Connect with us

Uncategorized

These five AI developments will shape 2021 and beyond

Published

on

The year 2020 was profoundly challenging for citizens, companies, and governments around the world. As covid-19 spread, requiring far-reaching health and safety restrictions, artificial intelligence (AI) applications played a crucial role in saving lives and fostering economic resilience. Research and development (R&D) to enhance core AI capabilities, from autonomous driving and natural language processing to quantum computing, continued unabated.  

Baidu was at the forefront of many important AI breakthroughs in 2020. This article outlines five significant advances with implications for combating covid-19 as well as transforming the future of our economies and society.

1. AI and vaccine development

The trend—and why it matters. It typically takes years, if not decades, to develop a new vaccine. But by March 2020, vaccine candidates to fight covid-19 were already undergoing human tests, just three months after the first reported cases. The record speed of vaccine development was partly thanks to AI models that helped researchers analyze vast amounts of data about coronavirus.

There are tens of thousands of subcomponents to the outer proteins of a virus. Machine learning models can sort through this blizzard of data and predict which subcomponents are the most immunogenic—i.e., capable of producing an immune response—and thereby guide researchers in designing targeted vaccines. The use of AI in vaccine development may revolutionize the way all vaccines are created in the future.

Baidu’s innovations. In February, Baidu opened its LinearFold AI algorithm for scientific and medical teams working to fight the virus. LinearFold predicts the secondary structure of the ribonucleic acid (RNA) sequence of a virus—and does so significantly faster than traditional RNA folding algorithms. LinearFold was able to predict the secondary structure of the SARS-CoV-2 RNA sequence in only 27 seconds, 120 times faster than other methods. This is significant, because the key breakthrough of covid-19 vaccines has been the development of messenger RNA (mRNA) vaccines. Instead of conventional approaches, which insert a small portion of a virus to trigger a human immune response, mRNA teaches cells how to make a protein that can prompt an immune response, which greatly shortens the time span involved in development and approval. 

To support mRNA vaccine development, Baidu later developed and released an AI algorithm for optimizing mRNA sequence design called LinearDesign, which aims to solve the problem of unstable and unproductive mRNA sequences in candidate vaccines.  

In addition to opening up access to LinearFold and LinearDesign for researchers around the world, Baidu also formed a strategic partnership with the National Institute for Viral Disease Control and Prevention, part of the Chinese Center for Disease Control and Prevention. Following an outbreak at Beijing’s Xinfadi market in June, Baidu’s AI technology allowed authorities to complete genome sequencing of the coronavirus strain within 10 hours, helping curb the outbreak. In December, Baidu unveiled PaddleHelix, a machine learning-based bio-computing framework aimed at facilitating the development of vaccine design, drug discovery, and precision medicine.

2. Fully automated driving and the rollout of robotaxis

The trend—and why it matters. Autonomous driving technology continued to mature in 2020, with the industry’s leading companies testing driverless cars and opening up robotaxi services to the public in various cities. Fully automated driving, which enables rides without a human safety driver on board, will be necessary for the scalability and commercialization of autonomous driving.

Baidu’s innovations. Over the past year, Baidu launched the Apollo Go Robotaxi service in the Chinese cities of Changsha, Cangzhou, and Beijing—including in busy commercial areas—becoming the only company in China to start robotaxi trial operations in multiple cities.

These developments are a result of Baidu’s continuous innovation in developing AI systems that can safely control a vehicle in complex road conditions and solve the majority of possible issues on the road, independent of a human driver.

At Baidu World 2020, its annual technology conference, Baidu also demonstrated its fully automated driving capability—where the AI system drives independently without an in-vehicle safety driver. To support fully automated driving, Baidu developed the 5G Remote Driving Service, a safety measure whereby remote human operators can take control of a vehicle in the event of an exceptional emergency. Baidu’s achievement of fully automated driving, and the rollout of its robotaxis, suggests a positive outlook for the commercialization of the technology in the near future.

Source: Baidu

3. Applied natural language processing

The trend—and why it matters. In 2020, natural language systems became significantly more advanced at processing aspects of human language like sentiment and intent, generating language that aligns with human speaking and writing patterns, and even visual understanding, meaning the capability to express understanding about an image through language. These natural language models are powering more accurate search results and more sophisticated chatbots and virtual assistants, leading to better user experiences and creating value for businesses.

Baidu’s innovations. Baidu released a new multiflow sequence framework for language generation called ERNIE-GEN. By training the model to predict semantically complete blocks of text, ERNIE-GEN performs at an elite level across a range of language generation tasks, including dialogue engagement, question generation, and abstractive summarization.

Baidu’s vision-language model ERNIE-ViL also achieved significant progress in visual understanding, ranking first on the VCR leaderboard, a dataset of 290,000 questions built by the University of Washington and the Allen Institute for AI, that aims to test visual understanding ability. ERNIE-ViL also achieved state-of-the-art performance on five vision-language downstream tasks. Visual understanding lays the foundation for computer systems to physically interact in everyday scenes, as it involves both understanding visual content and expressing it through language. It will be crucial for improving the quality of human-machine interaction.

4. Quantum computing

The trend—and why it matters. Quantum computing made significant inroads in 2020, including the Jiuzhang computer’s achievement of quantum supremacy. This carries significance for AI, since quantum computing has the potential to supercharge AI applications compared to binary-based classical computers. For example, quantum computing could be used to run a generative machine learning model through a larger dataset than a classical computer can process, thus making the model more accurate and useful in real-world settings. Advanced technologies such as deep learning algorithms are also playing an increasingly critical role in the development of quantum computing research. 

Baidu’s innovations. Baidu achieved a number of technical breakthroughs in 2020 that promise to bridge AI and quantum computing. In May, Baidu launched Paddle Quantum, a quantum machine learning development toolkit that can help scientists and developers quickly build and train quantum neural network models and provide advanced quantum computing applications. The open-source toolkit both supports developers building quantum AI applications, and helps deep learning enthusiasts develop quantum computing. In September, Baidu entered cloud-based quantum computing with the launch of Quantum Leaf, which provides quantum development kits such as QCompute, and can shorten the life cycle of quantum programming and help realize a ‘closed-loop’ quantum tool chain.

Source: Baidu

5. AI chips

The trend—and why it matters. AI hardware continued to develop in 2020, with the launch of several AI chips customized for specialized tasks. While an ordinary processor is capable of supporting AI tasks, AI-specific processors are modified with particular systems that can optimize performance for tasks like deep learning. As AI applications become more widespread, any increase in performance or reduction in cost can unlock more value for companies that operate a wide network of data centers for commercial cloud services, and can facilitate the company’s internal operations.

Source: Baidu

Baidu’s innovations. At Baidu World 2020, the company offered a glimpse into its next-generation AI processor, the Kunlun 2, which it plans to put into mass production in early 2021. The chip uses 7 nanometer (nm) processing technology and its maximum computational capability is over three times that of the previous generation, the Kunlun 1. The Kunlun chips are characterized by high performance, low cost, and high flexibility, which can support a broad range of AI applications and scenarios, helping foster greater AI adoption and reducing usage costs. More than 20,000 Kunlun 1 chips have now been deployed to support Baidu’s search engine and Baidu Cloud partners since they launched in 2018, empowering industrial manufacturing, smart cities, smart transportation, and other fields.

This content was produced by Baidu. It was not written by MIT Technology Review’s editorial staff.

Continue Reading
Comments

Uncategorized

Daily Crunch: A huge fintech exit as the week ends

Published

on

To get a roundup of TechCrunch’s biggest and most important stories delivered to your inbox every day at 3 p.m. PDT, subscribe here.

Our thanks to everyone who wrote in this week about the format changes to the newsletter! Feedback largely sorted into two themes: Some people really like the more narrative format, and some folks really want a more link-list styled missive. What follows is an attempt to balance both perspectives.

Starting today we’ll bold company names, so that you can more quickly pick out startups, add more bulleted points to sections, and, per a different piece of feedback, include more regular descriptors of companies that are not household names.

That said, we’re not going to abandon chatting with you every day, as TechCrunch is nothing if not full of things to say. So here’s a blend of what the new, updated Daily Crunch team had in mind, and your notes. A big thanks to everyone who wrote in!

Alex @alex on Twitter

A mega-exit for American fintech

The news that public fintech company Bill.com will buy Divvy, a Utah-based startup that helps small and midsized businesses manage their spend, was perhaps the biggest startup story of the week. Breaking late Thursday, the $2.5 billion transaction was long expected. Divvy had raised more than $400 million from PayPal Ventures, New Enterprise Associates, Insight Partners and Pelion Venture Partners.

TechCrunch covered the impending sale, rumors of which sprung up before Bill.com reported its Q1 earnings. To see the company drop the news at the same time as its earnings was not a surprise. For the burgeoning corporate payment space (more here on startups in the space like Ramp, Airbase and Brex).

I got to noodle on the financial results that Bill.com detailed regarding Divvy — they are pretty key metrics to help us value the startups that are competing to go public or find a similarly feathered corporate nest. In short, the corporate spend startup cohort is doing great. It’s even spawning new startups like Latin American-focused Clara, which raised $3.5 million earlier this year.

Broadly, the fintech market had a huge Q1 and is blasting its way toward a record venture capital year, like AI startups and the rest of the VC world.

Startups and venture capital

5 investors discuss the future of RPA after UiPath’s IPO

Much ink (erm, pixels) has been spilled about robotic process automation (RPA) recently, particularly in the wake of UiPath’s IPO last month.

But while some of the individuals Ron interviewed about the future of RPA believe the technology is in its “early infancy,” the pandemic increased attention toward things we can let robots handle for us. And it’s hard to argue that repetitive tasks like billing and spreadsheeting and paper-pushing should not be outsourced to robots.

“RPA allows companies to automate a group of highly mundane tasks and have a machine do the work instead of a human,” Ron writes. “Think of finding an invoice amount in an email, placing the figure in a spreadsheet and sending a Slack message to accounts payable. You could have humans do that, or you could do it more quickly and efficiently with a machine. We’re talking mind-numbing work that is well suited to automation.”

Although RPA is the fastest-growing category in enterprise software, the market remains surprisingly small. Ron spoke to five investors about where the sector is headed, where there are opportunities and the biggest threats to the RPA startup ecosystem.

(Extra Crunch is our membership program, which helps founders and startup teams get ahead. You can sign up here.)

The tech giants

It was a quieter day from the tech giants, who made plenty of news earlier in the week. The good news is that their relative calm means we can take a look at news from other Big Tech companies, those that don’t quite crack the $1 trillion market cap threshold yet:

Community

Some of us are mourning the shutdown of Nuzzel, so we asked … would you pay for it (and why)? Let us know what you think!

Continue Reading

Uncategorized

Tesla refutes Elon Musk’s timeline on ‘full self-driving’

Published

on

What Tesla CEO Elon Musk says publicly about the company’s progress on a fully autonomous driving system doesn’t match up with “engineering reality,” according to a memo that summarizes a meeting between California regulators and employees at the automaker.

The memo, which transparency site Plainsite obtained via a Freedom of Information Act request and subsequently released, shows that Musk has inflated the capabilities of the Autopilot advanced driver assistance system in Tesla vehicles, as well the company’s ability to deliver fully autonomous features by the end of the year. 

Tesla vehicles come standard with a driver assistance system branded as Autopilot. For an additional $10,000, owners can buy “full self-driving,” or FSD — a feature that Musk promises will one day deliver full autonomous driving capabilities. FSD, which has steadily increased in price and capability, has been available as an option for years. However, Tesla vehicles are not self-driving. FSD includes the parking feature Summon as well as Navigate on Autopilot, an active guidance system that navigates a car from a highway on-ramp to off-ramp, including interchanges and making lane changes. Once drivers enter a destination into the navigation system, they can enable “Navigate on Autopilot” for that trip.

Tesla vehicles are far from reaching that level of autonomy, a fact confirmed by statements made by the company’s director of Autopilot software CJ Moore to California regulators, the memo shows.

“Elon’s tweet does not match engineering reality per CJ,” according to the memo summarizing the conversation between regulators with the California Department of Motor Vehicles’ autonomous vehicles branch and four Tesla employees, including Moore.

The memo, which was written by California DMV’s Miguel Acosta, states that Moore described Autopilot — and the new features being tested — as a Level 2 system. That description matters in the world of automated driving.

There are five levels of automation under standards created by SAE International. Level 2 means two primary functions — like adaptive cruise and lane keeping — are automated and still have a human driver in the loop at all times. Level 2 is an advanced driver assistance system, and has become increasingly available in new vehicles, including those produced by Tesla, GM, Volvo and Mercedes. Tesla’s Autopilot and its more capable FSD were considered the most advanced systems available to consumers. However, other automakers have started to catch up.

Level 4 means the vehicle can handle all aspects of driving in certain conditions without human intervention and is what companies like Argo AI, Aurora, Cruise, Motional, Waymo and Zoox are working on. Level 5, which is widely viewed as a distant goal, would handle all driving in all environments and conditions.

Here is an important bit via Acosta’s summarization:

DMV asked CJ to address from an engineering perspective, Elon’s messaging about L5 capability by the end of the year. Elon’s tweet does not match engineering reality per CJ. Tesla is at Level 2 currently. The ratio of driver interaction would need to be in the magnitude of 1 or 2 million miles per driver interaction to move into higher levels of automation. Tesla indicated that Elon is extrapolating on the rates of improvement when speaking about L5 capabilities. Tesla couldn’t say if the rate of improvement would make it to L5 by end of calendar year.

Portions of this commentary were redacted. However, Plainsite was able to copy and paste the redacted part, which shows up as white space on a PDF, into another document.

The comments in the memo are contrary to what Musk has said repeatedly in the public sphere.

Musk is frequently asked on Twitter and in quarterly earnings calls for progress reports on FSD, including questions about when it will be rolled out via software updates to owners who have purchased the option. In a January earnings call, Musk said he was “highly confident the car will be able to drive itself with reliability in excess of a human this year.” In April 2021, during the company’s first quarter earnings call, Musk said “it’s really quite, quite tricky. But I am highly confident that we will get this done.”

The memo released this week provided other insights into Tesla’s push to test and eventually unlock greater levels of autonomy, including the number of vehicles testing a beta version of “Navigate on Autopilot on City Streets,” a feature that is meant to handle driving in urban areas and not just highways. Regulators also asked the Tesla employees if and how participants were being trained to test this feature, and how the sales team ensures that messaging about the vehicle capabilities and limitations are communicated.

As of the March meeting, there were 824 vehicles in a pilot program testing a beta version of “city streets.”  About 750 of those vehicles were being driven by employees and 71 by non-employees. Pilot participants are located across 37 states, with the majority of participants in California. As of March 2021, pilot participants have driven more than 153,000 miles using the City Streets feature, the memo states. The memo noted that Tesla planned to expand this pool of participants to approximately 1,600 later that month.

Tesla told the DMV that it is working on developing a video for the participants and that the next group of participants will include referrals from existing participants. “The new participants will be vetted by Tesla by looking at insurance telematics based on the VINs registered to that participant,” according to the memo.

Tesla also told the DMV that it is able to track when there are failures or when the feature is deactivated. Moore described these as “disengagements,” a term also used by companies testing and developing autonomous vehicle technology. The primary difference worth noting here is that these companies only use employees who are trained safety drivers, not the public.

Continue Reading

Uncategorized

Betting on upcoming startup markets

Published

on

Welcome back to The TechCrunch Exchange, a weekly startups-and-markets newsletter. It’s broadly based on the daily column that appears on Extra Crunch, but free, and made for your weekend reading. Want it in your inbox every Saturday? Sign up here.

Ready? Let’s talk money, startups and spicy IPO rumors.

Betting on upcoming startup markets

This week M25, a venture capital concern focused on investing in the Midwest of the United States, announced a new fund worth $31.8 million. As the firm noted in a release that The Exchange reviewed, its new fund is about three times the size of its preceding investment vehicle.

I caught up with M25 partner Mike Asem to chat about the round. Asem joined M25 in 2016 after partner Victor Gutwein spearheaded the effort with a small $1 million fund. Asem and Gutwein have led the firm since its first material, if technically second fund.

Asem said that his team had targeted a $25 million to $30 million fund three, meaning that they came in a bit higher than anticipated in fundraising terms. That’s not a surprise in today’s venture capital market, given the pace at which capital is both invested into VC funds and startups.

The investor told The Exchange that M25 has been investing out of its third fund for some time, including CASHDROP, a startup that I’ve heard good things about regarding its growth rate. (More here on the CASHDROP round that M25 put capital into.)

All that’s fine, but what makes M25 an interesting bet is that the firm only invests in Midwest-headquartered startups. Often when I chat to a fund that has a unique geographical focus, it’s merely that, a focus. As opposed to M25’s more hard-and-fast rule. Now with more capital and plans to take part in 12-15 deals per year, the group can double down on its thesis.

Per Asem, M25 has done about a third of its deals in Chicago, where it’s based, but has put capital into startups in 24 cities thus far. TechCrunch covered one of those companies, Metafy, earlier this week when it closed more than $5 million in new capital.

Why does M25 think that the Midwest is the place to deploy capital and generate outsize returns? Asem listed a number of perspectives that underpin his team’s thesis: The Midwest’s economic might, the network that his partner and him developed in the area before founding M25, and the fact that valuations can prove to be more attractive in the region at the stage that his firm invests. They are sufficiently different, he said, that his firm can generate material returns even with exits at around the $100 million mark, a lower threshold than most VCs with larger capital vehicles might find palatable.

M25 is not alone in its bets on alternative regions. The Exchange also chatted with Somak Chattopadhyay of Armory Square Ventures on Friday, a firm that is based in upstate New York and invests in B2B software companies in what we might call post-manufacturing cities. One of its investments has gone public, and the group’s latest fund is a multiple of the size of its first. Armory now has around $60 million in AUM.

All that’s to say that the venture capital boom is not merely helping firms like a16z raise another billion here, or another billion there. But the generally hot market for startups and private capital is helping even smaller firms raise more capital to take on less traditional spaces. It’s heartening.

On-demand pricing, and grokking the insurance game

This week The Exchange chatted with Twilio CFO Khozema Shipchandler about his company’s earnings report. You can read more on the hard numbers here. The short gist is that it was a good quarter. But what mattered most in our chat was Shipchandler riffing on where the center of gravity at Twilio will remain in revenue terms.

Briefly, Twilio is best known for building APIs that allow developers to leverage telecom services. Those developers and their employers pay for as much Twilio as they used. But over time Twilio has bought more and more companies, building out a diverse product set after its 2016-era IPO.

So we were curious: Where does the company stand on the on-demand versus SaaS pricing debate that is currently raging in the software world? Staunchly in the first camp, still, despite buying Segment, which is a SaaS service. Per Shipchandler, Twilio revenue is still more than 70% on-demand, and the company wants to make sure that its customers only buy more of its services as they sell more of their own.

Startups, then, probably don’t have to give up on on-demand pricing as they scale. Twilio is huge and is sticking to it!

Then there was Root’s earnings report. Again, here are the core numbers. The Exchange is keeping tabs on Root’s post-IPO performance not only because it was a company we tracked extensively during its late private life, but also because it is a bellwether of sorts for the yet-private, neoinsurane companies. Which matters for fellow neoinsurance player Hippo, as it is going public via a SPAC.

Alex Timm, Root’s CEO, said that his firm performed well in the first quarter, generating more direct written premium than anticipated, and at better loss-rates to boot. The company also remains very cash-rich post IPO, and Timm is confident that his company’s data science work has lots more room to improve Root’s underwriting models.

So, faster-than-expected growth, lots of cash, improving economics and a bullish technology take — Root’s stock is flying, right? No, it is not. Instead Root has taken a bit of a public-market pounding in recent months. The Exchange asked Timm about the disparity between how he views his company’s performance and future, and how it is being valued. He said that the insurance folks don’t always get its technology work and that tech folks don’t always grok Root’s insurance business.

That’s tough. But with years and years of cash at its current burn rate, Root has more than enough space to prove its critics wrong, provided that its modeling holds up over the next dozen quarters or so. Its share price can’t be great for the yet-private neoinsurance companies, however. Even if Next Insurance did just raise another grip of cash at another new, higher valuation.

Corporate spend’s big week

As you’ve read by now, Bill.com is buying corporate-spend unicorn Divvy for $2.5 billion. I dug into the numbers behind the deal here, if that’s your sort of thing.

But after collecting notes from the CEOs of Divvy competitors Ramp and Brex here, another bit of commentary came in that I wanted to share. Thejo Kote, the corporate spend startup Airbase’s CEO and founder did some math on Divvy’s results that Bill.com shared with its own investors, arguing that the company’s March payment volume and active customer account implies that the company’s “average spend volume per customer was $44,400 per month.”

Is that good or bad? Kote is not impressed, saying that Airbase’s “average spend volume per customer is almost 10 [times] that of Divvy,” or around “$375,000 per month.” What’s driving that difference? A focus on larger customers, and the fact that Airbase covers more ground, in Kote’s view, than Divvy by encompassing software work that Bill.com itself and Expensify manage.

I bring you all of this as the war in managing spend for companies large and small is heating up in software terms. With Divvy off the table, Ramp is now perhaps the largest player in the space not charging for the software it wraps around corporate cards. Brex recently launched a software product that it charges for on a recurring basis. (More on Brex at this link, if you are into it.)

Various and sundry

Two final notes for you, things that should make you either laugh, grimace, or howl:

  1. The Wall Street Journal’s Eliot Brown tweeted some data this week from the Financial Times, namely that amongst the roughly 40 SPACs that completed deals last year, a dozen and a half have lost more than half their value. And that the average drop amongst the combined entities is 38%. Woof.
  2. And, finally, welcome to peak everything.

More to come next week, including notes on the return of the Kaltura and Procore IPOs, and whatever it is we can suss out from the Krispy Kreme S-1 filing, as donuts are life.

Alex

Continue Reading

Trending