Adrenaline Shots for Apple AI

  • Apple has been criticized for not doing enough in AI. Two recent announcements show the company is closing the gap.
  • In the past two weeks, the company has announced the hiring of Google’s AI head, and an AI partnership with IBM.
  • Google’s AI head (John Giannandrea) brings credibility to Apple AI, critical in recruiting, and is likely work on AI-powered interfaces and Apple’s self-driving car program.
  • IBM partnership allows iOS developers access to IBM Watson’s enterprise machine learning, and use it to make smarter AI apps.

Core ML 101. At WWDC 2017 Apple unveiled Core ML, a platform that allows developers to integrate machine learning into an app. The AI model runs locally on iOS and does not need the cloud. At the time of the announcement, Apple outlined 15 domains for which they have created ML models, such as face detection, text summarization, and image captioning.

IBM Watson and Apple announcement. Two weeks ago Apple and IBM announced they will integrate IBM Watson with Apple Core ML. Previously, developers could convert AI models built on other third-party platforms, like TensorFlow (Google) or Azure ML (Microsoft) into Core ML, and then insert that model into an iOS app. Now developers will be able to use Watson to build the machine learning model, convert it to Core ML, and then feed the data back to Watson’s cloud. The reason why this is important is it allows iOS developers to leverage Watson’s capabilities and ultimately improve the AI in iOS apps.

Watson works locally on iOS and improves apps. What’s unique about Core ML is it runs locally on mobile devices, meaning it doesn’t have to send data back to a server. This is different than other mobile AI approaches. Running locally is an advantage when the speed of AI is important, like image recognition in AR or natural language processing. What’s new is Watson will be able to “teach” Core ML to run the AI model built with Watson. Basically, Watson does the hard work of getting a usable AI model built and then teaches it to Core ML, who can then run the model locally on its own. The app can then send data on the model’s performance back to Watson, at any time, to be analyzed for available improvements.

Recent history of Apple and IBM. In July 2014, Apple and IBM partnered to create enterprise applications on iOS devices, leveraging IBM’s big data and analytics and Apple’s hardware-software integration. IBM started selling iPhones and iPads to clients that came with software and applications for enterprise designed with Apple’s help.

Summary of big tech’s machine learning services. 

Disclaimer: We actively write about the themes in which we invest: artificial intelligence, robotics, virtual reality, and augmented reality. From time to time, we will write about companies that are in our portfolio.  Content on this site including opinions on specific themes in technology, market estimates, and estimates and commentary regarding publicly traded or private companies is not intended for use in making investment decisions. We hold no obligation to update any of our projections. We express no warranties about any estimates or opinions we make.

AI Is Pulling Off Its Own March Madness Upset

In what is becoming an annual tradition, we entered a bracket chosen by artificial intelligence into the Loup Ventures March Madness pool. (See our 2017 results here) Given this year has proven especially challenging to predict due to the uncanny number of upsets, Microsoft Bing Predicts’ bracket is in surprisingly good shape. Heading into the Final Four this weekend, it is tied for 2nd place out of 13 entrants. Here’s a look at its picks.

Results. To date, Bing has chosen 38 out of 64 games correctly, including the opening round. Bing was 3 of 4 in the opening round, 22 of 32 in the 1st Round, 8 of 16 in the 2nd Round, 3 of 8 in the Sweet Sixteen, and 2 for 4 in the Elite Eight.

It’s also important to note that if you look at Bing’s bracket now, it will show a different story because it re-picks winners for matches after each round. Even with this adjustment, it only picked 48 of 64 games correctly (75%, compared to 2017’s 72%) leading into tonight’s games.

Here are Bing’s results this year compared to last:

In 2018, Bing performed slightly worse in the 1st, 2nd, and Sweet Sixteen rounds. More importantly, Bing correctly picked 2 of the Final Four teams – no small feat for anyone, man or machine.

Methodology. The Bing Predicts algorithm factors in millions of data points in an effort to create the best predictive model. The algorithm looked at every college basketball game played in the last 16 years in an attempt to analyze correlations between measurable statistics and wins. The algorithm will give an output of the likelihood a team will win the game. It’s not meant to choose a certain winner, but the higher the percentage, the greater the disparity amongst the teams.

Walter Sun, an architect of the Bing Predicts algorithm, was asked by Wired Magazine about some of the important considerations in the algorithm. Defensive efficiency, strength of schedule, coaching rankings, and miles traveled were a few of the metrics that the algorithm measures.

Who will win the NCAA Championship? Below are the algorithm’s estimations for this weekend’s matchups.

Bing pits Michigan vs Villanova in the National Championship game on Monday night. This year has stumped even college basketball fanatics, and the AI bracket is no exception. At the end of the day, the algorithm is predicting the percentage chance each team has to win the game. It may be entirely accurate, but if Villanova has a 69% chance of winning, they also have a 31% chance of losing – thus is the nature of predicting something binary like the outcome of a sports game. What’s interesting, however, is that Bing picked these two teams to meet in this round at the beginning of the tournament with Kansas winning. Looking at it now, the algorithm gives Villanova the higher probability of winning. That’s quite the change of heart.

After a last-place finish in our pool last year, Microsoft’s algorithm has performed much better against us, whether by luck or by real improvement. If Kansas wins the NCAA tournament, there is a chance the AI bracket could win our pool. We’ll be watching this weekend to see if man will prevail over machine for another year.

Disclaimer: We actively write about the themes in which we invest: artificial intelligence, robotics, virtual reality, and augmented reality. From time to time, we will write about companies that are in our portfolio.  Content on this site including opinions on specific themes in technology, market estimates, and estimates and commentary regarding publicly traded or private companies is not intended for use in making investment decisions. We hold no obligation to update any of our projections. We express no warranties about any estimates or opinions we make.

Smart Speaker Satisfaction High, but It’s Early Days

We recently surveyed 520 US consumers about smart speakers and found that 89% of respondents were satisfied with them. A closer look at the results reveals the reason for this high satisfaction; early use cases are simple (Music, weather, general questions). While questions remain simple today, we expect what users demand from their smart speakers to become more complex. The survey covered smart speaker ownership, satisfaction levels, and common uses. Here are the key takeaways:

  • 31% of respondents own a smart speaker.
  • Amazon Echo dominates the market with 55% share, followed by Google Home at 23%. See more below.

  • 89% of smart speakers owners are either satisfied (59%) or very satisfied (30%) with their device.
  • Music, weather, and general knowledge questions dominate smart speaker usage. See more below.

In line with expectations. At roughly 1/3 of the U.S. population, smart speaker penetration is in line with our current estimations. Other than Cortana being slightly over-represented and Echo being slightly underrepresented, we believe the market share in the survey data also resembles the current landscape. In terms of smart speaker use cases, our survey finds the most common activities to be listening to music, getting the weather, and asking general questions. This is consistent with studies like the one from Quartz here.

It comes as no surprise, consequently, that 89% of respondents were satisfied or very satisfied with their smart speakers. This is due in large part to the relatively simple tasks that the majority of users demand of their devices. For example, Cortana scores a 57% on our comprehensive smart speaker test. On a standard report card, this is a failing grade, but Cortana is well suited to play your music from Spotify, tell you the weather, and answer any simple question you have, so it’s easy to see why the typical user would be plenty satisfied. Put simply, people aren’t using their smart speakers for anything all that smart. But we expect that to change.

Changing use over time. The top use cases for smart speakers today make sense because they are well defined and they work consistently. Benedict Evans put it well in a blog post early last year: “You can now use an audio wave-form to fill in a dialogue box – you can turn sound into text and text into a structured query, and you can work out where to send that query.” This works really well for simple ‘google-able’ questions or fetching info from a weather app, but as the use cases broaden, it is not always clear where to send a query. Just because calling up a Spotify playlist is a well built-out process doesn’t mean the same is true for a YouTube video or Podcast. It takes a huge amount of human time and energy to make these processes run smoothly. AI assistants are a new technology, so this is not a long-term concern, but until the voice ecosystem is more robust, users will have to settle for somewhat simple use cases.

The reason we are excited about smart speakers, however, involves the much wider use of voice as a computing input to remove friction. We believe the preferred interface for countless smart home devices and software services is not countless apps or small touchscreens, but your voice. This will involve drastically increasing the number of defined places you can send those queries and the number of connected devices in your life. Music, weather, and general questions won’t go away, but other activities will increasingly take place via voice. The desire for the voice interface is there. Smart speaker adoption is outpacing that of the smartphone, and the majority of users say they wouldn’t want to go back to their life without their smart speaker. We think it’s only a matter of time until voice cements itself into our everyday lives.

Disclaimer: We actively write about the themes in which we invest: virtual reality, augmented reality, artificial intelligence, and robotics. From time to time, we will write about companies that are in our portfolio.  Content on this site including opinions on specific themes in technology, market estimates, and estimates and commentary regarding publicly traded or private companies is not intended for use in making investment decisions. We hold no obligation to update any of our projections. We express no warranties about any estimates or opinions we make.

Face Off Part 3 – Echo vs. Google Home vs. Cortana

We tested the Amazon Echo, Google Home, and Harman Kardon Invoke (powered by Microsoft’s Cortana) smart speakers, each with 800 queries, and found that Google Home answered 81% of the questions correctly vs. Echo at 64% and Cortana at 56%.

Did Santa bring you a smart speaker? Smart speakers are quickly becoming one of the most common consumer uses of AI thanks to aggressive pricing and marketing from companies like Amazon and Google. We’ve conducted our latest round of queries, designed to test the full range of abilities and accuracy of a given AI assistant. Our conclusion: home assistants aren’t just gaining popularity, they’re making drastic improvements quickly. Here is what you can expect from the smart speaker that you unwrapped yesterday.

Methodology. Just as we have in February and August of this year, we asked 800 questions to the Amazon Echo (2nd generation this time around) and Google Home. We added the Harman Kardon Invoke speaker, coupled with Microsoft’s Cortana assistant, to our evaluation. The queries covered five categories: Local, Commerce, Navigation, Information, and Command. The smart speakers were graded on two metrics: did the device understand what was asked? (this can be seen on the device’s companion app), and did it answer or execute correctly? It is important to note that we have slightly modified our question set to be more reflective of the changing abilities of AI assistants. As voice computing becomes more versatile and smart speakers become more capable, we will continue to update our question set to be reflective of those improvements going forward. Our changes included questions around the use of smart home devices. We tested each of the speakers with the Philips Hue smart lighting and Wemo Mini smart plugs.

This round, Google Home was the decisive winner, answering 81% of questions correctly vs. the Echo’s 64% and the Invoke’s 56%.  On average, the speakers understood what was being asked 99% of the time. This is a jump up from 95% in August and means that these assistants’ natural language processing has improved to correctly understand, apart from a few anomalies, everything you say. Whether or not it correctly executes your request varies widely between devices and categories.

Results. Google Home continued its outperformance with the top score in each of the five categories. The Invoke, powered by Microsoft’s Cortana, scored more or less in line with the Amazon Echo – an impressive feat considering its tiny market share and comparatively short time in the hands of users. The Echo made considerable improvements in several categories, most notably navigation, while remaining relatively flat in areas like commerce and local.

Improvement across the board. We continue to be impressed with the rate at which these AI assistants are improving. Since our first test in February, the total number of correct responses has increased by 29% for the Echo and 42% for the Google Home. Cortana, which we tested in April on a Microsoft Surface Book, has improved 8%. In less than a year we have noted remarkable improvement in every category and can only expect this level of progress to continue as adoption grows and user base expands. Google Home, however seems to be on the steepest improvement curve. One would expect Google to dominate the information category with its unprecedented access to search data, but its outperformance in categories like commands, navigation, and even commerce furthers the case, in our minds, for Google Home to steal market share away from Alexa going forward.

New features emerge. During our testing, we noticed several minor, but encouraging enhancements to the voice computing experience, namely connectivity to both other devices and other services. Excluding the Echo, when you ask for navigation information, Google and Cortana will send directions to your phone either through a push notification or the companion app. Hailing a ride through Uber was smooth on both the Echo and Google Home. The Google Home could also hand you off to a designated EBay assistant to shop for used goods. We were impressed with improvements around media services like Spotify and smart home capabilities using both lights and smart switches.

Google takes a commanding lead in commerce. Google Home correctly answered 72% of commerce questions, as opposed to 42% and 15% for the Echo and Invoke respectively. This is an area that is thought to be dominated by Amazon. However, Google’s recent anti-Amazon partnerships with retailers like Target and Walmart to make voice-based commerce more accessible pushed the Google Home ahead of other assistants. This was the Google Home’s largest area of improvement, correctly answering 34% more queries than it did in August. 

Sound quality. Overwhelmingly, smart speaker owners use them to call up and listen to music. According to Quartz data, 74% of owners play music through their smart speakers, more than asking for weather and basic info. Of the devices we tested, the Harman Kardon Invoke sounded the best; the Google Home and Echo are very similar, but are not focused on premium sound quality. For this mid-tier price range (sale prices – Google Home: $79, Echo: $79, Invoke: $99) the Invoke sounds best. Great sounding music, however, is less defensible than the AI itself. A useful assistant can be loaded onto a tiny speaker like the Echo Dot and Home Mini, or a premium speaker as we have seen with the Alexa-enabled Sonos One, Google Home Max, and the incoming Apple HomePod. It becomes clear that you are paying almost exclusively for sound quality.

Still not mainstream. Despite our comprehensive question set, the everyday user rarely extends beyond simple queries like asking for the weather, general knowledge questions, or requesting a song. In fact, 65% of Alexa users have not even enabled a skill. In order for voice computing to take hold in the home and beyond, there needs to be improvement in terms of connectivity with other devices like phones, laptops, and televisions. We did note much better connection to your phone via companion apps, your television via devices like a Chromecast of Fire TV, and better control of smart home applications, which is an encouraging sign. A short list of persistent issues keeps smart speakers in the camp of tech gadget instead of essential device (like a phone or computer), but at the rate of change we are observing, we remain confident that AI assistants, in time, will become an integral part of how we interact with computers.

Disclaimer: We actively write about the themes in which we invest: artificial intelligence, robotics, virtual reality, and augmented reality. From time to time, we will write about companies that are in our portfolio.  Content on this site including opinions on specific themes in technology, market estimates, and estimates and commentary regarding publicly traded or private companies is not intended for use in making investment decisions. We hold no obligation to update any of our projections. We express no warranties about any estimates or opinions we make.

ARCore + ARKit Make Augmented Reality Ubiquitous

On Tuesday, Google released a developer preview of ARCore, a platform for developers to build augmented reality apps for Android. ARCore brings augmented reality to the world’s largest mobile operating system, starting with just the Pixel and Galaxy S8, but will expand to other existing and upcoming Android devices. The announcement comes on the heels of Apple’s ARKit, made available to developers in June. The introduction of both platforms blankets the mobile device market and will eventually make AR as ubiquitous as the devices themselves.

How does it work? Three basic components make up the technology that enables ARCore:

  1. Motion tracking – uses internal sensors and the phone’s camera to identify features and determine your position and orientation as you move through space
  2. Environmental understanding – detects the size and location of flat horizontal surfaces like tables and walls
  3. Light estimation – blurs the lines between the real and augmented world by helping virtual objects cast accurate shadows

What about Tango? Augmented reality is not a new area of interest for Google. Over three years ago, Google released Tango in an attempt to bring AR to smartphones and tablets.  Google’s custom hardware requirements, however, left Tango with little mainstream appeal. ARCore forgoes some of Tango’s power for increased accessibility. Fortunately for consumers, as AR becomes a core capability of devices going forward, hardware will catch up in the form of more sensors and better cameras, benefitting mobile AR as a whole.

Déjà vu. In July 2008, Apple opened the App Store with a total of 500 apps. One year later, it had seen over 2 billion downloads, and by 2011, the App Store was home to over 350,000 apps with 10 billion total downloads. The Android Market (later Google Play Store) was announced 3 months after the App Store, and although it had a slower takeoff due to smaller market share, it surpassed the App Store by 2014, both in terms of number of apps available and total downloads. The Google Play Store currently offers 2.8 million apps, compared to Apple’s 2.2 million. Just as smartphone apps erupted into existence, augmented reality will soon be a core technology available to millions of users. Google expects that by ARCore’s public launch, 100 million Android devices will support AR applications, and our research suggests over 200 million iPhones will become AR-enabled with the introduction of ARKit.

AR is here to stay. The two major device platforms are now wholeheartedly embracing and investing in augmented reality. Microsoft and Facebook have also heavily invested in AR’s future, further confirming AR’s position as a pivotal technology. We have previously written about the gold rush of AR applications on the App Store, which will only be amplified by the addition of ARCore. While the race for the pole position in AR heats up, there is one clear winner – the consumer. Aside from putting useful and fun new apps in our hands, expanding the user base of the underlying technology will accelerate the adoption of the next generation of computing.

Disclaimer: We actively write about the themes in which we invest: artificial intelligence, robotics, virtual reality, and augmented reality. From time to time, we will write about companies that are in our portfolio. Content on this site including opinions on specific themes in technology, market estimates, and estimates and commentary regarding publicly traded or private companies is not intended for use in making investment decisions. We hold no obligation to update any of our projections. We express no warranties about any estimates or opinions we make.