The Hard(ware) Part of AI - Awesome MLSS Newsletter 2nd Edition

AI is hard. Each day we're contending with hundreds of new models, "State of the Art" is now a bi-weekly update

But what about the hardware revolution? Check out our latest newsletter: newsletter.awesome-mlss.com/p/the-hard-w...

Join us as we explore "The Hard(ware) Part of AI" in our second issue

Interesting thread/piece.

One fun thing about this AGI-by-2027 stuff is, we get to find out of it's true, pretty soon!

💡 New ICLR paper! 💡
"On Linear Representations and Pretraining Data Frequency in Language Models":

We provide an explanation for when & why linear representations form in large (or small) language models.

Led by @jackmerullo.bsky.social, w/ @nlpnoah.bsky.social & @sarah-nlp.bsky.social

To all our followers, do remember to take time for yourself!

✈️ Headed to @iclr-conf.bsky.social — whether you’ll be there in person or tuning in remotely, I’d love to connect!

We’ll be presenting our paper on pre-training stability in language models and the PolyPythias 🧵

🔗 ArXiv: arxiv.org/abs/2503.09543
🤗 PolyPythias: huggingface.co/collections/...

Human-AI cooperation is important, but existing work trains on the same 5 Overcooked layouts, creating brittle strategies.

Instead, we find that training on billions of procedurally generated tasks trains agents to learn general cooperative norms that transfer to humans... like avoiding collision

Just shared a new article on "The State of Reinforcement Learning for LLM Reasoning"!
If you are new to reinforcement learning, this article has a generous intro section (PPO, GRPO, etc)
Also, I cover 15 recent articles focused on RL & Reasoning.

🔗 magazine.sebastianraschka.com/p/the-state-...

Digital Futures Institute announces eleven new Fellows for 2025-26 Eleven interdisciplinary projects awarded Fellowships from the Digital Futures Institute for the 2025-26 cohort.

👏 Meet our eleven new Digital Futures Institute Fellows!

Congratulations @lboungr.bsky.social, Rowan Boyson, Mark Cote, Amrita Dhillon, Alex Gould, Elisabeth Kelan, @niccoloridi.bsky.social, Gabriele Salciute Civiliene, Astrid Van den Bossche, @jamiewoodcock.bsky.social & @lorenzoz.bsky.social 🔽

awesome MLSS Newsletter Stay ahead with bi-weekly updates on new machine learning summer schools, upcoming deadlines, and latest research news.

SUMMER SCHOOL ANNOUNCEMENT

AI4Science Summer School deadline is in ONE DAY! Make sure you apply. For more information, go to awesome-mlss.com/summerschool...

We also have a biweekly newsletter for you. To subscribe, head on to newsletter.awesome-mlss.com/subscribe

QuestBench: Can LLMs ask the right question to acquire information in reasoning tasks? Recently, a large amount of work has focused on improving large language models' (LLMs') performance on reasoning benchmarks such as math and logic. However, past work has largely assumed that tasks a...

Test time inference has been the paradigm for 'reasoning' models. A new paper however asks if the models even know how to ask the right questions. arxiv.org/abs/2503.22674.

If anyone knows the author's @, please tag them

Speculating with a friend today about the @iclr-conf.bsky.social Test of Time award and we were trying to see if any papers could compete with Bahdanau et al (ie ATTENTION) but uh apparently we didn’t notice the just announced winner: Kingma & Ba (ADAM). Damn. ICLR 2015 was STACKED.

Very interesting findings, thank you @wissamantoun.bsky.social

SUMMER SCHOOL ANNOUNCEMENT

There are ML Summer Schools with deadlines less than ten days away! Follow deadlines here: awesome-mlss.com

Most urgent:

BAI Summer School on AI Agents and Agentic Systems 2025 - 1 day

AutoML School 2025 - 4 days

AI4Science Summer School 2025 - 6 days

Large language models are less effective at clinical prediction tasks than locally trained machine learning models AbstractObjectives. To determine the extent to which current large language models (LLMs) can serve as substitutes for traditional machine learning (ML) as

Large language models are less effective at clinical prediction tasks than locally trained machine learning models academic.oup.com/jamia/advanc... #llms #machinelearning #MLSky (not surprising but important to document)

The transformer was invented in Google. RLHF was not invented in industry labs, but came to prominence in OpenAI and DeepMind. I took 5 of the most influential papers (black dots) and visualized their references. Blue dots are papers that acknowledge federal funding (DARPA, NSF).

Just finished The AI Mirror by @shannonvallor.bsky.social. Very highly recommended. Transcending the stale accelerationist/doomer debate, Vallor reminds us what technology is for, and where (some of the) true dangers of AI lie. Powerful stuff, eloquently written global.oup.com/academic/pro...

Though to be fair, it didn’t do quatrains, so I asked for them

Now the paper is also on arxiv and can be easily cited:

arxiv.org/abs/2504.07128

Fun options: ask for an arcade game, a 3D game, a strategy game, etc.

And don’t be satisfied with the first result, push back and ask for improvements.

Paper page - SmolVLM: Redefining small and efficient multimodal models Join the discussion on this paper page

Today, we share the tech report for SmolVLM: Redefining small and efficient multimodal models.
🔥 Explaining how to create a tiny 256M VLM that uses less than 1GB of RAM and outperforms our 80B models from 18 months ago!
huggingface.co/papers/2504....

🤔 **Why join?**
1️⃣ Explore a rich, large-scale dataset for mobile sensing research.
2️⃣ Receive feedback for your work from an expert TPC.
3️⃣ Accepted papers will be invited to submit an extended version to @ieeepervasive.
4️⃣ Get the chance to win the Best Paper Award.

The 2025 AI Index Report | Stanford HAI

Stanford HAI releases 2025 AI index Report with cross domain data on how AI is impacting our lives hai.stanford.edu/ai-index/202... @stanfordhai.bsky.social

Brilliant work!

The deadline for the EoI is tomorrow (April 11th)!
We have:
✨6 confirmed speakers.
✨A session on interdisciplinary research.
✨A tutorial.
✨A panel discussion.
✨Poster and networking sessions.

AI Agents x Law Initiative A New Stanford and Industry Initiative Launched Yesterday

Great event on Tuesday at @stanfordhai.bsky.social convened by Dazza Greenwood and concluding talk by @tobinsouth.bsky.social around the legal landscape and challenges with #AIAgents. Recommend watching the recording and providing feedback here:
www.dazzagreenwood.com/p/ai-agents-...

Johns Hopkins APL Employing AI to Discover Materials for National Security Needs Using artificial intelligence and innovative materials science techniques, APL is leading an effort to rapidly discover revolutionary materials that allow critical operations in specific, extreme environments.

Using artificial intelligence and innovative materials science techniques, APL is leading an effort to rapidly discover revolutionary materials that allow critical operations in specific, extreme environments. jhuapl.link/6xf

Hi Fabian, Awesome MLSS is a non profit group that helps students find ML Summer Schools, and information on AI related events in general. Would love to speak with you!

Förderinitiative „Künstliche Intelligenz“

The DFG is funding independent research groups in AI (tinyurl.com/36f5m6kd). If you're interested in coming to Göttingen (Germany) by applying for one of these groups, we have an online event to inform you what the process is and how we could support you (tinyurl.com/yfpeambm). Check it out!

That is indeed good news.

When we read the news, images can convey different things than text itself.

Unlike other works which look at text, we study this as a “multimodal” framing problem & analyze where text and images communicate different “frames”.

Checkout our paper here: arxiv.org/abs/2503.20960

@aicentre.dk