10 ms·

There are surely some startups for which this is bullshit. But the good version of it is: - take some valuable task that's never been successfully automated be

by tlb 7y ago

There are surely some startups for which this is bullshit. But the good version of it is: - take some valuable task that's never been successfully automated before - do it manually (and expensively) for a while to acquire data - build an automated system with some combination of regular software and ML models trained on the data - now you can do a valuable task for free - scale up and profit The risk is that it's hard to guess how much data you'll need to train an accurate, automated model. Maybe it's very large, and you can't keep doing it manually long enough to get there. Maybe it's very small and lots of companies will automate the same task and you won't have any advantage. I think there'll be some big successes with this model, and many failures. So be skeptical -- ML isn't a magic bullet. But if a team has a good handle on how they're going to automate something valuable, it can be a good bet. As an investor, you may well face the situation down the line "We've burned through $10M doing it manually, and we're sure that with another $10M we can finish the automation." Then you have to make a hard decision. With some applications like self-driving cars, it might be $10B.

buboard 7y ago
This seems too easy a recipe to be worth it in the medium term - there is no moat. Better cover your data with very strict laws, like google does with its exclusive deals for medical data use.
- dfalfndfk 7y ago
  Like you say it's the access to data itself that is valuable. At risk of oversimplifying: building models is the easy part. Plenty of smart people who can do that. Data that is expensive to acquire is the best long-term play for an ML company. Either expensive due to regulations or expensive due to the quality of sources. Ideally both.
- paulddraper 7y ago
  Google didn't add any "laws" around that data though. The laws were already quite strict (HIPAA) and the data relatively inaccessible. Google jumped through all the hoops to get it. And an exclusive contract never hurts no matter what the service. (Logistics, payment processor, etc.)
  - buboard 7y ago
    > an exclusive contract never hurts Thats why this is unfair. The health industry incentivizes (and often mandates) open publishing of scientific results, but patient data is reserved with gold chains for the exclusive use of google
justapassenger 7y ago
> The risk is that it's hard to guess how much data you'll need to train an accurate, automated model. Maybe it's very large, and you can't keep doing it manually long enough to get there. Maybe it's very small and lots of companies will automate the same task and you won't have any advantage. I think you forgot most important option. It may not be a data problem. You may have all the data in the world and still not being able to solve the issue. Capabilities of ML are much more limited than hype makes people believe
- solidasparagus 7y ago
  I can't think of a company that wouldn't have valuable data. A core use of ML is reducing costs by making better spending decisions/reducing waste and that is relevant to almost every company I think.
  - variaga 7y ago
    Recurring expenses (i.e. the kind that will generate enough data to adequately train a ML system) are already handled pretty well by most businesses through traditional (non-ML) methods. Operations Research has been a thing for 60+ years. Businesses that don't handle their recurring expenses well now are likely to have organizational issues (e.g. senior management that ignores the advice of their reports on how to do things) that will prevent them from doing so even if ML is added to the mix. The place where businesses get in real trouble (and hence would see significant value from better decision making) is when doing things that they have _not_ done 10K+ times before. Things like "Should we expand into $NEW_MARKET?". ML isn't going to help them with that, because there will not be any useful historical data to train them with.
    - solidasparagus 7y ago
      I completely disagree with the assumption that because something has been done for 60+ years it won't be improved by new, extremely relevant technologies. Amazon is very effective at both OR and organizational dynamics, but they found massive savings by using ML to predict demand and thus inventory/costs (Research is here - https://arxiv.org/pdf/1711.11053.pdf https://arxiv.org/pdf/1711.11053.pdf).
nradov 7y ago
$10B won't be nearly enough to build level 4+ autonomous vehicle control software that will work in a wide variety of roads and weather. Probably too low by an order of magnitude or two.
sgt101 7y ago
There's another risk - which is rare events. Basically most of the reason why these workflows are not automated already is that there are a host of "once a week per operative" corner cases that don't show in the data for a long time, and when they do show they don't show as anything but noise. But this is where the human intelligence is spent and this is why the jobs are so hard to automate.

by tlb 7y ago

buboard 7y ago

This seems too easy a recipe to be worth it in the medium term - there is no moat. Better cover your data with very strict laws, like google does with its exclusive deals for medical data use.

dfalfndfk 7y ago
Like you say it's the access to data itself that is valuable. At risk of oversimplifying: building models is the easy part. Plenty of smart people who can do that. Data that is expensive to acquire is the best long-term play for an ML company. Either expensive due to regulations or expensive due to the quality of sources. Ideally both.
paulddraper 7y ago
Google didn't add any "laws" around that data though. The laws were already quite strict (HIPAA) and the data relatively inaccessible. Google jumped through all the hoops to get it. And an exclusive contract never hurts no matter what the service. (Logistics, payment processor, etc.)
- buboard 7y ago
  > an exclusive contract never hurts Thats why this is unfair. The health industry incentivizes (and often mandates) open publishing of scientific results, but patient data is reserved with gold chains for the exclusive use of google

justapassenger 7y ago

> The risk is that it's hard to guess how much data you'll need to train an accurate, automated model. Maybe it's very large, and you can't keep doing it manually long enough to get there. Maybe it's very small and lots of companies will automate the same task and you won't have any advantage. I think you forgot most important option. It may not be a data problem. You may have all the data in the world and still not being able to solve the issue. Capabilities of ML are much more limited than hype makes people believe

solidasparagus 7y ago
I can't think of a company that wouldn't have valuable data. A core use of ML is reducing costs by making better spending decisions/reducing waste and that is relevant to almost every company I think.
- variaga 7y ago
  Recurring expenses (i.e. the kind that will generate enough data to adequately train a ML system) are already handled pretty well by most businesses through traditional (non-ML) methods. Operations Research has been a thing for 60+ years. Businesses that don't handle their recurring expenses well now are likely to have organizational issues (e.g. senior management that ignores the advice of their reports on how to do things) that will prevent them from doing so even if ML is added to the mix. The place where businesses get in real trouble (and hence would see significant value from better decision making) is when doing things that they have _not_ done 10K+ times before. Things like "Should we expand into $NEW_MARKET?". ML isn't going to help them with that, because there will not be any useful historical data to train them with.
  - solidasparagus 7y ago
    I completely disagree with the assumption that because something has been done for 60+ years it won't be improved by new, extremely relevant technologies. Amazon is very effective at both OR and organizational dynamics, but they found massive savings by using ML to predict demand and thus inventory/costs (Research is here - https://arxiv.org/pdf/1711.11053.pdf https://arxiv.org/pdf/1711.11053.pdf).

nradov 7y ago

$10B won't be nearly enough to build level 4+ autonomous vehicle control software that will work in a wide variety of roads and weather. Probably too low by an order of magnitude or two.

sgt101 7y ago

There's another risk - which is rare events. Basically most of the reason why these workflows are not automated already is that there are a host of "once a week per operative" corner cases that don't show in the data for a long time, and when they do show they don't show as anything but noise. But this is where the human intelligence is spent and this is why the jobs are so hard to automate.