5 ms·

I also work in the field and while the work itself keeps me obsessed, similar to you, I keep a list of companies doing amazing things beyond "better ad personal

I also work in the field and while the work itself keeps me obsessed, similar to you, I keep a list of companies doing amazing things beyond "better ad personalization" so that I don't get a bit jaded. Some that might be interesting to you: PostEra - A medicinal chemistry platform that accelerates the compound development process in pharmaceuticals. They're running an open source moonshot for developing a compound for treating COVID, with tons of chemists around the world participating. Thorn - They use ML to identify child abuse content, both to help platforms filter it and to help law enforcement save children/arrest abusers. All of the healthcare startups (too many to list all), including Benevolent, Grail, Freenome, and more. Wildlife Protection Services - use ML to detect poachers in nature preserves around the world, and have already significantly increased the capture rate by rangers.

pdimitar 6y ago
Is there any open data that clearly demonstrates those companies have made a practical difference? Have their NNs been used at scale already?
- calebkaiser 6y ago
  Yep, if you Google Thorn you'll find a bunch of material on how law enforcement uses them and related statistics. Several of the other companies (I admittedly haven't done a deep dive on all of them) have similarly accessible material.
hooande 6y ago
my issue is that none of these things represent new capacities that humans didn't have before. a spreadsheet will help you to identify and track child abuse content. you could do that using paper if you committed enough resources My hope for machine learning was that it would allow people to see patterns that could not be seen before. While this does happen, most practical problems are driven by a few very obvious indicators. ML can identify them with a high consistency and low cost. This is a useful tool (like a hammer) much more than a super power (like iron man's suit)
- calebkaiser 6y ago
  I don't know about that. Before founding PostEra, its founders published research about their model, which significantly outperforms human chemists (if you're interested in ML, it's actually a fascinating use of an architecture commonly used in language tasks): https://www.chemistryworld.com/news/language-based-softwares-accurate-predictions-translate-to-benefits-for-chemists/4010437.article https://www.chemistryworld.com/news/language-based-softwares... Thorn's flagship tool, Spotlight, uses NLP to analyze huge volumes of advertisements on escort sites and flag profiles that have a higher risk of representing trafficking victims. You would need an enormous spreadsheet and near infinite supply of dedicated humans to manually review and refine some sort of statistical model for scoring ads, as the volume of advertisements produced is insane. The same for the deep genomics companies. The size of data generated by deep sequencing is beyond a person's ability to pattern match, and the patterns are potentially complex enough that they may never be noticed by human eyes. And, again, this is just a small list of startups in particularly moonshot-y spaces.
- philipkglass 6y ago
  Numerical weather forecasting enables genuinely new capabilities. Where is a hurricane going to make landfall a few days from now? In principle, all the arithmetic operations going in to the final forecast could be performed by hand. But then the "forecast" would be completed millennia after the hurricane arrived. The only way to get foreknowledge of weather from the model is to do the arithmetic at inhuman speeds. This isn't even AI by most people's intuitive notions of AI. AI is colloquially an artificial approach to intellectual tasks traditionally performed well by humans. Since humans were not very good at predicting hurricane tracks in the first place, the increasing capabilities of models to predict weather probably doesn't have as much wow factor. It's a new capability without any human-genius antecedents.

pdimitar 6y ago

Is there any open data that clearly demonstrates those companies have made a practical difference? Have their NNs been used at scale already?

calebkaiser 6y ago
Yep, if you Google Thorn you'll find a bunch of material on how law enforcement uses them and related statistics. Several of the other companies (I admittedly haven't done a deep dive on all of them) have similarly accessible material.

hooande 6y ago

my issue is that none of these things represent new capacities that humans didn't have before. a spreadsheet will help you to identify and track child abuse content. you could do that using paper if you committed enough resources My hope for machine learning was that it would allow people to see patterns that could not be seen before. While this does happen, most practical problems are driven by a few very obvious indicators. ML can identify them with a high consistency and low cost. This is a useful tool (like a hammer) much more than a super power (like iron man's suit)

calebkaiser 6y ago
I don't know about that. Before founding PostEra, its founders published research about their model, which significantly outperforms human chemists (if you're interested in ML, it's actually a fascinating use of an architecture commonly used in language tasks): https://www.chemistryworld.com/news/language-based-softwares-accurate-predictions-translate-to-benefits-for-chemists/4010437.article https://www.chemistryworld.com/news/language-based-softwares... Thorn's flagship tool, Spotlight, uses NLP to analyze huge volumes of advertisements on escort sites and flag profiles that have a higher risk of representing trafficking victims. You would need an enormous spreadsheet and near infinite supply of dedicated humans to manually review and refine some sort of statistical model for scoring ads, as the volume of advertisements produced is insane. The same for the deep genomics companies. The size of data generated by deep sequencing is beyond a person's ability to pattern match, and the patterns are potentially complex enough that they may never be noticed by human eyes. And, again, this is just a small list of startups in particularly moonshot-y spaces.
philipkglass 6y ago
Numerical weather forecasting enables genuinely new capabilities. Where is a hurricane going to make landfall a few days from now? In principle, all the arithmetic operations going in to the final forecast could be performed by hand. But then the "forecast" would be completed millennia after the hurricane arrived. The only way to get foreknowledge of weather from the model is to do the arithmetic at inhuman speeds. This isn't even AI by most people's intuitive notions of AI. AI is colloquially an artificial approach to intellectual tasks traditionally performed well by humans. Since humans were not very good at predicting hurricane tracks in the first place, the increasing capabilities of models to predict weather probably doesn't have as much wow factor. It's a new capability without any human-genius antecedents.