Who's Who of AI

Juan Diego Rodriguez

419 trust researcher @juand-r.bsky.social · 4,761 followers
CS PhD student at UT Austin in #NLP Interested in language, reasoning, semantics and cognitive science. One day we'll have more efficient, interpretable and robust models! Other interests: math, philosophy, cinema https://www.juandiego-rodriguez.com/
What they're sharing

Articles & links

↻ Juan Diego Rodriguez reposted
@zoltanv.bsky.social

wow, the Huggingface guys made an annotated version of Magnifica Humanitas huggingface.co/spaces/socie...

The Annotated Encyclical - a Hugging Face Space by society-ethics huggingface.co
AI Weekly's analysis →
  • Hugging Face's society-ethics team added 105 scholarly AI ethics citations to Pope Leo XIV's May 2026 encyclical Magnifica Humanitas.
  • The encyclical declares technology 'is never neutral' and warns that 'never has humanity had such power over itself.'
  • The annotation project is open for community contributions linking academic AI ethics research to the papal document.
Read full analysis →
 View on Bluesky →

2) Characterizing Narrative Content in Web-scale LLM Pretraining Data, by @teagrjohnson.bsky.social, @elliottash.bsky.social, @andrewpiper.bsky.social, @mariaa.bsky.social arxiv.org/abs/2606.19468 Why: annotation and analysis of narrative features across the pretraining data (…

Characterizing Narrative Content in Web-scale LLM Pretraining Data arxiv.org
AI Weekly's analysis →
  • A new arXiv preprint introduces NarraBERT, a RoBERTa-based classifier, and applies it to 3 million passages from the 3-trillion-token Dolma corpus.
  • The framework operationalizes three narrative elements, agency, setting, and events, across 11 interpretable dimensions, trained on 400 annotated passages.
  • The authors report narrative qualities are unequally distributed across pretraining sources and topics in ways current curation practices do not measure.
Read full analysis →
View on Bluesky · ♥ 4 ↻ 1 ↩ 1 · 3 from the directory shared this · 9d ago

An incomplete list: 1) Explaining Attention with Program Synthesis, by Amiri Hayes, @belindazli.bsky.social and arxiv.org/abs/2606.19317 Why: LMs can generate code that can explain/replace attention heads! 🤯 (Program synthesis 🤝 Interpretability)

[2606.19317] Explaining Attention with Program Synthesis arxiv.org
AI Weekly's analysis →
  • Under 1,000 LM-generated programs reproduce attention patterns in GPT-2, TinyLlama-1.1B and Llama-3B with above 75% average IoU on TinyStories.
  • Replacing 25% of attention heads with the synthesized programs raised average perplexity by only 16% while keeping downstream question-answering performance.
  • The pipeline computes a head's attention matrices, prompts a pretrained LM to write Python that reproduces them, then re-ranks by held-out accuracy.
Read full analysis →
View on Bluesky · ♥ 0 ↻ 0 ↩ 1 · 3 from the directory shared this · 9d ago

4) Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention arxiv.org/abs/2605.29548

Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention arxiv.org
AI Weekly's analysis →
  • The paper argues power-law scaling already implies a larger model will learn parts of the data distribution a smaller model cannot, even with infinite training data.
  • Pretraining experiments on OLMo models from 4M to 4B parameters found only the larger models learned the infrequent and complex tasks.
  • The proposed mechanism is reduced gradient interference: weaker common-task updates leave rare-task features intact in larger models.
Read full analysis →
View on Bluesky · ♥ 1 ↻ 0 ↩ 1 · 2 from the directory shared this · 9d ago

Does anyone know why MASS (arxiv.org/pdf/1905.02450) didn't take off, while BART did? (OK maybe no one remembers BART today, but it was popular around 2019-2022)

arxiv.org
View on Bluesky · ♥ 6 ↻ 1 ↩ 1 · 2 from the directory shared this · 10d ago
↻ Juan Diego Rodriguez reposted
@mariozechner.at

recommended reading von @mitsuhiko.at man kann den KI aspekt dabei vollends ignorieren, btw. lucumr.pocoo.org/2026/6/13/am...

Dangerous Technology For Americans Only lucumr.pocoo.org
AI Weekly's analysis →
  • Anthropic suspended Fable and Mythos access for foreign nationals, including its own foreign-national employees.
  • Ronacher argues the shift moved from barring hostile governments to using nationality itself as the access boundary.
  • Europe's structural dependency on US cloud, platforms, and AI feeds a talent drain Ronacher calls a self-reinforcing death spiral.
Read full analysis →
 View on Bluesky →
↻ Juan Diego Rodriguez reposted
@davidho.bsky.social

I think mathematics at this level is a human endeavor, and solving these problems is an exercise in creativity like working on a puzzle. We’re just losing our humanity by letting AI do all the fun stuff while we toil at our desks doing tedious stuff like filling out expense fo…

wsj.com View on Bluesky →

5) CREATE: Testing LLMs for Associative Creativity, arxiv.org/abs/2603.09970

arxiv.org
View on Bluesky · ♥ 0 ↻ 0 ↩ 1 · 9d ago
Their own posts

Recent commentary

Tip on reviewing ARR papers in 2026: - Check the references FIRST. I wasted 40 minutes reviewing a paper before realizing the references were LLM-generated.

View on Bluesky · ♥ 31 ↻ 12 ↩ 5 · 17d ago

I'm so tired of seeing LLM-generated reviews. They just wasting all our time. If this continues, "peer review" will mean nothing

View on Bluesky · ♥ 11 ↻ 1 ↩ 3 · 14d ago

Google AI: "oh I thought you meant Neuro-Linguistic Programming" I really thought we were over this! Apparently not

View on Bluesky · ♥ 5 ↻ 0 ↩ 1 · 10d ago

"Specification and Removal of Demons." Not what I expected to read in an NLP book! (Charniak and Wilks, Computational Semantics)

View on Bluesky · ♥ 7 ↻ 0 ↩ 0 · 11d ago

Claude is confidently telling me that Fable is a non-reasoning model “because it’s an Anthropic model” lol

View on Bluesky · ♥ 4 ↻ 0 ↩ 1 · 27d ago

Miyu Oba on “CxMP: A Linguistic Minimal-Pair Benchmark for Evaluating Constructional Understanding in Language Models”

View on Bluesky · ♥ 5 ↻ 0 ↩ 0 · 2d ago

Google AI: "I heavily hallucinated in that last response, completely made up details about the logo design and social events, and I am very sorry for spreading misinformation." I miss the old Google search

View on Bluesky · ♥ 4 ↻ 0 ↩ 0 · 1d ago

"An AI detector that actually works" lol

View on Bluesky · ♥ 1 ↻ 0 ↩ 1 · 16d ago

I tried to get Fable to make a system that turns textbooks into games. I think it's pretty funny that the demo with Patrick Murphy's "Probabilistic Machine Learning" was a RPG where the goal is to gather data and estimate P(doom)

View on Bluesky · ♥ 3 ↻ 0 ↩ 0 · 24d ago
Their network

In Juan Diego Rodriguez's orbit

Center = Juan Diego Rodriguez. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.