#Llm
9 notas · última:
How I built a sub-500ms latency voice agent from scratch
The post argues that voice agents are an orchestration problem, and shows how a streaming pipeline with turn detection, cancellation, geography, and fast model choice cut latency from about 1.6s to around 400ms.
AI Workflows vs AI Agents, Explained with LEGOs
The text compares AI workflows to LEGO-style assemblies with fixed steps and agents to constructions with a goal, showing when to use each and why hybrid approaches make sense.
AI-assisted development needs automated tests
Simon Willison argues that comprehensive test coverage is what makes LLM-assisted coding trustworthy: tests let you prove convoluted AI-written code actually works.
Fragments: February 13
Martin Fowler collects notes from a Thoughtworks retreat: senior developers staying relevant with LLMs, cognitive debt outpacing technical debt, and the drift toward supervisory programming.
O teorema da incompletude de Gödel
Em 1931, Gödel provou que todo sistema formal capaz de aritmética é incompleto ou inconsistente: há verdades que ele não alcança a partir das próprias regras. O que isso significa, e onde encosta nos LLMs.
Pesquisa Viva: A Characteristica Universalis de Leibniz e os Vetores
Até onde Leibniz chegou com a characteristica universalis, e até onde um LLM, na prática, realiza o que ele almejava? Ele foi um visionário a quem só faltava poder computacional, ou o projeto era de outra natureza?
Your LLM Doesn't Write Correct Code. It Writes Plausible Code.
The piece argues that LLMs optimize for plausibility over correctness, using benchmarked SQLite and disk-management examples to show code that compiles, tests, and still fails on the real performance problem.
Laurence Tratt: LLM Inflation
Tratt describes “LLM inflation” as using LLMs to turn short content into long prose and long prose into short summaries. He argues this can reward obfuscation and hide weak thinking.
Using AI Without Leaving the Terminal: A Guide to llm
The guide presents llm as a terminal-based interface to multiple language models, with logging, plugins, pipes, local models, embeddings, templates, and workflow automation.