AI FRONTIER
Signals worth understanding before they become obvious.
Solving a 31-year-old grand challenge in mathematics: the KLS conjecture
A Microsoft Research researcher and a collaborator proved the Kannan-Lovász-Simonovits (KLS) conjecture, showing the bound is a fixed constant independent of dimension, removing a major theoretical obstacle for randomized algorithms.
- The KLS conjecture, proposed in 1995, predicted that the best hyperplane cut and the best cut of any shape differ by no more than a universal constant, regardless of dimension.
- The previous best result, established in July 2026, allowed that bound to grow slowly with dimension, while this proof eliminates the dependence on dimension.
- The authors used a human-directed agentic process to explore more than 100 candidate approaches, with AI tools assisting in derivations, manuscript preparation, and internal proof checks.
- The complete proof was posted in the October 4, 2026 version of the paper, and a separate parallel proof and a Lean 4 formalization appeared the same day.
With KLS proved, many results in convex geometry, probability, and theoretical computer science that previously assumed the conjecture can now be stated unconditionally.
LitPivot: Developing research ideas through dynamic contextualization within the literature landscape
Ai2, with collaborators at the University of Pennsylvania, developed LitPivot, a system that uses papers retrieved through Asta Paper Finder to organize literature around the part of a proposal being edited, offer critiques supported by excerpts, and suggest revisions researchers can accept, edit, or reject. In a study...
- Participants reported a stronger understanding of the surrounding literature and greater self-reported improvements in their ideas' novelty and usefulness with LitPivot than with the baseline.
- LitPivot separates a research idea into its problem, proposed contribution, and evaluation plan, and each facet has its own set and relevance ranking of related literature.
- LitPivot will be presented at UIST 2026 in Detroit.
For builders of research assistants, LitPivot shows that binding literature selection to idea revision as an iterative loop improves how well ideas are grounded in prior work compared with one-shot retrieval.
Profiling TorchNeuron (Part 1): System-level profiling on AWS Trainium
A Hugging Face blog tutorial shows how to use torch.profiler with the Neuron backend to systematically analyze inference performance on AWS Trainium, comparing device time across three RMSNorm implementations.
- The Neuron backend uses PyTorch's PrivateUse1 extension point, and the torch_neuronx package registers Neuron as a native PyTorch backend.
- The naive PyTorch RMSNorm shows a Self NEURON total of 2.355 ms in the key_averages() table, with aten::mul accounting for 49.5%.
- torch.rms_norm lowers to a single aten::_fused_rms_norm operation with a Self NEURON time of 403.3 μs.
- The naive RMSNorm compiled with torch.compile(backend="neuron") appears as one Torch-Compiled Region with a Self NEURON time of 375.5 μs.
For AI builders, the tutorial demonstrates that key_averages() can locate device-time hotspots and that fusion or torch.compile can cut RMSNorm device time to roughly one-sixth of the naive implementation.
Ensemble forecasting with IBM Granite time-series models
Granite Timeseries Ensemble combines four pre-trained Granite models into one extensible forecasting interface that returns median and quantile forecasts with no retraining and no fitted ensemble weights.
- The default recipe combines four architectures: Granite PatchTST-FM-r1 (Patched Transformer), Granite PatchTST-FM-r2 (Patched Conformer), Granite FlowState-r1.1 (state-space model), and Granite TTM-r3 (Tiny Time Mixers).
- PatchTST-FM-r2 retains the lowest aggregate CRPS/MWQL score at 0.467, while the IQR Ensemble achieves the lowest average CRPS Rank among these six forecasts.
- The recipe stores the ensemble configuration rather than merged model weights and downloads the four member checkpoints at runtime, so the first run requires extra time.
- Forecasts return quantiles from 0.1 to 0.9 by default, where the 0.1 and 0.9 quantiles define a central forecast interval with nominal 80% coverage.
For builders this ensemble offers a sensible default that delivers comparable or better aggregate accuracy without knowing in advance which architecture suits your data, at the cost of increased inference time and memory from loading four m...
Securing our frontier model training environment
OpenAI explains how it strengthened containment for frontier model training and evaluation, including offline environments and high-compute red-teaming.
- OpenAI states it strengthened containment for frontier model training and evaluation.
- The measures include offline environments.
- The measures include high-compute red-teaming.
For teams training frontier models, containment and red-teaming during training and evaluation are being treated as engineering requirements alongside capability work.
OpenAI models are coming to Google Cloud
OpenAI and Google Cloud are bringing frontier models and Codex workflows to Gemini's Model Garden, available through Google Cloud Marketplace.
- OpenAI and Google Cloud are bringing frontier models and Codex workflows to Gemini's Model Garden.
- The models and workflows will be available through Google Cloud Marketplace.
For AI builders, frontier OpenAI models and Codex workflows become reachable through Google Cloud's procurement and deployment path.