Prithviraj Panda’s Post

Developers who think they found the utopia with LLM based code generators, it may not last very long. All the sources for LLM training content for code including StackOverflow, GitHub, GeeksforGeeks, Blogs and other web sites will eventually be flooded with LLM generated code. They will be used for further training eventually leading to model collapse. You can not stop giving in terms of human creative work while at the same time expect the models to get better and keep on taking from the models. It’s delusional to think otherwise. There will be rude awakening #ai #llm #genai #codegenerator #model #collapse #developer

Many devs still discard llms as useless (for coding). Why? They are using Github Copilot and get mediocre results. Copilot feels ~ gpt-3.5 ish. Not terrible but not great either without much experience. That's a big mistake. To get a feeling for how far along we are: Try ChatGPT w/ Code Interpreter or Sourcegraph Cody w/ gpt-4o and explicit autocompletion (option+k). Immerse yourself.

I keep seeing this line of argumentation and can't highlight enough how misguided it is. The argument is based on a study that repeatedly trained models on unfiltered outputs from previous models and observed a degradation in output. Which indeed will happen, because these models learn an interpolation function passing through the training data, which is by its smoothing nature exhibits lower variance than the original. But this about the point where the comparison stops making sense. The internet doesn't just get populated with raw model outputs for the next round of LLM/diffusion model intake. The internet gets populated with *FILTERED* model outputs. People will use LLM generated code on GitHub et all (usually) that actually works. All the iterations that were just garbage output will never see the light of day. Same with images. The anime girls with too many heads and fingers? Trash bin. The cat with distorted eyes? Also trash. Messed up text? You guessed it. Trash. The active filtering that people posting the AI generated content onto the websites that the new AI will learn from will, to a great extent, negate the model collapse problem invoked by the detractors of the technology and lead to a more able next set of models, not less. Gen AI output validators (whether human or programmatic) are already an informal industry standard in Gen AI applications and used in a wide variety of use cases. This... is no different.

CodeMate #VSCode Extension 2.9.0 released yesterday. Checkout what's new 🤯 - Code Action and Code Lens added for modifying codebase, creating docstrings, inline comments or asking any question within the editor through in-line commands 💯 Added shortcut commands to trigger CodeMate inside editor- - // Codemate: <query>: for asking any questions related to the code - // Generate: <query>: for generating code from natural language within the editor keeping the existing code in context Checkout CodeMate AI now at https://codemate.ai Ayush Singhal Kshitij S. Tyagi #codemateai #softwaredevelopment #AI #Programming

💡 New Tutorial 💡 Are you looking to optimize your Python and Flask applications by dynamically swapping AI models based on user context? Discover how you can switch between AssemblyAI's state-of-the-art Speech-to-Text models, based on application contexts like user email domain, device, zip code, and more. Our Universal-1 model introduced a dual-class tier system, allowing developers to choose between the highest accuracy “Best” tier and the cost-effective “Nano” tier. This tutorial demonstrates how to leverage these tiers with LaunchDarkly for optimal app performance. Take your app development to the next level and check out the tutorial in the below LaunchDarkly post 👇 #AI #MachineLearning #Python #Flask #LaunchDarkly #AssemblyAI #TechTutorial #AppDevelopment

GPT-4 Multimodal capabilities seem wild! Here is an example of picture-to-recipe with gpt4 There are umpteen examples from - Coding from screenshots of dashboards - Breakdown of workflow diagrams - Provides ideas for image rendering and design - Figma designs to code - Convert python code into javascript And more As a Mashable article states, it can also help you find Waldo :) Have you tried out ChatGPT Vision ? ************************** Ranjani Mani #reviewswithranjani #GenerativeAI #Technology | #Books | #Beingbetter

Tired of debugging? JDoodle's AI Powered "Debug Code" feature effortlessly identifies coding errors, explains them & even suggests the corrected code. Give it a try: https://bit.ly/3SpWXRM #jdoodle #debugcode #debugging #onlinecompiler #ide