When AI Fails & the Crashing Robotic Cars

Robotic Cars: More accidents (now With Less Injury & Damages!)

Google’s self-driving cars are apparently in 2x as many accidents as human drivers - If you think this is just buggy new tech, way too complex from the get-go - well, you’re partially right. An important detail I should share: as of Dec 2015 virtually all accidents were not the robots fault.

The accidents are caused by humans drivers unfamiliar with robotic drivers. Furthermore, Google has programmed the cars to obey the law in absolute terms - never speeding, difficulty merging in dense or fast freeway traffic. This opens up a bunch of legal, ethical questions -

  • Is it ever ok:
    • to avoid a rear-end collision if it means driving past the walkway at a red light assuming path is clear ?
    • to drive on the sidewalk to avoid a bicycle cutting across lanes? - is it ok for AI to accept this breach in ‘rules of the road’

I understand Google’s approach, especially when trying to minimize liability: always follow the rules - logically it follows that you cannot really be at fault if you always observe the law.

There would be massive liability if an accident happened because of intentionally designed ‘flexibility’ around the laws.

Don’t let the future escape us

The future will still arrive, even if the robots drive like octagenarians.

Perhaps a simple fix for now would be to use bright red flashing LEDs (think school buses) to warn human drivers they are about to rear-end an innocent robot.

I would be more comfortable with a car which had tiers of observance and rule adherance. To my mind this would be much closer to how humans drive.

Imagine 3-tiers of system-perception as follows: (decision & other layers omitted for simplicity)

  1. base: follows laws with annoying precision
  2. local: flexibile adjustments based on current traffic - to allow merge on the highway if say, 10MPH+ is needed. Conversely if the traffic is simply going too fast, the car should be smart enough to pull over to avoid being a nuissance to other drivers.
  3. 360: calculate ANY potential extreme collision risk/avoidance measures - driving on a shoulder or

This would likely require a smart balancing act - say tier 1 detects an immenent accident which cannot be avoided following the laws, it would then shift all processor power into tier 3 - hopefully finding a creative way to avoid harm.

Robotic cars are on the cusp of being technically smarter & faster than any human driver. Accept it. Welcome it.

References

  1. http://nn.cs.utexas.edu/pages/research/neat-warning/
  2. http://www.claimsjournal.com/magazines/idea-exchange/2014/09/29/255161.htm
Higher Order Programming

Array- or Set-based Coding Style

An Anti-Pattern?

This is an exploration of advantages gained when you code everything as an array. (Using Jedi concepts from SmallTalk)

Here’s some guiding principles:

  1. All input is array-like. Even if an array of 1.
  2. Higher level functions should generally accept AND return arrays. (Except for callback methods for loops: map/reduce/each/filter)
  3. 99 out of 100 devs code suffers from what I call acute schema surplusage syndrome -based models.
  4. Yes, beware fat clASS-backed models - with all the predictable trappings: fragile instance state - so many levers and knobs to mess with, DB transactions, sql locks, async/mutexing (that always works first time), using idiomatic property getter/setters, and your public/private/final/etc usage is solid, right?

  5. So let me take a common problem and shoehorn ~add~ some set-based musings.
  6. A hypothetical Blog Site has lots of Articles, and has even more Posts (Comments).
  7. Let’s add a delete method (below) - but with support for both singular OR arrays.
package net.danlevy.why.java___why.you.got.all.the.dots____it.must.be.all.the.factories;

public class Post {
  public String   title;
  public Date     created;
  public String   message;

  public Post(String title, String message) {
    this.title    = title;
    this.message  = message;
    this.created  = new Date();
  }

  public Date isArchived() {
    return this.created < new Date(2015, 0, 1);
  }

  // Post.delete` can be called w/ a singular Post or an array of Post[]
  public static int delete(Post post) {
    Post<List> posts = new Post<List>(post);
    return delete(posts)
  }

  public static int delete(Post<List> posts) {
    return posts.map(DB.cleanupPost);
  }
}



Forgive me if my Java is a little rusty.

Work-in-progress (updated Nov. 2015)

Beautiful Models (and Data)

Work-in-progress (updated Sep. 2015)

The issue we’ll examine is deceptively simple & subtle: Naming

I want to avoid the super-fancy-tech-lingo for this article; and hopefully I can illustrate the issue in a more useful fashion.

While covered in exhausting detail before, the subject matter often gets too technical for the novice programmer to draw any practical understanding. You probably don’t need to read this if the following makes sense: No-Sql denormalization strategy, or Boyce Codd Normal Forms

Recommended reading includes:

  1. Book: Code Complete
  2. http://phlonx.com/resources/nf3/
  3. https://en.wikipedia.org/wiki/Database_normalization

The Problem - by Example

Have you ever designed a data model (in code, Sql, or excel worksheets)? Does the following look familiar?

*** anti-pattern - don't copy-paste ***
* User
  - id
  - avatarUrl
  - email
  - passwordHash

* Agent
  - id
  - primaryPhoto
  - agentName
  - agentEmail
  - agentPhoneMain
  - agentEmailPrimary
  - agentPhonePrimary
  - agentAddressLine1
  - agentCompanyName
  - agentCompanyAddress
  - *userEmail* - 'Pointer' to User table ^^^

If this is familiar to you, I’ll bet you:

  1. Feel any change to your app will necessitate hours of arduous debugging.
  2. Fear ANY Changing Requirements schema refactor

The Cost of Bad (Naming) Habits

Let’s examine some of the subtle issues (probably familiar):

=======

Why is naming a field agentEmailPrimary the worst?

For starters, you are not creating an entirely new object unto the universe. Over-specificity has some traps:

  1. ‘Locked’ into highly specific name, means agentEmailPrimary probably make your views and related code 0% reusable, and featuring annoyingly recurring bugs like:
    • Data not syncing between tables (not obvious if user.email needs to propagate to agent.agentEmail or vice-versa - nevermind complexity of manually implementing where & how to enforce this ‘logic’ …)
    • Validation rules/logic are likely duplicated & inconsitent.
    • Increasingly, your project will resemble a shaky Jenga tower.
    • Fragility piles up with every single new file, as an extremely high attention to detail is required for even trivial changes

I know, you probably feel something like…

fuck this

A Solution

// Dan's Recommended Schema Consolidation:

User
  - id
  - role: ['agent', 'lead', 'admin']
  - name
  - phone
  - address
  - email
  - passwordHash
  - company
    - name
    - address

I removed the Agent table, as it didn’t contain fields which were uniquely related to Agents.

All changes were made with these general ideas in mind:

  1. Eliminate unessesary tables. If you have a few dozen tables, this step is mandatory.
  2. Try merge related tables. Important if you are coming from a SQL background to No-SQL
  3. Delete redundant data collection (e.g. remove ActivityLogs table if replaced by Google Analytics)
  4. Try keeping all field names to a single word/noun/pro-noun.
  5. There is no such thing as Agent.agentEmail or Agent.agentPhonePrimary. Period.
  6. By using Highly Specific Names, you cast-in-stone a specific level of code-reusability and durability, well, specifically ZERO %.
  7. Don’t think you are doing yourself any favors with crap like this User.profileSummaryEmail (where ‘profile’ could include contact details for a personal ads site) . This is probably a good point to create a new table, say Profiles which includes Profiles.email.

Work-in-progress (updated Sep. 2015)

AngularJS v2.0 and the Impending Schism

I think we are witnessing the Python 2->3 ‘Conversion’ all over again. AngularJS v2.0 introduces too many changes. Not least of which is TypeScript, which is a big ask amidst the finalization of JS’s latest version: ES6.

Let me say clearly: I love TypeScript. I seccretly wish the TC-39 meetings had produced it… They didn’t. However, They came up with another (totally different), also-awesome spec…

While TypeScript compiles to JavaScript, it doesn’t mean you blindly copy & paste ‘compiled’ TypeScript. It effectively becomes required learning, as to understand annotated AngularJS 2.0 TypeScript.

Now, newbies must climb ‘Mount TypeScript’ before they can even start assembling an Angular app (with some level of understanding).

I have a feeling how this might go…

endless loop

Oh well, I’ll add it to the Newbie training list: somewhere between Basic Shell Usage and Gulp or Grunt? Godsend+Misery!

Anyway, I hope this works out…

everything is going to be ok
Polyglot Redux

Programming Languages Notes

I’m sure my Miscellaneous Observations have been made before, but here is my list of most interesting languages:

JavaScript

My One True Love, supremely versatile & ubiquitous - the all-around, amazingly-powerful champ! It’s the #1 Most Active/Popular Language on GitHub.com for years running.

I hate to admit it, but for years I foolishly had nothing but scorn and derision for what is now, my favorite language.

ES6 has only increased my ~addiction~ love. While pure ES5 will always hold a special place in my heart, each time I use some ES6, I feel that radioactive spider-bite…

There were 4 factors which pushed me into the ES6 Camp:

  1. It’s fun. Seriously. There are tangible gains in beauty, clarity & productivity.
    • Subjective claims, you say? Let me show you a bit of ES6:
    • let expired = users.filter(u => Date.now() > u.trialDate)
    • Now you don’t have to pretend you know how to use Object.create or Object.defineProperty
    • See examples below
  2. As of July 2015, ES6 is an officially finalized standard now!
  3. Support is Effectively 100%*! … Ok, BabelJS is needed to patch your code so it’s ES5 compatible. Historically JS transpilers have been frowned upon. However, as of late (2014-15) things have changed as BabelJS has become a key enabler/driver of language advancement. Tons of companies including Microsoft & Facebook use it on some of the largest sites around.
  4. Latest versions of Node include the same V8 JS engine as Chrome v45, it’s v4.5

Examples

I’m going to show you what finally made me start drinking that ES6-flavoured KoolAid.

In my recent experience, ES6 helps you write code faster. To the point. Because code is more succinct, appreciably less brain power is needed to sift through and understand your old code (or that of a teammates).

I have regularly seen KLOC savings roughly of 20-50%. That’s like Kate Moss trim!

EcmaScript 5 vs ES 2016 - Demo: Classes, Destructuring, Sexiness

// /services/users.js
class Users {
  constructor(data) {
    this.users = data || [];
  }
  expired() {
    return this.users
      .filter(u => Date.now() > u.trialDate)
  }
}
  • No more tedious code to ‘extract’ and ‘check’ fields passed to a function. Cut to example add():
// /services/users.js
class Users {
  constructor(data) { this.users = data || []; }
  add({name, email, password}) {
    // Store pwd hash, We only need to define 1 explicit `var/let` - the other vars are 'defined' with the `{fields}` wizardry above ^^^
    let hash = getSha256(password);
    return http.post('/users', {
      'name': name,
      'email': email,
      'passwordHash': hash
    }).then(usr => this.users.push(usr)); // append user upon service response
  }
}

 

Jumping on ES6 can feel like going from:

huh

To

wtf

To

#winning

Just keep sifting through the new stuff. Check out string templates, auto this binding, more-sane inheritance…

Node.JS

Rust

Official Site

  • Pros
    • Imagine if there was a language as fast as C and as powerful as Python/C++, yet without the complexity/pitfalls that usually trap even the most skilled devs.
    • In fact I’d guess Rust is roughly as complex as the ES6 spec.
    • It includes a ton of extras:
      1. Essentially Rust transpiles from semi-dynamic syntax into pure C code!
      2. Including __all the best practices__ in C you would probably screw up on, I eventually always do.
        • Automatically you get:
        • Auto Memory management (no need for a slow garbage collector!)
        • Perfectly scoped Object ownership/locking (mutexting & context switching minimized)
        • Object lifetimes (automatically implemented*, and auto coded like you knew every edge case)
        • Prevent virtually all run time errors (seriously, your code-paths become explicit: you just can’t overlook a code-path)
    • Oh yeah, it throws in true language extensibility with a sensible ‘macro’ feature.
      • Need Comprehensions? Scala style? Done, and Like Python? Done.
        1. Too good to be true? Nah, It gets better:
        • Bleeding edge indicators (github.com stats) reveal Rust is highly competitive or even beating Go (Google’s hot-newish language)
          • About 4K More Stars than Go (currently around 12,200)
          • More total Contributors ( 2x! - 1,071 vs. Go’s 479 )
          • More forks ( 3X! - 2,343 vs. 765 )
          • Number of Open Issues, Loses by a hair ( 2,000 vs 1,730 from Go )
          • Pull Requests (Rust 70+ vs. Go’s 1)
        • I had to triple check the numbers too.
    • Other libraries are very stable due to the constructs & rules of rust.
    • Threading model usable by mere mortals
  • Cons
    • Decent web frameworks are relatively new, untested, and usually undocumented (though they are getting very impressive - as of March 2015).
    • Lots of early pre-1.0 breaking changes

Python

  • Pros
    • Overwhelmingly complete assortment of algorithms are already implemented in Python ( see: scilearnkit, numpy, matplotlib, pil/pillow, etc. )
    • Very Fun to write! Comprehensions and Decomposition are great features and make other languages seem just bloated!
    • Arrays, ‘Sequences’, Tuples etc. are relatively simple
# dummy code: defines a color + pixel-coord -
def pixel(x, y, r, g, b): return dict(x=x, y=y, r=r, g=g, b=b)
# Create a new pixel object and apply to set of vars
x, y, r, g, b = pixel(10, 20, 255, 255, 255)
# Now we can call pixel
  • Cons
    • Annoyingly, Python 2.x and 3.x are incompatible. The Great Schism continues, so many years later.
    • Some essential libraries are not nessacerily understood by some devs (numpy)

Haskell

  • Pros
    • Very rewarding when you finally memorize enough syntax to whip up comprehensions-based expressive patterns
    • You will learn mind-bending code patterns - often somewhat applicable to other languages.
  • Cons
    • Syntax & Patterns can be hard to get used to.
endless loop

SmallTalk-80

  • Pros
  • Cons
    • You will likely never use this language for anything. Zero projects. However it will have more of an impact on your coding style, faster than other functional languages… This should be in the pros list)

Work-in-progress (updated Dec. 2015)

Tag Cloud