3 ms·

I'm extremely skeptical that this works a fraction as well as it claims. Almost all of the information it provides is astrology-like well-meaning doublespeak th

by easong 11y ago

I'm extremely skeptical that this works a fraction as well as it claims. Almost all of the information it provides is astrology-like well-meaning doublespeak that applies to everyone at one point or another. There's even a name for it - http://en.wikipedia.org/wiki/Barnum_effect http://en.wikipedia.org/wiki/Barnum_effect. That probably won't stop businesses from wasting truckloads of money on their own automated soothsayer, though. Edit: Make sure you consult their physical terminal before emailing that big VC! http://imgur.com/UMIlcXz http://imgur.com/UMIlcXz

dchichkov 11y ago
Yes... To make a generative model that can outperform humans in a social setting? I think it is technically feasible to do basic sentiment analysis, but the generative model... sorry, I think that NLP tech is just not there yet. On the other hand, I think it is a good idea to have a sentiment analysis system in place that advices users on overly negative sentiment in their comments or e-mails. I wish there was one already integrated into YC comments and into Gmail. Potentially can make life quite a bit nicer actually. P.S. like your wording 'automated soothsayer'. Funny that the next logical step would be wasting money on 'automated soothslayers' ;)
- easong 11y ago
  Sentiment analysis to avoid overt negativity or rudeness is a great idea. It might prevent people like me from commenting too often, and I could see some perennially-depressed Reddit/OkCupid users getting a useful heads up from it as well.
SilasX 11y ago
But at least there's a way to test it: - Give some subjects the advice from this program, and some subjects advice from a knowledgeable co-worker of the person in question (PiQ). (Make sure they don't know which the advice is coming from.) - Find some way to measure the "success" of the subjects in question -- maybe have them and the PiQ give a subjective rating of the interactions, which should be good enough. - See if there's a significant difference between the two. If all such advice is unhelpful, or this service is no better than astrology, you'll see no difference.
- easong 11y ago
  An even easier test - give subjects the advice for a different person than the one they're writing to. If there's a similar success rate, then the personality advice is meaningless. Crystal does have a couple cool things going for it, though - for instance, they're in an excellent position to improve user writing quality by presenting standard stylistic advice (don't use passive voice, don't re-use words, etc) as advanced AI clairvoyance.
  - analog31 11y ago
    This is similar to a criterion that I once proposed when a personality test was given to the employees at my workplace: Would I recognize myself in the report if I found it on the mailroom floor without my name on it? You could take the reports, remove the names, and let people try to guess which person belonged to which report. To make an analogy, imagine finding a room full of un-labeled bottles of chemicals, and an inventory of chemical names. Could you match the chemicals to their names?
  - thaumasiotes 11y ago
    > give subjects the advice for a different person than the one they're writing to. If there's a similar success rate, then the personality advice is meaningless Not quite. This would mean that the personality advice is unrelated to the personality of the target. But it could still be good advice! It's quite possible that a machine that always gave you the same advice, no matter who you were writing to, would still improve your results. (What if the advice was "Are you drunk? Are you sure you want to send this?")

by easong 11y ago

dchichkov 11y ago

Yes... To make a generative model that can outperform humans in a social setting? I think it is technically feasible to do basic sentiment analysis, but the generative model... sorry, I think that NLP tech is just not there yet. On the other hand, I think it is a good idea to have a sentiment analysis system in place that advices users on overly negative sentiment in their comments or e-mails. I wish there was one already integrated into YC comments and into Gmail. Potentially can make life quite a bit nicer actually. P.S. like your wording 'automated soothsayer'. Funny that the next logical step would be wasting money on 'automated soothslayers' ;)

easong 11y ago
Sentiment analysis to avoid overt negativity or rudeness is a great idea. It might prevent people like me from commenting too often, and I could see some perennially-depressed Reddit/OkCupid users getting a useful heads up from it as well.

SilasX 11y ago

But at least there's a way to test it: - Give some subjects the advice from this program, and some subjects advice from a knowledgeable co-worker of the person in question (PiQ). (Make sure they don't know which the advice is coming from.) - Find some way to measure the "success" of the subjects in question -- maybe have them and the PiQ give a subjective rating of the interactions, which should be good enough. - See if there's a significant difference between the two. If all such advice is unhelpful, or this service is no better than astrology, you'll see no difference.

easong 11y ago
An even easier test - give subjects the advice for a different person than the one they're writing to. If there's a similar success rate, then the personality advice is meaningless. Crystal does have a couple cool things going for it, though - for instance, they're in an excellent position to improve user writing quality by presenting standard stylistic advice (don't use passive voice, don't re-use words, etc) as advanced AI clairvoyance.
- analog31 11y ago
  This is similar to a criterion that I once proposed when a personality test was given to the employees at my workplace: Would I recognize myself in the report if I found it on the mailroom floor without my name on it? You could take the reports, remove the names, and let people try to guess which person belonged to which report. To make an analogy, imagine finding a room full of un-labeled bottles of chemicals, and an inventory of chemical names. Could you match the chemicals to their names?
- thaumasiotes 11y ago
  > give subjects the advice for a different person than the one they're writing to. If there's a similar success rate, then the personality advice is meaningless Not quite. This would mean that the personality advice is unrelated to the personality of the target. But it could still be good advice! It's quite possible that a machine that always gave you the same advice, no matter who you were writing to, would still improve your results. (What if the advice was "Are you drunk? Are you sure you want to send this?")