Kernel Smoothing, Cargo Cults, & ChatGPT: Cosma Shalizi Takes on a Near-Impossible Teaching Task
This Seems to Me to Be Taking "Volunteer to Teach What You Want to Learn" to an Extreme: MAMLM Edition. Behind the paywall because I do not know what I think, and at the end I think all this piece demonstrates is my own great confusion. I bounce back and forth between the false promises & the real achievements of GPT LLMs, an so I go from Predict-o-Matic to Pagliacci, and back again…
Cosma Shalizi Volunteers to Explain the Unexplainable: Teaching LLMs Without a Net
Yes, Cosma Shalizi has assigned himself the task of, this fall, explaining how LLM GPTs really work. And he has assigned himself of explaining what they do. And he has done this in our current context, in which we really do not even know what they do. And he has committed to do so with neither gurus nor oracles to aid him. His only intellectual weapons are math, stat, humility, and a combination of wonder and confusion verging on despair.
Meanwhile, attempts to use GPT LLMs as real-world tools continue to reveal the gap between fluent mimicry and genuine understanding. True Believers crusade forward. They have armed themselvess with faith, hope, and enough NVIDIA GPUs that they can make this Clever Hans appear to actually do the math correctly, and with understanding.
So what happens next?
What happens when engineering triumphs outstrip epistemic foundations? Behind the magic of GPT LLMs lie the uncomfortable reality of the shoggothim, in which simple methods, scaled beyond comprehension, yield astonishing yet ungrounded results.
How little even the most knowledgeable of us truly know about what we have built! And do those more knowledgeable about the details and implementation know more, or less?
Pagliacci weeps, but do we weep or laugh?
Seeking a Guru, Finding None: Shalizi’s Forthcoming Course “Statistical Principles of Generative AI”
What has happened is that Cosma Shalizi has emailed “I am going to have to come up with an explanation of what the **** is going on, without the benefit of a guru…”
“What the **** is going on” is this: we now have GPT LLM (General-Purpose Transformer Large Language Model) form of MAMLMs (Modern Advanced Machine-Learning Models).
This is apropos of:
Cosma Shalizi: Statistical Principles of Generative AI <http://bactra.org/weblog/statsgen-f25.html>: ‘I should know better than to volunteer to do a new prep --- but I don't:
Special Topics in Statistics: Statistical Principles of Generative AI (36-473/36-673)
Description: Generative artificial intelligence… statistical models of text and images…. very new, but they rely on well-established ideas about modeling and inference…. Introduce students to the statistical underpinnings… emphasizing high-level principles over implementation details. It will also examine… the "artificial general intelligence" versus "cultural technology" debate, in light of those statistical foundations….
Expectations: All students can expect to do math, a lot of reading (not just skimming of AI-generated summaries) and writing, and do some small, desktop-or-laptop-scale programming exercises….
Topical outline (tentative): Data compression and generative modeling; probability, likelihood, perplexity, and information. LLMS are high-order parametric Markov models fit by maximum likelihood…. Estimating parametric Markov models…. Influence function…. Back-propagation…. Stochastic gradient descent…. Estimation and dynamics…. Prompting as conditioning…. Transformers; embedding discrete symbols into continuous vector spaces. Identification issues…. "Attention", a.k.a. kernel smoothing. State-space modeling. Generative diffusion… as a stochastic (Markov) process; a small amount of stochastic calculus. Learning to undo diffusion. Mixture models. Generative diffusion models vs. kernel density estimation. Information-theoretic methods for diffusion density estimation. Combining models of text and images. Prompting as conditioning, again.
All of this, especially the topical outline, is subject to revision as we get closer to the semester actually starting…