Lightnews — Scholar-powered news

Avijit Ghosh @evijit.io · 15h

Some of these new gpt/claude wrapper startups make me wonder how much the founders are paying themselves in salary because there is no way they expect their horrible idea to actually be sustainably profitable

2 8

Avijit Ghosh @evijit.io · 1d

HF papers link: huggingface.co/papers/2509....

Paper page - AI for Scientific Discovery is a Social Problem

Join the discussion on this paper page

huggingface.co

Avijit Ghosh @evijit.io · 1d

Discord Server: discord.com/invite/VYkdE...

Join the huggingscience Discord Server!

Hop in the #projects forum to get involved immediately!This is where we are going to change the game for open-source AI for science. No more things behind NDAs! Will launch projects soon. | 5704 membe...

discord.com

1

Avijit Ghosh @evijit.io · 1d

Hugging Science Initiative:
huggingface.co/hugging-scie...

hugging-science (Hugging Science)

Org profile for Hugging Science on Hugging Face, the AI community building the future.

huggingface.co

1

Avijit Ghosh @evijit.io · 1d

Paper: arxiv.org/abs/2509.06580

AI for Scientific Discovery is a Social Problem

Artificial intelligence promises to accelerate scientific discovery, yet its benefits remain unevenly distributed. While technical obstacles such as scarce data, fragmented standards, and unequal acce...

arxiv.org

1

Avijit Ghosh @evijit.io · 1d

We are launching Hugging Science: A global community addressing these barriers through:
✅ Collaborative challenges targeting upstream problems
✅ Cross-disciplinary education
✅ Recognition for data & infrastructure work
✅ Community-owned infrastructure

All links follow 🤗

1

Avijit Ghosh @evijit.io · 1d

4. Data and access inequities: Scientific data remain fragmented and difficult to harmonize. Compute and infrastructure gaps, especially in the Global South, extend iteration cycles and limit participation in data-intensive research.

1

Avijit Ghosh @evijit.io · 1d

3. Collaboration barriers: Machine learning and domain sciences operate with distinct goals and epistemic norms. Without shared success criteria or communication frameworks, many interdisciplinary projects fail before integration.

1

Avijit Ghosh @evijit.io · 1d

2. Misaligned incentives: Research culture rewards short-term model performance over durable infrastructure. High-quality datasets often have decades of impact, yet curation and maintenance work receive limited institutional recognition.

1

Avijit Ghosh @evijit.io · 1d

We find 4 main issues in AI4Science:

1. "AI Scientist": Framing discovery as a task for autonomous systems overlooks the interpretive and theoretical nature of science. Predictive success can mask conceptual errors, as models may fit data while learning incorrect mechanisms.

1

Avijit Ghosh @evijit.io · 1d

AI for scientific discovery is a social problem: In our new position paper, @cgeorgiaw.bsky.social and I show that culture, incentives, and coordination are the main obstacles to progress, and we are launching the Hugging Science Initiative to address this!

1 1 4

Avijit Ghosh @evijit.io · 5d

Unclear if this is necessarily true, a lot of tech companies are in the anti AI blocklist that people subscribe to without reading the nuances so bsky is less popular in the tech world than it should be (I wish it weren’t so because it’s so much less toxic than X)

3 3

Avijit Ghosh @evijit.io · 6d

At a panel a couple of weeks ago I exclaimed in despair: “who decided that the only mode of interacting with AI is via chatbots?” And this trend is relentless still.

www.theverge.com/news/787076/...

Microsoft launches ‘vibe working’ in Excel and Word

Vibe working is all about Office’s new Agent Mode.

www.theverge.com

3

Avijit Ghosh @evijit.io · 7d

Can price fall in meters tho

1 1

Avijit Ghosh @evijit.io · 8d

If you go into advanced settings there’s an update to the canvas feature that can directly edit documents and run code, maybe that

Reposted by Avijit Ghosh

Jordan! @jordanjamboree.bsky.social · 9d

Wife brought her 100 year old film camera to that pirates game from a few weeks ago and she got this shot that's just perfect

A black and white photo of a baseball game, you can see that it is mid pitch with a pirate at bat

140 1.3K 8.4K

Avijit Ghosh @evijit.io · 9d

Maybe if we could more permanently log in to arxiv and indicate our preferences there? I notice that I get logged out very frequently

1

Avijit Ghosh @evijit.io · 9d

cvpr.thecvf.com/Conferences/....

CVPR 2025 Changes

cvpr.thecvf.com

1 1

Avijit Ghosh @evijit.io · 9d

Right, or in other words, experts in a field need to document their tips and tricks so that others can follow. Gell-Mann amnesia otherwise would just decrease productivity: bsky.app/profile/amag...

Amazing Digital Carcass @amagire.bsky.social · 12d

"of course ChatGPT is full of errors when it comes to my specific area of professional or technical expertise, that's why I only use it for other stuff, where I can't tell if it's bullshit or not"

📰🗞💥

2

Avijit Ghosh @evijit.io · 9d

This works great! Wish there was a way to make it work on my phone as well

1 1

Avijit Ghosh @evijit.io · 9d

AI x Labor is such a nuanced topic that people refuse to meaningfully engage with. There's a lot to be covered between "ban AI" or "use AI otherwise you're fired"!

2

Avijit Ghosh @evijit.io · 9d

For my coding work, diplomatic word choices are less important, but then more important is to constantly remind the model to write unit tests and dummy functions, otherwise you end up with 3000 lines of code without knowing where the bug is. It comes from experience that needs to be shared.

1 2

Avijit Ghosh @evijit.io · 9d

[Also, people are not all the same - someone might produce better outputs with excel than with AI, and therefore success metrics should be individualized in the workplace by giving employees more agency within general parameters, but that is a more complex topic for another day].

2

Avijit Ghosh @evijit.io · 9d

It is extremely context and job dependent - for my own work I generally only use AI for creating the structure of documents but then have to go through stuff line by line as word choices are important in policy docs. For my coding work, less so.

Teams should document best practices over time, imo

2 2

Avijit Ghosh @evijit.io · 9d

The missing puzzle piece, imo, is training people to actually understand the limitations of models to use it as a productivity tool; even the CVPR guidelines specifically scope out what level of usage is allowed. One can only hope that there are internal team-level guidelines.

3 4