Struck Studio’s Post

Welcome to GibsonAI, your #ai backend engineer. Members of our beta group have been asking us if you can use GibsonAI with a pre-existing product. You can! We wanted to show a quick demo of that here. What's new? (1) We released v0.4.0 of our Code Composer. Code Composer is an intelligent system that combines the capabilities of an external LLM with GibsonAI to write code in a more efficient way with low latency. The goal is to make smart coding decisions for you but continue to give you software that can go directly to production. (2) Our AI team is working on our generic-v8 modeling system for creating even better, more consistent data models. We will keep you posted on all the progress there but expect a very excited Mike to show that off once it is ready. (3) We're making progress on our AI Pair Programmer. The functionality will include context about your database and coding environment and allow you to code in either SQL or Python. As always, additional languages are on the list. The goal of Pair Programmer is to keep you in the driver's seat while allowing you to code MUCH faster with more consistency. It's coming, I promise. (4) OK, and now for a big, audacious bet that I'm hoping to deliver on. I'm not a huge fan of Alembic, or migrations in general. Might AI be able to migrate your database automatically, consistently with each new release? If GibsonAI has it his way, then yes. Please don't hesitate to provide feedback in the comments or to reach out to me directly. Thanks for watching and stay tuned!

The rise of repository-to-prompt tools marks a significant shift in AI-assisted software development. These utilities, which convert entire codebases into prompts for large language models (LLMs), are reshaping how developers interact with AI. While promising enhanced productivity and code quality, these tools raise questions about security and the nature of software development itself. Are we sacrificing deep understanding of our codebases for quick AI-generated insights? The ability to feed entire projects to LLMs for comprehensive code reviews and refactoring suggestions is powerful, but it may lead to over-reliance on AI-driven solutions. Moreover, as these tools mature, we're seeing trends towards direct LLM integration and VRAM estimation. This push towards seamless AI integration in development workflows is both exciting and concerning. It could democratise access to advanced code analysis, but also potentially homogenise coding practices. The diversity of available tools - from simple Python scripts to web applications - highlights the growing importance of this niche. Yet, we must remain vigilant about the security implications, especially when dealing with proprietary or sensitive codebases. As developers, we need to strike a balance between leveraging these powerful AI tools and maintaining our core skills and understanding. What safeguards should we put in place to ensure responsible use of repository-to-prompt tools? How can we harness their potential without becoming overly dependent on AI-generated solutions? https://lnkd.in/emw9EVr2 #AIDevelopment #SoftwareEngineering #CodeSecurity

When it comes to using AI for coding on a large scale, there's much excitement and promising early examples. While we have real, practical tech today that can generate code at the method/snippet level quite well, tools for working at the application level, like Cognition Labs' Devin, are still on the horizon. One area that seems especially exciting in the shorter term is porting - rewriting an existing codebase with a new technology. For example, Amazon Q Developer purports to be able to port a codebase between major versions of Java. This is a compelling use case because porting a large codebase is (1) often a massive job with many details to keep track of, and (2) tedious work that nobody wants to do. Here's an example from Slack where they got creative and were able to automate much of a 10,000 hour job converting their automated testing suite from one framework to another. https://lnkd.in/eJyh33cd

A great post from the @Stackoverflow blog on how developers are using AI assistants for coding. ChatGPT is the most-used, followed by GitHub Copilot, then VSC IntelliCode and Codeium. Out of these, I've spent the most time with GitHub Copilot and found it very useful (specifically for Python & Bash), though I should note I don't consider myself a programmer 🙂 https://lnkd.in/gHJB2AXv

Why don't we use LangGraph or LangChain (anymore)? This question still occur from time to time. So I'm gonna explain it here. 1. LangChain besides the 'intregration' is basically an over-engineered python formatting engine. About the integration, once you have already chosen your stack for your project it gives very little benefit as you want to control and squeeze the most of your integration, you will have to re-implement things. 2. LangChain focused on React agents which never worked for non-common tasks because they stay in their data distribution without guidance. The hypothesis from the start was that scaling will solve this issue, hypothesis supported by the same big AI companies that falsify benchmarks and lie to people (and congress). We already knew 1 year ago that this would never work. Now people are realizing it, 1 year for such an evident thing (and how many $?). Apparently the so called "experts" were not even able to benchmark their solution and put into production something that they didn't test themselves. 3. LangGraph is an attempt to solve LangChain defaults (loops mostly), by adding another layer of over-engineered structure on top of it. The foundation is brittle and provide very little benefit for the user. The only thing that they provide is locking people into their architecture to sell them other services (LangSmith and Cloud now). 4. There is more value in designing systems with DSPy that provide evaluation of pipelines, automatic prompt optimization and finetuning of LoRa adapters than LangChain. Removing LangChain from your codebase to focus on DSPy will give you better optimization, control and tangible results! 5. If you choose LangChain/LangGraph because you can't build a proper software, we advise you to use HybridAGI that make the use of DSPy straightforward and include an hybrid vector/graph database + dynamic workflows made with Cypher (easier to program). The software is under free and open-source license, because we REALLY value the community and don't sell false promises. Yes you need to think about your problem/workflow maybe even your data structure if you use our fact memory. But the reality is that if you wasn't into AI the last 5 years, you are still learning and "zero-effort" solutions will fail as you will not be able to identify what went wrong to fix it. So better have the control over the LLM and build incrementally.

Is your agentic workflow stuck in pointless loops? What about reproducibility of results? Do you have any control over it? Here are some words of wisdom from Yoan Sallami (SynaLinks). TL;DR -- * 𝑩𝒖𝒊𝒍𝒅 𝒊𝒏𝒄𝒓𝒆𝒎𝒆𝒏𝒕𝒂𝒍𝒍𝒚 * 𝑼𝒔𝒆 𝒑𝒂𝒄𝒌𝒂𝒈𝒆𝒔 𝒕𝒉𝒂𝒕 𝒂𝒍𝒍𝒐𝒘 𝒇𝒍𝒆𝒙𝒊𝒃𝒊𝒍𝒊𝒕𝒚; 𝒂𝒗𝒐𝒊𝒅 𝒍𝒐𝒄𝒌-𝒊𝒏 * 𝑫𝒐𝒏'𝒕 𝒖𝒔𝒆 𝑨𝑰 𝒂𝒔 𝒂𝒏 𝒆𝒙𝒄𝒖𝒔𝒆 𝒕𝒐 𝒏𝒐𝒕 𝒖𝒏𝒅𝒆𝒓𝒔𝒕𝒂𝒏𝒅 𝒚𝒐𝒖𝒓 𝒅𝒂𝒕𝒂 𝒐𝒓 𝒕𝒉𝒆 𝒃𝒖𝒔𝒊𝒏𝒆𝒔𝒔 𝒍𝒐𝒈𝒊𝒄. #agents #ai #llm #genai #data

If you read only one thing about AI/LLMs this week, make it this: SWE-Bench Technical Report from Cognition Labs https://lnkd.in/e4sSWEwY Like most of you, I’ve been hearing a lot this week about Devin, a semi-autonomous AI software developer agent. Congrats to the team! It’s an impressive product that thinks about software engineering much like we do. Please do skim this report, it’s quick and great food for thought. Here are a few key insights: • Much of the improvement over previous LLM benchmark baselines comes from Devin’s ability to write and run code in its own local environment, which allows it to autonomously iterate by analyzing test results and attempting fixes. • Devin is also capable of making and executing multi-step plans. This enables it to decompose hard problems into easier ones, which are much easier for an LLM to solve on its own. • Devin doesn’t always get it right! The team shared several examples of failed understanding and broken code, often because of difficulties editing multiple files or understanding complicated logic. Some of this can be fixed via prompting or the environment harness, but as models improve, these problems may fix themselves! What a time to be alive. Excited to share soon what we’ve been working on at Stride.  We're focused on using AI to help dev teams do more of what they love, and less of what they don't. Bonus link: Check out this Figure 01 video. Remember: it’s only going to get better from here. https://lnkd.in/eiNphsyK

13 Hidden Open-Source Libraries to Level Up Your AI Game 🧙♂️🪄 If you're building AI applications, you NEED to check out this curated list of open-source tools and frameworks. Sunil Kumar Dash, an AI expert with 4+ years of experience, shares his favorite hidden gems to streamline your development process. 🚀 Highlights include: - Composio: Build reliable AI agents 10x faster - Julep: Create stateful AI agents with built-in memory - E2B: Securely execute code in AI apps - Camel-ai: Build scalable multi-agent systems - CopilotKit: Easily add AI capabilities to React apps ...and many more! These libraries can help you tackle everything from AI agent development and RAG applications to code execution and vector databases. Don't miss out on this valuable resource – explore the full list and give your AI projects a boost! #AI #OpenSource #MachineLearning #SoftwareDevelopment #Productivity https://lnkd.in/gZuGHwkH

Even Devin AI has a Girlfriend :) Meet : Devika - Agentic AI Software Engineer 👩💻 Devika is an advanced AI software engineer that can understand high-level human instructions, break them down into steps, research relevant information, and write code to achieve the given objective. Devika utilizes large language models, planning and reasoning algorithms, and web browsing abilities to intelligently develop software. Devika aims to revolutionize the way we build software by providing an AI pair programmer who can take on complex coding tasks with minimal human guidance. Whether you need to create a new feature, fix a bug, or develop an entire project from scratch, Devika is here to assist you. Key Features 🤖 Supports Claude 3, GPT-4, GPT-3.5, and Local LLMs via Ollama. For optimal performance: Use the Claude 3 family of models. 🧠 Advanced AI planning and reasoning capabilities 🔍 Contextual keyword extraction for focused research 🌐 Seamless web browsing and information gathering 💻 Code writing in multiple programming languages 📊 Dynamic agent state tracking and visualization 💬 Natural language interaction via chat interface 📂 Project-based organization and management 🔌 Extensible architecture for adding new features and integrations Read More : https://lnkd.in/gVxaqmXG

Replit is developing a Code Repair LLM, leveraging Replit's rich dataset of LSP diagnostics, which indicate errors in code but often don't suggest fixes. By synthesizing and verifying code error and repair pairs, they aim to build a model capable of suggesting repairs. This process includes distilling a more accurate model from synthesized data, avoiding mode collapse by starting from real error states, and ensuring the generated fixes are applicable. The model is fine-tuned on a dataset of these pairs without using natural language instructions, instead using a templated schema that mimics IDE interactions. They trained a 7B parameter model optimized for balance between capability and cost, using advanced techniques for efficient learning and inference. Initial evaluations show their Replit Code Repair model competes well with larger models on both their custom Replit benchmark and the Leetcode repair eval, highlighting the potential of specialized LLMs for improving coding efficiency. https://lnkd.in/g9xa6uj4