llm-guided autotuning speeds up pytorch kernel tuning
pytorch's helion dsl uses llms to propose kernel configurations, matching bayesian optimization performance while exploring 10x fewer configs and cutting tuning time by 6.7x.
aitag machine learning
summaries filed under machine learning
think beyond lora for fine-tuning
benchmarks show other parameter-efficient fine-tuning methods can beat lora on accuracy and memory, so it pays to test alternatives.
aishock-wave theory linked to symmetry-reduced sgd dynamics
a new mathematical framework connects shock-wave theory to the learning dynamics of stochastic gradient descent after removing parameter symmetries.
aitraining task generators without the solver bottleneck
a new method trains task generators using a lightweight probe instead of costly solver rollouts, making it practical to create frontier tasks for reinforcement learning.
aitrajectory-level supervision in offline rl
a statistical theory for offline policy optimization using only trajectory-level outcome labels instead of per-step rewards.
aigaussian mixture attention cuts transformer cost
a new attention method uses probabilistic routing through learned gaussian components to achieve linear-time sequence mixing, avoiding the quadratic cost of standard attention.
aipytorch foundation contributor awards open for 2026
nominations are open until july 17, 2026 for awards recognizing contributors to pytorch, vllm, deepspeed, ray, helion, and safetensors.
aibuilding a custom ai assistant for workflow control
a practical account of building a personal ai assistant with langchain and gpt-4o, focusing on memory, tools, and integration over off-the-shelf options.
aiagents can now search for tools at runtime
a new open specification lets ai agents discover tools, skills, and other agents dynamically instead of needing them pre-installed.
ailanguage-guided 3d motion forecasting with molmomotion
molmomotion predicts future 3d object motion from video, points, and text instructions, outperforming existing methods and aiding robotics and video generation.
aireplace pandas loops with 7 faster methods
learn seven faster alternatives to row-by-row loops in pandas for common data transformations.
aifive fun projects using openai codex
learn codex by building small, practical projects step by step, from a first app to a full-stack clone.
aidiffusion model generates irregular clinical time series with missingness
a diffusion-based method jointly models lab values and observation patterns to create realistic irregular clinical time series from mimic-iii data.
aifrom hub dataset to robot hardware with strands agents
strands robots sdk integrates lerobot tools into a single agent loop for recording, simulation, policy inference, and hardware deployment.
aiglm-5.2: open model for long coding tasks
glm-5.2 is an open-source model with a solid 1m-token context, improved coding, and flexible effort control, ranking highest among open models on long-horizon benchmarks.
aimodel choice matters in causal drug safety detection
a study compares classification models in the inferbert framework for pharmacovigilance, finding that simpler models can suffice and domain-specific pretraining helps.
aiediting multimodal models fails when inputs split
knowledge edits in multimodal large language models often revert to old facts when image-text pairs are separated, due to modality-specific knowledge storage.
aifixing shape-prior shortcuts in long-range fringe profilometry
a study uses mechanistic interpretability and uncertainty quantification to diagnose and repair neural network failures in single-shot 3d imaging at long distances.
aikv cache as editable, composable notes
prefix caching fails when one field changes, but treating the kv cache as a notebook of memoized conclusions enables editing and composing precomputed skills.
aillm engineer skills roadmap for 2026
a step-by-step guide covering foundations, prompting, retrieval, fine-tuning, and serving for building production llm applications.
ai