Hacker News new | past | comments | ask | show | jobs | submit login
Google ML/AI Comic (cloud.google.com)
331 points by jacquesm 29 days ago | hide | past | web | favorite | 39 comments

Annnnd Martha still doesn't have a ml-solvable business problem identified with a large enough curated dataset to actually create a useful model.

That's not a problem at all! In the fortune 500 there are plenty of companies with "problems" that used to be solved with boring things like "averages" and "business logic", but now you can replace those things with LSTMs and Deep NLP models and get half the performance with several orders of magnitude more complexity! The best part is the people building these systems have none of that annoying "engineering background" baggage that will mean they don't worry about stupid stuff like support and maintenance, or even basic debugging: if the model breaks you just build a brand new one!

As an added bonus, people with these skills cost x3 as much as a standard engineer. HR will love it!

Who needs a mean when you can train a model to calculate means....

This made me lol. The comic is seriously a great intro learning resource; it didn't occur to me the subtext might be "don't attempt ML if you're not at Google".

The message should be, whether intended or not, for Martha to send this comic to her boss and begin scheduling brainstorming meetings. With the vast plethora of introductory resources for ML today, there's no good excuse for both supervisors and engineers to understand what tools it might provide.

The phrase "we need to add AI to everything" should be met with as much consternation as "we need to promote synergy". Precision is important. It's the difference between spending 6 months developing an deploying trained neural nets and 60 minutes for linear regression.

You'd be surprised how many investment proposals still contain 'AI' just to sex them up. Even if it isn't applicable. Even if the application suggested is way beyond what current AI capabilities are. Case in point: received just such a proposal last week.

> The comic is seriously a great intro learning resource

No, it isn't. It's horrible. It read like advertising copy to sway dumb investors into pouring their life savings into Google stock.

Well she was warned not to open that door.

Managed healthcare companies have huge data and they have begun to use ML to classify their documents as medical and non-medical, as quality of care and quality of service etc., based on the text they extract from their members' health insurance claims. Generally the extracted text needs a lot of preprocessing.

... so that puts her about par in the industry, no?

Maybe by the end of part 3.

I just had a flashback and blacked out for 15 minutes in a spat of PTSD from your comment, please sir use your words more carefully they are powerful and over the target.

Comically absent in this description of ML which includes hard technology like NLP and actual use cases like self-driving cars is the elephant in the room of advertising and surveillance. It's just like Andrew Ng's machine learning course on Coursera - lists all these uses of machine learning....except mining user information for advertising and other purposes. If anything, it's buried under "image recognition" and "recommender systems". And yet it's what brings in the dough. Pretty telling that the overwhelming majority of this technology's current application is too unpalatable to acknowledge.

Looks like Scott McCloud worked on this. I highly recommend his book "Understanding Comics".

It appears that it is not the first time he works with them, he also worked on their chrome comics posted elsewhere in this thread.

Before clicking, I wondered if this was going to be a comic produced by ML. Has such a thing been done?

I wondered exactly the same thing and scrolled to the end for the human credits:

  Script by Dylan Meconis, Scott McCloud, Syne Mitchell
  Art by Dylan Meconis
  Color by Jenn Manley Lee
  Japanese localization by Kaz Sato, Mariko Ogawa
  Produced by the Google Comics Factory (Allen Tsai, Alison Lentz, Michael Richardson)

Even GPT-2 can’t create long, coherent stories; I doubt that such an AI which can explain things and draw useful pictures exists.

Ahh! I thought it looked and read like Scott McCloud[1]. He's got some great books on comic theory that I highly recommend.

[1] https://en.wikipedia.org/wiki/Scott_McCloud

Go Syne! Go Kaz! Woo!

I trained a model on Isaac Asimov stories to produce 3-panel webcomics that get served onto a Twitter feed a couple of times a day. If you'd like, I can share it and you can check it out!

Yes, please!

https://nightfall.glitch.me/ :)

It was inspired by Archillect and A Softer World [1].

[1] - http://www.asofterworld.com/

Not exactly what you're looking for, but related: Gwern makes anime faces with StyleGAN


https://federated.withgoogle.com/ this is another fun one

I'm picturing this comic as a tattered poster on the wall of an abandoned shell of a factory where the last human rebels live 200 years in the future after post-AI fallout.

With a bunch of 'wanted' posters from a.d. 2026 next to it.

If I would have seen this five years ago, I would find it cool but it is a hard sell now. So many similar projects are around now for conveying basics of ML mixed with a pinch of fun elements.

I don't think anyone is trying to sell you anything, friend.

Learning and educating come in different shapes, sizes, flavours, and many people learn in different ways and at different paces. This is just another way others can learn about a complex topic.

Ah yes, the common scenario of engineers getting time-and-a-half for their overtime work.

It's a good expectation to include subtly in something that will be read by a variety of people. The Overton window on this won't be moved quickly, but it's nice to think that it might be moved...

while the folks telling them what to do go home early every day.

Good explanation of machine learning somewhere under the layers of condescending rhetoric and marketing. I went into this expecting something like the chrome comic and boy was that underwhelming.

The Chrome comic is pretty iconic I think: https://www.google.com/googlebooks/chrome/

I like this, especially since it seems to make a fair case regarding CNN and Deep Learning (not solve-alls).. looking forward to the second part.

It's a great intro. It reminds me of the official scikit-learn tutorial I worked on a while ago.

Mel is Bezos?

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact