"We're not going straight to a world where AI does everything and engineers disappear. Instead, engineers are shifting rolesโfrom implementers to orchestrators."
"I think something people don't talk enough about when discussing AI agents and AI engineers doing all this stuff for youโฆ is basically we're all becoming engineering managers."
Just finished listening to an incredible podcast featuring an interview with Wu Yi โ a Tsinghua alum and former OpenAI researcher โ and his take on Reinforcement Learning (RL) was one of the clearest Iโve seen!
Wu Yi explains that RL is very different from traditional supervised learning (like image classification). In supervised learning, we train models using a fixed set of labeled data โ one-shot answers.
RL, on the other hand, is more like playing a game: you need to make a sequence of decisions (serve, move, react), and there's no single โcorrectโ path. The quality of your decisions is judged by the final outcome (win or lose). Itโs about multi-step decision-making โ much closer to how the real world works.
Key Takeaways:
Prompt engineering is an ๐ถ๐๐ฒ๐ฟ๐ฎ๐๐ถ๐๐ฒ ๐ฝ๐ฟ๐ผ๐ฐ๐ฒ๐๐ involving continuous ๐๐ฒ๐๐๐ถ๐ป๐ด, ๐บ๐ผ๐ฑ๐ถ๐ณ๐ถ๐ฐ๐ฎ๐๐ถ๐ผ๐ป, ๐ฎ๐ป๐ฑ ๐ผ๐ฝ๐๐ถ๐บ๐ถ๐๐ฎ๐๐ถ๐ผ๐ป. As models continue to evolve, prompts must adapt accordingly to maintain effectiveness.
Prompt engineering is about "communicating" with LLM in a way that maximizes the model's understanding and performance on a given task. At its core, prompt engineering involves designing, refining, and optimizing the text inputs (prompts) given to models to elicit accurate, relevant, and useful responses.
In todayโs fast-paced development landscape, automated browser testing and web scraping are more vital than ever. Three of the most prominent tools in this space are Selenium, Puppeteer, and Playwright. Each has its unique strengths and target use cases. In this post, weโll explore how they stack up in terms of browser support, language options, performance, ease of use, and more.
However, for the most powerful and reliable option, Playwright is the best of the three.
Define test cases to ensure you're actively improving your app & not causing any regressions.
Break down one LLM call into multiple
AI systems do a lot better when you have many LLM calls chained together. i.e, instead of sending an LLM call to a model to generate code, send it to a "architect" model to generate a plan, then a "coding" model to generate code, then a "reviewer" model to verify.
Start simple (with 1 LLM call)
Cursor is really dumb if not given enough context about your project. Here what you can do to improve your Cursor workflow
Claude/o1 are your best friends here. You should create a whole document containing every single detail of your project.
All this should be put into an instruction.md (name it however you want) so Cursor can index at any time.
Model Context Protocol (MCP) is an open protocol that standardizes how applications provide context to LLMs. It's a new standard for connecting AI assistants to the systems where data lives, including content repositories, business tools, and development environments. Its aim is to help frontier models produce better, more relevant responses.
As AI assistants gain mainstream adoption, the industry has invested heavily in model capabilities, achieving rapid advances in reasoning and quality. Yet even the most sophisticated models are constrained by their isolation from dataโtrapped behind information silos and legacy systems. Every new data source requires its own custom implementation, making truly connected systems difficult to scale.
MCP addresses this challenge. It provides a universal, open standard for connecting AI systems with data sources, replacing fragmented integrations with a single protocol. The result is a simpler, more reliable way to give AI systems access to the data they need. โ
RAG is a popular method that improves accuracy and relevance by finding the right information from reliable sources and transforming it into useful answers.
Large Language Models are trained on a fixed dataset, which limits their ability to handle private or recent information. They can sometimes "hallucinate", providing incorrect yet believable answers. Fine-tuning can help but it is expensive and not ideal for retraining again and again on new data. The Retrieval-Augmented Generation (RAG) framework addresses this issue by using external documents to improve the LLM's responses through in-context learning. RAG ensures that the information provided by the LLM is not only contextually relevant but also accurate and up-to-date.
There are four main components in RAG: