Live event today, MIT Technology Review will investigate how this technology could transform robotics and help unlock a new generation of intelligent machines. Join Will Douglas Heaven, our senior editor for AI, and Sam Sinha, founding AI researcher and head of world models at 1X Technologies, for the discussion.
I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology.
1 New York has become the first state to enact a data center moratorium Its governor banned large data-center construction for up to a year. (WSJ $) + A bill passed by state lawmakers could go even further. (Verge) + Everyone hates data centers. (MIT Technology Review)
2 Smartphone shipments have hit a 13-year low due to the memory crunch They fell 11% in the second quarter of 2026. (Reuters $) + The memory chip shortage has increased prices. (Gizmodo) + And threatens the promise of Moore’s Law. (MIT Technology Review)
3 Sugar molecules have been found in interstellar space for the first time It hints that life on Earth may have been seeded from space. (Nature) + And boosts the odds of living organisms existing elsewhere. (New Scientist $) + Researchers used radio telescopes and data to spot the molecules. (NYT $)
4 Nvidia has halved its Asia buyer list to stop AI chips reaching China It introduced a “white list” of companies that passed tougher checks. (FT $) + It moved amid tighter chip controls from the Trump administration. (Reuters $)
5 Russian state hackers are targeting routers to spy and steal, the US warns The government has warned users to secure their devices. (Ars Technica) + Now is a good time for doing crime. (MIT Technology Review)
6 Trump moved his crypto gains into stocks while urging people to buy more His crypto projects earned him a fortune—but steep losses for retail buyers. (Reuters $) + He’s called for Congress to pass a new crypto bill to honor Lindsey Graham. (CNBC)
7 A new cell therapy has saved four children with terminal brain cancer They were treated with an experimental immunotherapy. (New Scientist $) + Access for older children will also be limited. (Bloomberg $)
8 The LAPD has halted use of Flock surveillance cameras due to privacy issues Flock’s automated license plate readers have caused concerns. (LA Times $) + It’s also been criticized for sharing data with state and federal officials. (Engadget)
9 The US has approved launching a space mirror that reflects sunlight onto Earth As part of a controversial plan to power solar panels round the clock. (Wired $) + But geoengineering faces many practical challenges. (MIT Technology Review)
10 Anthropic says Claude’s values vary depending on your language It’s most cautious in English and most deferential in Arabic. (Gizmodo)
Quote of the day
“The age when humans are the highest life form on earth will end. For better or for worse, it will happen and it can’t be stopped.”
—SoftBank CEO Masayoshi Son predicts that AI will overtake human intelligence by 2040 in a speech at his company’s annual corporate conference in Tokyo, Reuters reports.
One More Thing
Inside the strange limbo facing millions of IVF embryos
Millions of embryos created through IVF sit frozen in time, stored in cryopreservation tanks around the world. Many are left in a peculiar limbo, with no clear path forward.
UK residents can discard them, make them available to other prospective parents, or donate them for research. People in the US can also opt for “adoption,” “placing” their embryos with families they get to choose. In Germany, people aren’t typically allowed to freeze embryos at all. And in Italy, unused embryos must remain frozen, ostensibly forever.
While these embryos remain in suspended animation, patients, clinicians, embryologists, and legislators must grapple with the essential question of what to do with them. What do these embryos mean to us? Who should be responsible for them?
The machine that could change the world will be housed in a room that looks like a data center crossed with an ice cream factory. Inside will be some 100 stainless-steel cabinets, each about six feet tall and connected to a supply of liquid helium that keeps them only a few degrees above absolute zero. Inside those cabinets will be hundreds of chips, and on those, thousands of particles of light flying through a maze of optical switches and beam splitters. Each photon must be accounted for, because precisely measuring where it ends up will help answer questions that current computers might take millions of years to solve.
This computer, as described, does not exist. It’s the brainchild of a company called PsiQuantum, founded in 2016 by four physicists from UK universities. In a crowded field of deep-pocketed competitors with similarly fantastical visions, the company aims to be first to fulfill its promise.
In the years since the physicist Richard Feynman first envisioned them in 1981, quantum computers have promised to speed up everything from medical research to AI by harnessing the qualities of quantum particles. Unlike normal computer bits, which can be either a 1 or 0, quantum bits can exist in multiple states at once. And combining enough of those quantum bits together could produce a computer capable of tasks well beyond the reach of today’s conventional machines. But even today’s best quantum prototypes are too small and error-prone to do anything useful.
That makes PsiQuantum’s promises for what its computers will ultimately do all the more bold. Consider the company’s hopes for predicting the effects of cytochrome P450 enzymes, which often break down drugs in the body. If pharma companies knew more precisely how they would work on a particular molecule, they could design more effective medications faster. Estimating this for a specific drug can take over 10 years with today’s methods, says Philipp Ernst, vice president of quantum applications for PsiQuantum, but “we aim to get it down to four minutes.”
The company’s chips will be contained in large cabinets. A quantum computer powerful enough to be commercially useful is expected to require roughly 100 of these cabinets connected together.
COURTESY OF PSIQUANTUM
In a field full of such claims, PsiQuantum has attracted unusual investment and scrutiny for two reasons: It is one of the few companies aiming directly at building a large and useful machine, and it is already working with a major chip manufacturer to build its systems using existing semiconductor fabs. Its vision has attracted momentum: Last year, PsiQuantum raised $1 billion in funding and broke ground in Chicago on a site it’s building in partnership with local governments. It also has a second site in the works in Australia, which it promises will be operational—meaning hardware-ready—in 2027. And it’s one of just two companies (along with Microsoft) to reach the third stage of an intensive government evaluation program to see which quantum companies might succeed.
Evaluating whether PsiQuantum will do what it says is harder than, say, judging a drugmaker by its clinical trial results: Advances in quantum computing are incremental, opaque, and tough to verify from the outside. But the company is now approaching its prove-it moment, when years of closed-door work and hundreds of millions in investment will either culminate in a useful quantum computer or fall short. We could start to know which as soon as next year.
A new kind of machine
Terry Rudolph, one of PsiQuantum’s four founders, is soft-spoken and shaggy-haired. He was born in Malawi and learned only after earning his first physics degree that he is a grandson of the famed physicist Erwin Schrödinger. He later self-published a 150-page book to explain quantum computing to teenagers (my PR contact gave me a signed copy with a wink that said “We never expect anyone to actually read this,” but I can report that it is a funny and helpful book).
Around 2014, Rudolph and his cofounders became increasingly convinced that the quantum breakthroughs they were finding to be possible in theory might also be possible in a real machine. They eventually left their academic positions and divided the tasks before them: Rudolph worked on theory, Mark Thompson on engineering, Pete Shadbolt on scaling the technology up, and Jeremy O’Brien on articulating the vision and finding investors (O’Brien served as CEO until February; he’s been replaced by Victor Peng, a veteran of the semiconductor industry).
To understand why the quantum computer the company is building would be a big deal, consider how imprecise much of modern science remains. We cannot reliably predict, for example, which lithium-ion battery will catch fire or how quickly a critical aircraft component will corrode.
This isn’t just because these systems are complex, though they are. It’s that, at their core, they are governed by quantum mechanics. Subatomic particles don’t have well-defined properties—this location and that velocity—but instead occupy quantum states spread across many possibilities. And that in turn influences a range of atomic and molecular behavior. Schrödinger (Rudolph’s grandfather, remember) showed how to describe this haziness mathematically a century ago this year, but precisely carrying out the calculations on real-world systems quickly becomes unfeasible even for the best computers. Scientists cope with this gap using approximations, imperfect simulations, or experiments on animals.
WINNI WINTERMEYER
WINNI WINTERMEYER
PsiQuantum co-founder and chief scientific officer Pete Shadbolt (left), and machinery the company has built to manufacture its own barium titanate, a material with the perfect qualities for routing light particles (right).
Feynman, David Deutsch, and other physicists in the 1980s wondered if we could do better. Maybe such complexity could instead be modeled using a new kind of machine. Rather than using transistors that are only ever on or off, this one would use particles held in quantum states, manipulate them to perform calculations, and then measure them at the end for an answer. Using quantum systems to simulate quantum systems would for the first time allow a simulation of physics and chemistry that directly reflected reality. It would be an invaluable tool for designing new drugs, materials, or really anything affected by quantum mechanics. Revolutionary, in other words.
Humankind’s leaps in understanding how nature works have often resulted in the invention of powerful new tools, Rudolph told me. “I don’t think it’s a coincidence that the Industrial Revolution coincided with our ability to calculate and simulate the laws of Newtonian mechanics, the laws of thermodynamics,…the laws of classical electromagnetism,” he says. “Whenever we have more power to calculate and simulate and understand things, we build incredible machines that come from it.” He sees something similar coming with quantum computers.
Chasing photons
One mystery has always been which quantum thing—ions, atoms, or something entirely new engineered with quantum properties—could be made stable and controllable enough to use as a qubit, the basic unit in the quantum computing world. Quantum systems are delicate, and observing any particular particle causes it to collapse into one state rather than a superposition of multiple states. If this happens during the computation rather than at the end, it produces an error that must be corrected for. Too many of these means the computer fails to produce a useful answer.
Just as engineers in the early days of aviation weren’t sure whether airplane wings would be fixed or flap like a bird’s, we’re not yet sure which of these quantum things will work best. Google and IBM are betting on superconducting qubits, superconducting circuits made of aluminum or other metals. Intel is using electrons. PsiQuantum is using photons, the particles that make up light.
“Photons have lots of nice things going for them,” Rudolph says. They can maintain quantum states for a long time; indeed, the photons in the universe’s cosmic microwave background may have done so for billions of years. But photons also move fast and scatter easily. More importantly, two photons are more likely to pass through one other than interact. That makes them a challenging candidate for quantum computation, in which qubits need ways to influence one another.
For a while, this last flaw seemed to doom the idea of quantum computing with light. But in 2001, researchers from the Los Alamos National Laboratory and the University of Queensland found a loophole. They discovered they could essentially fake interactions between photons by sending the light particles through a network of beam splitters and detectors. Their paper changed everything. PsiQuantum was created to make the theory a reality.
Size was the first problem; previous plans would have required a computer as large as California. Mercedes Gimeno-Segovia, who was a PhD student of Rudolph’s in the early 2010s (after almost becoming a professional violinist instead), thought of a way for the machine to be smaller.
The basic process since then has been this: First create photons with lasers and then “entangle” them, exploiting a quantum phenomenon in which the particles no longer have individual states but instead share one. Next, route them through a maze of gates that perform computations, and finally read out details of their quantum state at the end, all while tracking and correcting for the errors that occur. Succeeding at each of these steps millions of times is not so much an engineering hurdle as a brick wall. And building the supply chain—like manufacturing new materials with the qualities to route individual photons around—is arduous.
A sizable chunk of PsiQuantum’s funding is being spent on custom cooling machinery that uses tanks of liquid helium to cool the company’s chips. Shown here is part of the PsiQuantum’s cooling system at a facility in Milpitas, California.
COURTESY OF PSIQUANTUM
To get a sense of it all, last year I joined Shadbolt at the SLAC National Accelerator Laboratory, in Menlo Park, California. The center has helped produce several Nobel Prizes and played a role in the 1968 discovery of quarks, fundamental building blocks of matter that make up protons and neutrons. But PsiQuantum set up shop there essentially to siphon liquid helium from SLAC’s giant cryoplant. This is what the company uses to cool its computing cabinets down to deep-space temperatures.
Right now the cabinets operate at 2 K, or -456 °F, but the goal is to be able to run them slightly warmer—at a balmy -452 °F. Most quantum approaches require the whole machine to be cooled to superconducting temperatures, so that much of the expense in running it will actually be spent on refrigeration. But photonic computers require only one piece to be this cold—the detectors that measure single photons at the end of the computation. And the required temperature can be a bit higher. (PsiQuantum said in May that it will spend some of the $100 million award in CHIPS Act funding it’s slated to get on these detectors).
The siphoning setup was a temporary solution; PsiQuantum now has its own cooling system at its testing facility in Milpitas, California, and is setting up a larger one at its production site in Australia next year. These helium systems represent some of the biggest capital expenditures for any quantum company and will consume a significant chunk of PsiQuantum’s $1 billion funding round.
In the afternoon we drove to a lab in San Jose, where I donned a cleanroom suit—a head-to-toe covering that keeps dust at bay—to watch the manufacture of a blueish crystal called barium titanate.
It’s prized by PsiQuantum because it quickly and reliably routes light particles with very little electrical input, keeping the precious photons undisturbed as they move through the circuit. But for all barium titanate’s theoretical value to the company, its structure makes it a pain to manufacture, and the material wasn’t available at scale when PsiQuantum got its start. The company, in what Rudolph told me was an agonizing decision, opted to make it in-house, requiring a massive investment. I saw a technician—operating at what looked like a giant pressure cooker—adding the base elements to several hoppers; then I watched through a porthole as the elements got heated, vaporized, and finally crystallized into a thin layer on a wafer disc. At that time each disc took about 12 hours to make; the company now says several are produced each day. The discs then get shipped to the chipmaker GlobalFoundries in Malta, New York, where PsiQuantum’s chips are made.
WINNI WINTERMEYER
WINNI WINTERMEYER
The company has invested heavily in making its own barium titanate, a material whose delicate crystalline structure is tedious to manufacture.
PsiQuantum’s bet is that this entire supply chain, byzantine as it might sound, will make the company more efficient than its competitors. That’s because, if you squint, it looks like a souped-up and high-precision version of the existing supply chain for silicon photonic chips, another type of technology that transmits information with light—one that’s already used in data centers. If PsiQuantum produces its chips at scale, it can take advantage of tools and infrastructure that already exist.
But it’s not a given that one working chip can easily be wired up to thousands more. That’s why the company is testing in phases: Its Milpitas site has connected three cabinets together, with 250 chips in each, but the next step is to scale the systems up and see whether the company’s techniques for correcting errors can keep up. Once the cooling system arrives at the Australian site late next year, the company says, it aims to connect about 100 cabinets together. Then PsiQuantum will work up to running the world-changing algorithms it has promised.
The timeline for this, it’s worth noting, is up for debate. News articles have said that 2027 is the year that PsiQuantum aims to have its first full-scale quantum computer come online at its Australian site, but the company insists the deadline has been misread, and that it only intends for its facility to be “operational” by the end of next year. That means cooling systems in place and ready for hardware to be installed, but no promises about what size computer will be ready. In an industry where timelines are perpetually in flux yet central to how companies are judged, that distinction isn’t trivial.
Into the unknown
The outsider with perhaps the best guess of whether PsiQuantum will succeed is the Pentagon. The US Defense Advanced Research Projects Agency—the Pentagon’s research and development arm—has been running an initiative to determine which of the boastful quantum companies might actually deliver. In the last year and a half, the heads of the program have been sounding more confident. Joe Altepeter, who ran the program until last year and proudly described himself as a “quantum skeptic,” told me in March 2025: “I am more optimistic now than I have been at any point in the past 10 years.” And in a statement earlier this year, his successor, Micah Stoutimore, said “it now seems likely that someone will build a utility-scale quantum computer by 2033,” referring to a machine that generates more value from its calculations than it costs to build and operate.
The program has been scrutinizing PsiQuantum’s systems for over a year and putting them through the third stage of a benchmarking initiative meant to determine whether the technology will actually work. But to the rest of the industry, PsiQuantum is sort of a black box.
PsiQuantum has broken ground at the Illinois Quantum and Microelectronics Park outside Chicago, pictured here, and on another site in Moreton Bay, Australia. It aims to build large-scale quantum computers at each site.
COURTESY OF PSIQUANTUM
“It is very hard for an outsider to evaluate,” says Scott Aaronson, a theoretical computer scientist at the University of Texas at Austin who runs a popular blog that often covers the industry. Other companies, like Google and Quantinuum, have regularly published results over the years demonstrating chips and systems with incremental improvement, publicly laying the engineering groundwork needed to eventually build large machines.
PsiQuantum has instead focused squarely on a commercial goal—a computer with one million qubits, which is the scale that researchers expect to unlock research currently not possible on normal computers. PsiQuantum often differentiates itself with this industrial-scale goal, but IBM, which debuted a development road map in 2020, has been progressively building bigger and bigger systems. It initially targeted 2028 for a large-scale, error-corrected system, a deadline that now appears to have been pushed out to 2030.
Making it useful
On top of actually building the machine, a major focus for PsiQuantum is getting the rest of the world to develop a plan for how to use it. PsiQuantum has announced partnerships with customers including the defense giant Lockheed Martin, which intends to use it for materials design; the automaker Mercedes, which wants it for battery design; and the aerospace manufacturer Airbus.
That these companies don’t have a computer to experiment with is not a problem, according to Ernst at PsiQuantum. “There’s a PlayStation 6 probably coming up from Sony next year or the year after, and people are programming those games right now,” he says. “This is, in principle, very similar.” (It’s a glib analogy but not an entirely empty one; the quantum algorithms for solving a research problem can be cracked even if there is not yet hardware to run them on.)
The idea is that experts in quantum information from both PsiQuantum and its customers will be able to translate design problems—say, the requirements for a battery in a Mercedes electric vehicle—into algorithms the computer could solve. The company offers a software package called Construct, which companies can use to design their own algorithms that might one day run on the computer.
The future of quantum computing hinges on these algorithms. Quantum computers get painted as a speedup for everything, but in reality, they’re suited to a subset of problems, and answering a question with this sort of machine requires the question to be formulated with very specific types of algorithms. People spend entire careers working on such algorithms, even if the computers to run them don’t exist yet. At their core, they use the rules of quantum mechanics to manipulate probabilities in ways that ordinary computers can’t.
The most famous example, and a reason the government is so interested in quantum computers, is Shor’s algorithm. It was developed in 1994 by the theoretical computer scientist Peter Shor and could effectively break many forms of encryption used online, for everything from credit card numbers to military intelligence. The thing keeping the world together, for now, is that nobody has a computer to run the algorithm on (and security experts are already launching new encryption methods that could withstand attacks from a quantum computer). PsiQuantum is researching how long its systems might take to run Shor’s algorithm.
WINNI WINTERMEYER
WINNI WINTERMEYER
PsiQuantum’s chips are manufactured at GlobalFoundries in Malta, New York, and tested at company headquarters in California. Both PsiQuantum and GlobalFoundries have been awarded federal CHIPS Act funding.
The company also published a paper in December in collaboration with Airbus, essentially seeing if a new algorithm developed by the authors could beat a classical computer in modeling fluid dynamics, like the turbulence around an airplane wing. Andrew Childs, an expert in quantum simulation, told me PsiQuantum achieved only a moderate speed increase over what today’s computers can do. “It’s probably unlikely that speedups like this will have a significant practical impact until we have very large-scale quantum computers,” he said in an email. (When I asked Ernst, he agreed the improvement was modest.)
Some of the algorithms PsiQuantum is working on are not expected to be perfected or even used in the first applications of its computer. Instead, its initial tasks might be more along the lines that Feynman envisioned way back in 1981: simulating the smallest particles of our world.
The company’s most significant research in this realm is in modeling quantum chemistry. Take those pesky P450 enzymes. More precisely understanding how they operate, PsiQuantum says, would allow for faster drug development and testing.
Last year, PsiQuantum published methods for doing these sorts of chemistry calculations on a quantum computer, along with another paper demonstrating an algorithm that can simulate the collision of two molecules and estimate the likelihood of different outcomes femtosecond by femtosecond (there are one quadrillion femtoseconds in a second). It’s a remarkable amount of detail not currently possible with today’s technology, and it would allow drug and materials researchers to simulate new chemical interactions.
Dominic Berry, who developed some of the core techniques used in the collision paper but isn’t involved in PsiQuantum, says the company made impressive improvements, but to do the simulations scientists are most curious about would require the algorithm to be made even faster and PsiQuantum’s early computer to have fewer errors than currently expected.
Until PsiQuantum’s computers are up and running, the breakthroughs that these research papers tease remain in the realm of theory. It’s a space where Rudolph operates quite comfortably. He told me that Alan Turing created the theory of classical computing with pen and paper, imagining how the 1s and 0s would be represented in the machine, and how with the right approach to logic you could compute almost anything.
“But there is no way that by hand, with a pen and paper, Turing was ever going to produce—you know—Minecraft and Facebook,” he says. That took more than 70 years of tinkering (during which we fortunately created more useful things than Minecraft and Facebook).
For all the time Rudolph spends dreaming up things quantum computers might do, in other words, people working on those problems are still stuck with pen and paper for now: “Until you have the actual machine in hand, you don’t have the opportunity to really explore its potential.”
After more than a year of squabbling, a group of AIDS activists obtained an R&D agreement that was at the heart of a settlement between the U.S. government and Gilead Sciences over patents for HIV prevention drugs. But in their view, the deal shows the Biden administration missed a “historic” opportunity to invest in — and expand access to — HIV prevention tools.
As noted previously, the settlement resolved a lawsuit that was filed six years ago by the previous Trump administration after the Centers for Disease Control and Prevention maintained that Gilead infringed on its patent rights. The agency had helped fund academic research that later formed the basis for two Gilead HIV pills, Truvada and Descovy.
The administration had alleged that Gilead ignored the contributions by CDC scientists, exaggerated its own role in developing HIV prevention drugs, and refused to sign a licensing agreement despite “multiple attempts” at reaching a deal after unfairly reaping hundreds of millions of dollars from research funded by taxpayers.
This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.
Anthropic—currently the world’s most valuable AI company, with a nearly $1 trillion valuation—has a reputation for publishing strange and heady research. It’s looking into whether AI models can feel pain, for example, and will sometimes cut off chatbot conversations if it suspects users are “abusing” the model.
One niche that Anthropic spends more time and money on than other AI companies is called mechanistic interpretability, which means looking inside the complex math of an AI model to learn why it comes up with one particular output and not another. It’s complicated stuff; there are millions of data points that might contribute to any result, and wading through them can look more like word salad than anything useful. It’s also controversial. Describing AI models with terms borrowed from psychology and neuroscience can make their behavior seem more sophisticated than we might otherwise judge it to be.
That’s why, when Anthropic announced last week that it had found a new window into its models’ “internal thoughts” as they reason through answers, there was one colleague I had to talk to. Senior editor Will Douglas Heaven, aside from having a PhD in computer science, has spent a lot of time digging into what we can say about how AI models work. I spoke with him about what we should take from Anthropic’s new (and predictably quirky) research.
What did Anthropic learn here, exactly?
Anthropic has been trying to understand how large language models (LLMs) work for a few years now. Anthropic isn’t the only one looking at this, but I think the company has made it part of its core mission more than most. Anthropic’s CEO, Dario Amodei, has said we won’t be able to control LLMs fully unless we learn more about how they work.
So this new research is very much in that context. It goes deeper into the weird mechanisms inside LLMs than ever before. What Anthropic learned was that LLMs have a space inside them—which Anthropic calls the J-space—filled with words that don’t appear in their output but that seem to influence the way they puzzle through problems. All this was hidden until Anthropic developed a new technique to probe its model Claude, so it’s a genuine discovery.
Sometimes these words keep track of where the LLM has got to in a particular task, sometimes they look more like flashes of recognition (for example, “protein” might pop up when you give an LLM only the letters of a protein sequence), and sometimes they represent a kind of internal commentary on the model’s decision-making. In my favorite example, Claude decided to cheat on a coding test when the word “panic” appeared.
Anthropic also found that LLMs are able to describe and manipulate the words in this space. So somehow they seem to be making use of it.
Let’s step back for a second. I don’t think of large language models as simple, but they’re also not magic. There’s a bunch of math that learns relationships between words, right? So why is it so hard to “peer” into an LLM to know what’s going on?
Yeah, they’re not magic! I think the fact we don’t fully understand them plays into the mythmaking. And it’s worth noting that the whole narrative that Anthropic is leaning into here—that they’ve built this really mysterious technology, but don’t worry, because they’re also the ones to figure it out—very much fits with the company’s vibe. [See how Anthropic warned that its new models were so good at coding they posed a global cybersecurity risk, only for the US government to shut them down shortly thereafter.]
So yes: LLMs are just math. And yet it’s vastly complex math. Not only are today’s LLMs made out of hundreds of billions of numbers, but running them triggers a cascade of millions and millions of calculations. I wrote last year that if you printed out even a medium-size LLM on pieces of paper, it would cover a city the size of San Francisco.
It’s impossible to make sense of any of that math without specialist tools that highlight specific parts of an LLM at specific times. You need to know where to look and how to look. And building those tools requires understanding something of that complex math in the first place.
You’ve written elsewhere about this concept of studying LLMs the way one might study an organism’s brain. Is it fair to use “brain-like” terms when talking about how an LLM works?
I don’t love using those kinds of terms. LLMs are not brains. Talking like this is misleading because it can suggest that LLMs are capable of more human-like things than they are or that we can make assumptions about how they might behave that we shouldn’t. The whole anthropomorphization thing is also tied up with a bunch of strong ideological positions about what this technology is and what it’s going to be.
But at the same time, we lack a good alternative vocabulary for talking about what these models are doing. I can understand why people reach for words like “think” and “understand” and “brain-like”—they’re convenient shorthand.
Anthropic compares this new space it found inside LLMs to the space that some neuroscientists think our brains use to keep track of conscious thoughts. I asked the company how seriously we should take that comparison and it said in a statement: “Drawing these analogies was helpful to us in designing our experiments, as they allowed us to make many non-obvious experimental predictions about the J-space that turned out to be true. At the same time, it’s important to note that there are some important differences between the J-space (and language models in general) and the human brain, so we don’t mean to claim there’s a perfect correspondence.”
What’s a problem in AI that this new concept of the J-space might be used to solve?
Anthropic has said that monitoring the J-space could be a way to catch models doing something they shouldn’t. Because words pop up in this space that don’t appear in a model’s output, they can tell you things about its behavior that you might not have noticed otherwise—such as when it is giving biased responses or when it is weighing the pros and cons of cheating.
That’s the theory, at least. I think it’s better to think of this result as one more step on the path to understanding this technology overall than as something that will be useful by itself.
Good morning, everyone. Damian Garde here, filling in for Ed Silverman at Pharmalot’s satellite campus along the East River, where today’s cup of stimulation is filled not with coffee but rather a smoothie of curious color and questionable contents (what exactly is an “adaptogen”?). Anyway it’s Friday, as you’re almost certainly aware, and here are some tidbits to help you through the waning hours of another working week. …
German lawmakers passed a bill that would more than double the discount on branded medicines drugmakers must provide to the government,Reuters reports. The policy, part of an effort to plug a sizable budget gap in the country’s health insurance system, would increase the mandatory rebate from 7% to 15.5%. Industry groups have said the bill, if it clears Germany’s upper chamber, would deter investment and imperil the country’s access to new medicines.
The rapid rise of China’s biotech industry has led some American drug developers to do their work in near total secrecy,the Wall Street Journal observes. U.S. startups are increasingly loath to publish early data, disclose their scientific ambitions, or even publicize which diseases they hope to treat, all in fear that nimble Chinese firms will use that information to whip up competing drugs and beat them to the punch of starting clinical trials.
WASHINGTON — Health secretary Robert F. Kennedy Jr. is preparing to make it easier for people to claim that they were injured by a Covid-19 vaccine and receive compensation.
Kennedy is set to start the process of compiling a list of injuries that are presumed to be caused by Covid shots. People with those conditions could then ask for compensation from the government. It’s not clear what conditions may make the list — and that’s something that outside experts are keeping a close eye on.
Kennedy has long been critical of vaccines, saying none have been adequately safety tested. He’s said that he plans to overhaul a similar but separate program that provides compensation for individuals who claim injury by a vaccine recommended by the federal government.
Thousands of federal civil servants who academic researchers see as partners in conducting their work were fired. An unprecedented number of scientific projects funded by previous administrations were terminated. Universities were pressured to abandon diversity programs and work to curb health disparities. On a Friday evening, the government tried to push through a dramatic change to how it reimburses universities for research overhead.
All of these actions in the first year of the Trump administration were rapidly challenged in federal court, in many cases resulting in the administration having to walk back policies because they ran afoul of the Administrative Procedures Act, which governs how new policies and regulations are rolled out.
Andrew Twinamatsiko, who is director of the Center for Health Policy and the Law at Georgetown University and runs a health care litigation tracker, describes what happened last year as “tempests that we could weather” until there’s a new administration, when “there can be ways of reverting back to the baseline that we used to have.”
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology.
Your family’s $300 stake in OpenAI
Sam Altman’s proposal that Americans should share in the wealth created by AI is back in the spotlight, with reports that he is discussing giving the US government a 5% stake in OpenAI. At the company’s current valuation, that stake would be worth roughly $320 per American household.
The idea is meant to address concerns that AI companies are benefiting from human-generated work without compensating creators, while also easing fears that AI will cause a collapse of the labor market by providing a safety net.
The details, however, remain unclear. Indeed, the offer may be more powerful as a political narrative than as a policy plan.
This article is from The Algorithm, our weekly AI newsletter. Sign up to receive it in your inbox every Monday.
The must-reads
I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology.
1 A leaked Treasury report compares the AI market to the dotcom bubble Which contradicts the administration’s public optimism about AI. (NOTUS) + Fears that the market is overinflated are growing. (Reuters $) + And AI profits are hiding bigger risks in earnings reports. (FT $) + What even is the AI bubble? (MIT Technology Review)
2 Samsung profits have jumped 1,800% on booming AI chip sales It just reported its third consecutive record quarterly profit. (BBC) + But its shares slumped over fears that the AI boom will stall. (Reuters $) + That boom has turned Samsung into a $1 trillion company. (CNBC)
3 A US cyber agency is using Mythos to audit government code Sources say CISA is tapping Anthropic’s model to search for bugs. (Reuters $) + Agencies are using it despite Anthropic’s feud with the White House. (Axios)
4 Illinois’ governor has signed the nation’s strongest frontier AI law It’s designed to protect citizens from AI risks. (Gizmodo) + US lawmakers are clashing over AI rules. (MIT Technology Review)
5 A hidden tracker in Claude Code has been exposed and removed It secretly monitored users in China. (WP $) + Critics said it shows Anthropic’s willingness to surveil users. (Ars Technica) + The company has also found a hidden “thinking” space in Claude. (Axios)
6 Russia is suspected of flying drones over Europe from a shadow fleet The flights were reportedly launched from commercial ships. (Ars Technica) + Europe has a drone-filled vision for future wars. (MIT Technology Review)
7 A controversial AI “actor” is set to star in its first feature film Tilly Norwood will debut in a comedy-drama called “Misaligned.” (Variety) + A major actors union has lambasted the AI creation. (NBC News)
8 AI costs are driving US companies toward Chinese models Businesses are hunting for cheaper model alternatives. (CNBC) + Chinese AI labs are betting big on open source. (MIT Technology Review)
9 Researchers have shown quantum proofs can beat classical ones They found a problem that classical proofs can’t solve. (Quanta)
10 Earth will never be swallowed by the sun, according to new models But it probably won’t be much fun to live here by that point anyway! (Wired $)
Quote of the day
“The goal might be to make machines in our image. But what I fear is that—perhaps without even quite noticing—we remake ourselves in theirs.”
—Reporter Sarah O’Connor sounds a note of caution in her new book, We Are Not Machines, the Guardian reports.
One More Thing
KATE DEHLER
Adventures in the genetic time machine
Eske Willerslev, a specialist in recovering DNA from old bones and objects, has made numerous breakthroughs. These include recovering the first more or less complete genome of an ancient human and 2.4-million-year-old genetic material from Greenland, revealing that today’s Arctic desert was once a forest with poplar, birch, and mastodons.
These findings are part of a wave of discoveries from what’s being called an “ancient-DNA revolution.”
Beyond revealing stories of human migration and vanished ecosystems, scientists believe ancient DNA can unearth clues about modern diseases. It could even lead to a better food supply for our warming world. “And can we get that?” Willerslev asks. “Yes, I believe we can.”
