I like to think about whether a problem is inherently antagonistic when deciding whether data/models/etc. can automate it well ("eat it" in the words of the article).
Not antagonistic -- predicting whether someone will get lung cancer...they're probably not going to falsify the data from their physical, etc.
Antagonistic -- detecting network intrusions or predicting phishing attacks. These folks fight back, i.e. disguise, adapt, hide
In the latter case, a human component to a decision task (detection, labeling, etc.) will always remain.
So then when it comes to investment decisions, let's say data plus supervised ML becomes the name of the game. Start-ups will learn the model features of a business that gets investment dollars. They'll design themselves specifically to those features and in a sense dupe models for a while until the models are adapted. In this way, the human component can't be left out.
Of course, it doesn't take a model to get duped. A lot of companies might hire a bunch of phds and data scientists right before being acquired, because each hire bumps up the value of the company, etc. These moves are taken to increase a valuation rather than actual value and that's antagonistic in much the same way hiding spam is.
Yes, startups try to game VCs already. But if the 'features' that your business needs to achieve for VC are
1) hard to achieve and
2) generally drive you towards success
then what's the issue with that? We want more startups to succeed and VC investment is an important feedback mechanism that ought to drive more startups towards success or early discovery of failure.
I mean let's say startups try to game the VC game and they learn they need to demonstrate product market fit before seeking investment. Is this a bad thing? We'll try to be as transparent about our insights as possible so startups know what we think drives/indicates their success the best.
I think you are right in that final decision is certainly not going to be reduced to a supervised ML algorithm classifying investment opportunities. But I don't think this fundamentally has to do anything with antagonistic and non-antagonistic - you just need different types of methods
Not antagonistic -- predicting whether someone will get lung cancer...they're probably not going to falsify the data from their physical, etc.
Antagonistic -- detecting network intrusions or predicting phishing attacks. These folks fight back, i.e. disguise, adapt, hide
In the latter case, a human component to a decision task (detection, labeling, etc.) will always remain.
So then when it comes to investment decisions, let's say data plus supervised ML becomes the name of the game. Start-ups will learn the model features of a business that gets investment dollars. They'll design themselves specifically to those features and in a sense dupe models for a while until the models are adapted. In this way, the human component can't be left out.
Of course, it doesn't take a model to get duped. A lot of companies might hire a bunch of phds and data scientists right before being acquired, because each hire bumps up the value of the company, etc. These moves are taken to increase a valuation rather than actual value and that's antagonistic in much the same way hiding spam is.