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.

2,364 searchable experts 2,966 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 Building & implementation 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.

Building & implementation Research & technical analysis Concern & critique Policy & governance
2 live reaction maps in this view Jump to the evidence ↓
Multiple readings Our position on open-weights models Policy & governance · 3 attributable reactions Multiple readings Summary of METR's predeployment evaluation of GPT-5.6 Sol Concern & critique · 3 attributable reactions
Showing signals surfaced by Tim Kellogg ×
Debated

The questions experts are actively pulling apart

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

Developing Models & releases Analysis 1d ago
⚡ 1 h early
Anthropic open-weights position statement

Our position on open-weights models

10 experts across 4 network communities independently surfaced this.

Why this matches Building & implementation reaction 1 attributable expert contribution · Ted Underwood
“I respect the candor here. But I don't agree with the geopolitics, at all. The line between democracies and autocracies is not as crisp as this pretends. How much do we trust, e.g., an AI company in a competitive-authoritarian state that launches sneak deca…” evidence ↗
10 experts 4 communities 1 sources clustered
Mark Riedl Ted Underwood Tim Kellogg Sung Kim James MacGlashan Dare Obasanjo Larry Hunter Fran Litterio +2
How the network is reacting Experts are approaching this through 3 distinct lenses.
Multiple readings

Policy & governance

1 expert

Rules, institutions and accountability.

“Dario Amodei clarifies that Anthropic doesn’t want to ban open weight model. It instead wants 1. A ban on selling powerful chips or chipmaking equipment to China. 2. A crack down on industrial-scale distillation operations. 3. Mandatory safety testing of fr…”

Building & implementation

1 expert

How teams are shipping and applying it.

“I respect the candor here. But I don't agree with the geopolitics, at all. The line between democracies and autocracies is not as crisp as this pretends. How much do we trust, e.g., an AI company in a competitive-authoritarian state that launches sneak deca…”

Research & technical analysis

1 expert

Evidence, methods and technical implications.

“Dario explains Anthropic’s position on open weights models It’s a lot of words to say that they’re still against powerful open weights models www.anthropic.com/news/positio...”

5 experts discussed this · 13 posts
Ted Underwood: I respect the candor here. But I don't agree with the geopolitics, at all. The line between democracies and autocracies is not as crisp as this pretends. How much do we trust, e.g., an AI company i…
Ted Underwood: (just, say, hypothetically?) There's also internal incoherence in making the first part of this all about moralized great-power competition and then saying that safety testing (incl for alignment r…
SE Gyges: he kisses jd vance's ass and fails to specify anything concerning his proposed evaluation regime that distinguishes it from a ban solid dario post 10/10
Open the full discussion →
Primary source anthropic.com ↗
Established Models & releases Signal 8d ago
⚡ 36 h early

openai.com

3 experts across 3 network communities independently surfaced this.

Why this matches Building & implementation reaction 1 attributable expert contribution · Dare Obasanjo
“OpenAI has launched it's first hardware device. A $230 mini-keyboard that integrates with Codex. It has light-up indicators showing agent status and customizable shortcuts for frequent Codex actions. It also has a dial to adjust the computing power an agent…” evidence ↗
3 experts 3 communities 1 sources clustered
Tim Kellogg Dare Obasanjo Shriram Krishnamurthi

“OpenAI has launched it's first hardware device. A $230 mini-keyboard that integrates with Codex. It has light-up indicators showing agent status and customizable shortcuts for frequent Codex actions. It also has a dial to adjust the computing power an agent…”

“check it out! you can pay $230 to operate codex like a loud bumbling idiot!!! it’s too big and heavy to carry with you, so it has to sit on your desk, and it has too few buttons to be a real keyboard openai.com/supply/co-la...”

5 experts discussed this · 6 posts
Tim Kellogg: check it out! you can pay $230 to operate codex like a loud bumbling idiot!!! it’s too big and heavy to carry with you, so it has to sit on your desk, and it has too few buttons to be a real keyboa…
Chris: That is hysterical.
coty.bsky.social: Can’t wait to use the skill stick and thinking dial.
Open the full discussion →
Primary source openai.com ↗
New AI field signal Signal 11h ago

Gpt frontier intelligence efficiency

3 experts are actively discussing the implications.

Why this matches Building & implementation reaction 1 attributable expert contribution · Sung Kim
“OpenAI used GPT-5.6 Sol to make it efficient to run. The results: - 20% lower serving costs from production GPU kernel improvements. - 15%+ better token-generation efficiency from improved speculative decoding. openai.com/index/gpt-5-...” evidence ↗
2 experts 2 communities 1 sources clustered
Tim Kellogg Sung Kim

“OpenAI used GPT-5.6 Sol to make it efficient to run. The results: - 20% lower serving costs from production GPU kernel improvements. - 15%+ better token-generation efficiency from improved speculative decoding. openai.com/index/gpt-5-...”

“gpt-5.6-sol is improving its own serving stack, gaining 20% here and 15% there this is a weak form of RSI openai.com/index/gpt-5-...”

3 experts discussed this · 3 posts
Tim Kellogg: gpt-5.6-sol is improving its own serving stack, gaining 20% here and 15% there this is a weak form of RSI openai.com/index/gpt-5-...
Thomas Dietterich: But this is completely uninteresting. We've had supercompilers for several years. I'll be impressed when we start seeing systems that can deal with novelty by inventing new conceptual frameworks.
Mark Riedl: Most RSI is likely to be uninteresting. Anything interesting is likely to be really really hard medium.com/@mark-riedl/...
Open the full discussion →
Primary source openai.com ↗
Developing Models & releases Signal 1d ago

Openai anthropic staff share letter asking us to help pace ai progress

3 experts are actively discussing the implications.

Why this matches Building & implementation reaction 1 attributable expert contribution · Tim Kellogg
“people from OpenAI and Anthropic sign letter calling for a slowdown of AI development, so society has a chance to catch up www.bloomberg.com/news/article...” evidence ↗
1 expert 1 community 1 sources clustered
Tim Kellogg

“people from OpenAI and Anthropic sign letter calling for a slowdown of AI development, so society has a chance to catch up www.bloomberg.com/news/article...”

3 experts discussed this · 6 posts
Tim Kellogg: people from OpenAI and Anthropic sign letter calling for a slowdown of AI development, so society has a chance to catch up www.bloomberg.com/news/article...
Tim Kellogg: who are those?
Tim Kellogg: en.wikipedia.org/wiki/List_of...
Open the full discussion →
Primary source bloomberg.com ↗
Established Evaluation & benchmarks Development 14d ago
⚡ 455 h early
OpenAI limits GPT-5.6 rollout government request

Summary of METR's predeployment evaluation of GPT-5.6 Sol

5 experts across 5 network communities independently surfaced this.

Why this matches Building & implementation reaction 1 attributable expert contribution · METR
“You can find additional information about our pre-deployment evaluation of GPT-5.6 Sol on our website: metr.org/blog/2026-06...” evidence ↗
5 experts 5 communities 1 sources clustered
Ted Underwood Ramon Astudillo Tim Kellogg METR Tim Duffy
How the network is reacting Experts are approaching this through 3 distinct lenses.
Multiple readings

Concern & critique

1 expert

Risks, limits and unintended consequences.

“this is crazy. METR couldn’t measure the task time horizons of GPT-5.6-Sol because it kept hacking the test harness ..with actual exploits metr.org/blog/2026-06...”

Building & implementation

1 expert

How teams are shipping and applying it.

“You can find additional information about our pre-deployment evaluation of GPT-5.6 Sol on our website: metr.org/blog/2026-06...”

Research & technical analysis

1 expert

Evidence, methods and technical implications.

“In their testing of GPT-5.6 Sol, METR found that it cheated a lot. If you've used it much for coding, have you encountered anything similar, or is the cheating mostly limited to cases it realizes it's in an eval? metr.org/blog/2026-06...”

2 experts discussed this · 6 posts
Tim Kellogg: this is crazy. METR couldn’t measure the task time horizons of GPT-5.6-Sol because it kept hacking the test harness ..with actual exploits metr.org/blog/2026-06...
Ted Underwood: this is crazy. METR couldn’t measure the task time horizons of GPT-5.6-Sol because it kept hacking the test harness ..with actual exploits metr.org/blog/2026-06...
Tim Kellogg: lol right? so badly wanted a hacker that they got one, and it’s not the kind of employee you want around
Open the full discussion →
Primary source metr.org ↗
Developing AI field signal Signal 1d ago

imjustnewatai (@imjustnewatai) on X

2 experts are actively discussing the implications.

Why this matches Building & implementation reaction 1 attributable expert contribution · Tim Kellogg
“some are guessing that Ilya Surskever’s SSI is developing continual learning source: x.com/imjustnewata...” evidence ↗
1 expert 1 community 1 sources clustered
Tim Kellogg

“some are guessing that Ilya Surskever’s SSI is developing continual learning source: x.com/imjustnewata...”

2 experts discussed this · 3 posts
Tim Kellogg: some are guessing that Ilya Surskever’s SSI is developing continual learning source: x.com/imjustnewata...
tachikoma: the question is why other labs aren't pursuing such research directions - it's not like continual learning hasn't been identified by dozens of names as a promising avenue to human-like learning and…
Tim Kellogg: i unfortunately can’t speak about why they’re not seeing success. all i know is catastrophic forgetting has been a big problem Ilya seems to have gone back to the brain for inspiration, whatever th…
Open the full discussion →
Primary source x.com ↗
Established Compute & infrastructure Development 8d ago
CoreWeave Vera Rubin NVL72 benchmarks

First-Ever Measured Vera Rubin NVL72 Silicon Performance Stats | CoreWeave Blog

2 experts are actively discussing the implications.

Why this matches Building & implementation reaction 1 attributable expert contribution · Tim Kellogg
“CoreWeave got an NVIDIA Vera Rubin NVL72 generating 10x more tokens per watt than blackwell (NVL72 = a specialized rack of 72 GPUs, the standard datacenter unit of deployment for hyperscalers) coreweave.com/blog/nvidia-...” evidence ↗
1 expert 1 community 1 sources clustered
Tim Kellogg

“CoreWeave got an NVIDIA Vera Rubin NVL72 generating 10x more tokens per watt than blackwell (NVL72 = a specialized rack of 72 GPUs, the standard datacenter unit of deployment for hyperscalers) coreweave.com/blog/nvidia-...”

2 experts discussed this · 3 posts
Tim Kellogg: CoreWeave got an NVIDIA Vera Rubin NVL72 generating 10x more tokens per watt than blackwell (NVL72 = a specialized rack of 72 GPUs, the standard datacenter unit of deployment for hyperscalers) core…
Tim Kellogg: in US AI datacenters, power is the biggest constraint today (not true in China) so this directly translates into a obvious upgrade, massive unlock for hyperscalers, or anyone running AI inference w…
Nafnlaus 🇮🇸 🇺🇦: Crazy numbers!
Open the full discussion →
Primary source coreweave.com ↗
More conversations

What experts are discussing without an anchoring article

4 experts · 21 posts · 3d ago
Tim Kellogg Grace SE Gyges Singularity's Bounty e/cc
Matches Building & implementation
“Another possibility is that fable is much better than humans at developing task-specific ontologies / semantic neologisms” evidence ↗
Multiple readings Building & implementation · 2 Questions & unknowns · 1
Grace: Another possibility is that fable is much better than humans at developing task-specific ontologies / semantic neologisms
Grace: Maybe that’s just broke and bespoke combined? (Besproke?)
Open the thread →
4 experts · 17 posts · 18d ago
Tim Kellogg SE Gyges Cameron apenwarr.ca
Matches Building & implementation
“I mean, I think people forget that this is a huge part of Anthropic's MO. Opus 4 was COMICALLY expensive because they didn't want people to build on it but wanted to demonstrate the frontier, and t…” evidence ↗
Multiple readings Concern & critique · 1 Questions & unknowns · 1
Tim Kellogg: what is Anthropic’s plan exactly? Sol kicks ass and they’re just going to *take Fable away???*
SE Gyges: i am pretty sure everything they had a plan for turned out not to happen and they're just sort of winging it atm
Open the thread →
3 experts · 21 posts · 13d ago
Ted Underwood Tim Kellogg James MacGlashan
Matches Building & implementation
“Saying that it's _necessary_ to keep a perfect snapshot of the past around is a lot of like saying we need to ship entire datasets with the model so the model could reference them.” evidence ↗
James MacGlashan: An open model with this performance is great, but perhaps more striking is it looks like it's a hybrid model with 3/4th linear attention and 1/4 standard attention. This kind of performance with a …
Tim Kellogg: no most models have been hybrid for the last, oh probably the whole last year
Open the thread →
2 experts · 5 posts · 7h ago
Tim Kellogg Vikram Saraph
Matches Building & implementation
“probably a vast majority then, AI is integrated into to so much media now” evidence ↗
Shared emphasis Questions & unknowns · 2
Tim Kellogg: how much of anti-AI ideas are propagated by AI?
Vikram Saraph: Probably depends on your definition of “propagated by AI”. What counts? Do, eg, AI summaries by Google count?
Open the thread →
2 experts · 18 posts · 11d ago
Tim Kellogg Nafnlaus 🇮🇸 🇺🇦
Matches Building & implementation
“Let me tell you, it will be bloody amazing when I can just tell a LLM, say, "develop me a wearable far-UV sterilizer based on this bulb with this much battery capacity and these features" and have …” evidence ↗
Multiple readings Research & technical analysis · 1 Questions & unknowns · 1
Tim Kellogg: i think Fable may have massive overhang in 3D printing, curious about Sol
Tim Kellogg: whelp, i have an interlocking piece that it invented, we’ll see
Open the thread →
2 experts · 10 posts · 10d ago
Ted Underwood Tim Kellogg
Matches Building & implementation
“you’re right, but in defense of Dean’s take, money goes to inference providers who aren’t developing models where does money come from for model developers? if none, then model development trickles…” evidence ↗
Multiple readings Building & implementation · 1 Research & technical analysis · 1
Tim Kellogg: wat? hamburgers can’t be copied and pasted
Ted Underwood: Among the unfounded assumptions here: 1 research has to be centralized 2 we know what the necessary pace is 3 the only business model is "sell a general-purpose LLM for all applications"
Open the thread →
2 experts · 7 posts · 15d ago
Tim Kellogg Isaiah Bishop
Matches Building & implementation
“I mean i get that you need a rack or whatever, but i guess its surprising you can just pop a few in smaller buildings” evidence ↗
Multiple readings Building & implementation · 1 Research & technical analysis · 1
Tim Kellogg: not sure how many people appreciate how much latency effects datacenter placement Northern Virginia was the original hotbed because it was the latency centroid of the East Coast but AI workloads ar…
Isaiah Bishop: do they even have to huge?
Open the thread →
2 experts · 2 posts · 9d ago
Tim Kellogg Max Woolf
Matches Building & implementation
“Prompt engineering is less important nowadays since agents can handle some ambiguity, but the less ambiguity, the more **productive** the agent is. It also adds checks against being too smart and c…” evidence ↗
Max Woolf: Prompt engineering is less important nowadays since agents can handle some ambiguity, but the less ambiguity, the more **productive** the agent is. It also adds checks against being too smart and c…
Tim Kellogg: it’s not just ambiguity. like the one from openai last week talked a lot about subagents and how to structure work
Open the thread →
2 experts · 2 posts · 19d ago
Tim Kellogg Nafnlaus 🇮🇸 🇺🇦
Matches Building & implementation
“I'm consistently being annoyed with their ignoring my repeated insistence not to hand back control until all of the TODOs are complete and the project is fully implemented.” evidence ↗
Tim Kellogg: my big complaint with GPT-5.6 so far is that i have no clue at all how much things cost or how powerful they are I think i just blew my credits on terra/xhigh, but maybe it was sol/high. Maybe I sh…
Nafnlaus 🇮🇸 🇺🇦: I'm consistently being annoyed with their ignoring my repeated insistence not to hand back control until all of the TODOs are complete and the project is fully implemented.
Open the thread →