Cognitive Space’s Post

Are AI agents really writing production-quality code? 🤔 I recently explored how autonomous agents are tackling real-world software engineering tasks - and how one approach improved benchmarks by 3x. Curious to learn more? Check out my full blog post on AI made simple: https://lnkd.in/dKC-_8cj Key insights: • Using agent-computer interfaces (ACI) to enable AI-human-like interactions • Optimizing commands and feedback for language models • Implementing guardrails to prevent error propagation Thanks to Devansh Devansh for the guidance and featuring the article! Softude

Google DeepMind just unveiled Code Gemma, a groundbreaking collection of open code models. Gemma can now code! 🤯 Here's what you need to know: -- Code Gemma comes in 2B & 7B sizes, boasting an impressive 8192 context. -- These models are initialized from Gemma Base and further trained on a whopping 500B additional Tokens (web, code & math). -- Fine-tuned with SFT & RLHF for superior performance. -- The 2B model achieves 27% on HumanEval, while the 7B model surpasses expectations with a staggering 52%. -- Commercial use is allowed, making it perfect for industry applications. -- Optimized for on-device code completion, enhancing efficiency and productivity. -- You can access these cutting-edge models on Hugging Face. Get all the details: 🔗 Models: https://lnkd.in/gmsKgTWE 🔗 Google Blog: https://lnkd.in/gcFGufaf

🦙🎉 We've expanded our Prompt Engineering with Llama 2 course to include the new Llama 3 model by AI at Meta! Interact with these models through API calls and: 🌟 Learn prompt engineering best practices for Llama 2 & 3 🧪 Experiment with few-shot and chain-of-thought prompting 💻 Code alongside Code Llama and learn to use Llama Guard to check prompts and responses Get started and let us know what you think of this update! Sign up here: https://hubs.la/Q02vH6kr0

Introducing ExecuTorch Alpha ⚡ ExecuTorch Alpha is focused on deploying large language models and large ML models to the edge, stabilizing the API surface, and improving installation processes. Learn more in our latest blog: https://hubs.la/Q02vzv0k0

🚀 61% of engineering teams are already on board with Generative AI! Curious why and how to join them? 🤔 Check out our latest blog to see how AI is transforming software engineering and what it means for your team. Link in comments.

Running Inference on Edge just got a little bit easier. Check it out if you are or considering running an LLM on an edge device (Mobile, Desktop, IoT, etc)

Last one as we work through the modules of this interesting new course by Curtis Evans - https://lnkd.in/g5Q9xb-4 - Fundamentals of Generative AI for Beginners. Building large language models (LLMs) can be expensive and their output can be biased. In Part 4, you'll tackle some important design considerations for building LLMs with integrity, including: 🔧 how to build using quality data with optimized parameters 🔧 how to use suitable architectures to build 🔧 how to optimize model training 🔧 how to test models, evaluate them, and deploy them In addition, you'll learn about some important ethical considerations surrounding the building and use of these models to avoid harmful outputs. Put on your builder's hat, and let's go!

Inspired by the legendary Jon Krohn, I’ve signed up to teach an O'Reilly Pearson Live Event next Friday at 1pm ET on “Choosing the right LLM - how to select, train, and apply state-of-the-art LLMs to real-world business use cases.” My timing is impeccable: yesterday’s arrival of the astonishing Claude 3.5 Sonnet, breaking benchmark records including graduate-level reasoning, has only added to the DIZZYING array of LLM options. This class will get hands-on with 10 LLMs ranging from the largest Frontier models to the smallest Open Source alternatives. We’ll use Leaderboards and Arenas to compare specialist and generalist models. We’ll look at real world applications of LLMs, like Nebula.io, and discuss strategies for selecting and applying LLMs to commercial problems. Finally, we’ll work through 3 example business problems using techniques like RAG and QLoRA fine-tuning. Along the way, we’ll experiment with Hugging Face, LangChain, gradio, Weights & Biases, llama.cpp and more. The hardest part of preparing for this class is keeping up with the daily groundbreaking progress in Gen AI!! Please can I ask everyone working at frontier labs to take the next week off to give me a chance.. By the way - if you’re not already a member of O’Reilly, you can sign up for the 10 day free trial to take my class at no cost.. https://lnkd.in/guzcDA3x

Great talk by Prof. Subbarao Kambhampati, on LLM limitations and opportunities, laced with a lot of good humor ! His research team is doing incredible work to bring clarity on LLM potential vs hype. Else a lot of investment dollars will go wasted chasing shiny things which happens way too often in emerging tech ! I can resonate with a lot of the findings, and in some of the early work on building LLM based apps, we are seeing that an SME + prompt engineering (manual or programmatic) + QA skills are able to bring more stable outcomes. In fact, Alpha Code2 architecture from Google, is having a similar intiuition, of scaling generative modeling followed by discriminators. ps: Like the movie Salaar, first half of the lecture is easy, second half can be quite dense ! https://lnkd.in/gH55V5JA

ExecuTorch Alpha is here and it's bringing more AI #onArm! 🙌   Large language models already run on Arm CPUs, but with the release of ExecuTorch Alpha from PyTorch, developers now have additional support to ensure optimal performance of LLMs on the CPU alongside delegation to GPU and NPU.   See how you can get started with Meta’s Llama 2, 3 and other broad models in the link below.