Seventy percent is a startling growth forecast for an already enormous company. According to TechCrunch, Nvidia chief executive Jensen Huang said revenue could grow that much year over year in its next fiscal year. With analysts expecting roughly $400 billion in the current year, that outlook points toward something near $680 billion.

Keep one label attached to that number: Huang's outlook. It is a claim about what may happen next, made by the person with the strongest reason to believe in Nvidia's expansion. To understand why he thinks it is possible, look beyond the chip.

A fast chip can still be left waiting

Return to the busy kitchen. The cooks are processors, the ingredients are data, and the counter space is memory. A narrow pantry door represents a slow connection. Ventilation stands in for cooling. More cooks help only when ingredients, handoffs, electricity and heat removal all keep pace.

Computers are not kitchens, but the bottleneck is real. Doing calculations is only part of the job. Large AI models split work across many processors. Those processors repeatedly exchange intermediate results, so a delay between them can leave expensive hardware idle.

Nvidia's GB200 NVL72 makes the shift visible. It is not a graphics card for an ordinary desktop. It is a liquid-cooled rack containing 72 Blackwell graphics processors and 36 Grace central processors. Nvidia's NVLink connections join the 72 GPUs into what the company presents as one large computing resource.

Did you know?

One GB200 NVL72 rack contains 2,592 Arm CPU cores, according to Nvidia. “Rack-scale” is literal here: the buyer is acquiring a tightly connected building block for a data centre, not a card to slide into a home computer.

The product stretches beyond the rack

Huang's argument extends beyond hardware specifications. Nvidia hears from memory suppliers, equipment makers, cloud companies, specialist AI clouds and businesses building data centres. A proposed site may begin as land, access to power and a building shell—the unfinished structure waiting for computers. By following that chain, Nvidia gets an unusually broad view of projects that could become future orders.

He also offered two signs of momentum: GB200 NVL72 sales growing 27 percent month over month, and companies considered for Nvidia investment holding a combined $100 billion in customer contracts. The reporting does not show the contracts, identify every company or define the time window behind the monthly figure. They are evidence in Huang's case, not numbers readers can independently audit from the article.

Growth can create its own questions

TechCrunch raises a circular-financing concern: Nvidia invests in some businesses that also buy its equipment. Huang says Nvidia looks for companies with real, revenue-producing customer contracts before investing. Both things can be true—the customers may have genuine demand, while the financial relationship still deserves scrutiny.

Competition is widening too. Amazon, Google and Microsoft design their own accelerators. AI companies and chip specialists are trying different architectures, while software improvements may reduce the computing needed for some tasks. None of that makes Nvidia's systems irrelevant; it makes a precise growth percentage harder to treat as destiny.

The durable shift is simpler than the forecast. The unit of progress is becoming the whole system: processors, memory, networking, cooling, power and buildings designed together. The next leap may depend less on hiring one faster cook than on redesigning the entire kitchen.

Neurons communicate electrochemically. When many cells in a network become active, tiny voltage changes appear around them. Electrodes can detect those extracellular signals and reveal patterns in the conversation.

The instrument, however, changes what it measures. A penetrating probe can damage tissue. A flat electrode array asks a round, suspended organoid to rest against a rigid plane. Either approach can make long-term observation difficult.

Xiao Yang, a researcher at Johns Hopkins University featured in MIT Technology Review's biotechnology coverage, designs electronics that behave more like tissue. Her group's organoid device is called KiriE.

A trellis around a growing plant

An organoid begins as a small cluster grown from human stem cells. Given the right conditions, its cells organize into structures that model selected features of a developing brain. It is far simpler than a brain, but useful for watching cells mature, connect and react to experimental changes.

Imagine placing a young climbing plant inside a soft trellis. As it grows, more branches meet the frame without being pinned to a board. KiriE follows that idea. The organoid is placed on an open, ultrathin device and grows into contact with electrodes arranged around it.

The comparison has limits: cells are not vines, and the basket does not record every neuron. It samples network activity at multiple contact points while allowing the tissue to remain in suspension.

The clever part is the cut

The team uses photolithography—a chip-making process that patterns material with light—to build thin layers of metal and polymer. Carefully placed cuts let a flat sheet open into basket-like spiral or honeycomb designs. That three-dimensional shape is the useful transformation.

Did you know?

The team used Python geometry tools and a network-flow algorithm to route tiny electrode traces to their contact pads without letting paths cross and short-circuit. Software helped design the physical basket.

In the published study, KiriE recorded cortical organoids for up to 120 days. A 32-channel system sampled the signals, allowing the researchers to follow activity over time and observe responses to drugs and light stimulation.

The device also worked with assembloids: organoids representing different brain regions joined together so researchers can study communication between them. In one experiment, light activated cells in a cortical region and electrodes detected a response in a connected striatal region.

A signal worth following, carefully

To demonstrate disease modelling, the team studied an organoid line with one altered copy of DGCR8, a gene in the chromosome region associated with 22q11.2 deletion syndrome. At the ages measured, those organoids fired spontaneously at roughly three times the rate of controls. About 30 percent of channels exceeded one firing event per second, compared with 2 percent in controls.

That result shows the platform can detect a known activity difference in one laboratory model. It does not turn the basket into a diagnosis, a treatment or a miniature window onto a person's thoughts. Organoids lack the full anatomy, circulation, sensory input and life history of a human brain.

The achievement is quieter and more useful: researchers can now listen to three-dimensional neural tissue for months while asking less of it physically. Sometimes a better instrument comes not from making it stronger, but from knowing where to let it bend.

The video was ordinary: a mother and her daughter singing in a car. What appeared beneath it was not.

According to The Verge and Futurism, Kalie Robins cross-posted the Instagram video to Facebook. Beneath it, Meta AI showed her a suggested prompt asking who the child passenger was.

After she selected the prompt, Robins said the chatbot named both daughters and proposed more questions about their ages, home, activities, books and favourite places. Some answers appeared to draw on years of posts by Robins and relatives, including birthday messages and birth announcements.

The puzzle-piece problem

Imagine a librarian who lets you browse open shelves. Now imagine the librarian sees one family photograph, instantly gathers every index card connected to the people in it and suggests personal questions you could ask next. The shelves may be accessible in both cases. The second experience is still different.

This is the aggregation problem. A birthday in one post, a relative's name in another and a trail photograph somewhere else can seem unremarkable alone. Search and synthesis collapse the time and effort once needed to connect them. Information that felt like a family scrapbook can become a profile.

Futurism reports that Robins's accounts were public but followed by only a few hundred people. That detail matters because public does not always feel like frictionless and instantly summarized. Privacy is shaped not only by who can technically see something, but by how easily a system can retrieve and combine it.

Did you know?

Meta has embedded the same assistant across Facebook, Instagram, WhatsApp and Messenger. That makes prompt design a cross-product safety question, not merely a quirk of one standalone chatbot.

What Meta says happened

Meta told The Verge that an answer only contains content the person asking can already access. It suggested that videos where Robins named her children may have supplied the names. Robins told Futurism that her own existing posts did not publicize them. Without system logs, the reporting cannot settle that disagreement.

The company agreed on the larger design failure. It said these personal questions should never have been suggested, that the feature had “missed the mark,” and that it fixed the issue responsible for personal-topic prompts. The reports could not independently test how broad that fix was.

Robins also said the assistant displayed a photograph she remembered deleting years earlier. Several explanations remain possible: another visible copy may have existed, a relative may have reposted it, or Meta may have retrieved it some other way. The available accounts do not reveal which.

The question before the answer

This is one documented experience, supported by screenshots and company statements. It cannot tell us how often similar prompts appeared, whether every surfaced detail was accurate or exactly how the system selected them.

It does show why chatbot safety begins before an answer is generated. A button can steer curiosity as powerfully as an answer can. A useful assistant should recognize the boundary between explaining a post and encouraging someone to investigate a family.

Imagine a bridge collapses. Investigators inspect the blueprint, materials, permits, load tests and maintenance records. They ask who designed it, who checked it and whether anyone ignored a warning. They do not begin by wondering whether the bridge had bad intentions.

That is the comparison AI researcher Timnit Gebru uses in a WIRED interview. Software and bridges fail in different ways, but the procedural lesson travels well: complicated technology still comes with a chain of human decisions.

Follow the decisions

Some prominent AI-safety debates centre on a dramatic question: could a future system become powerful enough to threaten humanity? Gebru thinks framing the danger as a model independently “going rogue” starts in the wrong place.

The interview arrives amid a real disagreement. AI companies and researchers have warned about catastrophic future systems and malicious uses such as biological-weapons assistance. Gebru separates those concerns. People using tools dangerously, companies deploying them recklessly and a machine independently deciding to eliminate humanity are not the same claim.

She points instead to choices already visible: autonomous weapons used in war, growing energy and water demands from computing, and employers citing AI while eliminating jobs. These examples form her case. The interview itself does not measure their scale or establish one motive shared by every company.

The useful method is following a trail of decisions rather than searching for a machine's personality. Who chose the use? Who supplied the data and infrastructure? What tests were required? Who could halt deployment? Who pays when people are harmed?

Did you know?

The 1955 Dartmouth proposal that helped name the field imagined a two-month, ten-person summer study. Its agenda already included language, neural networks, machine self-improvement and creativity—questions that still sound remarkably current.

This argument has a history

Gebru connects her position to the 2021 paper “On the Dangers of Stochastic Parrots,” which she co-authored with Emily M. Bender, Angelina McMillan-Major and Margaret Mitchell. The paper challenged the assumption that bigger language models are automatically better. It asked researchers to account for environmental cost, embedded bias, opaque training data and the people most exposed when systems fail.

That history helps explain why she resists treating today's concerns as a separate, smaller conversation. From her perspective, scale, power and accountability have been linked throughout the modern language-model boom.

Her stronger claim—that doom-focused narratives are meant to distract—goes further. WIRED presents her interpretation, but an interview cannot establish a single intention across the many researchers, officials and companies discussing long-term risk.

Keep both horizons in view

We do not need to choose between preparing for future dangers and challenging harmful uses now. A distant risk may deserve research even when experts disagree about its probability. Present deployments deserve rules and scrutiny because people are already living with them.

The grounding question is the same at either horizon: who is making the decision, under what safeguards, and with what consequence? Before getting lost in what a machine might someday want, start with what people are choosing to do with it now.

A solid crystal looks still because its atoms hold an orderly average arrangement. At any instant, however, those atoms are vibrating around their positions. The motion carries heat, shapes electrical behaviour and changes as a material warms.

The difficult question is not simply how much an atom moves. It is whether nearby atoms tend to move together. A team at the University of Tokyo has found a way to infer that relationship with an atomic-scale version of the double-slit experiment.

Two paths make a pattern

In the familiar classroom experiment, waves pass through two openings and overlap. A detector records an interference pattern: bright bands where waves reinforce one another and darker regions where they cancel.

Pond ripples offer a picture, but electron waves are not little surfaces rising and falling. Their wave description gives the probability of detecting electrons at different positions.

Instead of cutting holes in a barrier, the researchers used a scanning transmission electron microscope. They aimed an electron probe about 1.1 ångströms wide between two neighbouring atomic columns in silicon. As the electrons travelled through the thin crystal, the columns channelled and scattered them into two coherent sources. The waves then overlapped on a detector.

The columns were 136 picometres apart—far less than a billionth of a metre. Isolating one pair mattered because signals from many columns would overlap and hide the local relationship. To make the weak pattern clearer, the team averaged 356 crystallographically equivalent measurements.

Heat did not erase everything

The researchers repeated the experiment from 300 kelvin, around room temperature, to 900 kelvin, hot enough for atomic vibration to become much stronger. If every atom wandered independently, those changing paths would quickly smear the fine bands.

The higher-order fringes did weaken as temperature rose, but the pattern did not vanish all at once. At 900 kelvin, a second-order maximum remained faintly visible while finer structure had faded. Simulations that included correlated atomic vibration reproduced the measurements far better than models of independent motion.

Imagine two dancers stepping sideways together on a moving stage. Both change position, yet the distance between them barely changes. A pattern sensitive to their separation can stay sharp. Atoms are participating in many overlapping vibrational modes, not following choreography, but the analogy captures why relative motion matters more than motion alone.

Did you know?

The atomic sources were roughly seven orders of magnitude closer together than millimetre-spaced openings in a classroom double-slit experiment. The silicon was only about 10 nanometres thick, while the detector sat roughly 10 centimetres away. Explore the open paper.

A clue about how heat travels

By comparing fringe visibility with phonon-based simulations, the team could extract information about directional correlations in atomic motion. Phonons are collective vibrations travelling through a crystal; they are one of the main ways heat moves through many solids.

Those correlations also reflect how strongly neighbouring atoms are bonded. That makes the interference pattern a possible local probe of bond stiffness, lattice dynamics and the pathways along which thermal energy travels.

The camera did not film two atoms dancing side by side. Their shared motion was inferred by matching measurements to detailed simulations, and the cleanest signal came from averaging equivalent sites in a simple silicon crystal. Defects, interfaces and more complicated materials will be harder tests.

For now, the achievement is wonderfully specific: a way to use interference to infer how atomic neighbours move together, in a space smaller than many molecules.

A Martian slope can look as though a rainstorm has just swept through. But a shape is not a weather report. Wind, tumbling grains, ancient water and seasonal ice can carve forms that resemble one another from orbit.

A team led by Apolline Leclef studied active gullies at Sisyphi Cavi, around 68 degrees south. Their question was deliberately narrow: what is changing these slopes today?

The team's paper is a preprint, so other specialists have not yet completed formal peer review. Its evidence chiefly concerns ongoing modification at this site, not the original formation of every Martian gully.

Two clues: what is there, and when

The researchers combined infrared observations from NASA's Mars Reconnaissance Orbiter and the European Space Agency's Mars Express with models of seasonal ice. Different materials absorb and reflect infrared light in characteristic ways, allowing scientists to map frost composition from orbit.

They found seasonal carbon-dioxide ice. Near the end of the frost season, they did not detect an independent surface deposit of water ice lingering after the carbon dioxide disappeared. Weak sulfate signatures were present, but their distribution did not line up convincingly with the active gullies.

Small amounts of water ice could still be buried, hidden in shade, mixed into carbon-dioxide frost or simply below an instrument's detection limit. The useful result is not “Mars has no water ice.” It is that the available maps did not show water ice behaving like the main seasonal driver at Sisyphi Cavi.

Then came the calendar. Dark spots linked to carbon-dioxide geysers appeared from late winter into early spring. The observed gully changes came later, in mid-to-late spring as frost thinned and vanished. Like a detective comparing an event with a suspect's timetable, the mismatch makes geysers a less convincing principal cause.

Gas can help grains move

The evidence instead favours avalanches or granular flows assisted by carbon-dioxide ice turning directly into gas. That change is called sublimation: the solid skips the liquid stage.

Picture a bag of sand with air pushed upward between the grains. As the gas reduces contact and friction, the sand can begin to behave more like a flowing material. On Mars, sublimating carbon dioxide may similarly help loosen a slope. The comparison describes a proposed physical mechanism; the orbiters did not film gas lifting individual grains.

Did you know?

Sisyphi Cavi lies near 68 degrees south, while most known Martian gullies occur between roughly 30 and 50 degrees latitude. Its unusually polar setting makes seasonal carbon-dioxide ice especially important to investigate. See the study site.

Present-day movement, ancient water

Mars really did host rivers, lakes and floods billions of years ago. Evidence for that wetter past appears across the planet. A dry-ice explanation for one modern slope does not rewrite ancient Martian history.

It also leaves room for better observations. A concealed water-ice component may have escaped detection, and a future mission could watch the surface process more directly. The current case is assembled from composition, seasonal timing and models—three clues that point in the same direction without giving us a close-up video.

That is still meaningful progress. The study separates two questions often blurred together: how Martian gullies first formed and what modifies some of them now. At Sisyphi Cavi, the latest changes may be telling a story about retreating frost and flowing grains, not a present-day river.

If you wanted to make something white, reaching for pigment would be a reasonable start. These researchers reached for structure instead.

A team led by Kyoto University has developed a way to turn selected parts of a polymer into microscopic foam. Its many internal boundaries scatter visible light, making the material look white without adding a white pigment.

Whiteness built from boundaries

Snow, clouds and some biological tissues share a broad optical trick. Light repeatedly meets boundaries between materials with different refractive properties. Each encounter changes its direction. After enough scattering, many visible wavelengths return toward our eyes together and we perceive white.

Imagine a crowded hall full of tiny windows, each angled differently. A beam entering the hall keeps changing direction as it crosses from one space to another. Light in the foam behaves as a wave rather than a traveller in a building, but the image captures the importance of many internal interfaces.

First light, then foam

The team starts with a light-sensitive polymer. Deeply penetrating ultraviolet-A light does two jobs in different parts of the material: it breaks some polymer chains and crosslinks others into a network. A weak solvent then enters, removes mobile fragments and helps pores nucleate and expand inside the crosslinked regions.

The researchers call the method deep-foam photolithography. In plain language, light writes a three-dimensional recipe into the material, and the solvent develops that recipe into foam. Varying the exposure changes how much the structure expands, which allows white and greyscale patterns rather than a simple on-or-off mark.

Did you know?

The process reached about 20,000 dots per inch—fine enough to create not only white and greyscale images but structures that separate light into diffractive colour. See the paper record.

The water trick is about the surface

The same foam can be collapsed in a controlled, viscoelastic way. That leaves roughness at several scales. A water droplet touches less solid surface, so it beads up and rolls more easily—an effect often compared with a lotus leaf.

The distinction matters. Pores inside the material scatter light; texture at its surface controls water. They are related design tools, not one mysterious property doing every job.

The platform reaches beyond appearance. The team demonstrated high-aspect-ratio foam structures, microfluidic patterns and tiny regions able to capture picolitre-scale droplets or colloidal particles. It also worked with several kinds of polymer and in both films and fibres.

From clever sample to useful material

The process could reduce reliance on white pigments or persistent fluorinated coatings in applications where structural effects are enough. That possibility is attractive, but a laboratory route is not automatically a greener product.

The final balance will depend on the chosen polymer, solvent recovery, energy use, manufacturing yield and service life. Outdoor coatings must survive abrasion, sunlight, dirt and years of weather; those tests are separate from producing an impressive sample.

For now, the advance is a versatile way to print function into a material's architecture. Sometimes the missing ingredient is not another substance. It is a different structure.

Most astronomy studies distant worlds by catching their light. In 2025, a small piece of another planetary system came through our own.

The object is called 3I/ATLAS. The “I” means interstellar, and the number tells us it is only the third visitor of this kind that astronomers have identified. Its path is hyperbolic: it was not born in orbit around the Sun and will not remain here. It swept through, curved under the Sun's gravity and continued outward.

That brief passage turned the Solar System into a laboratory for material assembled around an unknown star.

Reading a tail like a barcode

As a comet approaches the Sun, heat releases gas and dust from its nucleus. Sunlight and the solar wind then alter that material, creating a glowing atmosphere and tails. One tail is rich in ions—molecules carrying an electrical charge.

Astronomers cannot scoop those ions into a jar, but they can study their light. Pass light through an instrument that separates wavelengths and a pattern of lines appears. Each atom, molecule or ion contributes characteristic features. The result is often compared with a barcode.

The comparison is useful, with one caution: reading abundance from line brightness is not as simple as scanning groceries. Sunlight, collisions, chemical reactions and the solar wind all affect what glows and how strongly.

After 3I/ATLAS passed perihelion, its closest point to the Sun, researchers observed it on 30 November and 2 December 2025. They used the WEAVE instrument on the 4.2-metre William Herschel Telescope in the Canary Islands.

They detected five ions in the plasma tail at the same time: N₂⁺, CO⁺, CO₂⁺, H₂O⁺ and CH⁺. The clearest compositional clue came from comparing ionized nitrogen with ionized carbon monoxide.

Did you know?

WEAVE's Large Integral Field Unit uses 547 closely packed optical fibres. Instead of blending the entire comet into one measurement, it can compare spectra from different positions across the extended tail.

A chemical memory of deep cold

Nitrogen molecules are difficult to trap and preserve in cometary ice unless the environment is extremely cold. The team's lower limit for the nitrogen-to-carbon-monoxide ratio was high compared with measurements of Solar System comets.

Models connect that nitrogen-rich chemistry with formation below roughly −240 °C. Think of the spectrum as a luggage tag left on a traveller's bag. It cannot give us a street address, but it hints at the conditions where the bag was packed.

The temperature was not measured with a thermometer at the comet's birth. It is an interpretation based on which volatile molecules could have remained trapped as its material formed. The tail has also been processed by solar heating, outgassing, sunlight and charged particles. Researchers account for those effects, but they cannot rewind the comet perfectly.

The careful conclusion is that 3I/ATLAS is consistent with formation in an exceptionally cold outer region of another planetary system, far from its parent star.

One traveller, not a census

Interstellar objects are valuable precisely because ordinary comets belong to our own Solar System. Their chemistry reflects the disk of gas and dust that surrounded the young Sun. A visitor such as 3I/ATLAS offers a sample shaped elsewhere.

NASA's observations place the comet's nucleus somewhere between roughly 440 metres and 5.6 kilometres across—a wide range that shows how difficult it is to measure a bright, active object hidden inside its own haze. The spectrum reveals part of its volatile history, not a complete inventory of the nucleus.

Nor can one comet represent the variety of an entire galaxy. It may come from an unusual system, and current surveys notice only the visitors bright and close enough to detect. “Third known” describes our short observing history, not cosmic rarity.

Still, consider what happened. Material escaped an unknown planetary system, travelled through interstellar space and crossed the reach of our telescopes. For a few nights, light from its tail carried a chemical memory of another sun's cold outskirts. We could not follow it home, but we learned something about where it had been.