CEO of Conductrics. Inference is not a Technology.
Joined November 2010
- Tweets14.3K
- Following468
- Followers1.8K
- Likes4.3K
Hello @AmericanAir I think your website might not give correct info for Porter Air codeshare flights. I wound up with a Basic Econ Ticket from AUS to YYZ? But pretty sure only Main, Main Extra, and Premium are offered. I couldn't find any example dates where a Basic is offered.
Maybe someone at @AmericanAir can 'buy' a ticket from AUS to YYZ on the Porter air codeshare and see that the fair rules on the AA website don't match the rules at Porter (i.e no free checked bags, and no carry on bag allowed. I bought this test ticket - Porter rules don't match
Amusing podcast where @ezraklein plays Larry Gopnik to Mark Epstein's Nachtner
open.spotify.com/episode/1YW…
Good fun at #measurecamptoronto talking about Privacy by design and how applying data minimization gets you both K-anonymization AND ability to efficiently construct Gramian matrices for OLS.
Pure bias reinforcing dopamine right there. @mgershoff summarizes #PrivacybyDesign at #measurecampto
Off to London for @MeasureCampLDN
We have a new batch of Conductrics INC Select whiskey from @MilamandGreene over here from Texas and alfajores direct from Buenos Aires as prizes and give aways.
For talks I will do something on Privacy Engineering and AB Testing.
tl;dl Privacy by design & customer first philosophy, if not two sides of the same coin, are consistent with each other.
Collecting customer data is akin to asking for something. Customer first means asking from customers only when its a GOOD FAITH attempt to enrich THEIR lives.
"Collect everything" is easy to say. It can even be rationalized. Think a little deeper, though, and the dangers of this mindset are pretty easy to surface. That's the topic of our latest episode with @mgershoff! analyticshour.io/2024/09/03/…
I know people and I still haven't been able to get a bottle from the Wildlife collection. Gotta up my click submit game.
@MilamandGreene
prweb.com/releases/beep-beep…
Great. As an extension to her note about effects heterogeneity. I think it is more that contextual bandits/Targeting is unlikely to be valuable unless there is a theory why there should be heterogeneity. If you don't then just run a simple RCT or Epsilon-First bandit.
These recent slides from Susan Athey and Guido Imbens at NBER are a great recent review of the most valuable data science methods I'm aware of. They cover tons of ground with lots of pointers.
conference.nber.org/confer/2…
Do you mind if I DM you here on X @thegautamkamath ? If not no worries. Was looking for some links/follow up around a quick question on the dual view of the Diff Privacy and Pearson-Neyman Hypth Testing.
Oh! I have to read. I have been thinking informally about this for Tech/Data Science - that methodological preferences etc. are drawn from individuals who are clustered (in part by being socially connected in the valley via a small set of interconnected VCs)
1. osf.io/preprints/socarxiv/2a…
Meta-analyses combine trials that are presumably *independent*. But social relatedness between researchers --> more similar effect estimates, suggesting dependency via same "school of thought"
I had the honor to chat with Debra J. Farber (@privacyguru) a few weeks ago on the Shifting Privacy Left podcast.
We talked about #ABtesting, Privacy by Design, and acting intentionally.
You can listen here
shiftingprivacyleft.buzzspro…
I had assumed that @AmericanAir rewards work like other programs with a default option if you don't select before new year. But I just discovered for AA there is no default and you just lose the rewards. So they needlessly alienate their loyal customers (I'm 6 yr Ex Plat)🤣
Thanks @NandoDF ! BTW this 2009 talk covers much of what @ilyasut discusses in a bit of detail but uses GZIP as the compressor as the example algo rather than a NN/LLM
wiki.santafe.edu/images/9/9a…
Predicting the next word "only" is sufficient for language models to learn a large body of knowledge that enables then to code, answer questions, understand many topics, chat, and so on.
This is clear to many researchers now, and there are nice tutorials on why this works by @ilyasut resorting to compression ( youtube.com/watch?v=AKMuA_TV… ) and by @geoffreyhinton ( youtube.com/watch?v=iHCeAotH… ).
However, the emergence of types of understanding is not unique to language models. In arxiv.org/pdf/1804.06318.pdf by @notmisha and @brandondamos the authors trained models to predict the next few time stems of over a hundred robot hand sensors (Touch, Gyro, Accelerometer, Joint Info, Actuator Info, etc.). They ten found out that they could regress the shape of the thing the hand was touching from the activations of the neural networks using probes. That is, the model developed an internal representation of shapes even though it was simply used to predict "only" the next few senses. Awareness follows from simple predictions and interaction with the world.
I think it might be helpful if the first few sections, or first few minutes of each section, was dedicated to the history/context. So a little narrative including folks like Hume, Popper, Laplace (Bayes), Neyman/Pearson, maybe Stein etc. just to provide context. @PhDemetri
Post at @Conductrics discussing a some Privacy Engineering ideas in the context of AB Testing.
tl;dr
Many standard stats can be calculated on aggregated data without need to link to individuals. Useful when cardinality of feature set << Individuals.
blog.conductrics.com/privacy…
