Field Notes
Our Blog
Insights on AI, Machine Learning, Web Development, and emerging technologies from industry experts.
// jump to
PristrenPristrenInsights on AI, Machine Learning, Web Development, and emerging technologies from industry experts.
// jump to
Built by Pristren
Reading about AI tools? Run your team on Zlyqor — chat, meetings, projects, and time tracking in one workspace.
265–276 of 523
LLM costs scale from $0-50/month at pre-product to $500-5,000/month at growth stage. Here is what to expect, where to optimize, and the rule of thumb that keeps AI spend sustainable.
Mahmudul Haque Qudrati
CEO & ML Engineer
Three fast, cheap inference platforms for open source LLMs. Groq is the fastest, Together AI has the broadest model selection, Fireworks specializes in production-grade function calling.
Mahmudul Haque Qudrati
CEO & ML Engineer
Quantization reduces model weight precision from FP32 to INT4, cutting memory and compute by 4-8x. Q4_K_M is the sweet spot for most use cases - near full quality at a fraction of the size.
Mahmudul Haque Qudrati
CEO & ML Engineer
Flash Attention rewrites transformer attention to be IO-aware, reducing memory from O(n²) to O(n). It enables 128k context windows and cuts training costs by 2-4x. Here is how it works.
Mahmudul Haque Qudrati
CEO & ML Engineer
Benchmarks are gamed and vibes do not scale. Here is how to build real evaluations that tell you whether an LLM actually works for your specific use case.
Mahmudul Haque Qudrati
CEO & ML Engineer
A plain-English explanation of every major LLM benchmark: what each one tests, how it scores, and what a 1% difference actually means in practice.
Mahmudul Haque Qudrati
CEO & ML Engineer
LM-as-judge works well for relative preference ranking but breaks down for absolute quality scores. Here is how to set it up and avoid the major failure modes.
Mahmudul Haque Qudrati
CEO & ML Engineer
How to build an eval system that catches 80% of regressions with 20% of the effort. Start with real production examples, define clear scoring, and track it over time.
Mahmudul Haque Qudrati
CEO & ML Engineer
RAGAS gives you four metrics that cover every major failure mode in a retrieval-augmented generation pipeline. Here is what each metric measures and how to act on low scores.
Mahmudul Haque Qudrati
CEO & ML Engineer
SWE-Bench uses real GitHub issues from real projects to test whether models can write code that actually fixes software bugs. It is far more demanding than HumanEval.
Mahmudul Haque Qudrati
CEO & ML Engineer
Precision, recall, and F1 are the foundation of retrieval evaluation. Understanding the tradeoff between them tells you whether to optimize your RAG system for fewer wrong answers or fewer missed answers.
Mahmudul Haque Qudrati
CEO & ML Engineer
Chatbot Arena ranks LLMs through millions of real user preference votes rather than fixed benchmarks. It is the most contamination-resistant ranking system that exists today.
Mahmudul Haque Qudrati
CEO & ML Engineer