Matthew Lutze
@lutze.bsky.social
Independent researcher in AI safety and robustness.
Technology builder for public trust and security applications.
Technology builder for public trust and security applications.
The practice is called rhetorical analysis, you may enjoy learning about it:
miamioh.edu/howe-center/...
open.umn.edu/opentextbook...
miamioh.edu/howe-center/...
open.umn.edu/opentextbook...
Rhetorical Analyses
miamioh.edu
January 25, 2026 at 12:23 PM
The practice is called rhetorical analysis, you may enjoy learning about it:
miamioh.edu/howe-center/...
open.umn.edu/opentextbook...
miamioh.edu/howe-center/...
open.umn.edu/opentextbook...
And it is information about what a regime wants its population to think.
January 25, 2026 at 10:50 AM
And it is information about what a regime wants its population to think.
I would deeply love to see a visualization of the filtering process that this submission pool undergoes, it would be terribly fascinating!
January 24, 2026 at 6:23 PM
I would deeply love to see a visualization of the filtering process that this submission pool undergoes, it would be terribly fascinating!
It’s ok to be occasionally reminded of the cost of inaction, so that we remember how important it is to use our time and abilities to reinforce good things around us.
But if it starts to drag you down, it’s also ok to give yourself a break.
But if it starts to drag you down, it’s also ok to give yourself a break.
January 12, 2026 at 8:39 PM
It’s ok to be occasionally reminded of the cost of inaction, so that we remember how important it is to use our time and abilities to reinforce good things around us.
But if it starts to drag you down, it’s also ok to give yourself a break.
But if it starts to drag you down, it’s also ok to give yourself a break.
Ah sorry, I think you may be focusing on the face of the example I gave, instead of the analogy I intended to make between my convincing but shallow reproduction of a capability, and the general possibility of similar experience with current generative language models.
January 7, 2026 at 1:18 AM
Ah sorry, I think you may be focusing on the face of the example I gave, instead of the analogy I intended to make between my convincing but shallow reproduction of a capability, and the general possibility of similar experience with current generative language models.
It feels like when I was taught to sing French and Italian solos in high school. I won competitive awards because I reproduced the sounds convincingly well, consistently, but I didn’t learn either language and so never actually knew what I was singing.
January 7, 2026 at 12:02 AM
It feels like when I was taught to sing French and Italian solos in high school. I won competitive awards because I reproduced the sounds convincingly well, consistently, but I didn’t learn either language and so never actually knew what I was singing.
I guess I’ve had the impression that they produce this text gradually, a token at a time, because generating the idea and then self-talk confirmation or critique is rewarded.
But if the model doesn’t retain the new connection or reinforce, it’s again like, not actually introspecting.
But if the model doesn’t retain the new connection or reinforce, it’s again like, not actually introspecting.
January 7, 2026 at 12:02 AM
I guess I’ve had the impression that they produce this text gradually, a token at a time, because generating the idea and then self-talk confirmation or critique is rewarded.
But if the model doesn’t retain the new connection or reinforce, it’s again like, not actually introspecting.
But if the model doesn’t retain the new connection or reinforce, it’s again like, not actually introspecting.
But LRMs don’t create new connections, surface and restructure representation weights or spontaneously resolve misunderstandings.
The impression that might appear through selecting the best probabilistic self-talk “chain of thought” from perhaps a tree of sub-chains evaporates with the session.
The impression that might appear through selecting the best probabilistic self-talk “chain of thought” from perhaps a tree of sub-chains evaporates with the session.
January 6, 2026 at 10:35 PM
But LRMs don’t create new connections, surface and restructure representation weights or spontaneously resolve misunderstandings.
The impression that might appear through selecting the best probabilistic self-talk “chain of thought” from perhaps a tree of sub-chains evaporates with the session.
The impression that might appear through selecting the best probabilistic self-talk “chain of thought” from perhaps a tree of sub-chains evaporates with the session.
Twitter was a wild moment in time.
January 6, 2026 at 4:53 PM
Twitter was a wild moment in time.
My not great understanding is that Chatter’s approach hasn’t been validated, and didn’t he redefine thinking to be so narrow as to exclude unconscious cognition like procedural thoughts?
January 5, 2026 at 10:57 PM
My not great understanding is that Chatter’s approach hasn’t been validated, and didn’t he redefine thinking to be so narrow as to exclude unconscious cognition like procedural thoughts?
I wonder if it’s possible to intentionally trigger a (genuine?) aha moment.
I think that I think of them as an emergent experience, and the critical mass of potential nodes between which a new connection might manifest doesn’t feel deterministic.
But maybe that’s not the right framing?
I think that I think of them as an emergent experience, and the critical mass of potential nodes between which a new connection might manifest doesn’t feel deterministic.
But maybe that’s not the right framing?
January 5, 2026 at 10:32 PM
I wonder if it’s possible to intentionally trigger a (genuine?) aha moment.
I think that I think of them as an emergent experience, and the critical mass of potential nodes between which a new connection might manifest doesn’t feel deterministic.
But maybe that’s not the right framing?
I think that I think of them as an emergent experience, and the critical mass of potential nodes between which a new connection might manifest doesn’t feel deterministic.
But maybe that’s not the right framing?
I was using Quiet Posters more recently and found that nice, but this For You I bet is more engaging (not that I necessarily need more scrolling, but still)
Thanks x2 ❤️
Thanks x2 ❤️
January 5, 2026 at 10:06 PM
I was using Quiet Posters more recently and found that nice, but this For You I bet is more engaging (not that I necessarily need more scrolling, but still)
Thanks x2 ❤️
Thanks x2 ❤️
I’m glad you’ve found it!
I will unapologetically stalk through some of your network and see if I can’t improve my algorithm quality.
I will unapologetically stalk through some of your network and see if I can’t improve my algorithm quality.
January 5, 2026 at 8:59 PM
I’m glad you’ve found it!
I will unapologetically stalk through some of your network and see if I can’t improve my algorithm quality.
I will unapologetically stalk through some of your network and see if I can’t improve my algorithm quality.
My theory is that this is the natural state of these distributed asynchronous social networks.
I think Reddit’s sauce was in encouraging no-friction fission.
I think Reddit’s sauce was in encouraging no-friction fission.
January 5, 2026 at 8:12 PM
My theory is that this is the natural state of these distributed asynchronous social networks.
I think Reddit’s sauce was in encouraging no-friction fission.
I think Reddit’s sauce was in encouraging no-friction fission.
I honestly miss how informed I felt on Twitter. The people doing really cool stuff, the critical mass of cross-replying academics and scientists, I feel like did a touch-and-go 15 months ago and never came back.
January 5, 2026 at 8:12 PM
I honestly miss how informed I felt on Twitter. The people doing really cool stuff, the critical mass of cross-replying academics and scientists, I feel like did a touch-and-go 15 months ago and never came back.
I am not excited that another show which provided measured and impactful reporting has been destroyed, leaving yet fewer quality sources of journalism.
January 5, 2026 at 6:16 AM
I am not excited that another show which provided measured and impactful reporting has been destroyed, leaving yet fewer quality sources of journalism.
It’s the substance and purpose of the follow-up questions he asks, not just that he asks them, which I think we’re admiring here.
January 3, 2026 at 11:53 PM
It’s the substance and purpose of the follow-up questions he asks, not just that he asks them, which I think we’re admiring here.
Oh! Silly me, and yeah then an even greater signal for getting scikit-learn evangelized 😄
December 29, 2025 at 1:09 PM
Oh! Silly me, and yeah then an even greater signal for getting scikit-learn evangelized 😄
If you work in Python you should give it a look, you’ll absolutely love all of the tools it offers. This along with @matplotlib.org are a powerhouse of data science capabilities ❤️
December 29, 2025 at 11:43 AM
If you work in Python you should give it a look, you’ll absolutely love all of the tools it offers. This along with @matplotlib.org are a powerhouse of data science capabilities ❤️