Tobias Weber

KFAC is everywhere—from optimization to influence functions. While the intuition is simple, implementation is tricky. We (@BalintMucsanyi, @2bys2 ,@runame_) wrote a ground-up intro with code to help you get it right. 📖 arxiv.org/abs/2507.05127 💻 github.com/f-dangel/kfac-…

@deliprao Your X is Secretly a Y :D

For no good reason, I have made an exhaustive list of AI paper title patterns. Marti Hearst would be proud, but here goes nothing: Large Language Models are X X is All You Need Towards X Rethinking X Revisiting X On the X of Y Do Xs Really Y? What Do Xs Know? The X of Y Xs are All You Need Xs as Ys Xs are Zs Don’t X, Y! Beyond X Improving X with Y Understanding X via Y A Simple Approach to X A Survey of X Few-Shot X with Y Zero-Shot X with Y X: A Benchmark for Y X: Scaling Laws for Y Prompting X to Y Learning to X with Y Fine-Tuning X for Y Instructing Xs Evaluating Xs on Y Harnessing X for Y Using X to Solve Y X Meets Y X or Y? X without Y When X Fails Is X Enough? How Far Can X Go? Scaling Xs to Y Scaling Laws for X Unifying X and Y Bridging X and Y Demystifying X Analyzing X in Y X through the Lens of Y Mitigating X in Y X as a Service X as an Agent Simulating X with Y Yields of X in Z Limits of X

🎓 And that's a wrap on the LOGML Summer School! 🥰 It was an incredible experience organizing it. A huge thank you to all our speakers for their insightful talks, to our mentors for making it an unforgettable experience for the students, and to all of you for attending! 👏

@_onionesque Thank you for reading and sharing it :) Let me know, if you want to chat about FNOs at some point.

Enjoyed reading this, esp. as I try to educate myself on Fourier neural operators. arxiv.org/abs/2406.05072 Quite an elegant combination/development of ideas towards quantifying the uncertainty of FNOs. PS: I know one of the authors has a mysterious existence on twitter: @2bys2

@yaroslavvb Where is this replacement happening? I'm having trouble finding it.

@francoisfleuret I can't get rid of a feeling that backprop is a bug. During step computation, backprop people replaced A^{-1} with A^T .... I'm still trying to understand why that works (sometimes). For a system with infinite condition number, this approximation can be arbitrarily bad

TFW you find the bug, you fix the bug, but you really have no idea how it can have worked with it for half a week.

@naturecomputes Pmmp8mmmmimmmmmm

✈️Amsterdam bound. Thrilled to be participating in the Geometry and Shape Analysis for Neuroscience Minisymposium at #SIAMCSE23 this Monday. Please reach out if you are attending or in the area. Looking forward to great conversations

@MLSTjournal @gasteigerjo @jtmargraf @guennemann @fhi_mpg_de @TU_Muenchen @IOPmaterials @JPhysCM @OpenSciTalk At least 200 times are from me, it is just easier than finding it again in my Download folder.

'How robust are modern graph #neuralnetwork potentials in long and hot #moleculardynamics simulations?' by @gasteigerjo @jtmargraf @guennemann @fhi_mpg_de @TU_Muenchen hits 2000 downloads! bit.ly/3DMi4b5 #compchem #machinelearning #graphs #AI #materials #HPC #condmat