Mwape Chintu

Introducing GitHub Copilot agent mode (preview): the next evolution in AI-assisted coding. Copilot in agent mode is capable of running terminal commands, iterating on its own code to fix errors, and so much more. Available in VS Code Insiders today. Read more in our blogpost: https://lnkd.in/ggxf5xWH

3.305

55 Kommentare

StarVector is now on Hugging Face. You can transform images and text into high-quality SVG code with this new foundation model. The model uses Vision-Language architecture to process both visual and text inputs. It handles vectorization and text-guided SVG creation with impressive accuracy. Check out the models here: https://lnkd.in/gAU5j68t

623

28 Kommentare

🚨 OpenAI just released o3 and o4-mini! It's their most advanced reasoning models yet, now with full tool access and multimodal capabilities. Codex CLI, a terminal-based coding agent, is also part of this release. Key Features Smarter, Faster, More Capable: • o3 is OpenAI’s most powerful reasoning model to date, delivering best-in-class performance across coding, math, science, and visual perception tasks. • o4-mini is a smaller, cost-efficient reasoning model that still performs at elite levels, scoring 99.5% on AIME 2025 with tool access and significantly outperforming o3-mini. Full Tool Access in ChatGPT: • Both models can now search the web, run Python code, analyze uploaded files, interpret visual inputs, and generate images. • Trained to strategically decide when and how to use tools, enabling more thoughtful and complete answers, typically in under a minute. Multimodal Reasoning, Redefined: For the first time, OpenAI’s models can reason directly with images, interpreting whiteboards, diagrams, and even blurry sketches. • Combine visuals with search and code to generate dynamic, interactive outputs across domains like education, consulting, and science. Reinforcement Learning at Scale: • o3 and o4-mini benefit from massive reinforcement learning investments, allowing them to think longer and smarter with each inference step. • Reasoning accuracy increases the more time they’re given to think. Improved Safety and Refusals: • New safety datasets and LLM-powered monitors help prevent misuse, detect jailbreaks, and flag dangerous content (e.g., biorisk or malware generation). • Passes OpenAI’s updated Preparedness Framework, remaining below high-risk thresholds in all categories. Codex CLI: • OpenAI also released Codex CLI, a terminal-based coding agent that supports multimodal reasoning from your desktop. • Open-source and eligible for $1M in grants via OpenAI’s new fund. Access & Availability: • Available now in ChatGPT Plus, Pro, and Team, replacing o1, o3-mini, and o3-mini-high. • Free users can try o4-mini via the new "Think" mode. • Enterprise and Edu access begins in one week. • Also live in the Chat Completions API and Responses API, with support for custom tools and reasoning summaries. ⁠ Video Credit: OpenAI #openai #chatgpt #generativeai #ai #artificialintelligence

858

26 Kommentare

GitHub has unveiled Agent Mode for GitHub Copilot, enabling its AI assistant to iterate on its own code, detect errors, and fix them autonomously. Announcing the launch on X, GitHub CEO Thomas Dohmke said, "We are infusing the power of agentic AI into the GitHub Copilot experience, elevating Copilot from pair to peer programme." Alongside this, Copilot Edits is now generally available in Visual Studio Code, letting developers edit multiple files using natural language. GitHub also introduced Project Padawan, an AI-powered software engineering agent designed to handle coding tasks, review code, and automate workflows. At the Microsoft AI Tour in Bengaluru, GitHub’s Director of International Developer Relations, Karan MV, showcased how Copilot Workspace can generate code in Indian languages like Hindi and Kannada. Dohmke emphasized, "Every developer will conduct a symphony of AI agents in natural language." With AI-driven development becoming the future, GitHub Copilot is evolving from an assistant to an autonomous coding powerhouse. Read the complete article to know more. Link in the comment 🔗

292

5 Kommentare

→ What is Google Nano Banana? The name sounds strange. But behind it hides a tech journey worth exploring. • It starts with the Input Stage - capturing raw data. Could be text, images, or both. • Then Text Processing and Image Preprocessing prepare the inputs. Clean, structured, ready for the next steps. • Multimodal Alignment is where magic begins. It aligns different data types so they “talk” to each other. • Concept Understanding follows - the system grasps meaning and context. Not just words or pixels, but ideas. • Noise Initialization introduces controlled randomness. This stirs creativity and variation. • Guided Transformation shapes the raw data, driven by goals or instructions. • Attention Mechanism zooms in on what matters most - a focused lens for precision. • Iterative Refinement repeats transformations, polishing the output step by step. → But hold on. What ensures this process stays safe and consistent? • That’s the role of Safety & Consistency Check - zero tolerance for errors or biases slipping through. • Final Polishing adds finesse - smoothing edges, enhancing clarity. • Lastly, Output Delivery presents the polished product - ready to use, share, or build upon. 🔥 Join Our community for latest AI updates: https://lnkd.in/gNbAeJG2 🔥 Access to all popular LLMs from a single platform: https://thealpha.dev

65

18 Kommentare

The New Stack spotlights the PyTorch Foundation’s evolution into an umbrella foundation and the addition of vLLM and DeepSpeed as newly hosted platform projects. “Since its inception at the Linux Foundation two and a half years ago, the PyTorch Foundation has grown to include over 30 member organizations and 120 ecosystem projects. These newest moves are a natural evolution of its rapid growth and global momentum.” 🧠vLLM: Built for fast, efficient LLM inference with innovations like PagedAttention and deep PyTorch integration. ⚙️ DeepSpeed: A go-to library for scalable training, leveraging technologies like ZeRO and 3D parallelism—all built on PyTorch. 📖Read the full article from Steven J. Vaughan-Nichols: https://hubs.la/Q03lFYd00 #PyTorch #vLLM #DeepSpeed #OpenSourceAI #LLMInference

84

1 Kommentar

Neural embedding models have become a cornerstone of modern information retrieval. Today we introduce MUVERA, a state-of-the-art retrieval algorithm that reduces complex multi-vector retrieval back to single-vector maximum inner product search. More → http://goo.gle/4k8YRlN

454

4 Kommentare

Build a RAG system over your PDF documents! Dietmar (@docd in n8n) has built an n8n workflow that transforms PDF documents into searchable vector embeddings: 1️⃣ Preprocess PDF documents with Mistral AI OCR 2️⃣ Embed chunks with Cohere and store in a Weaviate vector database 3️⃣ Retrieve and rerank with Cohere’s reranker 4️⃣ Orchestrate everything with n8n n8n template: https://lnkd.in/gJucMmyh

93

1 Kommentar

CytoReason's RAG pipeline, utilizing NVIDIA NIM microservices, revolutionizes the curation of biological findings from scientific literature. ➡️ https://nvda.ws/3RdklC3 🧬 By automating the extraction of insights using large language models (LLMs), CytoReason has reduced manual curation time from days to hours while achieving 96% accuracy. #NVIDIAInception

74

2 Kommentare

This week, in The Batch, Andrew Ng discusses why AI-assisted coding makes it easier for developers to work across programming languages, but argues that understanding core concepts still matters. Plus: 📰 OpenAI launches five new models, will retire GPT-4.5 📰 Hugging Face debuts open robot 📰 U.S. escalates restrictions on AI chip exports to China Read The Batch: https://hubs.la/Q03jDzYz0

62

4 Kommentare

🚨 The latest episode of Gradient Dissent is here! One of our most thoughtful conversations yet. Lukas Biewald sat down with Thomas Dohmke, CEO of GitHub, to explore what software development looks like in an AI-powered world. They covered: – Why the GitHub + Microsoft acquisition actually worked – Why developers won't be replaced – The origin story of GitHub Copilot – Why “developer happiness” is a metric that moves the needle – And how systems thinking, not just coding—is the real skill of the future 💡 “If AI could build your app from a few prompts, your startup has no moat. Software still needs to be designed.” – Thomas Dohmke This is a grounded, insider’s look at how modern software engineering is evolving and what it means to lead in the age of AI. 🎧 Give it a listen and let us know what resonated most: https://lnkd.in/ghkvSfVk 📍 P.S. Fully Connected is happening June 17–18 in San Francisco. Join us for two days of hands-on workshops and talks from the top minds in AI. Reshare this post and you’ll be entered to win a free ticket.

36

1 Kommentar

📓 Fusion Diaries: Redshift support, clone, better stack traces, and a long-awaited bug squash. This was a big week for dbt Fusion, a mix of shipping new features and refining existing ones to improve everyday development. Here’s what’s new: 𝗥𝗲𝗱𝘀𝗵𝗶𝗳𝘁 𝘀𝘂𝗽𝗽𝗼𝗿𝘁 (𝗯𝗲𝘁𝗮) Fusion now supports Redshift via user+password authentication. More auth methods are coming later this year. 𝗡𝗲𝘄 𝗰𝗼𝗺𝗺𝗮𝗻𝗱: 𝗰𝗹𝗼𝗻𝗲 You can now create zero-copy clones (or fallback views) of production objects in your development schema. This is especially useful in cases where state deferral isn’t working as expected. 𝗖𝗧𝗥𝗟+𝗖 𝘁𝗼 𝗰𝗮𝗻𝗰𝗲𝗹 Change your mind halfway through a build? Fusion will now cancel in-progress queries when you hit CTRL-C instead of running them to completion. 𝗗𝗲𝘃 𝗲𝘅𝗽𝗲𝗿𝗶𝗲𝗻𝗰𝗲 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁𝘀 • Flags like --sample and --empty give you more control during model scaffolding • VS Code extension now includes “Refresh source schemas” and improved tooltips for multi-column expressions • Stack traces now preserve line-level context across packages, making it easier to debug and navigate errors •Improved introspection labeling and more precise error messages 𝗢𝗻𝗲 𝘀𝘁𝘂𝗯𝗯𝗼𝗿𝗻 𝗯𝘂𝗴, 𝗳𝗶𝘅𝗲𝗱 We finally tracked down the cause of a weird configuration bug - turns out Fusion was ignoring schema config overrides in packages with more than 40 models. A one-line fix once we found the problem. 𝗪𝗵𝗮𝘁’𝘀 𝗻𝗲𝘅𝘁 We’re finalizing the new authoring layer that will unify YAML validation and improve the dbt editing experience across Fusion and Core. We’re also continuing work on: • Introspective macro caching • Incremental compilation • Smarter deferral behavior This week we closed 30+ issues, released 5 new betas, and made meaningful progress toward a more reliable, powerful dbt experience. Thanks to everyone in the community who filed bugs, tested features, and helped shape the work. Read the full diary (link in comments) for more.

82

3 Kommentare

🎥 GraphRAG from Video with Qdrant and Neo4j In this project, Athos G built a multimodal retrieval pipeline that generates a knowledge graph from raw video. It combines object detection, vision-to-text grounding, temporal tracking, and combined storage in both graphs and vectors to enable retrieval that understands meaning, relationships, and context over time. 🔗 Check it out: https://lnkd.in/dcHKJYxy

253

6 Kommentare