likely llc

We’re calling her, “minnie timmie,” since, apparently even with that fro she still cannot reach the microphone when speaking at a conference. Here’s a reference:

So one thing, that pisses me off about Reinforcement Learning based technologies, is that it reinforces the things that influence it. Without a reference, I think this mathematical method/model’s initial application was to balance planes. So, if you have turbulence, the planes stabilizers could naturally adjust themselves. I’m guessing probability initially a military application, as was monte carlo simulation in the Manhattan project, which actually employed a bunch of women. They were actually the first computers and excellent arithmeticians, which is why I don’t understand why they don’t understand why I don’t have feelings have any feelings, but anyway…

But yeah, I think Google made a switch to the algorithm that generates search results. They made the switch from Page-Rank, which initially made google famous, alongside development of the first major search engine when the internet came around. An LLM based method who’s mathematical backend is a reinforcement learning (RL) model. In short, RL based methods re-inforce things that make designated positive decisions (a positive “self learning” feedbackloop). I think this is beneficial in an aircraft, and may be some other applications, but for LLM’s, may not be the best, especially search engines.

Let’s think about it. Suppose we have a positive feedback loop (not to be mistaken with things that can have a negative impact on society. this just means there’s some reward that’s continued to be rewarded. I don’t know, environmental applications, etc). So think, the gen pop, who are overwhelmingly uneducated, and not experts, for example an MD Gyno (I like Dr. Michelle Louie) or Cardiologist. These people are rigorously trained in order to make critical medical decisions that can possibly be life or death situations. If we trained a positive feedback loop based on these experts with historical (empirical validation) based on excellent case studies and experience, we might generate something useful. Instead, a lot of users of search engines and forums are people asking questions on reddit with no professional training. So, in short, I’ve noticed the search results after popularization of LLMs, on Google, are fucking nonsense. I’ve moved back to reliable resources (books/blogs by experts). And even Prof Andy Gelman, and Carpenter, a dynamic duo, recommended going back to the oral tradition (an African descent tradition) and blogs, who are small disconnected communities of people with a fascination and dedication to learning about expert topics.

But yeah, so we’re training LLMs on bullshit that generates more bullshit. In consequence, we get search engine bullshit. This is why I was writing about the blogosphere.

And, Minnie Timmy, is obsessed with data centers for some reason. I am not very moral, but someone’s gotta blow the whistle some times.

I am not an electrial engineer, but Timmy is. Computing is necessary in this world, now. But, given a group of expert electrical engineers and environmental scientists, and computer scientists with model validation skills, we could put together a more efficient data center model, instead of wasting resources. When the immigrants from Europe discovered the midwest, they had a theory that the trees would never end, and again, I am bad with names, and the search engines have turned to trash, so I cannot find the name, but it was in the history books in high school. But they were wrong. The trees are gone.

But yeah, group of experts make decisions as opposed to politicians that have no technical skillz and are only concerned with their political ideology.

Major Disagreements

Somewhere in this lecture, Minnie Timmy is quoted in saying, the continent in Africa has the best dance parties. This is an objectively incorrect statement.

Gebru says in the video below, not an exact quote, but that, countries in Africa have the best dance parties.

I thoroughly disagree. When traveling through Colombia I had the great pleasure, in Medellin (Escobar is dead, don’t worry, it’s mostly fine besides the basic tourist petty crime you find all over the world). I had the great pleasure of dancing at a salsa club with a professional dancer from Barranquilla, with a local, a Venezuelan guy, and then another American who came back from Carnival, in Barranquilla, that is a huge tourist destination. She let us trade off to practice and showed us some different steps, until sunrise. I wish I remembered her name. ;). Hey Shakira ;). Until sunrise.

I can’t find the quote, but here’s the lecture.

Finally

I want to talk about hierarchical models to more accurately estimate social science or demographics effects which reminded me of the character, Timmy, from south park, which lead me to calling her Minnie Timmy.

Hierarchical models have been studied largely and can reduce confirmation bias and add more insight and predict uncertainty better, and more accurately deduce causal inference which I think that community should learn more, but I don’t feel like writing anymore.

The go-to reference for me, was by Andrew Gelman and Lauren Hill, which uses lme4 in R, but Bayesian hierarchical models can be better applied using Stan. Here’s the books website, and you can find a copy the PDF but for legal reasons you need to contact the publisher

https://sites.stat.columbia.edu/gelman/arm

Likely LLC.

And then here’s Shakira, a dancer from Barranquilla, Colombia, throwing it down, objectively better dancing from the Latin American community, than any country in Africa, and we can accurately estimate this with a Bayesian hierarchical model: