Expert attention map

The Who's Who of AI

What credible people across AI noticed, why it matters, and where the field is converging or disagreeing.

2,397 searchable experts 4,280 tracked across all sources
Build my stream Create my edition
Filter the conversation Who is saying what?

Combine a professional role with a reaction lens. Both must match the same attributed contribution.

Clear all
Active evidence filter

Showing developments with attributable Opportunity & adoption reactions.

New Network reaction maps

See how experts are reacting—not just what they shared.

Posts are grouped by conversation and tone. Select a lens to filter the stream; these are never permanent labels on people.

Opportunity & adoption Building & implementation
2 live reaction maps in this view Jump to the evidence ↓
Reaction converging Verbalizable Representations Form a Global Workspace in Language Models Opportunity & adoption · 5 attributable reactions Shared emphasis On the Biology of a Large Language Model Opportunity & adoption · 2 attributable reactions
Showing signals surfaced by Nafnlaus 🇮🇸 🇺🇦 ×
Most important

The developments commanding sustained expert attention

One card per development. Sources are clustered; reaction bundles describe these posts, never the people behind them.

Developing Models & releases Research 2d ago
⚡ 3 h early
global workspace theory language models paper

Verbalizable Representations Form a Global Workspace in Language Models

6 experts across 4 network communities independently surfaced this.

Why this matches Opportunity & adoption reaction 4 attributable expert contributions · Mark Riedl, Matthew Kirschenbaum, Nafnlaus 🇮🇸 🇺🇦
“here is the Anthropic blog: transformer-circuits.pub/2026/workspa... I don't really like the framing of whether LLMs are conscious or not. It's completely unnecessary. As is the hand-waiving about whether LLMs have something equivalent to the (functional) g…” evidence ↗
6 experts 4 communities 1 sources clustered
Mark Riedl Tim Kellogg Matthew Kirschenbaum Edouard Machery Nafnlaus 🇮🇸 🇺🇦 david-p-reichert.bsky.social
How the network is reacting 4 experts are independently emphasizing opportunity & adoption.
Reaction converging

Opportunity & adoption

4 experts

New capabilities, benefits and practical upside.

“here is the Anthropic blog: transformer-circuits.pub/2026/workspa... I don't really like the framing of whether LLMs are conscious or not. It's completely unnecessary. As is the hand-waiving about whether LLMs have something equivalent to the (functional) g…”

Building & implementation

1 expert

How teams are shipping and applying it.

“Anthropic uses the word “conscious” 206 times in their latest research Specifically, they’re showing that LLMs implement one of the leading explanations of consciousness in humans (see pic) transformer-circuits.pub/2026/workspa...”

7 experts discussed this · 203 posts
Anil Dash: this is not an accurate description of what's happening. not least because I can write code and get them to output something different there. I can understand why it *seems* that way, but this is j…
Anil Dash: Who are you explaining this to?
Anil Dash: Do you see why explaining "All output from AI has to be validated" in a reply to me here is... not a useful contribution?
Open the full discussion →
Primary source transformer-circuits.pub ↗
Established Models & releases Research 12d ago
⚡ 1215 h early
Anthropic LLM biology interpretability paper

On the Biology of a Large Language Model

3 experts across 2 network communities independently surfaced this.

Why this matches Opportunity & adoption reaction 2 attributable expert contributions · rev. howard arson, Nafnlaus 🇮🇸 🇺🇦
“i think pre-planning rhymes is the best example of this behavior, where the model preactivates features associated with words which complete the rhyme and "stores" them as annotations on the newline token. transformer-circuits.pub/2025/attribu...” evidence ↗
3 experts 2 communities 1 sources clustered
rev. howard arson Mary Branscombe Nafnlaus 🇮🇸 🇺🇦
How the network is reacting 2 experts are emphasizing opportunity & adoption.
Shared emphasis

Opportunity & adoption

2 experts

New capabilities, benefits and practical upside.

“i think pre-planning rhymes is the best example of this behavior, where the model preactivates features associated with words which complete the rhyme and "stores" them as annotations on the newline token. transformer-circuits.pub/2025/attribu...”

6 experts discussed this · 124 posts
Nafnlaus 🇮🇸 🇺🇦: What "experts" are you talking about? Experts across the AI field have been freaking out about this. As far as I can tell, your "expert" seems to be Eryk Salvaggio, an artist pursuing a degree in d…
Nafnlaus 🇮🇸 🇺🇦: Wow, you really rebutted what I said there. You have certainly demonstrated that you are a person who should be taken seriously, about literally anything.
Doll: The same people who insist these machines are parrots that cant do the things they can demonstratably do??? Those scientists???
Open the full discussion →
Primary source transformer-circuits.pub ↗
Established Models & releases Resource 12d ago
Anthropic circuits interpretability October update

Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet

1 directory member surfaced this signal.

Why this matches Opportunity & adoption reaction 1 attributable expert contribution · Nafnlaus 🇮🇸 🇺🇦
“Which are built atop: transformer-circuits.pub/2024/scaling...” evidence ↗
1 expert 1 community 1 sources clustered
Nafnlaus 🇮🇸 🇺🇦

“Which are built atop: transformer-circuits.pub/2024/scaling...”

Primary source transformer-circuits.pub ↗
Debates gaining attention

What experts are discussing without an anchoring article

9 experts · 63 posts · 13d ago
Eugene Vinitsky Ted Underwood Cat Hicks Chelsea Parlett critical slop studies Adam L Mary Branscombe Amy Hoy
Matches Opportunity & adoption
“"why would I ever need a task to be done faster" come on, guys. Because we're dying. Because we are literally all dying all the time. Because we're running out of time. The fact that it's often har…” evidence ↗
Reaction converging Markets & investment · 4 Research & technical analysis · 1
Cat Hicks: "why would I ever need a task to be done faster" come on, guys. Because we're dying. Because we are literally all dying all the time. Because we're running out of time. The fact that it's often har…
Cat Hicks: because tasks gate other tasks. Because measuring the hours saved by washing machines is a relatively useful index for quality of life improvement. Because the time it takes for you to get from the…
Open the thread →
7 experts · 83 posts · 8d ago
Colin Mary Branscombe Doll Matt Darling Singularity's Bounty e/cc Nafnlaus 🇮🇸 🇺🇦 growth.wtf
Matches Opportunity & adoption
“Given that literally everything linguistic is thrown out before even the first Transformers layer, and even positioning is abstracted, and all logic is performed on latent (conceptual) states - and…” evidence ↗
Multiple readings Research & technical analysis · 2 Building & implementation · 1 Questions & unknowns · 1
Nafnlaus 🇮🇸 🇺🇦: LLMs are not databases. They do not contain their training data.
Nafnlaus 🇮🇸 🇺🇦: Base rate fallacy.
Open the thread →
4 experts · 34 posts · 4d ago
Gus Ravyn Nafnlaus 🇮🇸 🇺🇦 austegard.com
Matches Opportunity & adoption
“The problem with longer draft is that you stop with any speed up if the small model needs to generate too many tokens And thanks for the feedback. The diffusion Gemma team will be happy!” evidence ↗
Multiple readings Concern & critique · 1 Opportunity & adoption · 1 Questions & unknowns · 1
Gus: What would you like to see on Gemma 5?
Gus: Thanks Ana! Took note of everything.
Open the thread →
4 experts · 21 posts · 1d ago
Eugene Vinitsky Adam Wiemerslage Sam Barrett, PhD Nafnlaus 🇮🇸 🇺🇦
Matches Opportunity & adoption
“Removing voice is time-consuming for no benefit, and people vary greatly in how good they are in detecting it. Also, AI voice rubs off on frequent users, thus leading to actual human speech taking …” evidence ↗
Multiple readings Concern & critique · 1 Research & technical analysis · 1 Opportunity & adoption · 1
Eugene Vinitsky: It takes so long to write a paper because we are crazy about not making mistakes and checking as many edge cases as we can. This stressful responsibility is, I suppose, not obvious. so, if you’re n…
Eugene Vinitsky: Lots of new folks trying to write papers so it’s important to know: if there’s a mistake in the paper, the responsibility falls on you. You cannot offload it to the AI assistant.
Open the thread →