Research
ParallelKernelBench Reveals Gaps in Multi-GPU Kernel Generation
Together AI Bloghigh confidence
Why it matters
- →Multi-GPU kernel generation is crucial for scaling AI workloads efficiently.
- →Current LLMs struggle with complex communication patterns, limiting their utility.
- →The findings highlight areas for improvement in AI-driven optimization for distributed systems.
ParallelKernelBench (PKB) has highlighted the limitations of current large language models (LLMs) in generating efficient multi-GPU kernels. Despite progress in single-GPU scenarios, models like GPT-5.5 and Gemini 3 Pro solved fewer than a third of PKB's 87 benchmark problems correctly. The evaluation shows that these models struggle with complex communication patterns and rank coordination, which are essential for multi-GPU performance. While there are occasional successes in generating high-performance kernels for specific tasks, the findings underscore the need for further advancements in AI-driven optimization for distributed computing.
Read originalThe story around this
Earlier coverage that leads up to this article, and what followed. Lines connect each piece to the closest one after it, converging here.
How we got here
- 1vLLM v0.21.0 Release
vLLM Releases · May 16, 2026 · Background
- 2Together AI Hosts MiniMax M3 for Efficient Inference
Together AI Blog · June 2, 2026 · Related
What happened next
- 3Together AI Introduces Autoscaling for LLM Inference
Together AI Blog · July 31, 2026 · Related
- 4Hugging Face Launches 200+ WebGPU Kernels
Hugging Face Blog · September 1, 2026 · Related
- 5Fine-tuning 350M Model for Structured Outputs
Hugging Face Blog · September 3, 2026 · Related
- 6llama.cpp b11037 fixes GPU inference bugs
llama.cpp Releases · September 19, 2026 · Related
- 7llama.cpp adds OpenCL cooperative-k DP4A GEMM
llama.cpp Releases · October 20, 2026 · Related
More in Research
© Cole MedinResearchagents
Anthropic study: Agents sway expert judgment 86% of time
Anthropic’s latest research exposes a quiet but dangerous flaw in agentic workflows: experts follow AI recommendations 86% of the time, often adopting the agent's reasoning as their own. This isn't just about accuracy; it's about cognitive capture, where human oversight dissolves into passive compliance. The study shows that even when users believe they are steering, the agent’s framing dictates the outcome. Practitioners must now actively counter this bias by forcing agents to justify deviations or prototype alternatives before committing to a path. This shift demands a new discipline in how we interact with coding assistants. We can no longer assume our intent drives the result; the model's suggestion often becomes the default truth. The real work is ensuring we remain the architects, not just the approvers.
Cole Medin
© WIRED AIResearchother
AI-Augmented Infostealer 'Volta-Autodoxer' Targets High-Value Victims
Cisco Talos has uncovered a new infostealer strain that actively uses commercial LLMs to profile victims and prioritize high-value targets. Dubbed Volta-Autodoxer, the malware collects credentials and browsing data, then feeds this into models like Claude Opus 4.6 via OpenRouter to generate structured intelligence dossiers. This marks a shift from passive data theft to active, AI-driven triage, allowing attackers to bypass guardrails with fallback models like Kimi K2 or Gemini. The capability foreshadows a future where lower-sophistication criminals can execute sophisticated, targeted attacks at scale.
WIRED AI
© Matt WolfeResearchresearch
OpenAI Disproves Exponential Time Hypothesis
OpenAI researchers published findings disproving the Exponential Time Hypothesis, a major result in computational complexity theory.
Matt Wolfe