Angafirith
Ars TechnicaA jargon-free explanation of how AI large language models work
Want to really understand large language models? Here’s a gentle primer.
Becca 🏳️⚧️ 🇺🇦 🇸🇩 🇵🇸@arstechnica it's autocorrect, but more trained.
It doesn't "know", it doesn't "understand", it doesn't "think"..
It's just using probability (see statistics) to make sentences up (like autocorrect when it suggests other words as you type).
It doesn't "know", it doesn't "understand", it doesn't "think"..
It's just using probability (see statistics) to make sentences up (like autocorrect when it suggests other words as you type).
cetan@rebeccafinn @arstechnica several times the article says that researchers don't understand how the LLM "does" something. I find that very strange. How could they not know what it is doing? Is this just the author not differentiating between outside researchers that don't have access to proprietary information and internal researchers working to build the LLM?
Urethramancer🐀@cetan @rebeccafinn @arstechnica It's mathemagics created by neural networks, probably. They are known to create code which baffles humans, but somehow works :)
@Urethramancer @cetan @arstechnica this is a good video from 5 years ago... It explains this.
https://youtu.be/R9OHn5ZF4Uo
https://youtu.be/R9OHn5ZF4Uo
How AI Learns
@Urethramancer @cetan @arstechnica and then follow it with this video from the same guy:
https://youtu.be/wvWpdrfoEv0
https://youtu.be/wvWpdrfoEv0
How AI, Like ChatGPT, *Really* Learns
Conor O'Neill@arstechnica
This article about how Large Language models work is well worth reading.
This article about how Large Language models work is well worth reading.
Michael Gemar@arstechnica @dangillmor I would respectfully suggest that if an article uses “vector”, it’s not jargon-free (at least for most audiences).
@arstechnica
The article is 90% usable if one is trying to understand the LLM, By the end. When it starts to discuss GPT-3 performance the author radically pivots to examples of users anthropomorphizing the LLM model. One example even compares GPT-3 to 3 to 7/ year old kids. Kinda killed the whole thing for me.
A logical explanation of LLM structures suddenly morphs into witchcraft.
The article is 90% usable if one is trying to understand the LLM, By the end. When it starts to discuss GPT-3 performance the author radically pivots to examples of users anthropomorphizing the LLM model. One example even compares GPT-3 to 3 to 7/ year old kids. Kinda killed the whole thing for me.
A logical explanation of LLM structures suddenly morphs into witchcraft.
khays@arstechnica nice explanation