Research in software is difficult to trust most of the time, but some studies claim that type checking can reduce bugs by 15% in #python. This post covers advanced typing features such as never types, type guards, etc., that are often overlooked

martynassubonis.substack.com/p/advanced-o...

MCP improves interoperability across LLM systems and enables agentic behaviour. But it also introduces drawbacks: tool poisoning/shadowing, rug pulls, prompt injections, and scaling friction from session affinity and prompt bloat.

martynassubonis.substack.com/p/dissecting...

#mcp #llms #llm

Even at temperature 0, LLMs still exhibit randomness:
- 🔢 Floating point non-associativity: Tiny differences in computation order can change outputs.
- 🧩 Sparse Mixture of Experts: Inputs can interfere with each other during routing, depending on batch.

#llm #llms #ai #machinelearning #randomness

Empirical software laws:
1️⃣ Conway: Design mirrors communication.
2️⃣ Hyrum: All behaviors get depended on.
3️⃣ Goodhart: Metrics lose value as targets.
4️⃣ Jakob: Users prefer familiar patterns.
5️⃣ Linus: Open-source = safer, faster fixes.

martynassubonis.substack.com/p/5-empirica...

#software #design

Exploring distributed training: Tensor Parallelism (TP) and Fully Sharded Data Parallelism (FSDP). These are the strategies powering every massive trillion-parameter model training. 🚀

martynassubonis.substack.com/p/tensor-and...

#llm #machinelearning #gpt4 #distributed #ai

In distributed training, choosing the right communication algorithm for gradient synchronization can make a big difference. For example, on Summit’s 24,576 GPUs, two-tree all-reduce can cut latency by 180x and boost bandwidth 5x compared to the standard ring all-reduce.

If executing a typical CPU instruction took 1 second, then sending a data packet from Europe to US and back would take 4.75 years. Understanding different kind of latencies can help us design better systems.

#software #cpu #latency #systems

uv after 0.5.0—might be worth replacing Poetry/pyenv/pipx. It’s 10-100x faster 🚀, PEP-adhering, supports workspaces, and offers native Docker & CI/CD tools. A great tool for Python project management

#python #rust #uv #poetry

Why is the industry wasting money on slow PyTorch models for inference when they can have 9x speed on the same hardware?

#python #rust #pytorch #devops

Sharing key concepts for building effective ML workflows: reproducibility, artifact tracking, and automation, using a weather recognition project with Kubeflow Pipelines & Vertex AI as an example.

#python #devops #machinelearning #fine-tuning #ai

The “Python Project Management Primer” article explores how to manage Python project environments and dependencies, as well as how to structure projects effectively.

#python #devops #mlops