Latest

Plans as interactive HTML rather than walls of Markdown cut review time and make the feedback loop structured instead of ad-hoc, which could reshape how teams gate agent coding work. Developers running Claude Code on real projects can install the plugin and start routing plan approvals through numbered decision points. It comes from the Anthropic community plugin marketplace, so setup is a two-command install rather than custom tooling.

shao__meng@shao__mengSource

The lecture frames prompt-to-code generation as the newest rung on a compiler abstraction ladder, tracing criticism of vibe coding directly to 1957 arguments against FORTRAN. The nondeterminism point gives developers a principled way to reason about when AI-generated code needs verification versus when a compiler did not. Materials are openly available at agencyai.mit.edu/w4, so self-learners and educators anywhere can follow along without enrollment.

shao__meng@shao__mengSource

The post frames OCR as a routing and verification problem rather than a model-quality problem, which changes how teams should budget: cheaper models plus targeted compute on hard elements can beat one expensive frontier call. Practitioners building document pipelines will find a concrete pattern for handling tables, charts and multi-column pages. It also honestly covers where agentic OCR still fails.

Jerry Liu@jerryjliu0Source

DeepSpeed is often typed to ZeRO alone, but its automatic tensor, sequence and expert parallelism can reshape how teams scale training beyond a single strategy. Engineers planning multi-GPU or MoE training runs should review the session agenda and benchmarks before choosing between AutoTP, AutoSP and AutoEP versus alternatives like Megatron-LM.

Official accountPyTorch@PyTorchSource

The deck compresses hardware-level tradeoffs that usually take weeks of trial and error into one reusable resource, and attribution-based reuse makes it usable directly for internal team training or meetup talks. Anyone deploying local inference on constrained hardware can use it to pick between quantization levels or memory setups. Pair it with a llama.cpp install to follow the examples hands-on.

Merve Noyan@mervenoyannSource

Trainium becoming a standard PyTorch device means teams can apply familiar torch.profiler workflows to custom silicon instead of NVIDIA tooling. Practitioners running training or inference on AWS should note the session's path from profiler capture to hardware cycle debugging when deciding whether to adopt Trainium for cost-sensitive workloads.

Official accountPyTorch@PyTorchSource