For we are mistaken when we look forward to death; the major portion of death has already passed. Whatever years be behind us are in death’s hands.
So, actually 10 years ago already. Time flies.
Pretty universal set of things to check for in new ideas.
Not knowing the answer of this question = they do not know they are travelling on Titanic at best, if not a boat full of holes.
I recently joined a seed-stage startup solving a business problem via audio analysis in the manner I described above. I'm not spending much time doing ML yet, but I'm banking on my belief that we're solving a valuable problem (customers want to buy our hacky MVP) and that ML can and will be needed to scale our solution. By deeply understanding the customer as a first step, I think the ML systems we build will be business critical and enduring. Time will tell
The most successful models I've ever built have been logistic regression models. If you can rephrase your problem in a way that's amenable to run-of-the-mill statistical techniques, you can frequently achieve much better results than you can with 'deep learning'.
I had a pretty good regression model but it was not taken seriously. So I wrote it using "a neural network in TensorFlow" and the next thing you know the whole company is asking me how it works and what it does.
This product is certified buzzword compliant.
I think this overstates network effects and under-appreciates branding and simply reinvesting significant capital into continued R&D.
Google got great because of PageRank, but it stayed great due to AdWords.
Google got great because of PageRank, but it stayed viable due to AdWords.
Fixed that for ya
> How do you make money? Be on the lookout for what I call “multistage rockets”: “Today, we’re doing X. But our grand plan is to do Y, which will be really profitable”. These usually fail.
But, it is also completely non-informative. All startups usually fail. All (completely) new endeavors, even in established companies, usually fail. That does NOT mean though that it's not worth a try. You cannot generally anticipate apriori what will work and what not - and you CERTAINLY cannot anticipate it from the business plan (what was the business plan of Facebook? Instagram?). If you stand any chance to anticipate "what will work" - it's from the team, not the business plan.
I may be wrong, I'm (mostly) a DL hardware guy, this is just my two cents :)
Even for the long tail? I'm pretty skeptical of this.
If you do your research on google and then buy from amazon or somewhere else and don't use the adword link google's business strategy fails.
Right now I'm at the Web Summit, in Lisbon. I saw a few ML/AI startups here, but I was surprised there weren't a lot more. There was some imbalance. There were many startups trying to create new online social networks for niche groups -- business professionals, women, entrepreneurs, sales professionals, etc. I do think there is some innovation that can happen in social networks, but generally I regard it as an over-crowded space. For the most part, the benefits offered by the Internet seemed to have been mostly absorbed by the 1994-2008 cycle, and what's left is fairly minor compared to what happened previously.
One of the few areas where I still see the possibility of significant traction (the creation of large companies) is with information gathering and a combination of ML and NLP. We've already seen a wave of startups which were no more than pretty GUIs over existing technology. At the risk of being unfair, BigML.com is just a nice interface over existing ML tools, and API.ai is just a nice interface over some NLP tools. AmenityAnalytics goes a bit further combining their web scraping and NLP scripts with a nice interface that customers can use to filter the incoming data. But there is still a wide space for companies to go much further in this field.
All the same, I agree this is a good question: "Why does anyone need this?" You should really ask it whenever you are joining any startup.
Where do you get your data from?
If your data is sourced from users of your product, do you tell them what you’re collecting?
There many other factors to optimize for, like the people you’ll work with, the technologies you’ll work on, commute, etc.
1) Is a Non-Compete Agreement required and what are the terms (time and geographic coverage; compensation during the non-compete time frame)?
2) Is the company going to assert ownership over IP created before employment began?
3) Is the company going to assert ownership over IP created that is unrelated to current responsibilities (or IP in a area unrelated to the company's current business), especially IP created on the employee's own time, not using any company resources?
4) What is the companies policy on employees moonlighting / doing side contract work for other companies not related to it's business?
Inquire about this before you get your offer. Ask me how I know...
EDIT: Another one, which is kind of risky but I personally think is important and would ask any small company:
5) How transparent/open are you about the company's financial health? Are the books open or semi-open?
Keeping in mind that some companies will not share this information and may even find the question offensive. One time I asked the owner/founder this, and he would not share--I found out later through a side channel that he was very offended that I asked for this information (but hired me anyway because he needed someone like me).
My experiences have never been great when asking such loaded questions. To be clear, I interviewed extensively at small startups and I'm only taking those interviews into consideration for this comment.
The HR and technical interviewers dedicate a large amount of time for questioning you, and reserve the last few minutes to answer your questions. I lost count of how many times the interviewers would just blabber something meaningless in a rush, rather than answer a questions with patience and honesty. Of course, I made my decisions on whether or not to join the team based on such experiences. However, I cannot think of any remarkable instances where the interviewers answered such questions without getting impatient.
Any chats with directors/PMs/C-level executives after receiving an offer were also not very informative. I walked out several times going in and coming back from such chats with no questions properly answered because: (a) they are still figuring out, or (b) they cannot discuss certain details because you haven't said yes to them.
For us, a great interview is one where the candidate teaches us and thereby demonstrates the potential to be a force multiplier for the team. Also, someone asking these kinds of questions suggests we may be able to put them in front of a customer!
If, let's say, you have 0.01% of a Series A that exits at a billion in five years. You'd expect to be diluted by half and there is around a 20% chance of exit that high so roughly ((0.01%billion)/2).20 /5 or about 2k a year in 'expected value'(or why, numerically at least, it's likely not worth it to take a paycut to work at a startup unless you're seeing substantial equity).
That figure probably oughta be at least one order of magnitude less.
I now recommend students to find out whether companies are just name dropping ML or doing serious ML work. Typically, the best way is to ask former employees/interns. But I question directed at this also would not hurt.
 There is nothing wrong with that, but it sets the wrong expectations.
One of the reasons for this is that the people designing the systems don't know how to do anything else.
There's such a huge, huge gap between using ML to do the same thing everyone else is doing (say... image classification, or maybe regression on volume vs sales...) and solving unique business problems using ML.
As a wanna be startup founder, I found my ideas have bad answers to at least 2-3 of these questions.
Any bad answer to these questions is a showstopper, yes.
It's an indicator that this company is definitely going to fail.
In your example, not having talked to many customers is the classical Juicero approach: Build a solution, then search for a problem that this solution is solving.
Not talking to customers means not understanding if there is demand for something, and if there are customers who would like to pay for a solution. (Problems exist in markets where target groups are not used or not willing to pay for digital solutions)
If you're starting out, I highly recommend reading the blog of Amy Hoy and Alex Hillmann (https://stackingthebricks.com/) - they defend the idea to first go on a "sales safari" where you simply obvserve your target group, and then find a solution for their pains.
As a founder, you want to continuously talk to potenttial customers about your idea or your prototype: make them use it and perform thinking-aloud tests.
What about a product targeting the mass market, if I'm building a mobile app, is talking to a dozen friends enough? or should I pay for a market research which can reach hundreds of strangers.
Do you talk to customers with prototypes of the features or do you just have a basic skeleton and say 'what if I have this and that'?
I actually subscribed to Amy Hoy's mailing list but haven't checked it out for a while...
That "10X" sounds like a bit of a heuristic, but is it a good one? Surely, what
you're looking for is some assurance that the rewards from the use of machine
learning justify its cost and that the profit the company can make out of it
is higher than the profit they would be (or were) making without it?
It doesn't even matter even the profit is really much higher. In business, as
in war (and, er, board games) any edge you can get is enough to push ahead. If
a company can make even 1% more of what it did before, thanks to some new
technology or whatever other trick, then that trick is worth pursuing in
I'm thinking it's more like this TED talk about something being remarkable:
Many ML/AI companies especially if funded and therefore highly visible may get acquired as the market is very hot. Isn't that good for the engineers?
Only if the potential acquirers are happy with the answers to the same questions. As a percentage of net worth, individual engineers have a lot more at stake than VC's or potential acquirers.
General it's always legitimate to ask these questions. Why? Because they cut straight to how the company will be valued, by a customer or by a potential acquirer.
Best thing to do is try to get a contact inside the company, then when you get the offer, call back and start asking those questions.
Even better still - don't agree to an interview unless you know in advance that the job will move your career in a good direction, then don't even bother asking questions. It's not like they have to tell you the truth anyway.
Only if you are in a situation where you need this job, right now. Try to avoid being in that situation. It isn't always possible, no, but there are reasonably steps you can take to try, such as prioritizing saving up a cushion over creature comforts and applying for multiple jobs.
And if you are good at a hot tech, which ML is right now, you probably do have some leverage just from rarity.
Yes, the perfectly free market of labor may be a theoretical construct, but don't give in to the despair that you have no control over your employment and that they have all the power. It'll become a self-fulfilling prophecy.
Generally, their advantage is in the way they use data to solve a vertical specific problem. They more "bolt on" AI.
AI isn't really the focus of most of these companies. This is just what investors want to hear about.
The article here isn't calling them out specifically, but AI infra startups are what are being discussed here as the "hammer looking for a nail" startup.
YC funded a few of them including Skymind (my company) and deepgram (mainly audio). There are a lot more of these startups such as vicarious that mainly publish papers but raise tons of funding hoping for a deepmind style exit (they didn't really have a business model).
With that context out of the way, I'll also maybe just state some learnings from the other side of the table.
Skymind has typically made money trying to reduce dependencies on cloud providers by decoupling the AI infra from google cloud and co from the cloud provider itself.
A horizontal play like this is by its nature very hard. We maintain the whole stack including our own framework. We also started in 2014 and have a decent foothold in enterprise. I wouldn't recommend trying to do this in 2017.
That being said, one other semi similar horizontal AI startup that is being started are the chip companies. Those are significantly harder to run than even what we're doing.
The most common "failed" type of startup that I think of when we think "horizontal" plays are Machine Learning as a service, which is hugely a loss leader for the various cloud providers (some companies are building on top of these providers though).
This is where you see Metamind, Alchemy API, Scaled Inference (founded in 2015), even Nervana before they sold were trying to a "nervana asic cloud" among others.
Maybe a lot of investors from SV view this as a waste of time dissecting, but I would personally love to see a bit more content on acknowledging some of these trends in the market.
The allure of "AI as a service" and the horizontal dev tools infra play is that you can try to build the next AWS similar to what the container and database companies are trying to do. Execution is definitely key for this to work though. Research also can't be the primary focus.
I won't comment on what will or won't work there
There's only one question they need to ask:
"How much are you going to pay me?"
Edit: The reason anyone would wax poetically about everything else is because he or she realizes that they are not being paid enough. It is like day trading - the moment it goes against a bad trader he or she calls it "investment"
Edit: Thanks for the down votes! It makes me feel warm and fuzzy inside to know that it is still possible to swindle people!
But mostly because it really sucks to spend a couple of years building stuff that never gets used. There are a few people who seem happy to get paid even if nothing is achieved. But most engineers I know do the work because the like building things that get used, things that are valuable.
First off, the people you're negotiating with see it as valuable, so it's a very bad negotiating move to say, "Keep your dumb equity and just give me cash." You'd be slightly better off just pointing to their family photo and saying, "Wow, your kids are ugly." Second, going through the work of valuing equity is necessary to see how the investors (that is, they people actually paying your salary) look at the company. Third, on the off chance you are successful, you're going to really wish you took the equity part of the negotiation seriously.
The down votes on calling spade a spade demonstrate it quite well
Sure, but the 'correct' answer to that question depends on a lot of other factors related to the job. If you want me to work long hours, doing dull, morally dubious, work in a high stress environment you're going to have to pay me a lot more.
Sleep on that.
Absolutely. Money is just one of many factors I consider when choosing a job. I see no 'shame' in that. I honestly have/earn enough money to cover all my expenses, so getting a bit more isn't that super important to me. As long as you meet my floor I'm willing to negotiate away money for other perks. I'd rather have time to spend my money rather than just more money.
If the technology is creating some value for someone, and
if you have only one smart business dude that knows how to approach the right people and validate markets,
reverse engineering from technology to market _does_ work.
Source: Did that myself.
Countless examples of where it fails.
Unfortunately, sometimes the hype is so strong, the cart comes before the horse and gets acquired for quite a lot - which is why these companies exist.
The trick is to find out all these answers WITHOUT asking the interviewer.