Inside | Real news, curated by real humans
Inside AI

Inside AI (Aug 27th, 2017)

Happy Sunday and welcome to the latest edition of Inside AI! For those of you who are new, I'm Rob May, CEO of Talla, and active investor in the A.I. space. If you have an early stage A.I. startup looking for investment, send me a note.  Our goal in this newsletter is not to be too technical, but to give you a very wide ranging overview of what is happening in A.I.


If you like the newsletter, I hope you will forward this to a friend so they can subscribe too.  

Or if you really want to support us, go Premium and get a special Wednesday edition each week, plus extra analysis and reports. Our August report is on sales and marketing A.I. startups. Thank you to Ascent,  Kylie and Pillar, for being sponsors. (email austin@inside.com if you want to become a sponsor.)

-- Big Idea --

The most interesting thing I read this week was this article about how you could build a backdoor into an A.I. system as a way to hack it.  The idea is that, a hacker could train a model to do something unusual when it identifies a specific use case.  For example, train a visual sign identification model that, when a stop sign has graffiti that says "HEY!" shut down the car.  Now, if you want to stop someone, and you know their route to work, you just spray paint "HEY!" on the sign the an hour before they drive by.  Their car sees it and shuts down.  You can now harass them or whatever.  The use case doesn't matter.  What matters is that these types of exploits are difficult to find because of now neural nets work.  If you deploy a pretrained model, you can't catch it unless you happen to test the model on a use case that triggers the hack.  And given the nearly infinite space of input options, that is unlikely.  And you can't inspect the code and understand it because there is no rule in the network that is readable to humans that says "when a stop sign says 'HEY' stop the car."  Neural nets are just a bunch of nodes with weights.  They don't mean anything to us.  My prediction is that model certification, possibly via blockchain signatures, is one way to solve this long term.

  • Email gray
  • Permalink gray

-- Commentary --

There is a problem I'm seeing across the A.I. software landscape that is delaying the deployment of A.I. software at companies.  It's not a long term problem, because we all assume A.I. is going to upend every industry eventually, but it's more a short term value problem that might make A.I. difficult to deploy today.

SaaS was an easy winner.  Moving from on-premise software to the cloud was a no brainer:  pay monthly, access it using your web browser, easy onboarding and setup, never install updates.  Done.  But A.I. doesn't provide the out-of-box value, compared to traditional SaaS, as SaaS did compared to on-premise software.

The reason is that making existing software concepts "AI driven" is hard.  Let's take an example.  Say you wanted to build an A.I. version of Microsoft Word.  The technology isn't there yet to have the tool write an article for you so, how would you approach it? Let's try going menu item by menu item to see how we would apply A.I.

"File-> New"  - how would you make an A.I. version of this?  Maybe you could look at the last action someone did before opening Word, and assume a template based on that.  Probably not that useful though.  Let's try a better menu item.  How about "Insert ->  SmartArt"?  Maybe you could use a natural language description of what someone wanted, or analyze the text written so far in the Word document, and auto-generate an image for it.  That could be cool.  

I don't want to go through all the menu items on Microsoft Word but, you get the point.  Some menu items could be great for A.I. use cases, others not so much.  But here is the bigger problem - each menu item that could benefit from A.I. has to be separately trained, using it's own data set.  So if you were going to do an A.I. "insert -> smartart" feature, that's one data set and one training approach, and if you were going to do an A.I. version of "format -> columns" that would be another data set and possibly different approach.  The problem then, is that every feature of Microsoft Word that you want to make intelligent might be it's own 3-4 month project.

Now let me ask you - if I come to you with an A.I. competitor to MS Word that has 3 menu items A.I. enabled, are you going to buy it?  No.  It's not better, in total, than MS Word.  So I have to resort to selling it as an add-on.  How much will you pay for that add-on?  Probably not that much.  

The problem with A.I. in enterprise software is that enterprise software does A LOT.  To attack it with an A.I. driven approach you need one of three things:

1)  Raise a ton of money so that you can take several years to build out all the A.I. versions of features you need to have a truly A.I. driven competitive product.  This would be like not launching your MS Word competitor until it had 20 of the menu items A.I. enabled.  That might be interesting, but it takes a lot of dollars to get there.

2) Sell small A.I. add-ons in the early days while you lay the groundwork to compete on a full A.I. enterprise solution.  This takes a lot of time, and has slowly scaling revenue.

3)  Take a vertically targeted approach and look for one feature, or one customer type that needs a certain feature really badly, that can make your A.I. play singularly focused at first.  This is the only realistic approach that is working well now, which is why you hear so much chatter about vertical focus in A.I.

It's coming for sure.  I firmly believe everything will be upended by A.I., but, adoption patterns matter.  Companies and market changes are path dependent.  When you look at A.I. opportunities, don't just look at the tech.  Don't just look at the team.  Look at the path to deployment and adoption because, it's different than SaaS.

  • Email gray
  • Permalink gray

That's all for this week. Thanks again for reading. Please send me any articles you find that you think should be included in future newsletters. I can't respond to everything, but I do read it all and I appreciate the feedback.   

@robmay

-- ABOUT ME --
For new readers, I'm the co-founder and CEO of Talla. I'm also an active angel investor in A.I.  I live in Boston, but spend significant time in the Bay Area, and New York.  

  • Email gray
  • Permalink gray

Subscribe to Inside AI

MORE NEWSLETTERS

News, people, culture, events and the trends shaping the Bay area

Inside San Francisco

News, people, culture, events and the trends shaping the Bay area

DAILY
News, people, culture, events and the trends shaping the Bay area

Inside San Francisco

DAILY

SUBSCRIBED!

Share via

The news, trends and tech that is reshaping the rapidly changing world of retail

Inside Retail

The news, trends and tech that is reshaping the rapidly changing world of retail

DAILY
The news, trends and tech that is reshaping the rapidly changing world of retail

Inside Retail

DAILY

SUBSCRIBED!

Share via

The present and future of AR news and technology

Inside AR

The present and future of AR news and technology

DAILY
The present and future of AR news and technology

Inside AR

DAILY

SUBSCRIBED!

Share via

Drone news for hobbyists, professionals, and investors

Inside Drones

Drone news for hobbyists, professionals, and investors

WEEKLY
Drone news for hobbyists, professionals, and investors

Inside Drones

WEEKLY

SUBSCRIBED!

Share via