2025-06-03 13:11:00
I've been a ludite on the subject of applied AI, such as LLM (large language models). Like cryptocoin and blockchain stuff, I've been avoiding the subject as I've been sceptical as to the actual value of the applied products.
Now I'm finally coming around and I've decided I need to learn how do these things work? Not as in "how can I use them?", but "what makes them tick?".
There's a few maths and programming YouTubers whose work I really appreciate, insofar that they've clarified for me how LLM work internally.
If you have no more than one hour, watch Grant Sanderson's talk where he visualizes the internal workings of LLM in general.
Then if you have days and days at your disposal:
I've learned a lot so far! I'm happily reassured that I haven't been lieing to students about the capabilities and impossibilities involving general purpose LLM and IT work.
3Blue1Brown's series on neural networks, starting with video 1 here, helped me really understand the underlying functions and concepts that show how "AI" currently works. The video series goes over the classical "recognition of handwritten numbers" example and explains in a series of videos what the neurons in a net do and how. And more importantly it clarifies why, to us humans and even the AI's creators, it's completely invisible WHY or HOW an AI comes to certain decissions. It's not transparent.
What got my rabbit holing started? A recent Computerphile video called "The forbidden AI technique". Chana Messinger goes over research by OpenAI that talks about their new LLM "deciding", or "obfuscating", or "lieing" and "getting penalized or rewarded" which had me completely confused.
Up until this week I saw LLM as purely statistical engines, looking for "the next most logical word to say". Which they indeed kind of are. But the anthropomorphization of the LLM is what got me so confused! What did they mean by reward or punishment? In retrospect, literally a trainer telling the LLM "this result was good, this result was bad". And what constitutes "lieing" or "obfuscation"? The LLM adjusting its weights, bias and parameters during training, so certain chains of words would no longer be given as output.
It's like Hagrid's "Shouldn't've said that, I should not have said that..." realization.
Now, to learn a lot more!
kilala.nl tags: work, studies,
View or add comments (curr. 0)
All content, with exception of "borrowed" blogpost images, or unless otherwise indicated, is copyright of Tess Sluijter. The character Kilala the cat-demon is copyright of Rumiko Takahashi and used here without permission.