A Terrifying Theory of Language - Dr. Elan Barenholtz, DemystifySci #413
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Related videos
Things Are Looking BAD for Big Orange | Secret Podcast
The Bulwark
129.3k views
Markdown is a terrible language
Theo - t3․gg
155.4k views
Trump Is Getting a TERRIBLE Deal (w/ Ben Rhodes) | The Bulwark Podcast
The Bulwark
110.4k views
Vikings, Ragnar, Berserkers, Valhalla & the Warriors of the Viking Age | Lex Fridman Podcast #495
Lex Fridman
124.1k views
The language holding our agents back.
Theo - t3․gg
89.3k views
THIS STORY IS TERRIFYING
Timcast
74.6k views
What’s the best programming language for AI?
Theo - t3․gg
75.8k views
We Need to Delete 500 Years of Fake History - Lance Weaver, DemystifySci #370
The DemystifySci Podcast
28.5k views
Where Biology Goes Off the Rails - Dr. Michael Levin, DemystifySci #369
The DemystifySci Podcast
27.4k views
Hidden Payoff of Civilizational Ruin - Dr. Dani Sulikowski, DemystifySci #360
The DemystifySci Podcast
40.7k views
Top Comments (10)
I am so tired of people misrepresenting how LLMs work. They don't even store words. They run on vectors/activations which encapsulate ideas or the gist of things. Words are just the input and the output. In reality LLMs work because ideas make sense and are related. These relationships are mapped in multidimensional space as geometry. More specifically, LLMs operate on token sequences of words but internally transform them into high-dimensional activation patterns. Meaning is represented not as discrete symbols but as distributed patterns across layers, where relationships between concepts emerge as geometric structure in activation space. Words serve only as the interface; the underlying computation captures contextual relationships between ideas.
But I think in "ideas" not words; often I have to struggle to get the right word for my idea. I don't just think in words.
One of the major ways language "goes wrong" is the use of personification such as "science tells us!" As though there's only monolithic ideas regarding science, rather than an evolving multiplicity of disagreement!
Listen on the go at all podcast locations: https://anchor.fm/demystifysci Pick up our music on vinyl or stream on bandcamp, spotify, etc.. https://demystifysci-shop.fourthwall.com/products/vinyl-lp-secretary-of-nature-everything-is-so-good-here Paradox Lost is now on pre-sale! https://buy.stripe.com/7sY7sKdoN5d29eUdYddEs0b
This is making me think of the self-transforming machine elves that Terence McKenna mentions.
Reality does not change, our perception of it is what is a shadow.
Also…whenever someone starts to pontificate about language, always listen to how they themselves use language. That gives you a better insight into their real beliefs…
This is fascinating but mostly because it is something I have always been conscious of but without giving it any specific thought or attention.
How can you possibly want to come back? I'm always disappointed that it's over, no matter how harsh or pleasant conditions are. Always pulled around the next bend in the trail, what's at the top of that hill? Yeah, all this stuff is kinda nice, but out there every step is deeply interesting. And will never cease to be.
11:34 wittgenstein's language games?
Unlock the Data Inside
Turn Videos into Knowledge
- Get FREE 10/day: transcripts, summaries, chats
- Chat with videos, export text & PDF
- $1 free API credit for RAG, chatbots & research
Free forever plan • All features unlocked
Top Comments (10)
I am so tired of people misrepresenting how LLMs work. They don't even store words. They run on vectors/activations which encapsulate ideas or the gist of things. Words are just the input and the output. In reality LLMs work because ideas make sense and are related. These relationships are mapped in multidimensional space as geometry. More specifically, LLMs operate on token sequences of words but internally transform them into high-dimensional activation patterns. Meaning is represented not as discrete symbols but as distributed patterns across layers, where relationships between concepts emerge as geometric structure in activation space. Words serve only as the interface; the underlying computation captures contextual relationships between ideas.
But I think in "ideas" not words; often I have to struggle to get the right word for my idea. I don't just think in words.
One of the major ways language "goes wrong" is the use of personification such as "science tells us!" As though there's only monolithic ideas regarding science, rather than an evolving multiplicity of disagreement!
Listen on the go at all podcast locations: https://anchor.fm/demystifysci Pick up our music on vinyl or stream on bandcamp, spotify, etc.. https://demystifysci-shop.fourthwall.com/products/vinyl-lp-secretary-of-nature-everything-is-so-good-here Paradox Lost is now on pre-sale! https://buy.stripe.com/7sY7sKdoN5d29eUdYddEs0b
This is making me think of the self-transforming machine elves that Terence McKenna mentions.
Reality does not change, our perception of it is what is a shadow.
Also…whenever someone starts to pontificate about language, always listen to how they themselves use language. That gives you a better insight into their real beliefs…
This is fascinating but mostly because it is something I have always been conscious of but without giving it any specific thought or attention.
How can you possibly want to come back? I'm always disappointed that it's over, no matter how harsh or pleasant conditions are. Always pulled around the next bend in the trail, what's at the top of that hill? Yeah, all this stuff is kinda nice, but out there every step is deeply interesting. And will never cease to be.
11:34 wittgenstein's language games?