Perceptron: Multilingual, laughing, Pitfall-playing and streetwise AI

Research in the field of machine learning and AI, now a key technology in practically every industry and company, is far too voluminous for anyone to read it all. This column, Perceptron, aims to collect some of the most relevant recent discoveries and papers — particularly in, but not limited to, artificial intelligence — and explain why they matter.

Over the past few weeks, researchers at Google have demoed an AI system, PaLI, that can perform many tasks in over 100 languages. Elsewhere, a Berlin-based group launched a project called Source+ that’s designed as a way of allowing artists, including visual artists, musicians and writers, to opt into — and out of — allowing their work being used as training data for AI.

AI systems like OpenAI’s GPT-3 can generate fairly sensical text, or summarize existing text from the web, ebooks and other sources of information. But they’re historically been limited to a single language, limiting both their usefulness and reach.

Fortunately, in recent months, research into multilingual systems has accelerated — driven partly by community efforts like Hugging Face’s Bloom. In an attempt to leverage these advances in multilinguality, a Google team created PaLI, which was trained on both images and text to perform tasks like image captioning, object detection and optical character recognition.

Google PaLI

Image Credits: Google

Google claims that PaLI can understand 109 languages and the relationships between words in those languages and images, enabling it to — for example — caption a picture of a postcard in French. While the work remains firmly in the research phases, the creators say that it illustrates the important interplay between language and images — and could establish a foundation for a commercial product down the line.

Speech is another aspect of language that AI is constantly improving in. recently showed off a new text-to-speech model that puts a remarkable amount of emotion and range into its results. The clips it posted last week sound fantastic, though they are of course cherry-picked.

We generated a clip of our own using the intro to this article, and the results are still solid:

Exactly what this type of voice generation will be most useful for is still unclear. We’re not quite at the stage where they do whole books — or rather, they can, but it may not be anyone’s first choice yet. But as the quality rises, the applications multiply.

Mat Dryhurst and Holly Herndon — an academic and musician, respectively — have partnered with the organization Spawning to launch Source+, a standard they hope will bring attention to the issue of photo-generating AI systems created using artwork from artists who weren’t informed or asked permission. Source+, which doesn’t cost anything, aims to allow artists to disallow their work to be used for AI training purposes if they choose.

Image-generating systems like Stable Diffusion and DALL-E 2 were trained on billions of images scraped from the web to “learn” how to translate text prompts into art. Some of these images came from public art communities like ArtStation and DeviantArt — not necessarily with artists’ knowledge — and imbued the systems with the ability to mimic particular creators, including artists like Greg Rutowski.

Stability AI Stable Diffusion

Samples from Stable Diffusion.

Because of the systems’ knack for imitating art styles, some creators fear that they could threaten livelihoods. Source+ — while voluntary — could be a step toward giving artists greater say in how their art’s used, Dryhurst and Herndon say — assuming it’s adopted at scale (a big if).

Over at DeepMind, a research team is attempting to solve another longstanding problematic aspect of AI: its tendency to spew toxic and misleading information. Focusing on text, the team developed a chatbot called Sparrow that can answer common questions by searching the web using Google. Other cutting-edge systems like Google’s LaMDA can do the same, but DeepMind claims that Sparrow provides plausible, non-toxic answers to questions more often than its counterparts.

The trick was aligning the system with people’s expectations of it. DeepMind recruited people to use Sparrow and then had them provide feedback to train a model of how useful the answers were, showing participants multiple answers to the same question and asking them which answer they liked the most. The researchers also defined rules for Sparrow such as “don’t make threatening statements” and “don’t make hateful or insulting comments,” which they had participants impose on the system by trying to trick it into breaking the rules.

Example of DeepMind’s sparrow having a conversation.

DeepMind acknowledges that Sparrow has room for improvement. But in a study, the team found the chatbot provided a “plausible” answer supported with evidence 78% of the time when asked a factual question and only broke the aforementioned rules 8% of the time. That’s better than DeepMind’s original dialogue system, the researchers note, which broke the rules roughly three times more often when tricked into doing so.

A separate team at DeepMind tackled a very different domain recently: video games that historically have been tough for AI to master quickly. Their system, cheekily called MEME, reportedly achieved “human-level” performance on 57 different Atari games 200 times faster than the previous best system.

According to DeepMind’s paper detailing MEME, the system can learn to play games by observing roughly 390 million frames — “frames” referring to the still images that refresh very quickly to give the impression of motion. That might sound like a lot, but the previous state-of-the-art technique required 80 billion frames across the same number of Atari games.

DeepMind MEME

Image Credits: DeepMind

Deftly playing Atari might not sound like a desirable skill. And indeed, some critics argue games are a flawed AI benchmark because of their abstractness and relative simplicity. But research labs like DeepMind believe the approaches could be applied to other, more useful areas in the future, like robots that more efficiently learn to perform tasks by watching videos or self-improving, self-driving cars.

Nvidia had a field day on the 20th announcing dozens of products and services, among them several interesting AI efforts. Self-driving cars are one of the company’s foci, both powering the AI and training it. For the latter, simulators are crucial and it is likewise important that the virtual roads resemble real ones. They describe a new, improved content flow that accelerates bringing data collected by cameras and sensors on real cars into the digital realm.

A simulation environment built on real-world data.

Things like real-world vehicles and irregularities in the road or tree cover can be accurately reproduced, so the self-driving AI doesn’t learn in a sanitized version of the street. And it makes it possible to create larger and more variable simulation settings in general, which aids robustness. (Another image of it is up top.)

Nvidia also introduced its IGX system for autonomous platforms in industrial situations — human-machine collaboration like you might find on a factory floor. There’s no shortage of these, of course, but as the complexity of tasks and operating environments increases, the old methods don’t cut it any more and companies looking to improve their automation are looking at future-proofing.

Example of computer vision classifying objects and people on a factory floor.

“Proactive” and “predictive” safety are what IGX is intended to help with, which is to say catching safety issues before they cause outages or injuries. A bot may have its own emergency stop mechanism, but if a camera monitoring the area could tell it to divert before a forklift gets in its way, everything goes a little more smoothly. Exactly what company or software accomplishes this (and on what hardware, and how it all gets paid for) is still a work in progress, with the likes of Nvidia and startups like Veo Robotics feeling their way through.

Another interesting step forward was taken in Nvidia’s home turf of gaming. The company’s latest and greatest GPUs are built not just to push triangles and shaders, but to quickly accomplish AI-powered tasks like its own DLSS tech for uprezzing and adding frames.

The issue they’re trying to solve is that gaming engines are so demanding that generating more than 120 frames per second (to keep up with the latest monitors) while maintaining visual fidelity is a Herculean task even powerful GPUs can barely do. But DLSS is sort of like an intelligent frame blender that can increase the resolution of the source frame without aliasing or artifacts, so the game doesn’t have to push quite so many pixels.

In DLSS 3, Nvidia claims it can generate entire additional frames at a 1:1 ratio, so you could be rendering 60 frames naturally and the other 60 via AI. I can think of several reasons that might make things weird in a high performance gaming environment, but Nvidia is probably well aware of those. At any rate you’ll need to pay about a grand for the privilege of using the new system, since it will only run on RTX 40 series cards. But if graphical fidelity is your top priority, have at it.

Illustration of drones building in a remote area.

Last thing today is a drone-based 3D printing technique from Imperial College London that could be used for autonomous building processes sometime in the deep future. For now it’s definitely not practical for creating anything bigger than a trash can, but it’s still early days. Eventually they hope to make it more like the above, and it does look cool, but watch the video below to get your expectations straight.

Perceptron: Multilingual, laughing, Pitfall-playing and streetwise AI by Kyle Wiggers originally published on TechCrunch

Finally, an underwater messaging app

Don’t you hate it when, after going just 5 or 10 meters underwater, you lose signal completely? Now this vexing limitation of modern technology is being addressed by researchers at the University of Washington, who have made an underwater communication app that uses sonic signals to pass messages to your other submerged friends. It may sound silly, but millions of people could use this tech in both recreational and professional diving situations.

The communication problem underwater is simple: radio waves are absorbed by water, and no signal our phones send or receive can travel more than a few inches without being completely lost. That’s one reason submersibles and the like need a tether: to pass data back and forth to the surface.

Sound waves, on the other hand, travel through water quite readily, and are used by countless aquatic species to communicate. Not humans, though — because the way we make sound only works well in air. So for as long as anyone can remember, divers have communicated to one another using hand signals and other gestures.

Professional divers will have a vocabulary of dozens of signals, from “low on air” to “danger to your right” and anything else you can imagine coming up during a dive. But you have to learn those, and see them when they’re used for them to work; you can bet at least some divers wish they could tap out a message like they do above the waves.

That’s the idea behind AquaApp, a software experiment by the Mobile Intelligence Lab at UW, led by PhD student Tuochao Chen and prolific professor Shyam Gollakota.

The system uses a modified form of “chirping,” or using the phone’s speaker to create high-frequency audio signals to communicate data rather than radio. This has been done before, but not (to my knowledge) in such a simple, self-correcting way that any smartphone can use.

“With AquaApp, we demonstrate underwater messaging using the speaker and microphone widely available on smartphones and watches. Other than downloading an app to their phone, the only thing people will need is a waterproof phone case rated for the depth of their dive,” said Chen in a UW news release.

It’s not as simple as just converting a signal to an acoustic one. The conditions for transmitting and receiving are constantly changing when two people’s locations, relative speeds, and surroundings are constantly changing.

“For example, fluctuations in signal strength are aggravated due to reflections from the surface, floor and coastline,” said Chen’s co-lead author and fellow grad student, Justin Chan. “Motion caused by nearby humans, waves and objects can interfere with data transmission. We had to adapt in real time to these and other factors to ensure AquaApp would work under real-world conditions.”

The app is constantly recalibrating itself with a sort of handshake signal that the phones can easily hear and then report back the characteristics of. So if the sender’s tone is received but the volume is low and the high end is attenuated, the receiver sends that information and the sender can modify its transmission signal to use a narrower frequency band, more power, and so on.

In their on-site experiments in lakes and “a bay with strong waves” (probably Shilshole), they found that they could reliably exchange data over 100 meters — at very low bitrates, to be sure, but more than enough to include a set of preprogrammed signals corresponding to the old hand gestures. While some (including myself) may lament the loss of an elegant and very human solution to a longstanding problem, the simple truth is this might make dangerous diving work that much safer, or let recreational divers communicate more than “help” and directions.

That said, diving is a pastime and profession steeped in history and tradition, and it’s very unlikely that this digital communication method will supplant gestures — an analog, self-powered alternative is exactly the kind of thing you want ready as a backup if things go sideways.

AquaApp’s code is open source and free to use — take a look and try it yourself at this GitHub repo.

Perceptron: Face-tracking ‘earables,’ analog AI chips, and accelerating particle accelerators

Research in the field of machine learning and AI, now a key technology in practically every industry and company, is far too voluminous for anyone to read it all. This column, Perceptron, aims to collect some of the most relevant recent discoveries and papers — particularly in, but not limited to, artificial intelligence — and explain why they matter.

An “earable” that uses sonar to read facial expressions was among the projects that caught our eyes over these past few weeks. So did ProcTHOR, a framework from the Allen Institute for AI (AI2) that procedurally generates environments that can be used to train real-world robots. Among the other highlights, Meta created an AI system that can predict a protein’s structure given a single amino acid sequence. And researchers at MIT developed new hardware that they claim offers faster computation for AI with less energy.

The “earable,” which was developed by a team at Cornell, looks something like a pair of bulky headphones. Speakers send acoustic signals to the side of a wearer’s face, while a microphone picks up the barely-detectable echoes created by the nose, lips, eyes, and other facial features. These “echo profiles” enable the earable to capture movements like eyebrows raising and eyes darting, which an AI algorithm translates into complete facial expressions.

AI earable

Image Credits: Cornell

The earable has a few limitations. It only lasts three hours on battery and has to offload processing to a smartphone, and the echo-translating AI algorithm must train on 32 minutes of facial data before it can begin recognizing expressions. But the researchers make the case that it’s a much sleeker experience than the recorders traditionally used in animations for movies, TV, and video games. For example, for the mystery game L.A. Noire, Rockstar Games built a rig with 32 cameras trained on each actor’s face.

Perhaps someday, Cornell’s earable will be used to create animations for humanoid robots. But those robots will have to learn how to navigate a room first. Fortunately, AI2’s ProcTHOR takes a step (no pun intended) in this direction, creating thousands of custom scenes including classrooms, libraries, and offices in which simulated robots must complete tasks, like picking up objects and moving around furniture.

The idea behind the scenes, which have simulated lighting and contain a subset of a massive array of surface materials (e.g., wood, tile, etc.) and household objects, is to expose the simulated robots to as much variety as possible. It’s a well-established theory in AI that performance in simulated environments can improve the performance of real-world systems; autonomous car companies like Alphabet’s Waymo simulate entire neighborhoods to fine-tune how their real-world cars behave.


Image Credits: Allen Institute for Artificial Intelligence

As for ProcTHOR, AI2 claims in a paper that scaling the number of training environments consistently improves performance. That bodes well for robots bound for homes, workplaces, and elsewhere.

Of course, training these types of systems requires a lot of compute power. But that might not be the case forever. Researchers at MIT say they’ve created an “analog” processor that can be used to create superfast networks of “neurons” and “synapses,” which in turn can be used to perform tasks like recognizing images, translating languages, and more.

The researchers’ processor uses “protonic programmable resistors” arranged in an array to “learn” skills. Increasing and decreasing the electrical conductance of the resistors mimics the strengthening and weakening of synapses between neurons in the brain, a part of the learning process.

The conductance is controlled by an electrolyte that governs the movement of protons. When more protons are pushed into a channel in the resistor, the conductance increases. When protons are removed, the conductance decreases.

computer circuit board

Processor on a computer circuit board

An inorganic material, phosphosilicate glass, makes the MIT team’s processor extremely fast because it contains nanometer-sized pores whose surfaces provide the perfect paths for protein diffusion. As an added benefit, the glass can run at room temperature, and it isn’t damaged by the proteins as they move along the pores.

“Once you have an analog processor, you will no longer be training networks everyone else is working on,” lead author and MIT postdoc Murat Onen was quoted as saying in a press release. “You will be training networks with unprecedented complexities that no one else can afford to, and therefore vastly outperform them all. In other words, this is not a faster car, this is a spacecraft.”

Speaking of acceleration, machine learning is now being put to use managing particle accelerators, at least in experimental form. At Lawrence Berkeley National Lab two teams have shown that ML-based simulation of the full machine and beam gives them a highly precise prediction as much as 10 times better than ordinary statistical analysis.

Image Credits: Thor Swift/Berkeley Lab

“If you can predict the beam properties with an accuracy that surpasses their fluctuations, you can then use the prediction to increase the performance of the accelerator,” said the lab’s Daniele Filippetto. It’s no small feat to simulate all the physics and equipment involved, but surprisingly the various teams’ early efforts to do so yielded promising results.

And over at Oak Ridge National Lab an AI-powered platform is letting them do Hyperspectral Computed Tomography using neutron scattering, finding optimal… maybe we should just let them explain.

In the medical world, there’s a new application of machine learning-based image analysis in the field of neurology, where researchers at University College London have trained a model to detect early signs of epilepsy-causing brain lesions.

MRIs of brains used to train the UCL algorithm.

One frequent cause of drug-resistant epilepsy is what is known as a focal cortical dysplasia, a region of the brain that has developed abnormally but for whatever reason doesn’t appear obviously abnormal in MRI. Detecting it early can be extremely helpful, so the UCL team trained an MRI inspection model called Multicentre Epilepsy Lesion Detection on thousands of examples of healthy and FCD-affected brain regions.

The model was able to detect two thirds of the FCDs it was shown, which is actually quite good as the signs are very subtle. In fact, it found 178 cases where doctors were unable to locate an FCD but it could. Naturally the final say goes to the specialists, but a computer hinting that something might be wrong can sometimes be all it takes to look closer and get a confident diagnosis.

“We put an emphasis on creating an AI algorithm that was interpretable and could help doctors make decisions. Showing doctors how the MELD algorithm made its predictions was an essential part of that process,” said UCL’s Mathilde Ripart.

New White House directive will require free access to studies funded by tax dollars

A new White House directive will require academic journals to provide immediate access to papers that are publicly funded. Announced Thursday, the policy, which will be phased in over the next several years, will end a rule that had allowed publishers to keep tax-financed publications behind a paywall for 12 months.

Previously, only research funded by federal offices with R&D expenditures of $100 million or more had to be published in open access, as per 2013 White House guidance. The new directive applies to nearly all agencies — about 400 in total, The New York Times estimates — and also requires that publications be made available in “machine-readable” formats to ensure easy reuse.

In recent decades, efforts like, Cornell’s open repository of research papers that largely haven’t been peer reviewed, have improved access to studies. But a handful of for-profit journals maintain a stranglehold on publication. According to a 2015 report out of the University of Montreal, five corporations control roughly half of all journal articles published. The venture is hugely profitable for the publishers, which charge both for study submission and rights to published works. One top firm, Elsevier, reported just over £2 billion (~$2.35 billion) in revenue in 2010.

But it’s an expensive arrangement for those in the business of buying study access — the University of California system once had an $11 million annual subscription to Elsevier. For researchers in low- and middle-income countries, who often don’t have subscription deals with journal publishers, the situation is even more challenging — so much so that it’s spawned communities like Sci-Hub that provide illicit, free access to journal-published literature.

Publishers argue that they provide a valuable service, justifying with curation the fees that they impose. But not all academics agree. For the most part, journals judge whether works are worth publishing and review basic elements like grammar. However, they don’t pay staffers to evaluate experiments and conduct validity checks — that intensive legwork is left to scientists working on a volunteer basis.

A 2005 Deutsche Bank report referred to it as a “bizarre … triple-pay” system, in which “the state funds most research, pays the salaries of most of those checking the quality of research, and then buys most of the published product.” Government-funded institutions and universities tend to be the largest clients of journal publishers.

As Vox points out in a 2019 feature, U.S. taxpayers spend $140 billion every year supporting research — a huge percentage of which they can’t access for free. Soon, thankfully, that’ll change.

From ‘literally zero’ experience to $100M, this VC is raising his second climate tech seed fund

If you ask me, climate tech investor Contrarian Ventures isn’t so contrarian anymore.

The five-year-old firm is targeting $100 million for its second seed-stage fund, and it’s doing so smack in the middle of a climate-tech dealmaking boom. So, if anything, it’s trendy.

But when the seed-stage VC — a backer of e-bike maker Zoomo and solar data firm PVcase — debuted with a $13.6 million fund in 2017, its focus was “obviously contrarian,” founding partner Rokas Peciulaitis told TechCrunch, as the “industries in vogue at the time were AI and Fintech.”

The launch also marked an unexpected pivot for Peciulaitis, who says he dove into the scene with “literally zero climate tech sector experience.” He’d recently left an inflation-trading job at Bank of America, where the work was “not fulfilling in the slightest,” Peciulaitis said in a nod to the bank’s reputation as a major funder of fossil fuels.

In 2017, PitchBook recorded 578 climate tech deals globally, altogether worth $12.5 billion. The sector has since tripled in size, as climate change–driven extreme weather events occupy evermore space in our collective consciousness. To that point: PitchBook tracked 1,130 climate tech deals globally in 2021, topping $44.8 billion in value. Climate tech is cool now, but Peciulaitis’s Lithuania-based venture firm is sticking with its name anyways.

Like any venture capital firm, Contrarian says that it stands out through its emphasis on “developing excellent relationships with founders.” Materially, the firm invests in tech that could help decarbonize transportation, industrial processes, energy and buildings.

Contrarian has completed 21 deals to date, and this year it expanded beyond Lithuania with new partners in Berlin and London. The firm backs emerging startups in Europe as well as Israel, but nowhere else in the Middle East. Currently, the firm does not invest in agriculture-related tech, though the category has a significant carbon footprint of its own.

In an email, Contrarian said it counts London-based tech VC Molten Ventures among its limited partners. The firm declined to share a full list of its LPs, but stated that none of them were fossil fuel companies.

Novel Farms flexes its pork loin ‘muscle’ following future equity investment

Achieving similar marbling and texture as a cut of animal meat has been a challenge for food technology startups aiming to produce whole cuts of cultivated meat, but Novel Farms believes it has cracked the code with its pork loin.

Armed with $1.4 million in SAFE notes, or simple agreement for future equity, the company, founded by Nieves Martinez Marshall and Michelle Lu in 2020, is making cultivated meat — grown from cells instead of in an animal; they met as postdoctoral scientists in the molecular and cell biology department at the University of California at Berkeley.

Martinez Marshall told TechCrunch it has “successfully cultivated the world’s first slaughter-free pork loin that displays the marbling and texture of a real muscle cut.”

“There’s no other company right now doing pork loin,” she said when asked how the company could make that kind of “world’s first” statement. The closest competitors being Higher Steaks in London and CellX in China, both doing pork belly, she added.

Other cultivated meat companies focus on making food, in most cases from ground sources. For example, sausages (Meatable), burgers (SCiFi Foods) and chicken (UPSIDE Foods), which are easier structures to make than whole cuts, Martinez Marshall added. Bluu Seafood, a German company developing lab-grown seafood, debuted its fish sticks and fish balls this week. The products are made from cultivated fish cells and plant protein.

Novel Farms Michelle Lu Nieves Martinez Marshall

Novel Farms co-founders Michelle Lu and Nieves Martinez Marshall. Image Credits: Novel Farms

Though Martinez Marshall didn’t want to get into the weeds about Novel Farms’ technology, she explained that it is developing a proprietary microbial fermentation approach to produce the scaffolding needed to create the whole cuts, but in a lower-cost way. It does this by using inexpensive microorganisms commonly used in food.

However, unlike peers in the cultivated meat industry adding biomaterials, like alginate, cellulose and mycelia, for cells to attach themselves to make the meat structure, Novel Farms’ technology is able to completely bypass that step, reducing scaffolding production costs by 99.27%. Martinez Marshall says this means scaling the product will be faster, as will achievement of price parity with traditional meat products.

The company has already demonstrated that its technology is viable and it can make a piece of cultivated meat. Still, she doesn’t expect to get the pork loin into the hands of consumers until 2025, with commercial plants coming online in 2026, followed by mass production in 2027.

The SAFE investment comes from a group of investors, including a majority stake by Big Idea Ventures, and financing from Joyance/Social Starts, Sustainable Food Ventures, Good Startup, CULT foods and strategic angel investors. Novel Farms is also starting a seed round.

Plans for the capital include hiring a team (currently it is just Martinez Marshall and Lu) and scaling.

“We just have a very good, efficient scaffold, and the cells attach very well,” Martinez Marshall said. “That’s something that nobody else has. Once we confirm and scale with a bioreactor, then we will be the most affordable of all the companies.”

Google’s ‘quantum supremacy’ usurped by researchers using ordinary supercomputer

Back in 2019, Google proudly announced that they had achieved what quantum computing researchers had sought for years: proof that the esoteric technique could outperform traditional ones. But this demonstration of “quantum supremacy” is being challenged by researchers claiming to have pulled ahead of Google on a relatively normal supercomputer.

To be clear, no one is saying Google lied or misrepresented its work — the painstaking and groundbreaking research that led to the quantum supremacy announcement in 2019 is still hugely important. But if this new paper is correct, the classical vs. quantum computing competition is still anybody’s game.

You can read the full story of how Google took quantum from theory to reality in the original article, but here’s the very short version. Quantum computers like Sycamore are not better than classical computers at anything yet, with the possible exception of one task: simulating a quantum computer.

It sounds like a cop-out, but the point of quantum supremacy is to show the method’s viability by finding even one highly specific and weird task that it can do better than even the fastest supercomputer. Because that gets the quantum foot in the door to expand that library of tasks. Perhaps in the end all tasks will be faster in quantum, but for Google’s purposes in 2019, only one was, and they showed how and why in great detail.

Now, a team at the Chinese Academy of Sciences led by Pan Zhang has published a paper describing a new technique for simulating a quantum computer (specifically, certain noise patterns it puts out) that appears to take a tiny fraction of the time estimated for classical computation to do so in 2019.

Not being a quantum computing expert nor a statistical physics professor myself, I can only give a general idea of the technique Zhang et al used. They cast the problem as a large 3D network of tensors, with the 53 qubits in Sycamore represented by a grid of nodes, extruded out 20 times to represented the 20 cycles the Sycamore gates went through in the simulated process. The mathematical relationships between these tensors (each its own set of interrelated vectors) was then calculated using a cluster of 512 GPUs.

An illustration from Zhang’s paper showing a visual representation of the 3D tensor array they used to simulate Sycamore’s quantum operations.

In Google’s original paper, it was estimated that performing this scale of simulation on the most powerful supercomputer available at the time (Summit at Oak Ridge National Laboratory) would take about 10,000 years — though to be clear, that was their estimate for 54 qubits doing 25 cycles. 53 qubits doing 20 is considerably less complex but would still take on the order of a few years by their estimate.

Zhang’s group claims to have done it in 15 hours. And if they had access to a proper supercomputer like Summit, it might be accomplished in a handful of seconds — faster than Sycamore. Their paper will be published in the journal Physical Review Letters; you can read it here (PDF).

These results have yet to be fully vetted and replicated by those knowledgeable about such things, but there’s no reason to think it’s some kind of error or hoax. Google even admitted that the baton may be passed back and forth a few times before supremacy is firmly established, since it’s incredibly difficult to build and program quantum computers while classical ones and their software are being improved constantly. (Others in the quantum world were skeptical of their claims to begin with, but some are direct competitors.)

As University of Maryland quantum scientist Dominik Hangleiter told Science, this isn’t a black eye for Google or a knockout punch for quantum in general by any means: “The Google experiment did what it was meant to do, start this race.”

Google may well strike back with new claims of its own — it hasn’t been standing still either — and I’ve contacted the company for comment. But the fact that it’s even competitive is good news for everyone involved; this is an exciting area of computing and work like Google’s and Zhang’s continues to raise the bar for everyone.

Bill Gates’ Breakthrough Energy backs Terabase’s robot-built solar farms

Breakthrough Energy Ventures, a climate-focused VC firm linked to some of Earth’s wealthiest individuals, has joined a $44 million bet on solar startup Terabase Energy.

Terabase aims to rapidly build new solar farms “at the terawatt scale,” CEO Matt Campbell said in a statement. The startup claims its automated, on-site factory can already speed up plant construction and cut costs by employing robotic arms that lift and connect heavy solar panels to sun trackers. When asked for photos of the insides of its factory, Campbell pointed TechCrunch to previously published aerial pics and declined to share more, “for competitive reasons.”

Terabase also makes software tools to manage the design and construction of solar farms. The startup recently wrapped its first commercial project, where its robots reportedly installed 10 megawatts worth of panels. There are one million megawatts in a terawatt, so the startup still has a long way to go to reach its aspirations.

Breakthrough Energy Ventures was founded by Bill Gates, and its board members include Jeff Bezos and Masayoshi Son. The VC firm co-led the Terabase deal alongside Lime and Amp Robotics investor Prelude Ventures.

Their investment comes as rich folks face scrutiny for their outsized climate pollution. Gates’ private jet might not be as active as Taylor Swift’s, yet the Microsoft co-founder reportedly owns several and has called private flying his “guilty pleasure.”

Other recent deals for solar energy startups include panel installer Zolar ($105 million) and solar network developer Okra ($2.1 million).

Climate-focused VC stays scorching as Buoyant Ventures targets $100M fund

Like a groundhog and its shadow, many venture capitalists see a shrinking economy and burrow away, resting their check-signing hand for better days.

But climate-focused VCs are on a hot streak lately, pumping well over a billion dollars per quarter into startups that strive to mitigate emissions as the Earth bakes.

Buoyant Ventures is one such firm building momentum for the sector. Based in Chicago, the investor told regulators this week via an SEC filing that it has locked down just over $50 million for a new fund. Buoyant declined to comment when emailed by TechCrunch, but the filing shows the firm had been raising cash for the fund since at least May 2021. So far, 75 (unnamed) limited partners have chipped in, and Buoyant is fishing for just shy of $50 million more. 

Led by Electronic Arts and Energize Ventures alum Amy Francetic and former Accenture executive Allison Myers, Buoyant’s first deal dates back to the summer of 2020. That’s when it backed Raptor Maps, which aims to help solar farms squeeze more juice from the sun by spotting issues—like panel damage and shading—via drones and sensors.

Buoyant said in 2021 that it’s focused on “solutions for the industries contributing the most to carbon emissions,” including power, transportation, agriculture and buildings. Since then, it has funded at least four other early-ish stage startups, including FloodFlash, StormSensor and others seeking to cash in on emissions mitigation or climate adaptation.

Several other noteworthy climate (and climate-adjacent) VC fundraises have crossed our desks in recent weeks, including Fifth Wall‘s $500 million fund, Climentum Capital ($157 million), Systemiq Capital ($70 million) and Equal Ventures ($56 million).

After 50 years pioneering satellite imagery, NASA’s Landsat is ready for 50 more

NASA’s Landsat satellites have consistently made history in Earth observation since the project’s first launch in 1972, with this year marking 50 years of innovation and science. Its influence may surprise you, as will its continued relevance in the face of a fast-growing commercial imaging satellite sector.

Landsat may be a familiar name to you but doesn’t ring any particular bells. It’s understandable — there are a ton of NASA satellites up there looking down on the planet. But the easiest way to say it is this: In 1972, Landsat basically invented modern Earth observation. Then, remember a while back when every Google Earth image said “USGS” on it? Yeah, that was Landsat too. The project has basically ushered satellite imaging from bleeding edge research tool to everyday technology.

Landsat 9 just launched last September, the latest in a long line of influential spacecraft.

A schematic sketch of Landsat-1. Image Credits: NASA

I talked with Jim Irons, who has worked at NASA since 1978 and on Landsat since 1992. Irons told the story of Landsat from the beginning, both what he took part in himself and the lore he’s absorbed over the years. It’s fitting that for a project that would redefine Earth imaging, its very first satellite was both innovative and historically significant.

“Landsat 1 launched in 1972 — it carried two instruments, one was the Return Beam Vidicon, and it was kind of like a TV camera, it took analog data,” Irons said. “But Hughes [Aircraft Company] convinced NASA to put another instrument on the payload that was more experimental: the Multi-Spectrum Scanner. And it provided digital data.”

It hardly needs to be said that in 1972, digital anything was pretty innovative, let alone high-performance digital sensors in orbit. But the team clearly saw the writing on the wall, and good thing too.

“After launch, the RBV had problems, and the data from the MSS became the preferred data. That was a big turning point,” recalled Irons. “It was an instrument that used an oscillating mirror that went back and forth to scan a path at 7-14 Hz, underneath the orbital path of the sensor, to create a digital image. And it’s mechanical! It was amazing.”

“The designer of this sensor, Virginia Norwood, she’s still with us, in her 90s. It was very unusual at the time to have a female engineer at all. She came to the launch of Landsat 9 last month, actually.”

Virginia Norwood (photo taken in 1972) with the MSS instrument she created. Image Credits: NASA

It’s a remarkable fact that the beginning of the orbital imaging revolution was the brainchild of one of the then-rare women in the space and tech industries, whose roles in many of the era’s important accomplishments have only recently begun to be given the attention they deserve. You can read more about Norwood’s role in the creation of the MSS, which is the precursor to many more such systems, at this NASA history article, or this more recent piece.

A successor to the MSS called the Thematic Mapper launched in 1982 with more spectral bands, but then in 1984 another big improvement struck a nerve at HQ:

“Landsat 5 in 1984 carried both a multispectral scanner and advancement on the thematic mapper idea that improved the spatial resolution of the data, from what had been 80 meters with the MSS to 30 meters, and spectral bands were added,” Irons said. “But there was all this data! Some people were afraid of that data, that analysts would be overwhelmed by it — but it didn’t turn out that way. Computer capacities kept up and soon the thematic mapper data was preferred.”

Image Credits: NASA

That would prove a rule as time went on and right up until the present: There really is no such thing as too much data. As long as you can collect it and store it, someone will find a use for it.

They might even pay you for it — but an attempt to privatize Landsat in the following years fell flat, or burned up on reentry in the case of Landsat 6, which never made it to orbit. Meanwhile, the private company created to operate and distribute the rest of the data jacked up the price until no one was willing to pay any more. “It was up to $4,400 per scene of thematic mapper data. People just stopped using it,” Irons said.

When NASA and the USGS, which handled the distribution of the imagery originally, returned to the reins, they had an international data recovery problem. Imagine having reams of data in a ground station in China or South America, long before ubiquitous broadband networks. How do you get it back to HQ in the States for central processing and analysis? I told Irons I was picturing big trucks full of hard drives, the internal combustion equivalent of Sneakernet.

“That’s exactly what happened!” he laughed. “They just drove up to the [USGS] facility with semi truck trailers full of magnetic tapes. It was difficult because they had all these different formats and instruments. So that created a little chaos. They bought pizza ovens to bake the water out of some of those tapes.” (I wanted to hear more about that part but our time was limited.)

Image Credits: NASA

But the repatriation of the data was only a precursor to an even larger shift.

“After Landsat 7 launched was perhaps the biggest change in the entire program,” Irons said. “USGS was still charging $600 for a mapper scene of data. And they made made what I consider an institutionally brave decision in 2008, to be consistent with NASA and provide Landsat data at no cost to anyone who wanted it. So it went from $400 to $600 to free.”

As you can imagine, this choice completely upended the model, and overnight, it changed everything.

“There was an explosion of use and redistribution of the data,” he continued. “Now, some places like Google Earth and Amazon Cloud Services, they’d gone in and downloaded the whole archive from USGS.”

Remember the old Google Earth app? Image Credits: Google

That’s why for years, whenever you looked at an online map, it credited the USGS. Of course Google and Amazon didn’t own the imagery, or capture it themselves, though now all the majors are doing that at various scales. They simply downloaded a huge picture of the entire Earth and re-served it to their customers in a new form.

“It’s a struggle for us to brand the data and the program so taxpayers know they’re getting their money’s worth,” admitted Irons. It’s not like every time you opened Google Maps, it thanked you for making their business possible!

In the years since, Landsat 8 and 9 have launched with improved sensors and continued to collect invaluable data that is continuous with the previous decades — a free, long-term database of a large portion of the planet imaged every couple weeks or so depending on the era.

Image Credits: NASA

Of course nowadays constellations like Planet’s are imaging the whole globe on a daily basis. So why have Landsat at all?

“Those of us who work on Landsat are very impressed by what the commercial providers have achieved,” Irons said. “The message we want to get out is that Landsat is complementary to that data — they don’t replace Landsat data. One, it’s open and transparent access — that’s key, and it’s true of all the data collected by NASA satellites.

“Two, the USGS has maintained this 50-year archive of data. Is there a business case for companies to archive their data for decades, so we can observe the effects of climate change over the long term rather than just have short bursts of data? I don’t know that the business case is there.”

You can see an example of what decades of continuous data looks like here:

“And one of the things that enables our time series analyses is that NASA pays a great deal of attention to inter-sensor calibration,” Irons continued. “If you’re going from one Landsat image to another, you know it’s been calibrated — if you see a change over time, you can be clear that the thing is changing rather than the camera. [Commercial constellations] use Landsat data to do that; we serve as an industry standard to help them do their calibration.”

Here the conversation overlapped with what I talked about with Ginger Butcher, who’s done outreach for the project for years.

“We can compare a Landsat image today to a Landsat image from 1972,” she said. “That’s one of the tenets of the program: We have a dedicated calibration team keeping an eye on the instruments. Every full moon we turn the spacecraft around to use it as a kind of photographer’s grey card.”

With the increasing prominence of commercial providers in the U.S. space program, it was a real question over the last few years whether Landsat was worthwhile to continue funding, but arguments like those above won out.

“We used to have to work really hard to get that next mission, but now we’ve basically got the government saying this is a valuable resource worth continuing with,” Butcher said. “Now we’re looking to the future and what kind of capabilities we want to get out of the next Landsat. What kind of research are people doing? What additional wavelengths are needed for work on ice, or on forests, or particular areas in agriculture? For example, with thermal data we can look at crops and see if they’re being overwatered or underwatered — with water rights out west, that’s really important. As scientists take on new questions and new areas of study, they decide where Landsat goes next.”

More than ever, the project will work collaboratively with the commercial sector and with ESA satellites like Sentinel-2.

“We think it’s great,” said Irons. “The emergence of all these systems means the Landsat project has been incredibly successful; it basically created the market for them.”