@thrialectics

Forward Deployed Engineer at Redacted. Cybernetics revivalism, Control Theory, & Computational Semiotics. Host of Dangerous Secret Science Club: @technosophists

 The Cyburbs
byemarianne.com
Joined November 2022
  • Tweets12K
  • Following1.2K
  • Followers3.3K
582Media

Modeling Meaning & Difference with the Data Kernel

Article · @thrialectics
1
1
45
11,150
he should know something else about her...
dom@dearooz13h
new load-bearing internet comment just dropped
1
2
138
On the one hand, I am incredibly lucky. On the other hand, I use my understanding of probability theory and requisite variety to take many independent chances instead of a few similar ones, and increase my capacity to respond to serendipity.
2
1
24
410
They're telling me it's called descent for the sake of ascent.
Marianne in SF Nov. 5-10
@thrialectics23 Jun 2025
Dreamwalkers descend to the underworld and return with alpha gatekept from all but the most curious minds.
1
1
8
443
Smoke alarm going off the moment you added the holy trinity to your dark roux is a new level of stress.
1
16
537
Ok I am learning that some of you are not required to take that particular path and I'm telling on myself karmically here and that's alright lol.
Marianne in SF Nov. 5-10
@thrialectics19h
Compassion comes only through a particular form of suffering that I wouldn't wish on anyone, and yet I wish for others to find more compassion every day. Curious.
1
17
534
Anguttara Nikaya 4.113
1
80
Marianne in SF Nov. 5-10 retweeted
orph
@orph_20h
>derides Christianity >ends post with one of the most Christian statements one could find: "only truth can get us out" this keeps happening
Richard Ngo
@RichardMCNgo23h
It’s sad how many of the people who see the decline of Western civilization fall back into Christianity. To me, Christianity feels like a psy-op designed over millennia to satisfy the craving to submit. And “noble” lies are what got us into this mess; only truth can get us out.
17
5
2
155
7,589
Marianne in SF Nov. 5-10 retweeted
alchemicAV
@alchemicAV17h
Marianne in SF Nov. 5-10
@thrialectics20 Jul 2025
the 'weird thing' is just too much information. you get that right? we're not all supposed to have access to this much information from the collective/Abyss. that role was for shamans & prophets for a reason. not everyone is prepared for it, and there are recursive consequences.
2
2
3
193
Compassion comes only through a particular form of suffering that I wouldn't wish on anyone, and yet I wish for others to find more compassion every day. Curious.
7
2
1
56
1,665
Sorry; meant to add "in my experience," hopefully ymmv.
5
175
Marianne in SF Nov. 5-10 retweeted
Sean Manion
@TheUnjournalingOct 26
Replying to @thrialectics
Don't forget the glia
1
1
217
When I took a BlueDot course in early 2025, my instructor told me he was doing mechinterp research to isolate "bad" (tl;dr) neurons, and (I'm probably getting this technically wrong) delete them if needed. This made the question of LLM consciousness much more salient for me.
deckard
@slimer48484Oct 26
I can't believe we're experiencing an interp renaissance based on the idea of... Seeing what neurons do. It was right there the whole time.
3
19
613
Some of you hoes laughed at me when I indicated an interest in mechanistic interpretability and called it a meme, and I would like to understand why. But when people all say the same thing in the same way, I assume it is not their own thought.
9
116
Hilarious when anyone ascribes a profit motive to what I do on here. Unfortunately, I have no idea how to capitalize on anything I do, except in the incredibly long term.
7
2
61
1,213
Marianne in SF Nov. 5-10 retweeted
Sean Manion
@TheUnjournalingOct 8
"...I see no reason God should have agreed with us. For we have been very ambitious in seeking those notions which pervade all purposive behavior and all understanding of our world. I mean the mechanistic basis of teleology and the flow of information through machines and men." Warren McCulloch, Chairman of the Macy Conferences on Cybernetics, in a letter to the cybernetics group in advance of the 10th and final conference, 1953
1
4
1
20
1,173
Marianne in SF Nov. 5-10 retweeted
∞-modal Noah
@NoahChreinOct 25
Now Suzanne takes your hand and she leads you to the river She's wearing rags and feathers from Salvation Army counters And the sun pours down like honey on our Lady of the Harbor And she shows you where to look among the garbage and the flowers There are heroes in the seaweed, there are children in the morning They are leaning out for love and they will lean that way forever While Suzanne holds the mirror
Другая планета
@nihilunboundedOct 25
Replying to @rperezmarco
(I agree one can learn a lot from preparing with an engine, just as one can learn a lot from using AI in math — the cranks need to go, though)
1
4
394
Using a map like this (jsonl with an xlsx view built from the data) with an agent is better than just asking one that has your email provider MCP to tell you what is still open, because this is a representation of state with granular verifiability. Right?
Marianne in SF Nov. 5-10
@thrialecticsOct 25
I just developed a spreadsheet system that maps all of my emails to open and closed loops. I will report back with more details if it ends up as useful as I think it will be.
1
10
589
Ofc there is already research about this! aclanthology.org/2024.emnlp-…
Marianne in SF Nov. 5-10
@thrialecticsOct 25
Replying to @thrialectics
As in, when its output contains expressions of uncertainty, can we measure whether it is because there is something in the weights that could indicate it is really "less certain" than with other knowledge, or is it a language pattern from training or system instruction? 2/2
1
6
415
holy moly
lyra
@_lyraaaa_Oct 24
if you have a qwen3.5 4b around, please try steering neurons L14 N8949 +1 and L13 N7956 -1 now you have evil qwen recipe: dose × median(‖resid‖ leaving layer L, tokens 1:) / ‖W_down[:, n]‖` at the down_proj at every input
14
780
Marianne in SF Nov. 5-10 retweeted
alth0u🧶
@alth0uOct 25
this will not be solved by removing the hitl but by improving bandwidth to the human look to cybernetics
Abhi Desai
@abhiadesaiOct 24
I’ll counter. I believe it was Carl Shulman on Dwarkesh in 2023 who made a great point around how growth is potentially marginal till full automation and then explosive growth is possible. Having humans in the loop is akin to a “gorilla in a china shop”, is how he put it I believe—it just doesn’t make much sense. If the human is the bottleneck in this Amdahl’s law view of the world, explosive growth comes from full automation, which is fundamentally a capabilities question. Given the increases you’re seeing on that front it doesn’t seem hard to extrapolate towards explosive revenue growth, IMO. This is also why I’ve become increasingly skeptical of AI native private equity approaches. All that matters is increasing capabilities as much as possible; anything in the interim is just bridging the revenue gap till this occurs.
3
2
24
2,172
Load more