Curious about how fast model inference can really get? Our CTO, Fergus Finn, breaks down the maths behind inference speed in his latest post: Inference Arithmetic. Try adjusting the configurable parameters - tensor parallelism, concurrent users, sequence length, accelerator type, and precision - to see how different hardware setups and workload patterns shift real-world performance. Key takeaways: 1️⃣ Transformer inference has two main bottlenecks - memory bandwidth and compute capacity 2️⃣ The prefill phase (processing input tokens) is typically compute-bound, while the decode phase (generating tokens) often becomes memory-bandwidth-bound 3️⃣ There’s a threshold in batch or token size where inference flips from bandwidth-bound to compute-bound - below it, you’ll hit diminishing returns Full article linked in the comments 👇

Yongchao Chen with Harvard University and Massachusetts Institute of Technology covered the Google paper "TUMIX: Multi-Agent Test-Time Scaling with Tool-Use Mixture" at our latest community paper reading! Watch: https://lnkd.in/gn9yuQR6 The paper proposes Tool-Use Mixture (TUMIX), an ensemble framework that runs multiple agents in parallel, each employing distinct tool-use strategies and answer paths. Agents in TUMIX iteratively share and refine responses based on the question and previous answers. In experiments, TUMIX achieves significant gains over state-of-the-art tool-augmented and test-time scaling methods.

Juho Timonen just wrote a post about the challenges of using numerical solvers in (probabilistic) computing. Solving differential equations numerically is hard work; doing inference over the parameters of ODEs is even harder. It's a minor miracle that modern MCMC samplers can do this at all for moderately sized datasets, so long as the number of equations is not too large. Of course, the complexity of each equation matters as well. In the post, Juho digs into to some of the core issues and offers suggestions for improvement. https://lnkd.in/gmnqGnUx

This notebook explores two mathematical constants related by their series and integral representations: the MRB constant, denoted as S, and its integrated analog, denoted as c. The MRB constant, discovered by Marvin Ray Burns, is defined by the series: ... The integrated analog, ... is defined through integration over the complex plane: ... Both constants exhibit interesting behavior when expanded through series involving logarithmic terms raised to integer powers. This notebook numerically investigates these constants, providing insights into their properties and differences.

Project Spotlight ✨ Our member Jeevan Rao has developed a Computer Vision algorithm to accurately count the stars on a flag. This project highlights the practical use of image processing and pattern recognition in real-world applications. #EyeQClub #SIMATS #ComputerVision #AI #ImageProcessing #StudentProjects #Innovation

Another big step toward practical quantum computing! Muggle translation: By combining a high-performance supercomputer with a quantum processor, researchers have now achieved energies for a challenging chemistry problem that surpass the accuracy of strong classical approximations (like CISD and CCSD), and get remarkably close to the best classical method available today (DMRG). What blows my mind is the progress since last year. If you look at the pink dotted line in the plot (the 2024 quantum SOTA), we’ve already stepped far beyond it — in just a few months. Honestly, the pace of improvement over the last two years feels exponential. Every few months, the ceiling of what’s possible shifts upward… and with this experiment we just raised it again. Real utility for quantum computing doesn’t feel theoretical anymore. It feels near.

I’m eager to announce today our new arXiv https://lnkd.in/e9xc6YwE where we demonstrate the first closed-loop workflow between a quantum computer and an entire classical supercomputer. Earlier this year we have shown that a quantum computer coupled with classical HPC can perform chemistry calculations beyond the exact diagonalization regime. We received a lot of feedback and one of the natural questions was if one could close a loop between quantum and classical compute at large scale. We answered those questions and more with this work in collaboration with RIKEN. We optimize molecular energies obtained from quantum data, using a cost function computed on 152,064 classical nodes of the supercomputer Fugaku. We dramatically improve on our benchmark Fe4S4 result of last year, reaching extrapolated energies approaching the best DMRG calculations. There’s a lot of technical work behind this scientific result, including optimal orchestration of quantum and classical computational resources and a new hybrid C++ interface to the high-performance computing workloads. This charts a path forward for how quantum and classical HPC can interact together. Congratulations and thanks again to the amazing team of scientists and engineers at IBM Quantum and RIKEN that made this possible.

Tracking an object over time (a sequence of frames) is a challenge in computer vision. All approaches I know of are lost as soon as the tracked object transforms into a next stage (for example a caterpillar that develops into a butterfly, water that hits the ground, or apples that are cut into slices).  A research team at Cornell proposed a new approach that seems to solve exactly that. They published a paper “Tracking and Understanding Object Transformations” which was accepted at NeurIPS 2025. 𝗧𝗿𝗮𝗰𝗸 𝗔𝗻𝘆 𝗦𝘁𝗮𝘁𝗲 continues tracking an object through each transformation and explains what changed along the way.  Track Any State shows good results with: - Maintaining object tracking even after significant transformations    - Identifying and describing newly emerging object states - Reconstructing missing segments using TubeletGraph - Achieving state-of-the-art performance under transformative conditions See for yourself in the video. Especially the water hitting the bowl is impressive.  Find all relevant information here: https://lnkd.in/euXBqgVQ

Just came up with an approach to tackling the second hardest problem in computer science -- naming things: 1. Come up with a name (for a variable, method, parameter, etc.). 2. Give a think-mode LLM a brief high-level context about your system, show it the name, and ask it to guess how the system is expected to behave given that variable/method/parameter. 3. Compare your intuition with what the LLM came up with.