Profile avatar
gregoryfaletto.com
Statistician, Data Scientist.
252 posts 270 followers 602 following
Active Commenter
comment in response to post
Yes and unit economics will start to become more of the ballgame than they already are, especially if "reliably replacing a human end-to-end" is looking unachievable for the most part
comment in response to post
Yeah reasoning is kind of the biggest lift in the last couple of years I guess. I guess we'll see as really huge models get trained, maybe scaling laws will keep up... maybe they won't
comment in response to post
Yeah we'll see what happens, maybe a year from now we'll look back on current capabilities and feel like they're way worse. But when I look back at say last fall, the improvements feel noticeably slower than the improvements in the 9 months before that.
comment in response to post
(People keep saying "you can code things without even having to know how to code!" and that keeps... not really being true. Can code in languages that you don't really know if you're already a coder, sure.)
comment in response to post
One thing I've been thinking about: let's say the transformer architecture kind of plateaus at a point where they're v useful if you have pro-level knowledge in your field (coding etc.), but can't reliably replace a human unsupervised on end-to-end tasks. Could be a pretty nice sweet spot for labor
comment in response to post
Which model?
comment in response to post
Let's at least all be clear about who we're arguing about deserving more sympathy.
comment in response to post
I wish we'd acknowledge more that the biggest "losers" in the 2022 - 2024 economy was about the 70th - 90th percentile of income, who experienced rapid wage growth in the bottom quartile of earners as price increases. (chart is unfairly negative IMO since includes 2020, but couldn't find better one)
comment in response to post
Citizens have a civic duty to vote for whichever of the major-party candidates they think would be a better leader. Not doing so is a moral failing, and it's good for the country to discourage it. If the candidate you dislike less doesn't win, then the candidate you dislike more will.
comment in response to post
Vanilla Least Squares (VLS)
comment in response to post
Pretty unlikely it happened today
comment in response to post
It is. Try it yourself right now
comment in response to post
And I didn't bring up the topic, but I also think it's kind of not okay for news outlets to prioritize this topic (which it really feels like we have already gotten to the bottom of) over harm being done in the present that elected officials have the power to stop, but don't.
comment in response to post
I just disagree that the criticism is "proportional" in any reasonable sense. I guess my starting point for thinking about proportions is units like lives lost that can be directly tied to a specific policy choice, number of people wrongfully detained, number of laws broken, etc.
comment in response to post
Biden's age issues were the main news topic for almost a full month after the June 27th debate. There were certainly many other days where it was the leading story. In your view, how much additional criticism is needed to be proportional?
comment in response to post
The "why" is potentially more complicated depending on how deep you want to get into "why." (Looking at the system prompt would probably help.) But "what" is going on is perfectly clear, and IMO the less we lose sight of that (and the more the general public understands that) the better
comment in response to post
You push specific dough through this special machine, you get noodles. You push specific strings through those matrices, you get some text. Both machines can be quite useful for certain purposes
comment in response to post
Big numbers are getting pushed through a bunch of very particular matrix multiplications
comment in response to post
realizing I'm way behind for the day, making myself 9 full bowls of plain spaghetti, then interlacing my fingers and cracking my knuckles as I prepare to get started
comment in response to post
A personal assistant who is 80% as smart as you and can get stuff done while you get a glass of water is a massively useful tool
comment in response to post
Not that specific one, but some of these posts seem photoshopped, or at the very least not reproducible. When I see these posts sometimes I'll try the exact query on the exact model and get back a sensible response
comment in response to post
Let he who has never typed a function from memory and then realized they misremembered the exact name cast the first stone
comment in response to post
something I think about all the time is how many people there are whose only voluntary interaction with LLMs was trying whatever free OpenAI model was available months/years ago and getting frustrated
comment in response to post
That's very helpful, thank you!
comment in response to post
It's very important that we either forbid developers from making bad investments that would lose them money or forbid them from investments that would make them too much money, depending on which one of those would result in housing getting built
comment in response to post
☹️
comment in response to post
I can invert a 3 x 3 matrix by hand, doesn't mean it makes sense for me to actually do it!
comment in response to post
middle manager is one of the most underrated jobs. it is hard, people who are really good at it are rare, if you could snap your fingers and make every middle manager twice as good the world would be noticeably better.
comment in response to post
Yes, and precise positioning of text, images etc. using the "arrange" tab on the right side of the screen
comment in response to post
Coriolis hive activated
comment in response to post
To me it's like, sure calculators shouldn't be allowed on all tests, but it would also be insane to insist not only that kids should never learn how to use calculators, it's in fact shameful if they ever did
comment in response to post
It's kind of fascinating to me that some people think this is obviously cheating 100% of the time and there's never a place for it. These kids are going to graduate and the ones who didn't learn how to "cheat" like this are going to have employers who are pissed off at them for it.
comment in response to post
if you have a pressure cooker, this is really good. i substituted tempeh for mushrooms cooking.nytimes.com/recipes/1021...
comment in response to post
I think the difference to the argument is that instead of this term disappearing asymptotically because it's mean 0, it just wouldn't appear at all
comment in response to post
Indeed, and I believe the same argument in the post works for ANCOVA I too
comment in response to post
And how much will the resulting estimator look like vanilla ANCOVA II? It turns out they're asymptotically equivalent--makes sense since they have the same asymptotic properties. In this blog post, I walk through the math & supporting simulations. Check it out! gregoryfaletto.com/2025/05/03/e...
comment in response to post
So here's a question: what if we use ANCOVA II as the outcome model in a cross-fit AIPW estimator? Since treatment assignment is randomized, the propensity scores are known--errors are already o(1/√n). So even if ANCOVA II is misspecified, we should get consistency and asymptotic normality.
comment in response to post
If we use the AIPW estimator with cross-fitting, Theorem 5.1 of Chernozhukov et al (2018) guarantees consistent and asymptotically normal estimation of the ATE under unconfoundedness and overlap if the product of the rates of convergence of the propensity and outcome models are o(1/√n).
comment in response to post
Another setting where we can get consistent and asymptotically normal estimation of the ATE with a misspecified model is the doubly-robust AIPW estimator.
comment in response to post
In randomized experiments, the OLS coefficient on the treatment dummy is a consistent and asymptotically normal estimator for the average treatment effect (ATE). This holds regardless of whether the linearity assumption is true.
comment in response to post
The ANCOVA II model is a linear regression outcome model with covariates, a treatment dummy, and interactions between them. (Here W_i is the treatment dummy.) (From Imbens and Rubin 2015.)
comment in response to post
Maybe for the slice of situations that require knowledge of what to do but not really skills or special expensive tools, plumbers get called less and people take care of it themselves more
comment in response to post
Not even sure I agree it's NBD for plumbers. Could be a substitute for plumbers in the sense that it allows anyone to take some photos and describe a situation and get specific-to-your-situation instructions on how to fix it.
comment in response to post
It's not okay, and this is basically the reason. If you want to be able to peek at the results, then increase the sample size, and repeat, you have to use anytime-valid inference techniques e.g. arxiv.org/abs/2210.01948
comment in response to post
It's a tough competition
comment in response to post
People are too loosey-goosey about it though. Just because there's a word for it doesn't mean it's okay!
comment in response to post
I assume the thought process to invent it was "hmm I want to be emphatic but there's nothing in front of me to bang on and I don't want to point or wag my finger (too lecture-y) or make a gesture that looks violent. Guess I'll just have to do something human has ever done before (to be relatable)"
comment in response to post
(emphasizing important syllables with the weird politician thumb-on-top closed fist) The tariffs are the problem. And that's why we're quite pleased with them