iiinova — where your digital self keeps evolving

Paul Hager @paulhager.bsky.social · 12/09/2026

AGI is when a robot can help you move apartments

Reposted by Paul Hager

Ethan Mollick @emollick.bsky.social · 27/04/2026

Very cool analysis of the submissions to a major management journal that shows how much the system of science, built for humans, is under strain as a result of AI. AI can be used to do better science or it can be used to just do more stuff. The danger is that "more" is winning due to incentives.

36018

Paul Hager @paulhager.bsky.social · 22/04/2026

Will be at #iclr2026 in Rio this week! If anyone wants to catch up or chat about tabular data, LLM benchmarking, or histopathology send me a message. Also if anyone is interested in learning more about QuantCo and the great work we are doing on in-vivo microscopy at VIRDX LMK

010

Paul Hager @paulhager.bsky.social · 11/04/2026

Discussing a book with someone is such a deep way of connecting. Your reaction to specific passages provides such a personal understanding of how you think about and see the world

010

Paul Hager @paulhager.bsky.social · 10/04/2026

Maybe this is AGI? When you start arguing with a model and it checks your post history to try and win the argument like a lot of real people online like to do.

010

Paul Hager @paulhager.bsky.social · 10/04/2026

The only issue with post-session hooks is that you delegate the decision making to Claude what gets remembered and what does not which can quickly pollute your context over time. Being very explicit with what is worthy of being remembered maybe is necessary as long as our longterm taste is superior

110

Paul Hager @paulhager.bsky.social · 10/04/2026

I solved this by having a lessons-learned.md in my global context and everytime I tell it to remember something and not make the mistake again it saves it there. Works very well so far after ~3 months of use.

030

Paul Hager @paulhager.bsky.social · 09/04/2026

Many should prevail against one but one can often prevail against many because he is not only one. One can bribe, appeal, cajole, threaten, blackmail, and force many to one’s side and thus maintain power. It is greed, selfishness, and fear that gives one power over many by turning one into many

000

Reposted by Paul Hager

Florian Hölzl @hlzl.eurosky.social · 08/04/2026

While 2) is with a focus on post-training and reasoning jobs.plattnerfoundation.org/HPI/job/Pots... Come check them out and apply!

jobs.plattnerfoundation.org

PhD Candidate (f/m/x) - Post-Training, Reasoning, and Multimodal Health AI

011

Reposted by Paul Hager

Florian Hölzl @hlzl.eurosky.social · 08/04/2026

👋 We have two fully-funded PhD positions open at @hpi.bsky.social with the Chair of Human-Centered Transformative AI. 1) On efficient transformer architectures with a focus on recurrent approaches jobs.plattnerfoundation.org/HPI/job/Pots... 1/2

jobs.plattnerfoundation.org

PhD Candidate (f/m/x) - Efficient Reasoning Architectures for Medical AI

154

Reposted by Paul Hager

Julian Suk @sukjulian.bsky.social · 27/03/2026

One month left 'til @iclr-conf.bsky.social, about time to launch our ✨@gram-org.bsky.social Competititon✨ The theme is geometry x AI4science with a dataset kindly provided by BeyondMath. Deadline: April 22, 2026 (AoE) 🔗 gram-competition.github.io

163

Paul Hager @paulhager.bsky.social · 25/03/2026

London is a lot greener than I thought it would be. The sky is exactly as gray as expected though

010

Paul Hager @paulhager.bsky.social · 25/03/2026

Coding agents are creatures of context

010

Paul Hager @paulhager.bsky.social · 04/03/2026

Shouldn’t it always technically be two new numbers if they automatically add the negative 🤔?

000

Paul Hager @paulhager.bsky.social · 20/01/2026

Is it just me or has the quality of shoe laces gone down in the last couple of years? All new shoes I buy need a double knot not to open during the day

000

Reposted by Paul Hager

TUM AI in Medicine Lab @tum-aim-lab.bsky.social · 02/12/2025

It’s this time of the year again! #NeurIPS2025 If you are in San Diego, make sure to check out 2 works from our lab this week 📝📝 #AIMresearch 1/3

133

Reposted by Paul Hager

TUM AI in Medicine Lab @tum-aim-lab.bsky.social · 29/10/2025

Quick look back at an insightful day yesterday at the Bavarian Conference on AI in Medicine, where @paulhager.bsky.social , @luciehuang.bsky.social, Alina Dima, and Vasiliki Sideri-Lampretsa were representing us. Team, thank you for being such good ambassadors! 👏 #AIMNews #AIinMedicine

021

Reposted by Paul Hager

Derek Thompson @dkthomp.bsky.social · 10/10/2025

1. Social media is now 5% following friends and 95% watching videos made by strangers. 2. Every podcast is turning into a YouTube show. 3. AI companies can't stop building TikTok clones. Everything is becoming television. I wrote about why that matters. www.derekthompson.org/p/why-everyt...

derekthompson.org

Everything Is Television

A theory of culture and attention

1315047

Paul Hager @paulhager.bsky.social · 10/10/2025

Commuting in cities without headphones make you feel more like you're part of society and not just on your own lonely island. Much nicer than the 1000th podcast or song

000

Paul Hager @paulhager.bsky.social · 10/10/2025

The speed at which LLMs allow you to try new research ideas is astounding. Especially for the sweet spot where the logic is not fantastically complex but the implementation is laborious.

000

Paul Hager @paulhager.bsky.social · 23/09/2025

Few things as satisfying as the moment when your paper finally is within the page limit after hours of rewriting, figure optimizations, and moving things to the appendix.

020

Reposted by Paul Hager

NeurIPS Europe @neuripseurope.bsky.social · 19/09/2025

Congratulations to everyone who got their @neuripsconf.bsky.social papers accepted 🎉🎉🎉 At #EurIPS we are looking forward to welcoming presentations of all accepted NeurIPS papers, including a new “Salon des Refusés” track for papers which were rejected due to space constraints!

15016

Reposted by Paul Hager

NeurIPS Europe @neuripseurope.bsky.social · 19/09/2025

And lastly, if @neuripsconf.bsky.social would choose to reverse the decisions on the papers affected by space constraints, we would be happy and able to accommodate their presentation

02611

Paul Hager @paulhager.bsky.social · 19/09/2025

The policy of rejecting accepted papers by NeurIPs due to "space issues" will affect so many conferences going forward that now need to also deal with papers that should have already been published competing with new research. These conferences have structural issues that must be fixed.

010

Paul Hager @paulhager.bsky.social · 29/08/2025

People who try to start „crowd clapping on beat“ at live music performances should be thrown in jail

000

Reposted by Paul Hager

TUM AI in Medicine Lab @tum-aim-lab.bsky.social · 26/08/2025

🚀 Towards cardiac MRI foundation models: Comprehensive visual-tabular representations for whole-heart assessment and beyond. We introduce ViTa, a multi-view, multi-modal, multi-task model for cardiac MRI. 🧵 1/3

121

Reposted by Paul Hager

TUM AI in Medicine Lab @tum-aim-lab.bsky.social · 20/08/2025

🔃 We're moving beyond traditional benchmarking for LLMs. 🚩 Meet Dynamic, Automatic & Systematic (DAS) Red-Teaming from Jiazhen Pan and Bailiang Jian together with great collaborators! #AIMresearch 🧵 1/4

Outline of the Dynamic, Automatic & Systematic (DAS) Red‑Teaming method.

131

Paul Hager @paulhager.bsky.social · 01/08/2025

Hold on a second, summer hasn’t even begun. Three weeks of rain and 18 degrees does not a summer make.

030

Paul Hager @paulhager.bsky.social · 24/06/2025

Considering the current meta of Benchmark -> Problem -> Dataset -> Retrain ... This seems like a free paper.

010

Paul Hager @paulhager.bsky.social · 24/06/2025

So after spending a couple of days at CVPR I always get really hype about how far CV has come and then I ask 4o and Imagen to generate a map with 4 pins in it and I'm quickly thrown back to reality

130

Paul Hager @paulhager.bsky.social · 12/06/2025

I will never get used to how cold American buildings are. I love it when it’s over 30 degrees outside and then less than 15 inside. So much fun being permanently too cold or too hot.

010

Paul Hager @paulhager.bsky.social · 12/06/2025

The ideal modern AI research workflow is: 1. Develop a new benchmark to show a weakness of SOTA models. 2. Create a new (or cleverly adapt an old) dataset to address this weakness. 3. Retrain a transformer.

010

Paul Hager @paulhager.bsky.social · 12/06/2025

Honestly pretty happy about this trend. Architectural modifications are contrived 99% of the time to try and check some notion of novelty for reviewers. Hopefully this will not be as important going forward.

000

Paul Hager @paulhager.bsky.social · 12/06/2025

Biggest takeaway from CVPR so far: datasets 👏🏼 datasets 👏🏼 datasets 👏🏼

120

Paul Hager @paulhager.bsky.social · 11/06/2025

Going to be at CVPR the next couple of days presenting our paper „A Tale of Two Classes: Adapting Supervised Contrastive Learning to Binary Imbalanced Datasets“. arxiv.org/abs/2503.17024 Always happy to meet anyone working on representation learning or tabular DL and medical data

arxiv.org

A Tale of Two Classes: Adapting Supervised Contrastive Learning to Binary Imbalanced Datasets

Supervised contrastive learning (SupCon) has proven to be a powerful alternative to the standard cross-entropy loss for classification of multi-class balanced datasets. However, it struggles to learn ...

073

Paul Hager @paulhager.bsky.social · 11/05/2025

Wolt and Uber Eats should start sponsoring conferences. Send them a screenshot of your NeurIPS abstract submission confirmation and get your first delivery free.

020

Paul Hager @paulhager.bsky.social · 05/05/2025

Deutsche Bahn top form continues with a 45 minute delay on a 3 hour „sprinter“ route…. At this point I’m more mad that it wasn’t at least an hour, then I would have at least gotten compensation

020

Paul Hager @paulhager.bsky.social · 01/05/2025

Praying for some heavy rain and plummeting temperatures soon or this paper isn’t going to be as polished as it could be 😅

010

Paul Hager @paulhager.bsky.social · 01/05/2025

Few things more depressing than staying inside all day and working on a paper for a conference deadline in two weeks when it’s a holiday and absolutely amazing weather outside. The other side of the coin is the freedom to spontaneously take random days off in winter when it snows a lot to go ski!

000

Paul Hager @paulhager.bsky.social · 29/04/2025

Do animals other than humans also appreciate beauty? Is a falcon flying above the Grand Canyon happier than one flying above Gary, Indiana?

000

Paul Hager @paulhager.bsky.social · 29/04/2025

After being inspired by this quote by Francis Bacon I have decided to post more online.

020

Paul Hager @paulhager.bsky.social · 28/04/2025

The co-option of the brightest and most hard working individuals by private enterprise away from government is the biggest loss to civil society in modern history.

000

Paul Hager @paulhager.bsky.social · 28/04/2025

Forget about Corona, does anyone remember the Lesesucht pandemic?

000

Paul Hager @paulhager.bsky.social · 28/04/2025

Reading is conversing with the ideas and sentiments of authors across time and space. It taps into the fact that so much of what happens to us is not unique. There is so much shared experience across history. That’s what makes reading such a pleasure. Its like chatting with friends you can’t see.

010

Paul Hager @paulhager.bsky.social · 28/04/2025

Considering how young experimental psychology and even historicism is, it make you wonder if there are any core disciplines that exist that we could be studying right now but no one has thought to try (or at least hasn’t gained much traction).

000

Paul Hager @paulhager.bsky.social · 01/04/2025

Interested in our work? Want to chat about supervised learning & how data distributions influence it? Be sure to visit our poster in Nashville! Big thanks to my co-first author @milddave.bsky.social and our supervisors @danielrueckert.bsky.social & @martinmenten.bsky.social !

020

Paul Hager @paulhager.bsky.social · 01/04/2025

Based on this, we also suggest two simple but highly performative methods. These outperform SupCon and other leading methods for long-tailed datasets by up to 35% on binary tasks.

110

Paul Hager @paulhager.bsky.social · 01/04/2025

We found that leading methods like Supervised Contrastive Learning (SupCon) completely collapse at high binary imbalance. To understand why, we developed two new metrics analyzing representation space quality – finding they correlate much better with downstream performance than canonical metrics.

110

Paul Hager @paulhager.bsky.social · 01/04/2025

If you’ve ever worked with imbalanced datasets, you know how difficult training deep learning models can be. There’s lots of literature on multi-class imbalance, but surprisingly little for binary distributions, which are very common (e.g., in medicine).

110