OTC CatchUp #296 Summary

He finished baseline cache benchmarks (pointer chasing and stride tests) and published results on GitHub.

He is moving on to core-to-core latency measurements across CPU core complexes, using ping-pong style approaches and existing Rust tools as references before writing custom code.

He is also continuing with Onur Mutlu’s lectures.

He described LangChain as a framework for LLM-based apps with pipelines over text, image, or audio inputs, commonly using models such as Gemini or OpenAI.

He noted that these apps are usually packaged and deployed on cloud platforms such as GCP or AWS.

LangChain as the base framework, LangGraph for stateful workflows, LangSmith for tracing and debugging, and LangFlow as a visual builder.

Darshan noted that LangSmith is hard to self-host and that he used it for debugging and model access.

He described Hyper-V as a Type 1 hypervisor that boots before guest OSes and virtualizes CPU core state.

SNP/TDX encrypt and isolate guest memory so the host or other guests cannot freely read it, and they matter mainly in virtualization rather than bare-metal single-OS setups.

The discussion covered multi-level page tables, machine pages, 4 KB base pages, pinned pages, STLB/TLB behavior, shadow paging, and CR3 as the page-table base.

The group recommended project-based learning, studying systems topics, and trying languages such as Go or C for deeper understanding.

Advice included focusing on strong engineering fundamentals and building more complex systems, rather than chasing a single shortcut to keep up with AI.

Jaden Furtado suggested longer internships in Java/Spring to compare skills against industry expectations.

Harsh said he would post recommended back-end learning links in the chat.

The group also cautioned against overusing AI while learning, and treated it as a guide rather than a substitute for understanding.

He expects to finish CPTS in September and is considering OSWE or OSAI instead of OSCP.

Suggested prep paths included PentesterLab and practice on open-source projects.

The frontend is a static site on Cloudflare Pages, calling a FastAPI backend on Railway.

The ingestion pipeline chunks blogs, PDFs, and websites, embeds them, stores vectors in Chroma, and uses Neon (serverless Postgres) for chat and relational storage.

He described a harness that defines the model’s world and exposes planning and execution tools (including read/write access).

The group discussed similarity search over embeddings (natural-language style matching, not keyword-only search).

Alternatives such as Pinecone and LanceDB came up; Darshan chose Chroma for developer experience.