Alexa and exoskeletons show how AI isn’t going to remake industries overnight
This is a reprint (more or less) of the ARCHITECHT newsletter from Aug. 7, 2018. Sign up here to get new issues delivered to your inbox.
First, a mea culpa: In yesterday’s issue, I referred to a Gartner Magic Quadrant without specifying that it was for public cloud storage, not IaaS over all — a result of me not reading the cited article close enough. I apologize for any confusion this caused, although the change of focus for the MQ doesn’t change my analysis of Alibaba’s cloud opportunity. For what it’s worth, it’s moving toward the top-right in the overall IaaS MQ, too.
The biggest news for me today had to be the report that only about 2 percent of Amazon Echo users have made a purchase via the device this year. It sounds like this is in part because they don’t trust the device’s accuracy, and possibly because they don’t even have that feature turned on. If memory suits me (I disconnected my three Echoes last year), that option is buried somewhere in a clunky app, and there’s also a real fear of accidental or unauthorized purchases.
But I suspect the real reason, beyond anything to do with the Echo devices or Alexa app specifically, is just that people have decided online shopping is easy enough. At least for now. Sure, it’s not as easy as saying, “Alexa, buy me a laptop,” but the added inconvenience gets you important benefits such as price comparisons, reviews, specs and, perhaps most importantly, more opportunities to talk yourself out of yet another whimsical purchase.
I really do believe we’re going to see consumers put the brakes on their smart device habits pretty soon, or at the very least stop gobbling up every new gadget and feature that’s released. Some things just don’t need to be automated, especially at the risk of opening up privacy or security risks. Also, I assume there’s a limit on people’s ability to keep up with a constant stream of new stuff when they’ve just gotten used to the last new thing. Machines don’t care about constant change, but people do.
Doing continuous delivery on your web app is one thing, but trying to do it inside people’s homes, on their ingrained habits and household budgets, is something else altogether. It won’t make shareholders or investors happy, but it really is OK to s l o w d o w n on smart-home innovation for the sake of innovation. The huge winners in this space will be those that nail the experience of optimizing and automating tasks that people really want help with — you’d have to pry my not-intelligent Roomba from my cold, dead hands — not that make it easier to buy more stuff or avoid turning a thermostat dial.
And on a related note, Ford announced on Tuesday that it’s rolling out lightweight “exoskeletons” that will help reduce shoulder stress for autoworkers who have to continuously lift their arms above their heads. This reminds me of the podcast discussion I had last year with Brian Ballard of Upskill, about the role of Google Glass and similar technologies in industry. It’s possible I’m just being naive, but it appears there’s a possible future where rather than spending billions of dollars retrofitting factories and risking economic upheaval by laying off workers, manufacturers spend markedly less money augmenting employees so they can do their jobs better and more safely.
You might also recall that Elon Musk recently blamed an over-reliance on automation for slowing down production, and vowed to remedy it by hiring more people.
I think the moral of the story is that AI (and automation, in general) is a powerful tool, but there are technological, cultural and psychological limitations to its application. We all want to live in a better, safer, simpler world, but rushing to get there before we’re ready is a recipe for disappointment and, potentially, disaster.
AI and machine learning
5 findings from O’Reilly’s machine learning adoption survey companies should know
I only read this summary, not the whole report, but the finding about “low single-digit” use of cloud-provider models is notable. For a variety of reasons, companies doing AI today seem to prefer — maybe out of necessity — building models themselves or working with consultants. That’s part of the reason we don’t see too many horizontal AI startups, because it’s not easily scalable work.
Artificial intelligence continues its fundraising tear In 2018
But you probably didn’t need an article to tell you that. This does, however, include short profiles of a few promising AI startups that you might not have heard of.
Scale API wants self-driving cars to share data
It also raised $18 million, a fact that this story buries. This is a good idea in theory — we want all cars to have access to good data — but it’s hard to see it catching on en masse. Not only is data is a pretty big moat for autonomous car makers right now, but they also need to exercise control over their systems to avoid accidents, liability, etc.
Chinese AI startup Tianrang raises a $26M funding round, launches new project to apply ML to cities
Not a lot of details here, but the founder is a former Alibaba exec and the initial application appears to be analyzing Alibaba data.
The Defense Department has produced the first tools for catching deepfakes
In this example, it probably doesn’t take an algorithm to spot Nicholas Cage as Donald Trump. But, yes, more realistic deepfakes will eventually become a real problem.
Hard to imagine Google, Facebook building AI without (checks notes) Dell EMC’s Data Science Provisioning Portal
The headline really is funny. But on a serious note, I think the big business in AI will be selling boxes that can sit on the edge and run inference tasks. Some companies will train their own models and perhaps do so regularly, but many of those will use the cloud and many more might use prebuilt models.
The future of clinical trials: The promise of AI and the role of big tech
Probably the most in-depth look you’re going to see (outside of the pharma industry) of how AI could improve a process that typically costs billions and takes years.
MnasNet: Towards automating the design of mobile machine learning models
Google is continuing its mission to optimize AI models on mobile devices, and based this research on its AutoML project.
Large scale language modeling: Converging on 40GB of text in four hours
Nvidia researchers pushing the limits of NLP training on a task they say used to take a month. Albeit, they were using 128 V100 GPUs, the latest version of which appears to have a list price of about $8,700.
Cloud and infrastructure
Larry Ellison delivers Oracle’s next autonomous database tool, more AWS trash talk
Does anybody have experience with Oracle’s existing autonomous database? It’s easy enough to take shots at Oracle-as-cloud-provider, but a truly self-healing, self-optimizing database would be a really big deal, right?
Oracle challenges Pentagon’s multibillion-dollar cloud computing contract before bids are even submitted
Autonomous database or not, this is some classic Oracle. For starters, it seems to be conceding defeat even before the process begins, perhaps hoping to delay until it gets more features online, or just to stick it to AWS. Also, the argument seems flawed, in the sense that sticking with a single provider for a single project might actually be the best choice, especially if a goal is minimizing complexity. Further, this is where third-party and open source platform software come into play to simplify the process of moving to other cloud providers if need be.
RiskRecon’s security assessment services for third-party vendors raises $25 million
In an age of GDPR liability and numerous companies getting breached via their SaaS providers, this sounds like a great idea.
Helm, the package manager for Kubernetes
If you’re thinking about Kubernetes but don’t know what Helm is, you probably should read this. For now, it’s a pretty important piece of the Kubernetes application puzzle.
Security differences: Containers vs. serverless vs. virtual machines
This might not be too much new information, but any reminders to think about security at every layer of the application stack seem like a public service.
Anchore: Container security starts with the images
Speaking of container security, here’s a profile of a startup that’s focused on it. At the moment, it’s tooling is for ensuring containers are compliant and properly configured before they’re running.
Horrors of using Azure Kubernetes Service in production
I don’t claim to know how the exact truth here, but the situation — including the Hacker News discussion — is a good reminder of the complexity of distributed systems (especially multi-tenant, as-a-service versions) and the importance of customer support. Microsoft obviously has some very smart engineers and dev advocates, but even it can’t afford to slip on product experience and support in a world where expectations are high, moving platforms is easy enough, and everyone has a voice.
M3: Uber’s open source, large-scale metrics platform for Prometheus
More open source engineering from Uber, this time around Prometheus — the tool of choice for monitoring cloud-native environments. I’m not certain if Uber has offered up any of its tech as CNCF projects, but that seems the place where they’d have the best chance at growing.
Data and analytics
Nuage: Making data systems management scalable
This seems like a clever solution to bridging the gap between data engineers and application developers. Sometimes, after I log into LinkedIn after a prolonged absence, I mourn how much good engineering was wasted on a platform that appears to keep needlessly expanding.
StreetLight Data raises $10 million for real-time traffic analytics software
How has no one solved traffic yet? Seriously. We have a lot of data, and congestion is a major problem that could at least be mitigated with intelligent streetlights and maybe more turn lanes. Maybe we also need more A/B testing on light cadence?
Taxonomy of time series forecasting problems
I certainly learned a lot from reading this. It’s probably common knowledge among most data scientists, but could be useful for other folks thinking about their companies can take advantage of time-series data.
