Gen AI is magical, it makes stuff appear out of thin air!
And it's limited, everything it makes kinda looks the same
And it's forgetful, it doesn't remember what it just did
And it's dangerous! It can make things that never happened
Starting with theory might be the simplest way to explain, but it leaves out the hook. Why should they care?
It's also critical not to leave out the ethical topics (resource consumption, e-waste production, concerns about how the source data is harvested - both how it DDoS websites and how authors are not necessarily happy with their work ending in the models)
Brilliant's AI course has step-by-step interactive textgen LLMs trained on TS (Swift) lyrics and terms of services with quizzes for comprehension and gamified points.
Here's a quick take:
LLM AI are really good at generating bytes that are similar to other bytes, but aren't yet very good at caring whether what they've generated is wrong or incorrect. Reinforcement Learning is one way to help prevent that.
AI Agents are built on LLMs. An LLM (Large Language Model) is a trained graph of token transition probabilities (a "Neural Network" (NN), a learning computer (Terminator (1984))). LLMs are graphical models. Clean your room. The grass is green and the sky is blue. Clean it well
AI Agents fail where LLMs fail at "accuracy" due to hallucinations even given human-curates training data.
There are lots of new methods for AI Agents built on LLMs which build on "Chain of Thought"; basically feeding the output from the model back through as an input a bunch of times. ("feed-forward")
But if you've ever heard a microphone that's too close to a speaker, you're already familiar with runaway feedback loops that need intervention.
There are not as many new Agentic AIs built on logical reasoning and inference. There are not as many AI Agents built on the Scientific Method that we know to be crucial to safety and QA in engineering.
That said, the hands on approach here is great and also foundational in my experience.
Some of the takeaways feel over-reliant on implementation details that don’t capture intent. E.g. something like “the LLM is just trying to predict the next word” sort of has the explanatory power of “your computer works because it’s just using binary”—like yeah, sure, practically speaking yes—but that’s just the most efficient way to lay out their respective architectures and could conceivably be changed from under you.
I wonder if something like neural style transfer would work as a prelude. It helps me. Don’t know how you’d introduce it, but with NST you have two objectives—content loss and style loss—and can see pretty quickly visually how the model balances between the two and where it fails.
The bigger picture here is that people came up with a bunch of desirable properties they wanted to see and a way to automatically fulfill some of them some of the time by looking at lots of examples, and it’s why you get a text box that can write like Shakespeare but can’t tell you whether your grandma will be okay after she went to the hospital an hour ago.
Are there newer changes that are actually doing prediction of tokens out of order or such, or are this a case of immense internal model state tracking but still using it to drive the prediction of a next token, one at a time?
(Wrapped in a variety of tooling/prompts/meta-prompts to further shape what sorts of paragraphs are produced compared to ye olden days of the gpt3 chat completion api.)
What I'd add: the lesson about hallucinations should come early, not just in the RAG module. Kids (and adults) need to internalize "confident-sounding doesn't mean correct" before they get too comfortable. The gap between fluency and accuracy is the thing that trips everyone up.
Us millennials didn’t do so great with native access to the internet, zoomers got wrecked growing up with social media. . . It’s rough.
The simplest explanation I can give is we have a machine that you feed it some text from the internet, and you turn the crank. Most machines we've had previously would stop getting better at predicting the next word after a few thousand cranks. You can crank the crank on an LLM 10^20 times and it will still get smarter. It will get so smart so quickly that no human can fit in their mind all the complexity of what it's built inside of it except through indirect methods but we know that it's getting smarter through benchmarks and some reasonably simple proofs that it can simulate any electronic circuit. We only understand how it works by the smallest increment of its intelligence improvement and through induction understanding that that should lead to further improvements in its intelligence.
No, it won't.
Training/learning is separate from the execution of the model. Takes megadollars to train and kilodollars to effectively run.
Its basically a really complicated PID loop. You can test and 'learn' the 3 neuron function, and then you can put it into execution. Can't do both.
Sure, theres context length and fine tuning to slightly alter a model.
But theres no adaptive, self-growing LLM. Probably won't be for a long time.
They should learn to think for themselves about the whole picture, not learn about 'awesomeness'.
