I think "recreated a small portion of their core research" would be a fairer title. The current title makes it sound like he's actually built a competitor to the company, when in fact, all he's proved is that basic tech is often an almost trivial part of a company, compared with all of the housekeeping and other busywork required to actually provide a service to customers.
I have to add there is zero reverse-engineering here. The main purpose of this article seems to be name-dropping for the 500M company while trying to generate some 'thoughtleader'-cred
When you're peddling knowledge, knowing the formula is a big deal.
Most utilities have existing load reduction programs in place that do this kind of analysis via subcontractors who build energy efficiently plans. In my state, you can get substantial energy discounts if you agree to let the system operator control a portion of your electric supply and shut stuff off when there is a peak usage event.
All of these contractors bid on this stuff. If I can pay this guy or some other random consultant guy to do this analysis with more modern techniques, I can get rid of 3-5 energy engineers @150k of payroll and benefits each. That means I win the bid.
I can guarantee you that paying this guy money that will make him do backflips will cost maybe 10% of what it would cost if I went to Oracle or IBM.
You folks who work with this stuff don't understand how poor the tooling is in verticals like this. If you ever have an opportunity to see IBM Watson people in the room with CIOs, take it and try not to shake your head too much. They demonstrate analysis that is much less robust than this of things like help desk tickets, and CIOs think they've met HAL.
As someone who does ML for a living, couldn't agree more. Usually your initial research only constitutes a tiny portion of your spent resources, it's putting it in production that makes you a 500 mil company. As always the devil is in the details, and it's unreal how much can go wrong once you actually have to ACT using the model and not just theorize. Sounds like someone hasn't left academia yet...
It's a great effort here and I appreciate the post, but I think the title is a little misleading. There's a whole lot more to a company, let alone a $500MM company, than its tech/product. It's like saying you reverse engineered a kettle and now you're Breville. Consider team, brand, adoption, supporting infrastructure/marketing channels/processes/etc etc etc.
I apologise if this is nitpicky, but I think a more accurate description would be "I reverse-engineered a $500M AI company's algorithm in one week".
I really don't like the arrogance coming through the writing style, especially given how clear it is the author has at best basic working knowledge of ML.
Never mind the fact that k-means is ML 101 and the $500M company is likely using more sophisticated ones, the fact that he says the following tells me he's just reading tutorials and plugging data into libraries (which is fine but not with this tone of know-it-all writing):
"I played with the number of clusters, and the one that allowed me to get the most significant clusters was 6 (this was a trial and error approach, for brevity I’ll report just the final outcome)."
Anyone who has studied clustering knows you would at the very least do a scree plot here. You can defer to intuition but there's more to it than running kmeans and claiming you've reverse-engineered a $500M company.
To be fair, he did say that he was just reporting the final outcome. It's unclear exactly what was involved in his trial and error approach or how he judged "significant clusters".
Opower is most certainly not an 'Artificial Intelligence company'. Opower's foundations are in Behavioral Science, literally applying the results of an experiment performed by Robert Cialdini to electrical utilities all over the world. (http://www.slate.com/articles/technology/the_efficient_plane...). If being a 'Big Data company' is a thing, that would better describe Opower - they ingested massive amounts of utility customer usage data, and from that were able to find similar house holds, rank them, and produce a behavioral effect that worked. It's an impressive feat but not AI.
Also, I really hope "I recreated this AI company in a weekend" isn't the new "I recreated Twitter in a weekend." No, you didn't.
Posting an outrageous title to get attention, thereby triggering a discussion entirely focused on the outrageous title is such an anti-pattern that we've demoted this post.
Normally we can find representative language in the body of an article to serve as a substantive title, but I tried and came up empty in this case. That can't be a good sign.
Unfortunately, this type of headline is less clickbaity than other headlines I've seen recently relating to Data Science/Machine Learning, even on Hacker News. (although in this case, the clickbaityness is more deliberate)
It is honestly one of the reasons I am cutting down on producing blog articles on those topics, because I can't compete with clickbait-articles-which-peddle-machine-learning-as-magic-when-it-is-not, and it is beginning to get frustrating.
Yes, this genre is well along on the hype curve. We may not always be able to assess when articles are doing this but we're definitely open to suggestions for better (= more accurate and neutral) titles.
Finding problem space in haystack is more important than building solution. Because that would mean founders' know why a solution exists as it exists in problem space and how to mould/pivot it down the line and build company around it.
I perfectly know that the title is an exaggeration, but choosing an headline for medium is a tough job :)
The reason why I chose to call it like this, is not just to get a couple of clicks, but because while talking to companies that are not ML-aware I often get the question: "how the hell does X do that?!?!?", and the answer 80% of the time could be 4 lines of sklearn.
My goal is to spread awareness on the potential of ML among companies: my intended audience was not ML engineers, to whom this article looks more like "I spent 6 days cleaning data, 20 minutes plotting different clusters representations, and the rest of the day writing an article", but the business person that is not fully aware of what it means to use ML today, and to whom it looks like something amazing and extremely valuable.
As the top comment on this post (https://news.ycombinator.com/item?id=13965043) demonstrates, the title is misleading at best, and I do not believe you want to intentionally mislead your readers. Rhetorical smiley faces do not change that.
Keep in mind that clickbait titles do get penalized on Hacker News, as dang notes.
I didn't even post it on HN, you guys aren't my target as I explained.
I'm trying to make business execs more eager to experiment ML in their companies, and less afraid about the years of R&D and skynet scenarios they currently relate AI to.
The title was an hyperbole? Yes, but it worked in getting the attention of my target audience. Anyone trying to work as an ML engineer knows that what I did is simple and far from being worth $500M, but should still thank me for spreading awareness on the potential of this technology among who's still scared.
Ex-Opower Data Scientist here. Cool to see this done. Just wanted to add a link to the Opower blog post describing the original work this was based on, done by Erik Shilts: https://blog.opower.com/2014/10/load-curve-archetypes/.
26 comments
[ 2.8 ms ] story [ 81.5 ms ] threadMost utilities have existing load reduction programs in place that do this kind of analysis via subcontractors who build energy efficiently plans. In my state, you can get substantial energy discounts if you agree to let the system operator control a portion of your electric supply and shut stuff off when there is a peak usage event.
All of these contractors bid on this stuff. If I can pay this guy or some other random consultant guy to do this analysis with more modern techniques, I can get rid of 3-5 energy engineers @150k of payroll and benefits each. That means I win the bid.
I can guarantee you that paying this guy money that will make him do backflips will cost maybe 10% of what it would cost if I went to Oracle or IBM.
You folks who work with this stuff don't understand how poor the tooling is in verticals like this. If you ever have an opportunity to see IBM Watson people in the room with CIOs, take it and try not to shake your head too much. They demonstrate analysis that is much less robust than this of things like help desk tickets, and CIOs think they've met HAL.
I apologise if this is nitpicky, but I think a more accurate description would be "I reverse-engineered a $500M AI company's algorithm in one week".
Never mind the fact that k-means is ML 101 and the $500M company is likely using more sophisticated ones, the fact that he says the following tells me he's just reading tutorials and plugging data into libraries (which is fine but not with this tone of know-it-all writing):
"I played with the number of clusters, and the one that allowed me to get the most significant clusters was 6 (this was a trial and error approach, for brevity I’ll report just the final outcome)."
Anyone who has studied clustering knows you would at the very least do a scree plot here. You can defer to intuition but there's more to it than running kmeans and claiming you've reverse-engineered a $500M company.
Opower is most certainly not an 'Artificial Intelligence company'. Opower's foundations are in Behavioral Science, literally applying the results of an experiment performed by Robert Cialdini to electrical utilities all over the world. (http://www.slate.com/articles/technology/the_efficient_plane...). If being a 'Big Data company' is a thing, that would better describe Opower - they ingested massive amounts of utility customer usage data, and from that were able to find similar house holds, rank them, and produce a behavioral effect that worked. It's an impressive feat but not AI.
Also, I really hope "I recreated this AI company in a weekend" isn't the new "I recreated Twitter in a weekend." No, you didn't.
Normally we can find representative language in the body of an article to serve as a substantive title, but I tried and came up empty in this case. That can't be a good sign.
It is honestly one of the reasons I am cutting down on producing blog articles on those topics, because I can't compete with clickbait-articles-which-peddle-machine-learning-as-magic-when-it-is-not, and it is beginning to get frustrating.
On the other hand, this is excellent PR for Opower.
I perfectly know that the title is an exaggeration, but choosing an headline for medium is a tough job :)
The reason why I chose to call it like this, is not just to get a couple of clicks, but because while talking to companies that are not ML-aware I often get the question: "how the hell does X do that?!?!?", and the answer 80% of the time could be 4 lines of sklearn.
My goal is to spread awareness on the potential of ML among companies: my intended audience was not ML engineers, to whom this article looks more like "I spent 6 days cleaning data, 20 minutes plotting different clusters representations, and the rest of the day writing an article", but the business person that is not fully aware of what it means to use ML today, and to whom it looks like something amazing and extremely valuable.
Does it make more sense now? :D
Keep in mind that clickbait titles do get penalized on Hacker News, as dang notes.
I'm trying to make business execs more eager to experiment ML in their companies, and less afraid about the years of R&D and skynet scenarios they currently relate AI to.
The title was an hyperbole? Yes, but it worked in getting the attention of my target audience. Anyone trying to work as an ML engineer knows that what I did is simple and far from being worth $500M, but should still thank me for spreading awareness on the potential of this technology among who's still scared.