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.

2,365 searchable experts 2,967 tracked across all sources
Clear
Showing signals surfaced by Leshem (Legend) Choshen @EMNLP ×
Debated

The questions experts are actively pulling apart

One card per development. Sources are clustered; expert reactions remain attributable.

Developing AI research Research 1d ago
⚡ 26 h early
automated discovery no universal harness paper

Automated Discovery Has No Universally Superior Harness

2 experts are actively discussing the implications.

2 experts 2 communities 1 sources clustered
Leshem (Legend) Choshen @EMNLP arxiv cs.CL

“Paper : arxiv.org/abs/2607.18235 Github : github.com/akshat57/har...”

“Akshat Gupta, Jermaine Lei, Alexander Lu, Gopala Anumanchipalli, Leshem Choshen Automated Discovery Has No Universally Superior Harness https://arxiv.org/abs/2607.18235”

2 experts discussed this · 15 posts
Leshem (Legend) Choshen @EMNLP: OpenEvolve underperforms simple autmated discovery harnesses the rest are insignificant from each other. The best choice changed across model–problem pairs. We ran a controlled study (3m+ rollouts)…
Leshem (Legend) Choshen @EMNLP: Automated discovery has high run-to-run variance, yet harnesses are often evaluated with only 3–5 runs. When a harness performs better, how do we know it is genuinely better—and not simply lucky?
Leshem (Legend) Choshen @EMNLP: To answer this, we systematically evaluated 30 budget-matched harnesses across 12 model–problem pairs, using repeated-trial statistical analysis. We found :
Open the full discussion →
Primary source arxiv.org ↗
Established AI field signal Resource 18d ago
vamsin07 Whisper ASR tools release

GitHub - vamsin07/multilingual-bpe-tokenizers: Whisper-compatible per-lang byte-level BPE tokenizers (recipe + samples) · GitHub

2 experts are actively discussing the implications.

2 experts 1 community 2 sources clustered
Leshem (Legend) Choshen @EMNLP austegard.com

“Website: vamsin07.github.io/buzzasr-docs/ Tokenizers: github.com/vamsin07/mul... SFT models: github.com/vamsin07/whi...”

2 experts discussed this · 4 posts
Leshem (Legend) Choshen @EMNLP: Sota on 89 languages speech models. There's plenty of speech data it appears, so the simplest fine-tuning plus tokenization on public data just improves everything substantially. And it's not even …
Leshem (Legend) Choshen @EMNLP: What they did? Cleaned data, took a big open speech model (whisper) changed the tokenizer and fine-tuned per data.
Leshem (Legend) Choshen @EMNLP: Let's help them: there's a lot more data that needs cleaning, if we do we can train on a scale larger of models and create much much better models (GPU we'll figure out). Ones that are likely to be…
Open the full discussion →
Primary source github.com ↗ vamsin07.github.io ↗
More conversations

What experts are discussing without an anchoring article

3 experts · 5 posts · 8d ago
Leshem (Legend) Choshen @EMNLP David Marx The Data Therapist in the Blue Sky
Leshem (Legend) Choshen @EMNLP: I've been told multilingual models are worse in Arabic, but not that they are worse at understanding a simple (English?) table, or understand columns or strategize. It is like all I've been told ab…
Leshem (Legend) Choshen @EMNLP: Is Gemma 4 multilingual? Not really🤖 A true multilingual LLM should share language-invariant skills like spatial understanding across languages. It doesnt😱 We had LLMs play 2D board games against t…
Open the thread →