Lightnews — Scholar-powered news

Jess Hamrick @jhamrick.bsky.social · 6d

Also, some people don't have mental imagery at all (aphantasia)! My conclusion based on the evidence is that it's we do some form of latent and/or piecemeal simulation but it's definitely not pixel perfect.

2

Jess Hamrick @jhamrick.bsky.social · 6d

I used to research intuitive physics. There's examples where people's predictive ability seems quite good (e.g. www.pnas.org/doi/10.1073/...) but also many mistakes and reported limitations (e.g. www.nature.com/articles/s41..., www.sciencedirect.com/science/arti...)

Simulation as an engine of physical scene understanding | PNAS

In a glance, we can perceive whether a stack of dishes will topple, a branch will support a child’s weight, a grocery bag is poorly packed and liab...

www.pnas.org

2 3

Reposted by Jess Hamrick

Kunal Jha @kjha02.bsky.social · 9d

Forget modeling every belief and goal! What if we represented people as following simple scripts instead (i.e "cross the crosswalk")?

Our new paper shows AI which models others’ minds as Python code 💻 can quickly and accurately predict human behavior!

shorturl.at/siUYI%F0%9F%...

3 14 36

Jess Hamrick @jhamrick.bsky.social · 10d

This is so so cool! I tried to build an AI system to do a variation of the Finke task waaay back in... 2016 or 2017? It didn't work very well, hah. (It was a combination of Bayesian inference over the structured representation with a CNN recognition model). Amazing that LLMs are able to do this.

1 4

Reposted by Jess Hamrick

Jorge Morales @jorge-morales.bsky.social · 11d

Imagine an apple 🍎. Is your mental image more like a picture or more like a thought? In a new preprint led by Morgan McCarty—our lab's wonderful RA—we develop a new approach to this old cognitive science question and find that LLMs excel at tasks thought to be solvable only via visual imagery. 🧵

Artificial Phantasia: Evidence for Propositional Reasoning-Based Mental Imagery in Large Language Models

This study offers a novel approach for benchmarking complex cognitive behavior in artificial systems. Almost universally, Large Language Models (LLMs) perform best on tasks which may be included in th...

arxiv.org

5 36 110

Jess Hamrick @jhamrick.bsky.social · 11d

I think your links got messed up, the paper is here: github.com/NVlabs/RLP/b...

1

Reposted by Jess Hamrick

Sung Kim @sungkim.bsky.social · 11d

Nvidia's RLP: Reinforcement Learning Pretraining—information-driven, verifier-free objective that teaches models to think before they predict

🔥+19% vs BASE on Qwen3-1.7B
🚀+35% vs BASE on Nemotron-Nano-12B

2 1 26

Reposted by Jess Hamrick

Mark Riedl @markriedl.bsky.social · 12d

Final blog post on the visualization medium.com/@mark-riedl/...

Visualizing the Influence of Federal Funding on the AI Boom

The Transformer was invented in Google. Reinforcement Learning with Human Feedback (RLHF) was not invented in industry labs, but is most…

medium.com

1 2 10

Reposted by Jess Hamrick

Ted Underwood @tedunderwood.com · 12d

Resharing this, because it's proving valuable enough that I spent 10 minutes looking it up. TLDR: It's true that some famous recent papers in AI were produced in the private sector. But they *cite* lots of papers with academic authors and federal funding.

Mark Riedl @markriedl.bsky.social · Apr 15

Continuing to fiddle. I've discovered that data visualization is extremely susceptible to "procrastiwork". I should just stop and release my notebook.

1 4 31

Reposted by Jess Hamrick

Nathan Lambert @natolambert.bsky.social · 12d

Nice to see another fully open, multimodal LM released! Good license, training code, pretraining data, all here.
LLaVA-OneVision-1.5: Fully Open Framework for Democratized Multimodal Training

Slowly, the community is growing.
arxiv.org/abs/2509.236...

9 50

Reposted by Jess Hamrick

Case🍂 @case.cryptoanarchy.network · 13d

Claude Sonnet 4.5 dropped www.anthropic.com/news/claude-...
assets.anthropic.com/m/12f214efcc...

Introducing Claude Sonnet 4.5

Claude Sonnet 4.5 is the best coding model in the world, strongest model for building complex agents, and best model at using computers.

www.anthropic.com

1 2 8

Reposted by Jess Hamrick

Andy Tseng @andytseng.bsky.social · 13d

Interesting deep dive for anyone curious about how #ClaudeCode is built.
#Anthropic #Claude #GenAI #DevOps #VibeCoding

How Claude Code is built

A rare look into how the new, popular dev tool is built, and what it might mean for the future of software building with AI. Exclusive.

newsletter.pragmaticengineer.com

1 6 9

Reposted by Jess Hamrick

Ozlandclone 🌪️ 🩵💛🩵 @ozlandclone.bsky.social · 13d

It's aster time! I have never seen so many monarchs, bumble bees and other pollinators in my yard. I had 6 monarchs on my New England aster at one time! Maybe they are on their migration. Anyway, shows how important late season natives are. 🌱 #nativeplants, #pollinators

3 19 110

Reposted by Jess Hamrick

Alexander Doria @dorialexander.bsky.social · 13d

I know there is a lot of competition today, but this might be most consequential release for people training models: in-depth exploration of full-finetuning, lora, RL efficiencies by John Schulman (ThinkingMachines). thinkingmachines.ai/blog/lora/

3 7 54

Reposted by Jess Hamrick

The Green Party of England & Wales @greenparty.org.uk · 14d

🎉 80,000 Green Party members!

📈 But we're not stopping there.

💚 We have no time to waste. Join the Green Party today ⤵️

19 230 630

Reposted by Jess Hamrick

Sam Gershman @gershbrain.bsky.social · 14d

I was part of an interesting panel discussion yesterday at an ARC event. Maybe everybody knows this already, but I was quite surprised by how "general" intelligence was conceptualized in relation to human intelligence and the ARC benchmarks.

2 3 23

Reposted by Jess Hamrick

Alonso Silva @alonsosilva.bsky.social · 26d

Want to visualize the response format constraints on the LLM when working in a Jupyter notebook?
Then you might be interested in my new project `litelines`.
Litelines lets you visualize the selected path by the LLM.
It supports a Pydantic schema as a response format, as well as regular expressions.

1 4 8

Reposted by Jess Hamrick

Tim Kellogg @timkellogg.me · 15d

sheesh! AI bluesky has arrived

not just good content, there’s more and more original work, people from labs, and people with genuinely interesting perspectives

when i joined, it was so painful trying to find even traces

6 9 140

Reposted by Jess Hamrick

Shawn Simister @narphorium.com · 15d

Teaching LLMs to Plan: Logical Chain-of-Thought Instruction Tuning for Symbolic Planning
arxiv.org/abs/2509.13351

2 8

Jess Hamrick @jhamrick.bsky.social · 15d

I was wondering about that too...

1

Reposted by Jess Hamrick

Mor Naaman @informor.bsky.social · 15d

For instructors out there: a very cool set of AI Course Policy Icons by Cornell's GenAI taskforce, to be used in combination for your syllabi or assignments.

Inspired by @creativecommons.bsky.social and available w/ a CC license. Sample icons attached here.

teaching.cornell.edu/generative-a...

ANY-AI icon: (Any Tool, Any Use, Any Time) to use in policy of GenAI use in a course

AT (Approved Tools Only) icon to use in policy of GenAI use in a course

UA (Use with Attribution) icon to use in policy of GenAI use in a course

AS (Assignment-Specific) icon to use in policy of GenAI use in a course

2 5

Jess Hamrick @jhamrick.bsky.social · 15d

For others like me who might be unsure what this is about www.eu-inc.org has some further details

Why the EU–INC?

Europe has the talent, ambition, and ecosystems to create innovative companies, but fragmentation between European nations is holding us back.

"A startup from California can expand and raise money all across the United States. But our companies still face way too many national barriers that make it hard to work Europa-wide, and way too much regulatory burden."

– Ursula von der Leyen, Oct 2024

1 2

Reposted by Jess Hamrick

Thomas Wolf @thomwolf.bsky.social · 15d

EU–INC is the single best thing Europe could do to catch-up in the AI race

A simple unified pan-European startup structure, with modern employee ownership and simple access to capital, able to tap into Europe’s full talent pool.

‼️ but it’s at high risk of not seeing the light of day. You can help👇

3 13 37

Reposted by Jess Hamrick

Alexander Doria @dorialexander.bsky.social · 15d

And new paper out: Pleias 1.0: the First Family of Language Models Trained on Fully Open Data

How we train an open everything model on a new pretraining environment with releasable data (Common Corpus) with an open source framework (Nanotron from HuggingFace).

www.sciencedirect.com/science/arti...

8 50 170

Reposted by Jess Hamrick

Scott McGrath @smcgrath.phd · 15d

This is a really good thread about forecasting too far into the future for medical AI, and what the “we should stop training doctors/lawyers” crowd is missing.

Tagging for #MedSky #MLSky

Deena Mousa @deenamousa.com · 17d

In 2016 Geoffrey Hinton said “we should stop training radiologists now" since AI would soon be better at their jobs.

He was right: models have outperformed radiologists on benchmarks for ~a decade.

Yet radiology jobs are at record highs, with an average salary of $520k.

Why?

1 3 12