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
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.

Research & technical analysis
1 live reaction map in this view Jump to the evidence ↓
Reaction converging Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights Research & technical analysis · 7 attributable reactions
Showing signals surfaced by Gus ×
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 AI research Research 2d ago
LLMs read write neural network weights

Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights

16 experts across 6 network communities independently surfaced this.

16 experts 6 communities 1 sources clustered
Mark Riedl Ted Underwood Ethan Mollick Ramon Astudillo Tim Kellogg Gus Alejandra Caraballo David Marx +8
How the network is reacting 7 experts are independently emphasizing research & technical analysis.
Reaction converging

Research & technical analysis

7 experts

Evidence, methods and technical implications.

“This is fascinating. Very large models can read the weights of small models and intuit what those models were trained to do, and make direct weight changes to improve them. rikiyat.github.io/blog/weight-...”

4 experts discussed this · 5 posts
Ethan Mollick: If this holds up, it is big: "We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x …
Antonino Greco: If this holds up, it is big: "We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x …
Thomas Dietterich: Can someone provide a plausible mechanism for this? Has GPT-6 Astra been trained on neural network weight (or activation) interpretation?
Open the full discussion →
Primary source rikiyat.github.io ↗
Developing Models & releases Resource 2d ago
⚡ 187 h early
Google LiteRT-LM on-device AI release

LiteRT-LM Overview  |  Google AI Edge  |  Google for Developers

3 experts across 2 network communities independently surfaced this.

3 experts 2 communities 1 sources clustered
Google Gemma Gus Evangelos Kazakos

“Access the updated Gemma 4 12B file: https://t.co/N8QOgLNc5C Quickly try Gemma using LiteRT-LM CLI: https://t.co/E4e6Quyu9D Learn more about building with LiteRT-LM: https://t.co/dmBCWFdbEB”

“🤖💎 Reachy on Wheels runs a 100% local, battery-powered robotics loop on LiteRT: Gemma 4 E4B handles voice chat + start/stop tool calls, while EmbeddingGemma 2 steers on an Arduino Ventuno Q from live camera input with zero cloud calls! https://developers.go…”

2 experts discussed this · 4 posts
Gus: 🤖💎 Reachy on Wheels runs a 100% local, battery-powered robotics loop on LiteRT: Gemma 4 E4B handles voice chat + start/stop tool calls, while EmbeddingGemma 2 steers on an Arduino Ventuno Q from li…
Evangelos Kazakos: 🤖💎 Reachy on Wheels runs a 100% local, battery-powered robotics loop on LiteRT: Gemma 4 E4B handles voice chat + start/stop tool calls, while EmbeddingGemma 2 steers on an Arduino Ventuno Q from li…
Michiel Bontenbal: Love it. I have used a lot of embedding models but did not see that it can be used for steering. :-)
Open the full discussion →
Primary source developers.google.com ↗
Established AI field signal Resource 16d ago
Bluesky image search tool launched

Bluesky Image Search

2 experts are actively discussing the implications.

2 experts 1 community 1 sources clustered
Gus austegard.com

“Ok - the fruits of my latest vibe AI engineering are ready: An all in-browser vector index (<4MB) + custom 4-layer EG2-compatible query embedder (~25MB) + custom WGSL runtime (KBs) of 16,878 Bluesky images, weighing in at a combined < 30MB INSTANT search as…”

2 experts discussed this · 6 posts
austegard.com: FInal writeup: github.com/oaustegard/o...
austegard.com: We COULD (and briefly did) use the standard (and generally excellent) onnxruntime-web to run the loaded query model, but why not go for broke with a custom WGSL runner? Results: github.com/oaustega…
austegard.com: The (relatively) hard part: Tearing EG2 apart and trying to minimize the size without losing too much accuracy in the process. Ended up with a 4-layer text-only version of the model distilled with …
Open the full discussion →
Primary source austegard.com ↗
Established Models & releases Resource 15d ago
vLLM project pull requests activity

Pull requests · vllm-project/vllm

2 experts are actively discussing the implications.

1 expert 1 community 2 sources clustered
Gus

“👀💎 A Google engineer paired with Gemini 4 Argon in an iterative profiling and kernel-optimization loop to pinpoint the exact vLLM scheduling bottlenecks and land all 8 upstream PRs in days. Full PR list on GitHub: https://goo.gle/4zNtANl”

2 experts discussed this · 5 posts
Gus: 🚀🛠️ Huge vLLM throughput jump on Gemma 4 E2B: 2.85x faster on a single A100, climbing from ~37,900 to 108,000+ tok/s with 100% token parity and zero quality drop! Landed across 8 upstream PRs in da…
Gus: 👀💎 A Google engineer paired with Gemini 4 Argon in an iterative profiling and kernel-optimization loop to pinpoint the exact vLLM scheduling bottlenecks and land all 8 upstream PRs in days. Full PR…
spacecowboy17.bsky.social: Is there a way to tell how much more room is left for optimization with zero quality tradeoffs and not counting speculative decoding? Is the theoretical limit how long it takes to feed network weig…
Open the full discussion →
Primary source github.com ↗ goo.gle ↗
More conversations

What experts are discussing without an anchoring article

3 experts · 10 posts · 19d ago
Gus austin Nafnlaus 🇮🇸 🇺🇦
Gus: I replied to a Gemma user asking what they would like to see in future Gemma models, and somehow this became my most popular post ever! With great insights and requests. Thanks everyone! 🤩🙏🏼 I trie…
Nafnlaus 🇮🇸 🇺🇦: Something dense in the ~8-9B range would be great. The best thing out there right now is Qwen 3.5 9B, which is several generations old. Builtin speculative decoding would be great :)
Open the thread →
2 experts · 7 posts · 9d ago
Tim Kellogg Gus
Tim Kellogg: yes, its a qwen fine tune. yes, LLMs and dLLMs are both language models, so there’s no reason you can’t cutover from LLM->dLLM
Tim Kellogg: which is quite a thing to grapple with if you’re in the “it’s just predicting the next token” camp if you can just switch it over to generating all tokens at once, doesn’t that demonstrate a deeper…
Open the thread →
2 experts · 2 posts · 2d ago
Gus Dare Obasanjo
Dare Obasanjo: OpenAI just poached the product lead for Claude Code. I was wondering why a Claude Code product leader would leave pre-IPO Anthropic shares on the table. The answer was pre-IPO OpenAI shares.
Gus: ok, now it makes sense!
Open the thread →
2 experts · 2 posts · 3d ago
Gus Dare Obasanjo
Dare Obasanjo: Following the departure of two engineers from the Claude Code team in October, product lead Cat Wu has also left Anthropic ahead of its IPO. The exodus from pre-IPO frontier labs this year was not …
Gus: Why would they leave? I can't understand that
Open the thread →
2 experts · 2 posts · 12d ago
Sung Kim Gus
Sung Kim: alexandr wang said that for every three “self-driving” cars, there is one human operator overseeing them. So, the idea that AI will simply take our jobs is a fallacy. AI may eliminate some jobs, bu…
Gus: I hope so too. I think AI is pretty helpful, but if we push for taking people's job only it will be bad for everyone, even the big companies laying people off.
Open the thread →