Abhishek Sharma’s Post

Melty is a groundbreaking open-source AI code editor designed for 10x engineers. It uniquely understands your workflow, from terminal to GitHub, acting as a collaborative partner in crafting production-ready code. Key Features: - Better Code Understanding: Clarifies your code instead of complicating it. - Real-time Collaboration: Monitors changes like a pair programmer. - Adaptive Learning: Learns and personalizes to your codebase. - Seamless Integration: Works with your compiler, terminal, debugger, and tools like Linear and GitHub. Check out Melty • https://melty.sh! #opensource #ai #connections

Back in My Day: A Programmer's Tale "Back in my day, we didn’t have AI. We had Stack Overflow!" And yes, we copy-pasted errors into search bars like true adventurers. 🧙♂️ AI might write code for you now, but it won’t give you the same satisfaction as finding a golden answer buried in a comment thread from 2010. 🤓 Back then, every solved bug felt like winning a mini lottery (after sifting through 20 irrelevant answers). Sure, AI is fast, but can it remind you to “RTFM”? Didn’t think so. Let’s not forget our roots—because Stack Overflow was the lifeboat for developers drowning in syntax errors. Respect the past, embrace the future. And never forget to upvote. 👍 #ProgrammingHumor #AIvsStackOverflow #CodeMemories

Code completion using AI agents, showcasing the rapid evolution of AI in software engineering. Have you experimented with AI coding assistants? What's your take on their potential impact? Drop your thoughts below and let's spark a conversation about the AI revolution in software development! 🔥💬 #AIinCoding #SoftwareEngineering #TechInnovation CodeStory (YC S23) https://lnkd.in/gv9YewtZ

Big tech is aggressively promoting AI code assistants. All the tools that I have used fall short in significant ways. In addition to their shortcomings, they also lock you into their ecosystem and are not transparent about how they use your data or how the LLM works. An open-source project called Aider (https://aider.chat/) is miles ahead of the commercial alternatives. It has an interface and features that match what it is like to pair a program with a colleague. The key differences are that Aider can: - Map your whole repository. - Make complex changes across multiple source files. - Run commands and use the output in subsequent steps. - Allows you to use many LLM models, both proprietary and open-source. - Updates the context it uses for prediction while updating it. - Store the entire chat history of each repo you work on as a markdown file for later reference. I was fascinated to learn that Aider wrote 82% of the last release itself! The project maintains an LLM leaderboard: https://lnkd.in/dJWJxN2S

A global survey by Stack Overflow reveals that 81% of developers expect #AI to take over code documentation, 80% trust it for refining code testing, and 76% believe it will reshape code writing. It's clear that AI is revolutionizing every stage of the software development lifecycle. As we push the boundaries of application modernization, AI is transforming how developers work and innovate. Explore more here: https://bit.ly/3XJWA8O  

New Tools Brings New Type of Full Stack read more at Yaniv Shalev Blog

So as the tech and business world increasingly 'tests in prod' this whole $1T shift to #LLMs and Software 2.0, I'm increasingly noticing this gap: 1. Practice: Much of our genAI product & consulting work goes similarly. As LLM prompts/agent prototypes become production deliverables, initial vibes-based testing breaks down on quality and we switch to evals-based. It's the Software 2.0 way to add CI tests for systems of models vs systems of code. 2. Gap: And yet, so much LLM usage is as new low-code 2.0 tools like Github Copilot and Cursor, and where the biggest ultimate societal win is helping folks who are neither statistically-rigorous data scientists nor 24/7 software engineers. Today's genAI low-coding tools fail to help with evals. As soon as folks want LLMs to go into production, vs generated JS/Python, think dashboards or alert analyzers that use prompt-based LLM agent flows to think at runtime.... we have a major functionality gap. "Hallucinations" are a symptom of this very wide gap. An analog here is when programming moved from assembly to structured programming with type checking compilers and garbage collection runtimes to automate a lot of the low-level quality work for you. Anyone who has used dspy, langchain, etc knows LLMs are still far from the equivalent notion of 'safe' structured programming for RAG/agents/etc.

I have written a new article that explores the development of an AI agent capable of generating code directly from GitHub issues and automating the pull request process. By integrating Large Language Models (LLMs) with GitHub, this approach aims to enhance software development efficiency.

📢 𝗜𝘀𝘀𝘂𝗲 𝟰𝟯𝟮 𝗶𝘀 𝗢𝘂𝘁! In today's edition, you can find: 📰 𝗡𝗲𝘄𝘀: - Nvidia ousts Intel from Dow Jones Index after 25-year run - Perplexity CEO offers to replace striking NYT staff with AI - ChatGPT Search is not OpenAI's 'Google killer' yet ... and more! 💡📚 𝗔𝗿𝘁𝗶𝗰𝗹𝗲𝘀: - OOPS in JS - Ultimate (by Subham M.) - 🔥12 Best AI Coding Assistant Tools for Devs🧑💻 (by Kiran Naragund) - Re-discovering the Joy of Building (by Arpit Nath) - Mastering Clean Code with SOLID, DRY, KISS, and YAGNI Principles (by Himanshu Singour) ☠️ 𝗣𝗼𝘀𝘁𝗺𝗼𝗿𝘁𝗲𝗺: PagerDuty's third-party SMS and voice service failed due to AWS outage. 👨💻 𝗥𝗲𝗽𝗼𝘀𝗶𝘁𝗼𝗿𝗶𝗲𝘀: From Binance Spot API docs to high-performance HTML/CSS renderer. (by Binance Paradigm Dioxus Labs) 🔗 𝗖𝗵𝗲𝗰𝗸 𝗶𝘁 𝗼𝘂𝘁: https://lnkd.in/eVhwm-rV 💌 Subscribe to get fresh news every day: https://0xcafe.news

I've been exploring how to work with large codebases using AI for a while. With larger context windows, it works really well to pack your entire repo—or at least a module or slice—and share it as context, especially for o1 pro. I've seen it create an entire feature for an app's TypeScript backend, only requiring some adjustments. O1 pro spends about 3 or 4 minutes thinking and, because I've given thousands of tokens worth of code, also follows that code's patterns. Very powerful pattern overall but I'm sure it's a fairly crude instrument. I hope to see more powerful just-in-time, say, exploration of a codebase by the AI itself for example, as it reasons, so that it picks just the necessary amount of context.