Skip to content

Commit a404f71

Browse files
committed
updated lecture 20
1 parent c02ba7f commit a404f71

3 files changed

Lines changed: 19 additions & 16 deletions

File tree

slides/week6/lecture20.html

Lines changed: 9 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -721,10 +721,10 @@ <h1 id="what-language-actually-does">What language actually does</h1>
721721
<p>Stephens et al (2010) found that successful communication <strong>replicated the speaker's brain patterns in the listener's brain</strong>, with the listener's responses temporally coupled to — and sometimes <em>anticipating</em> — the speaker's. Language doesn't just convey ideas: <em>it programs one brain to enter the same state as another</em>.</p>
722722
</div>
723723
<div class="note-box" data-title="Further reading">
724-
<p><a href="https://doi.org/10.1073/pnas.1008662107"><strong>Stephens, Silbert &amp; Hasson (2010, <em>PNAS</em>)</strong></a> &quot;Speaker–listener neural coupling underlies successful communication&quot; — Recorded brain activity from speakers and listeners during natural storytelling.</p>
724+
<p><a href="https://doi.org/10.1073/pnas.1008662107"><strong>Stephens, Silbert, &amp; Hasson (2010, <em>PNAS</em>)</strong></a> &quot;Speaker–listener neural coupling underlies successful communication&quot; — Recorded brain activity from speakers and listeners during natural storytelling.</p>
725725
</div>
726726
</section>
727-
</foreignObject></svg><svg data-marpit-svg="" viewBox="0 0 1280 720"><foreignObject width="1280" height="720"><section id="4" data-class="scale-70" data-theme="cdl-theme" lang="en-US" class="scale-70" style="--class:scale-70;--theme:cdl-theme;" data-transition-back="{&quot;name&quot;:&quot;fade&quot;,&quot;duration&quot;:&quot;0.25s&quot;,&quot;builtinFallback&quot;:true}" data-transition="{&quot;name&quot;:&quot;fade&quot;,&quot;duration&quot;:&quot;0.25s&quot;,&quot;builtinFallback&quot;:true}">
727+
</foreignObject></svg><svg data-marpit-svg="" viewBox="0 0 1280 720"><foreignObject width="1280" height="720"><section id="4" data-class="scale-65" data-theme="cdl-theme" lang="en-US" class="scale-65" style="--class:scale-65;--theme:cdl-theme;" data-transition-back="{&quot;name&quot;:&quot;fade&quot;,&quot;duration&quot;:&quot;0.25s&quot;,&quot;builtinFallback&quot;:true}" data-transition="{&quot;name&quot;:&quot;fade&quot;,&quot;duration&quot;:&quot;0.25s&quot;,&quot;builtinFallback&quot;:true}">
728728
<h1 id="language-transfers-memories-between-brains">Language transfers memories between brains</h1>
729729
<div class="definition-box" data-title="Language reconstructs experience">
730730
<p>When person A watches a movie and then <em>tells</em> person B the story, person B's brain activity during listening resembles person A's brain activity during <em>watching</em> — not during speaking. Language doesn't just transmit verbal patterns. It reconstructs the speaker's <strong>perceptual experience</strong> in the listener's brain. You don't just hear a story — your brain <em>relives</em> it.</p>
@@ -733,8 +733,8 @@ <h1 id="language-transfers-memories-between-brains">Language transfers memories
733733
<p>The brain constructs the same <strong>amodal narrative representation</strong> whether the story is read, heard, or watched as a movie. Language isn't special because of sound or letters — it programs a specific neural state regardless of input channel.</p>
734734
</div>
735735
<div class="note-box" data-title="Further reading">
736-
<p><a href="https://doi.org/10.1093/cercor/bhw351"><strong>Zadbood et al. (2017, <em>Cerebral Cortex</em>)</strong></a>: &quot;How we transmit memories to other brains&quot; — Narrative reconstructs the speaker's perceptual experience in the listener.</p>
737-
<p><a href="https://doi.org/10.1523/JNEUROSCI.1580-13.2013"><strong>Regev, Honey &amp; Hasson (2013, <em>J. Neurosci.</em>)</strong></a>: Neural patterns are modality-independent.</p>
736+
<p><a href="https://academic.oup.com/cercor/article/27/10/4988/4080827"><strong>Zadbood et al. (2017, <em>Cerebral Cortex</em>)</strong></a>: &quot;How we transmit memories to other brains&quot; — Narrative reconstructs the speaker's perceptual experience in the listener.</p>
737+
<p><a href="https://doi.org/10.1523/JNEUROSCI.1580-13.2013"><strong>Regev, Honey, &amp; Hasson (2013, <em>J. Neurosci.</em>)</strong></a>: &quot;Selective and invariant neural responses to spoken and written narratives&quot; — The same brain regions respond to stories regardless of modality.</p>
738738
</div>
739739
<div class="tip-box" data-title="Discussion">
740740
<p>If language reconstructs the speaker's perceptual experience in the listener's brain, is there a meaningful difference between &quot;experiencing something&quot; and &quot;hearing a vivid enough description of it&quot;? Where does the line fall? (Note: &quot;hearing&quot; could also mean &quot;reading&quot; or &quot;watching&quot; — the specific input channel isn't the point here.)</p>
@@ -749,8 +749,8 @@ <h1 id="llms-speak-the-brains-language">LLMs speak the brain's language</h1>
749749
<div class="note-box" data-title="Further reading">
750750
<p><a href="https://doi.org/10.1016/j.neuron.2024.06.025"><strong>Zada et al. (2024, <em>Neuron</em>)</strong></a>: &quot;A shared model-based linguistic space for transmitting our thoughts from brain to brain&quot; — Demonstrated that LLM embeddings can track real-time speaker–listener neural coupling during natural conversation.</p>
751751
<p><a href="https://rdcu.be/cpMwZ"><strong>Heusser et al. (2021, <em>Nature Human Behaviour</em>)</strong></a>: &quot;Geometric models reveal behavioral and neural signatures of transforming naturalistic experiences into episodic memories&quot; — Brain activity during movie watching predicts how you <em>recount</em> it later.</p>
752-
<p><a href="https://doi.org/10.31234/osf.io/dh3q2_v2"><strong>Fitzpatrick et al. (2026, <em>Nature Commumnications</em>)</strong></a>: &quot;Text embedding models yield high-resolution insights into conceptual knowledge from short multiple-choice quizzes&quot; - Text embeddings can be used to accurately model (an approximation of) <em>everything</em> you know!</p>
753-
<p><a href="http://caligari.dartmouth.edu/~jmanning/pubs/MannKaha12.pdf"><strong>Manning &amp; Kahana (2012, <em>Memory</em>)</strong></a>: &quot;Interpreting semantic clustering effects in free recall&quot; - When we use the &quot;wrong&quot; embedding model to model our thoughts and memories, how misleading is it?</p>
752+
<p><a href="https://doi.org/10.31234/osf.io/dh3q2_v2"><strong>Fitzpatrick et al. (2026, <em>Nature Communications</em>)</strong></a>: &quot;Text embedding models yield high-resolution insights into conceptual knowledge from short multiple-choice quizzes&quot; Text embeddings can be used to accurately model (an approximation of) <em>everything</em> you know!</p>
753+
<p><a href="http://caligari.dartmouth.edu/~jmanning/pubs/MannKaha12.pdf"><strong>Manning &amp; Kahana (2012, <em>Memory</em>)</strong></a>: &quot;Interpreting semantic clustering effects in free recall&quot; When we use the &quot;wrong&quot; embedding model to model our thoughts and memories, how misleading is it?</p>
754754
</div>
755755
<div class="tip-box" data-title="Discussion">
756756
<p>Think about all <em>possible</em> concepts that &quot;text&quot; could express. LLMs cannot possibly learn this infinite space. Rather, they learn a <strong>much</strong> lower-dimensional subspace that captures something akin to &quot;the concepts that are expressed in most text <em>in practice</em>.&quot; What are the implications of this idea? Does it speak to limitations of humans? Of language? Of text as a medium for thought and/or communication?</p>
@@ -820,7 +820,7 @@ <h1 id="what-llms-have-inside">What LLMs have inside</h1>
820820
<p>Anthropic shows LLMs build structured concept maps. LeCun argues they lack grounding. Is a map of other people's concepts a form of understanding? Twist: as the Internet becomes &quot;polluted&quot; with LLM-generated text, how will this affect what LLMs can learn?</p>
821821
</div>
822822
</section>
823-
</foreignObject></svg><svg data-marpit-svg="" viewBox="0 0 1280 720"><foreignObject width="1280" height="720"><section id="10" data-class="scale-75" data-theme="cdl-theme" lang="en-US" class="scale-75" style="--class:scale-75;--theme:cdl-theme;" data-transition-back="{&quot;name&quot;:&quot;fade&quot;,&quot;duration&quot;:&quot;0.25s&quot;,&quot;builtinFallback&quot;:true}" data-transition="{&quot;name&quot;:&quot;fade&quot;,&quot;duration&quot;:&quot;0.25s&quot;,&quot;builtinFallback&quot;:true}">
823+
</foreignObject></svg><svg data-marpit-svg="" viewBox="0 0 1280 720"><foreignObject width="1280" height="720"><section id="10" data-class="scale-70" data-theme="cdl-theme" lang="en-US" class="scale-70" style="--class:scale-70;--theme:cdl-theme;" data-transition-back="{&quot;name&quot;:&quot;fade&quot;,&quot;duration&quot;:&quot;0.25s&quot;,&quot;builtinFallback&quot;:true}" data-transition="{&quot;name&quot;:&quot;fade&quot;,&quot;duration&quot;:&quot;0.25s&quot;,&quot;builtinFallback&quot;:true}">
824824
<h1 id="are-llms-thinking-or-performing">Are LLMs &quot;thinking&quot; or &quot;performing&quot;?</h1>
825825
<div class="definition-box" data-title="LLMs as role-play engines">
826826
<p>LLMs are best understood as <strong>&quot;engines for role-play&quot;</strong> — they simulate the behavior of a plausible speaker, drawing on the vast repertoire of speakers in training data. When GPT writes a poem, it's not expressing itself — it's simulating someone who would write that poem. The performance can be indistinguishable from the real thing.</p>
@@ -829,7 +829,8 @@ <h1 id="are-llms-thinking-or-performing">Are LLMs &quot;thinking&quot; or &quot;
829829
<p>LLMs have <strong>attention</strong> (the mechanism) but may lack an <strong>attention schema</strong> — an internal model of their own attentional states. Humans don't just attend to things; we are <em>aware that we are attending</em>. LLMs process information without modeling the fact that they're processing it. They are &quot;looking&quot; but not &quot;aware of looking.&quot;</p>
830830
</div>
831831
<div class="note-box" data-title="References">
832-
<p><a href="https://doi.org/10.1145/3624724"><strong>Shanahan (2024, <em>CACM</em>)</strong></a>: &quot;Talking about large language models&quot; — LLMs as role-play engines simulating plausible speakers. <a href="https://arxiv.org/abs/2411.00983"><strong>Farrell, Graziano et al. (2025, <em>arXiv</em>)</strong></a>: &quot;Attention schema in LLMs&quot; — LLMs lack an internal model of their own attentional states.</p>
832+
<p><a href="https://doi.org/10.1145/3624724"><strong>Shanahan (2024, <em>CACM</em>)</strong></a>: &quot;Talking about large language models&quot; — LLMs as role-play engines simulating plausible speakers.</p>
833+
<p><a href="https://arxiv.org/abs/2411.00983"><strong>Farrell, Graziano et al. (2025, <em>arXiv</em>)</strong></a>: &quot;Attention schema in LLMs&quot; — LLMs lack an internal model of their own attentional states.</p>
833834
</div>
834835
<div class="tip-box" data-title="Discussion">
835836
<p>When you read a novel and feel sad for a character, are you &quot;really&quot; feeling sadness, or performing a simulation of sadness triggered by text? If it's genuine for you, what would make it not genuine for an LLM?</p>

slides/week6/lecture20.md

Lines changed: 10 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -55,12 +55,12 @@ Stephens et al (2010) found that successful communication **replicated the speak
5555

5656
<div class="note-box" data-title="Further reading">
5757

58-
[**Stephens, Silbert & Hasson (2010, *PNAS*)**](https://doi.org/10.1073/pnas.1008662107) "Speaker–listener neural coupling underlies successful communication" — Recorded brain activity from speakers and listeners during natural storytelling.
58+
[**Stephens, Silbert, & Hasson (2010, *PNAS*)**](https://doi.org/10.1073/pnas.1008662107) "Speaker–listener neural coupling underlies successful communication" — Recorded brain activity from speakers and listeners during natural storytelling.
5959

6060
</div>
6161

6262
---
63-
<!-- _class: scale-70 -->
63+
<!-- _class: scale-65 -->
6464

6565
# Language transfers memories between brains
6666

@@ -78,9 +78,9 @@ The brain constructs the same **amodal narrative representation** whether the st
7878

7979
<div class="note-box" data-title="Further reading">
8080

81-
[**Zadbood et al. (2017, *Cerebral Cortex*)**](https://doi.org/10.1093/cercor/bhw351): "How we transmit memories to other brains" — Narrative reconstructs the speaker's perceptual experience in the listener.
81+
[**Zadbood et al. (2017, *Cerebral Cortex*)**](https://academic.oup.com/cercor/article/27/10/4988/4080827): "How we transmit memories to other brains" — Narrative reconstructs the speaker's perceptual experience in the listener.
8282

83-
[**Regev, Honey & Hasson (2013, *J. Neurosci.*)**](https://doi.org/10.1523/JNEUROSCI.1580-13.2013): Neural patterns are modality-independent.
83+
[**Regev, Honey, & Hasson (2013, *J. Neurosci.*)**](https://doi.org/10.1523/JNEUROSCI.1580-13.2013): "Selective and invariant neural responses to spoken and written narratives" — The same brain regions respond to stories regardless of modality.
8484

8585
</div>
8686

@@ -109,9 +109,9 @@ Text embedding spaces provide a **shared numerical coordinate system** for track
109109

110110
[**Heusser et al. (2021, *Nature Human Behaviour*)**](https://rdcu.be/cpMwZ): "Geometric models reveal behavioral and neural signatures of transforming naturalistic experiences into episodic memories" — Brain activity during movie watching predicts how you *recount* it later.
111111

112-
[**Fitzpatrick et al. (2026, *Nature Commumnications*)**](https://doi.org/10.31234/osf.io/dh3q2_v2): "Text embedding models yield high-resolution insights into conceptual knowledge from short multiple-choice quizzes" - Text embeddings can be used to accurately model (an approximation of) *everything* you know!
112+
[**Fitzpatrick et al. (2026, *Nature Communications*)**](https://doi.org/10.31234/osf.io/dh3q2_v2): "Text embedding models yield high-resolution insights into conceptual knowledge from short multiple-choice quizzes" Text embeddings can be used to accurately model (an approximation of) *everything* you know!
113113

114-
[**Manning & Kahana (2012, *Memory*)**](http://caligari.dartmouth.edu/~jmanning/pubs/MannKaha12.pdf): "Interpreting semantic clustering effects in free recall" - When we use the "wrong" embedding model to model our thoughts and memories, how misleading is it?
114+
[**Manning & Kahana (2012, *Memory*)**](http://caligari.dartmouth.edu/~jmanning/pubs/MannKaha12.pdf): "Interpreting semantic clustering effects in free recall" When we use the "wrong" embedding model to model our thoughts and memories, how misleading is it?
115115

116116
</div>
117117

@@ -248,7 +248,7 @@ Anthropic shows LLMs build structured concept maps. LeCun argues they lack groun
248248
</div>
249249

250250
---
251-
<!-- _class: scale-75 -->
251+
<!-- _class: scale-70 -->
252252

253253
# Are LLMs "thinking" or "performing"?
254254

@@ -266,7 +266,9 @@ LLMs have **attention** (the mechanism) but may lack an **attention schema** —
266266

267267
<div class="note-box" data-title="References">
268268

269-
[**Shanahan (2024, *CACM*)**](https://doi.org/10.1145/3624724): "Talking about large language models" — LLMs as role-play engines simulating plausible speakers. [**Farrell, Graziano et al. (2025, *arXiv*)**](https://arxiv.org/abs/2411.00983): "Attention schema in LLMs" — LLMs lack an internal model of their own attentional states.
269+
[**Shanahan (2024, *CACM*)**](https://doi.org/10.1145/3624724): "Talking about large language models" — LLMs as role-play engines simulating plausible speakers.
270+
271+
[**Farrell, Graziano et al. (2025, *arXiv*)**](https://arxiv.org/abs/2411.00983): "Attention schema in LLMs" — LLMs lack an internal model of their own attentional states.
270272

271273
</div>
272274

slides/week6/lecture20.pdf

-4.03 KB
Binary file not shown.

0 commit comments

Comments
 (0)