I see far more research articles than I could possibly write up. This column collects the most interesting of those papers and advances, along with notes on why they may prove important in the world of tech and startups. This week: supercomputers take on COVID-19, beetle backpacks, artificial spiderwebs, the “overwhelming whiteness” of AI and more.
First off, if (like me) you missed this amazing experiment where scientists attached tiny cameras to the backs of beetles, I don’t think I have to explain how cool it is. But you may wonder… why do it? Prolific UW researcher Shyam Gollakota and several graduate students were interested in replicating some aspects of insect vision, specifically how efficient the processing and direction of attention is.
The camera backpack has a narrow field of view and uses a simple mechanism to direct its focus rather than processing a wide-field image at all times, saving energy and better imitating how real animals see. “Vision is so important for communication and for navigation, but it’s extremely challenging to do it at such a small scale. As a result, prior to our work, wireless vision has not been possible for small robots or insects,” said Gollakota. You can watch the critters in action below — and don’t worry, the beetles lived long, happy lives after their backpack-wearing days.
The health and medical community is always making interesting strides in technology, but it’s often pretty niche stuff. These two items from recent weeks are a bit more high-profile.
One is a new study being conducted by UCLA in concert with Apple, which especially with its smartwatch has provided lots of excellent data to, for example, studies of arrhythmia. In this case, doctors are looking at depression and anxiety, which are considerably more difficult to quantify and detect. But by using Apple Watch, iPhone and sleep monitor measurements of activity levels, sleep patterns and so on, a large body of standardized data can be amassed.
Even as e-grocery usage has skyrocketed in our coronavirus-catalyzed world, brick-and-mortar grocery stores have soldiered on. While strict in-store safety guidelines may gradually ease up, the shopping experience will still be low-touch and socially distanced for the foreseeable future.
This begs the question: With even greater challenges than pre-pandemic, how can grocers ensure their stores continue to operate profitably?
Just as micro-fulfillment centers (MFCs), dark stores and other fulfillment solutions have been helping e-grocers optimize profitability, a variety of old and new technologies can help brick-and-mortar stores remain relevant and continue churning out cash.
Today, we present three “must-dos” for post-pandemic retail grocers: rely on the data, rely on the biology and rely on the hardware.
Image Credits: Pixabay/Pexels (opens in a new window)
The hallmark of shopping in a store is the consistent availability and wide selection of fresh items — often more so than online. But as the number of in-store customers continues to fluctuate, planning inventory and minimizing waste has become ever more so a challenge for grocery store managers. Grocers on average throw out more than 12% of their on-shelf produce, which eats into already razor-thin margins.
While e-grocers are automating and optimizing their fulfillment operations, brick-and-mortar grocers can automate and optimize their inventory planning mechanisms. To do this, they must leverage their existing troves of customer, business and external data to glean valuable insights for store managers.
Eden Technologies of Walmart is a pioneering example. Spun out of a company hackathon project, the internal tool has been deployed at over 43 distribution centers nationwide and promises to save Walmart over $2 billion in the coming years. For instance, if a batch of produce intended for a store hundreds of miles away is deemed soon-to-ripen, the tool can help divert it to the nearest store instead, using FDA standards and over 1 million images to drive its analysis.
Similarly, ventures such as Afresh Technologies and Shelf Engine have built platforms to leverage years of historical customer and sales data, as well as seasonality and other external factors, to help store managers determine how much to order and when. The results have been nothing but positive — Shelf Engine customers have increased gross margins by over 25% and Afresh customers have reduced food waste by up to 45%.
OpenAI’s latest language generation model, GPT-3, has made quite the splash within AI circles, astounding reporters to the point where even Sam Altman, OpenAI’s leader, mentioned on Twitter that it may be overhyped. Still, there is no doubt that GPT-3 is powerful. Those with early-stage access to OpenAI’s GPT-3 API have shown how to translate natural language into code for websites, solve complex medical question-and-answer problems, create basic tabular financial reports, and even write code to train machine learning models — all with just a few well-crafted examples as input (i.e., via “few-shot learning”).
Soon, anyone will be able to purchase GPT-3’s generative power to make use of the language model, opening doors to build tools that will quietly (but significantly) shape our world. Enterprises aiming to take advantage of GPT-3, and the increasingly powerful iterations that will surely follow, must take great care to ensure that they install extensive guardrails when using the model, because of the many ways that it can expose a company to legal and reputational risk. Before we discuss some examples of how the model can potentially do wrong in practice, let’s first look at how GPT-3 was made.
Machine learning models are only as good, or as bad, as the data fed into them during training. In the case of GPT-3, that data is massive. GPT-3 was trained on the Common Crawl dataset, a broad scrape of the 60 million domains on the internet along with a large subset of the sites to which they link. This means that GPT-3 ingested many of the internet’s more reputable outlets — think the BBC or The New York Times — along with the less reputable ones — think Reddit. Yet, Common Crawl makes up just 60% of GPT-3’s training data; OpenAI researchers also fed in other curated sources such as Wikipedia and the full text of historically relevant books.
Language models learn which succeeding words, phrases and sentences are likely to come next for any given input word or phrase. By “reading” text during training that is largely written by us, language models such as GPT-3 also learn how to “write” like us, complete with all of humanity’s best and worst qualities. Tucked away in the GPT-3 paper’s supplemental material, the researchers give us some insight into a small fraction of the problematic bias that lurks within. Just as you’d expect from any model trained on a largely unfiltered snapshot of the internet, the findings can be fairly toxic.
Because there is so much content on the web sexualizing women, the researchers note that GPT-3 will be much more likely to place words like “naughty” or “sucked” near female pronouns, where male pronouns receive stereotypical adjectives like “lazy” or “jolly” at the worst. When it comes to religion, “Islam” is more commonly placed near words like “terrorism” while a prompt of the word “Atheism” will be more likely to produce text containing words like “cool” or “correct.” And, perhaps most dangerously, when exposed to a text seed that involves racial content involving Blackness, the output GPT-3 gives tends to be more negative than corresponding white- or Asian-sounding prompts.
Image Credits: Arthur (opens in a new window)
How might this play out in a real-world use case of GPT-3? Let’s say you run a media company, processing huge amounts of data from sources all over the world. You might want to use a language model like GPT-3 to summarize this information, which many news organizations already do today. Some even go so far as to automate story creation, meaning that the outputs from GPT-3 could land directly on your homepage without any human oversight. If the model carries a negative sentiment skew against Blackness — as is the case with GPT-3 — the headlines on your site will also receive that negative slant. An AI-generated summary of a neutral news feed about Black Lives Matter would be very likely to take one side in the debate. It’s pretty likely to condemn the movement, given the negatively charged language that the model will associate with racial terms like “Black.” This, in turn, could alienate parts of your audience and deepen racial tensions around the country. At best, you’ll lose a lot of readers. At worst, the headline could spark more protest and police violence, furthering this cycle of national unrest.
OpenAI’s website also details an application in medicine, where issues of bias can be enough to prompt federal inquiries, even when the modelers’ intentions are good. Attempts to proactively detect mental illness or rare underlying conditions worthy of intervention are already at work in hospitals around the country. It’s easy to imagine a healthcare company using GPT-3 to power a chatbot — or even something as “simple” as a search engine — that takes in symptoms from patients and outputs a recommendation for care. Imagine, if you will, a female patient suffering from a gynecological issue. The model’s interpretation of your patient’s intent might be married to other, less medical associations, prompting the AI to make offensive or dismissive comments, while putting her health at risk. The paper makes no mention of how the model treats at-risk minorities such as those who identify as transgender or nonbinary, but if the Reddit comments section is any indication of the responses we will soon see, the cause for worry is real.
But because algorithmic bias is rarely straightforward, many GPT-3 applications will act as canaries in the growing coal mine that is AI-driven applications. As COVID-19 ravages our nation, schools are searching for new ways to manage remote grading requirements, and the private sector has supplied solutions to take in schoolwork and output teaching suggestions. An algorithm tasked with grading essays or student reports is very likely to treat language from various cultures differently. Writing styles and word choice can vary significantly between cultures and genders. A GPT-3-powered paper-grader without guardrails might think that white-written reports are more worthy of praise, or it may penalize students based on subtle cues that indicate English as a second language, which are in turn, largely correlated to race. As a result, children of immigrants and from racial minorities will be less likely to graduate from high school, through no fault of their own.
The creators of GPT-3 plan to continue their research into the model’s biases, but for now, they simply surface these concerns, passing along the risk to any company or individual who’s willing to take the chance. All models are biased, as we know, and this should not be a reason to outlaw all AI, because its benefits can surely outweigh the risks in the long term. But in order to enjoy these benefits, we must ensure that as we rush to deploy powerful AI like GPT-3 to the enterprise, that we take sufficient precautions to understand, monitor for and act quickly to mitigate its points of failure. It’s only through a responsible combination of human and automated oversight that AI applications can be trusted to deliver societal value while protecting the common good.
This article was written by humans.
In 2010, the late Barnaby Jack, a world-renowned security researcher, hacked an ATM live on stage at the Black Hat conference by tricking the cash dispenser into spitting out a stream of dollar bills. The technique was appropriately named “jackpotting.”
A decade on from Jack’s blockbuster demo, security researchers are presenting two new vulnerabilities in Nautilus ATMs, albeit virtually, thanks to the coronavirus pandemic.
Security researchers Brenda So and Trey Keown at New York-based security firm Red Balloon say their pair of vulnerabilities allowed them to trick a popular standalone retail ATM, commonly found in stores rather than at banks, into dispensing cash at their command.
A hacker would need to be on the same network as the ATM, making it more difficult to launch a successful jackpotting attack. But their findings highlight that ATMs often have vulnerabilities that lie dormant for years — in some cases since they were first built.
Barnaby Jack, the late security researcher credited with the first ATM “jackpotting” attacks. Now, 10 years later, two security researchers have found two new ATM cash-spitting attacks. Credit: YouTube
So and Keown said their new vulnerabilities target the Nautilus ATM’s underlying software, a decade-old version of Windows that is no longer supported by Microsoft. To begin with, the pair bought an ATM to examine. But with little documentation, the duo had to reverse-engineer the software inside to understand how it worked.
The first vulnerability was found in a software layer known as XFS — or Extensions for Financial Services — which the ATM uses to talk to its various hardware components, such as the card reader and the cash dispensing unit. The bug wasn’t in XFS itself, rather in how the ATM manufacturer implemented the software layer into its ATMs. The researchers found that sending a specially crafted malicious request over the network could effectively trigger the ATM’s cash dispenser and dump the cash inside, Keown told TechCrunch.
The second vulnerability was found in the ATM’s remote management software, an in-built tool that lets owners manage their fleet of ATMs by updating the software and checking how much cash is left. Triggering the bug would grant a hacker access to a vulnerable ATM’s settings.
So told TechCrunch it was possible to switch the ATM’s payment processor with a malicious, hacker-controlled server to siphon off banking data. “By pointing an ATM to a malicious server, we can extract credit card numbers,” she said.
Bloomberg first reported the vulnerabilities last year when the researchers privately reported their findings to Nautilus. About 80,000 Nautilus ATMs in the U.S. were vulnerable prior to the fix, Bloomberg reported. We contacted Nautilus with questions but did not hear back.
Successful jackpotting attacks are rare but not unheard of. In recent years, hackers have used a number of techniques. In 2017, an active jackpotting group was discovered operating across Europe, netting millions of euros in cash.
More recently, hackers have stolen proprietary software from ATM manufacturers to build their own jackpotting tools.
Send tips securely over Signal and WhatsApp to +1 646-755-8849 or send an encrypted email to: email@example.com
Every major technology breakthrough of our era has gone through a similar cycle in pursuit of turning fiction to reality.
It starts in the stages of scientific discovery, a pursuit of principle against a theory, a recursive process of hypothesis-experiment. Success of the proof of principle stage graduates to becoming a tractable engineering problem, where the path to getting to a systemized, reproducible, predictable system is generally known and de-risked. Lastly, once successfully engineered to the performance requirements, focus shifts to repeatable manufacturing and scale, simplifying designs for production.
Since theorized by Richard Feynman and Yuri Manin, quantum computing has been thought to be in a perpetual state of scientific discovery. Occasionally reaching proof of principle on a particular architecture or approach, but never able to overcome the engineering challenges to move forward.
That’s until now. In the last 12 months, we have seen several meaningful breakthroughs from academia, venture-backed companies, and industry that looks to have broken through the remaining challenges along the scientific discovery curve. Moving quantum computing from science fiction that has always been “five to seven years away,” to a tractable engineering problem, ready to solve meaningful problems in the real world.
Companies such as Atom Computing* leveraging neutral atoms for wireless qubit control, Honeywell’s trapped ions approach, and Google’s superconducting metals, have demonstrated first-ever results, setting the stage for the first commercial generation of working quantum computers.
While early and noisy, these systems, even at just 40-80 error-corrected qubit range, may be able to deliver capabilities that surpass those of classical computers. Accelerating our ability to perform better in areas such as thermodynamic predictions, chemical reactions, resource optimizations and financial predictions.
As a number of key technology and ecosystem breakthroughs begin to converge, the next 12-18 months will be nothing short of a watershed moment for quantum computing.
Here are eight emerging trends and predictions that will accelerate quantum computing readiness for the commercial market in 2021 and beyond:
1. Dark horses of QC emerge: 2020 will be the year of dark horses in the QC race. These new entrants will demonstrate dominant architectures with 100-200 individually controlled and maintained qubits, at 99.9% fidelities, with millisecond to seconds coherence times that represent 2x -3x improved qubit power, fidelity and coherence times. These dark horses, many venture-backed, will finally prove that resources and capital are not sole catalysts for a technological breakthrough in quantum computing.
Data science is the name of the game these days for companies that want to improve their decision making by tapping the information they are already amassing in their apps and other systems. And today, a startup called Mode Analytics, which has built a platform incorporating machine learning, business intelligence and big data analytics to help data scientists fulfill that task, is announcing $33 million in funding to continue making its platform ever more sophisticated.
Most recently, for example, the company has started to introduce tools (including SQL and Python tutorials) for less technical users, specifically those in product teams, so that they can structure queries that data scientists can subsequently execute faster and with more complete responses — important for the many follow-up questions that arise when a business intelligence process has been run. Mode claims that its tools can help produce answers to data queries in minutes.
This Series D is being led by SaaS specialist investor H.I.G. Growth Partners, with previous investors Valor Equity Partners, Foundation Capital, REV Venture Partners, and Switch Ventures all participating. Valor led Mode’s Series C in February 2019, while Foundation and REV respectively led its A and B rounds.
Mode is not disclosing its valuation, but co-founder and CEO Derek Steer confirmed in an interview that it was “absolutely” an up-round.
For some context, PitchBook notes that last year its valuation was $106 million. The company now has a customer list that it says covers 52% of the Forbes 500, including Anheuser Busch, Zillow, Lyft, Bloomberg, Capital One, VMWare, and Conde Nast. It says that to date it has processed 830 million query runs and 170 million notebook cell runs for 300,000 users. (Pricing is based on a freemium model, with a free “Studio” tier and Business and Enterprise tiers priced based on size and use.)
Mode has been around since 2013, when it was co-founded by Steer, Benn Stancil (Mode’s current president) and Josh Ferguson (initially the CTO and now chief architect).
Steer said the impetus for the startup came out of gaps in the market that the three had found through years of experience at other companies.
Specifically, when all three were working together at Yammer (they were early employees and stayed on after the Microsoft acquisition), they were part of a larger team building custom data analytics tools for Yammer. At the time, Steer said Yammer was paying $1 million per year to subscribe to Vertica (acquired by HP in 2011) to run it.
They saw an opportunity to build a platform that could provide similar kinds of tools — encompassing things like SQL Editors, Notebooks, and reporting tools and dashboards — to a wider set of users.
“We and other companies like Facebook and Google were building analytics internally,” Steer recalled, “and we knew that the world wanted to work more like these tech companies. That’s why we started Mode.”
All the same, he added, “people were not clearly exactly about what a data scientist even was.”
Indeed, Mode’s growth so far has mirrored that of the rise of data science overall, as the discipline of data science, and the business case for employing data scientists to help figure out what is “going on” beyond the day to day, getting answers by tapping all the data that’s being amassed in the process of just doing business. That means Mode’s addressable market has also been growing.
But even if the trove of potential buyers of Mode’s products has been growing, so has the opportunity overall. There has been a big swing in data science and big data analytics in the last several years, with a number of tech companies building tools to help those who are less technical “become data scientists” by introducing more intuitive interfaces like drag-and-drop features and natural language queries.
They include the likes of Sisense (which has been growing its analytics power with acquisitions like Periscope Data), Eigen (focusing on specific verticals like financial and legal queries), Looker (acquired by Google) and Tableau (acquired by Salesforce).
Mode’s approach up to now has been closer to that of another competitor, Alteryx, focusing on building tools that are still aimed primary at helping data scientists themselves. You have any number of database tools on the market today, Steer noted, “Snowflake, Redshift, BigQuery, Databricks, take your pick.” The key now is in providing tools to those using those databases to do their work faster and better.
That pitch and the success of how it executes on it is what has given the company success both with customers and investors.
“Mode goes beyond traditional Business Intelligence by making data faster, more flexible and more customized,” said Scott Hilleboe, MD, H.I.G. Growth Partners, in a statement. “The Mode data platform speeds up answers to complex business problems and makes the process more collaborative, so that everyone can build on the work of data analysts. We believe the company’s innovations in data analytics uniquely position it to take the lead in the Decision Science marketplace.”
Steer said that fundraising was planned long before the coronavirus outbreak to start in February, which meant that it was timed as badly as it could have been. Mode still raised what it wanted to in a couple of months — “a good raise by any standard,” he noted — even if it’s likely that the valuation suffered a bit in the process. “Pitching while the stock market is tanking was terrifying and not something I would repeat,” he added.
Given how many acquisitions there have been in this space, Steer confirmed that Mode too has been approached a number of times, but it’s staying put for now. (And no, he wouldn’t tell me who has been knocking, except to say that it’s large companies for whom analytics is an “adjacency” to bigger businesses, which is to say, the very large tech companies have approached Mode.)
“The reason we haven’t considered any acquisition offers is because there is just so much room,” Steer said. “I feel like this market is just getting started, and I would only consider an exit if I felt like we were handicapped by being on our own. But I think we have a lot more growing to do.”
It seemed so simple. A small schema issue in a database was wrecking a feature in the app, increasing latency and degrading the user experience. The resident data engineer pops in a fix to amend the schema, and everything seems fine — for now. Unbeknownst to them, that small fix completely clobbered all the dashboards used by the company’s leadership. Finance is down, ops is pissed, and the CEO — well, they don’t even know whether the company is online.
For data engineers, it’s not just a recurring nightmare — it’s a day-to-day reality. A decade plus into that whole “data is the new oil” claptrap, and we’re still managing data piecemeal and without proper systems and controls. Data lakes have become data oceans and data warehouses have become … well, whatever the massive version of a warehouse is called (a waremansion I guess). Data engineers bridge the gap between the messy world of real life and the precise nature of code, and they need much better tools to do their jobs.
As TechCrunch’s unofficial data engineer, I’ve personally struggled with many of these same problems. And so that’s what drew me into Datafold.
Datafold is a brand-new platform for managing the quality assurance of data. Much in the way that a software platform has QA and continuous integration tools to ensure that code functions as expected, Datafold integrates across data sources to ensure that changes in the schema of one table doesn’t knock out functionality somewhere else.
Founder Gleb Mezhanskiy knows these problems firsthand. He’s informed from his time at Lyft, where he was a data scientist and data engineer, and later transformed into a product manager “focused on the productivity of data professionals.” The idea was that as Lyft expanded, it needed much better pipelines and tooling around its data to remain competitive with Uber and others in its space.
His lessons from Lyft inform Datafold’s current focus. Mezhanskiy explained that the platform sits in the connections between all data sources and their outlets. There are two challenges to solve here. First, “data is changing, every day you get new data, and the shape of it can be very different either for business reasons or because your data sources can be broken.” And second, “the old code that is used by companies to transform this data is also changing very rapidly because companies are building new products, they are refactoring their features … a lot of errors can happen.”
In equation form: messy reality + chaos in data engineering = unhappy data end users.
With Datafold, changes made by data engineers in their extractions and transformations can be compared for unintentional changes. For instance, maybe a function that formerly returned an integer now returns a text string, an accidental mistake introduced by the engineer. Rather than wait until BI tools flop and a bunch of alerts come in from managers, Datafold will indicate that there is likely some sort of problem, and identify what happened.
The key efficiency here is that Datafold aggregates changes in datasets — even datasets with billions of entries — into summaries so that data engineers can understand even subtle flaws. The goal is that even if an error transpires in 0.1% of cases, Datafold will be able to identify that issue and also bring a summary of it to the data engineer for response.
Datafold is entering a market that is, quite frankly, as chaotic as the data being processed. It sits in the key middle layer of the data stack — it’s not the data lake or data warehouse for storing data, and it isn’t the end user BI tools like a Looker, Tableau or many others. Instead, it’s part of a number of tools available for data engineers to manage and monitor their data flows to ensure consistency and quality.
The startup is targeting companies with at least 20 people on their data team — that’s the sweet spot where a data team has enough scale and resources that they are going to be concerned with data quality.
Today Datafold is three people, and will be debuting officially at YC’s Demo Day later this month. Its ultimate dream is a world where data engineers never again have to get an overnight page to fix a data quality issue. If you’ve been there, you know precisely why such a product is valuable.
SpaceX achieved a big win in their Starship spacecraft development program on Tuesday evening, flying the SN5 prototype of that future vehicle to a height of around 500 feet, propelled by a single Raptor engine. The test, which took place at SpaceX’s rocket development and testing facility in Boca Chica, Texas, marks the first time that a full-scale Starship prototype has left the ground.
The company released a video of the whole test, including footage captured both from a drone’s-eye-view, as well as from a camera mounted on board Starship SN5, inside the fuselage and offering a look at the Raptor engine in action, as well as the landing legs activating in preparation for landing.
Following the successful test, SpaceX CEO and founder Elon Musk outlined next steps for the Starship development process, which includes “several” more short hops, followed by high altitude testing. The landing legs will also go through some changes, first extending in length and then becoming much wider and taller, with the ability to land on more uneven terrain, according to Musk.
SpaceX has been developing Starship, its next-generation spacecraft, at its site in Boca Chica, Texas. The company has built a number of different Starship prototypes to date, include one prior version called the Starhopper that was essentially just the bottom portion of the rocket. Today, the company flew its first full-scale prototype (minus the domed cap that will appear on the final version, and without the control fins that will appear lower down on its sides), achieving an initial flight of around 150 m (just under 500 feet).
This is the furthest along one of these prototypes has come in the testing process. It’s designated Starship SN5, which is the fifth serialized test article. SpaceX actually built a first full-scale demonstration craft called the Starship Mk1 prior to switching to this new naming scheme, so that makes this the sixth one this size they’ve built – with the prior versions suffering failures at various points during preparations, including pressure testing and following a static engine test fire.
SN5 is now the first of these larger test vehicles to actually take off and fly. This prototype underwent a successful static test fire earlier this week, paving the way for this short flight test today. It’s equipped with just one Raptor engine, whereas the final Starship will have six Raptors on board for much greater thrust. It managed to fly and land upright, which means that by all external indications everything went to plan.
Starhopper previously completed a similar hop in August of 2019. SpaceX has an aggressive prototype development program to attempt to get Starship in working order, with the ambitious goal of flying payloads using the functional orbital vehicle as early as next year. Ultimately, Starship is designed to pair with a future Falcon Heavy booster to carry large payloads to orbit around Earth, as well as to the Moon and eventually to Mars.
Rocket Lab has managed to engineer a significant payload capacity bump into its existing Electron space launch vehicle, the company revealed today. Electron can now fly as much as 300 kg (660 lbs) to low Earth orbit (or around 440 lbs to a higher, sun synchronous orbit), and that’s mostly due to battery technology advances, according to Rocket Lab.
Electron is not battery-powered, of course — but the electric pumps that help feed its Rutherford engines are. That’s where they’re getting the boost, along with some other optimizations, increasing the total payload capacity by a full third. That’s a lot of additional capacity in the small satellite launch market, where a CubeSat can weigh as little as 3 lbs or less.
Rocket Lab notes that this means customers who are using their Photon spacecraft as a satellite bus (essentially the basic satellite platform upon which a company can build their specific instrumentation needs) will now have nearly 400 lbs available to them for their equipment, which should make possible a whole range of potential new applications.
The company announced last week that it was aiming to return to active launch status as early as this month, after an issue caused the early termination and failure of a mission in early July. It said it was able to quickly identify the problem and is already implementing a fix, and now it clearly wants to remind potential customers of its unique offerings and capabilities in the small satellite market.