RA_Fisher

“We’re holding our own” - Captain McSorley., 7:10pm, 5 minutes before contact was reported lost. Seemingly more a story of patriarchal arrogance than force of nature.

For statistical AI models, we can use out of sample prediction error as an objective measure to compare models. What makes evaluating LLMs difficult is that comparisons are inextricable from utility (whereas statistical AI models do have a pre-utility step wherein it can be shown out of sample prediction epsilon is minimized).

Or, maybe you don’t understand why it’s rational?

Yes, for sure! And being able to orchestrate AI to use that knowledge provides leverage for fulfilling tasks.

Eventually, yes, I think we'll delegate to AI in more and more complete ways, but it's a process that takes some time.

We only need to do that when it’s practical for the task at hand. Some tasks are life-and-death, but many have much lower stakes.

By moving up a level in the abstraction layer similar to moving from Assembly to C++ to Python (to LLM). There’s speed in delegation (and checking as beneficial).

I chose it long ago and it was a tribute to the man. Then I learned about support for eugenics, and how he shilled for big tobacco by repeatedly insisting a causal relation between smoking and cancer was unidentifiable. He did certainly have some advanced statistical ideas. These days I keep it more as a tribute to the field of statistics. Fisher is worth learning— with context.

Politics-wise, I benefit from learning from Fisher (but obviously the reverse isn’t true). So it’s reasonable to say my views are more informed.

RTO seems like a bound-to-fail last gasp to enforce the patriarchy at work (and things that flow from it, sexism, racism, ableism, etc.).

This claim's so big that it requires theoretical proof, empirical analysis isn't convincing (given the size of the claim). Causal inference experts have long known that many inputs map to outputs (that's why identification of the inputs that actually caused a given output is a never-ending task).

For sure! Measuring parameters given data is central to statistics. It’s a way to concentrate information for practical use. Sufficient statistics are very interesting, bc once computed, they provably contain as much information as the data (lossless). Love statistics, it’s so cool!

Yes, ChatGPT can do it using Python today (the statsmodels library). I use it all the time (I’m a statistician).

LLMs can use Python to simulate from probability distributions. Though, admittedly they have to code and use their own MCMC samplers (and can’t yet utilize Stan and PyMC directly).

> are we going to include the water used to grow the almonds you ate?

That’s a good idea. Like nutrition facts but for everyday economic climate decisions.

Competition is good and makes us all better off, don’t be afraid.

Bc open source models aren’t as good and it’d cost a lot more than $100k to self-host given the same load.

That’s what people said about SaaS. People pay good money for things they could theoretically do themselves with more time and money.

Also, according to LLM Arena, all the best models are proprietary (bc trade secrets can be very powerful).

> too far off the data distribution.

I guess his prompts couldn’t provide sufficient information either (there’s no limit). Sounds more like a user issue to me. :) I don’t think there’s anyone that can type faster than ChatGPT.

It’s a giant leap of technology? Investing in that is the perfect usage of debt.

I'm a scientist and I'd never work that way. I'm methodical, because I've learned it's the fastest and highest-ROI approach.

Guessing without understanding is extremely unlikely to produce the best results in a repeatable manner. It's surprising to me when companies don't know that. For that reason, I generally want to work with experts that understand what they're doing (otherwise is probably a waste of time).

Why can’t they release it and let people decide for themselves?

True, aside from the exceptional employers that’ll actually provide a good career.

I like simplicity and I love a good base R idiom, but there's a lot less consistency in base R compared to the tidyverse (and that comes with a productivity penalty).

Lattice is really low-level. It's like doing vis with matplotlib (requires a lot of time and hair-pulling). Higher level interfaces boost productivity.

Why? The tidyverse is so readable, elegant, compositional, functional and declarative. It allows me to produce a lot more and higher quality than I could without it. ggplot2 is the best visualization software hands down, and dplyr leverages Unix’s famous point free programming style (that reduces the surface area for errors).

Doesn’t the term have contractual obligations for OpenAI and Microsoft?

He uses it instead of Google search.

My dentist said he uses it near constantly, and so do I.

Also, when asked the pilot said they didn’t switch the fuel control off.

That seems like an important clue, bc it points at a possible glitch that could affect other aircraft.

Another horrible cost of militarism.

Location: Michigan, US

Remote: Yes

Willing to relocate: No

Skills: extremely advanced Bayesian statistics / modeling and econometrics, causal inference (experimental and cases where exp are too costly)

Resume / CV: 9th employee and founding data scientist at Zapier (2014-2023), Master’s statistician, B.S. economist / econometrician

Email: cpeter9@gmail.com

I wonder how they’ll get around the model producing “average,” (mediocre) outputs? Perhaps quantity?