Microsoft Research’s Post

Some really exciting results from Microsoft Research demonstrating that we are at the absolute cutting edge of AI/LLM research. Full disclosure, get ready for some complex and deep detail (!!) but the spoiler is this provides a clear path to revolutionize the efficiency, including cost and energy consumption, of future LLMs. #Microsofteducation #highereducation #Microsoftedu

[arXiv] Microsoft Research introduces Q-Sparse: a breakthrough in training fully sparsely-activated LLMs. Q-Sparse supports both full-precision and 1-bit LLMs. Its synergy with BitNet b1.58 advances LLM efficiency, including cost and energy use. https://msft.it/6019lumcJ

More on BitNet, the 1.58 bit LLM paper by Microsoft, which has become one of the most upvoted papers since it came out. This one: https://lnkd.in/eFcyUUhb Looks like there have been more than one independent replication of the result in the paper. One is a 1B model trained on 60B tokens, another is a 3B model trained on 100B tokens. Here: https://lnkd.in/eQyx_9_c And here: https://lnkd.in/e6a2vixH And a blog post on 1bit ML, it may be generalizable across different domains and modalities (with a Colab notebook to test out yourself): https://lnkd.in/ehZDiEKi There’s even talk of using 1-bit embedding vectors for efficient RAG. It’s a new world. Not to sound like a broken record, this looks like a promising direction for further democratizing LLMs. Models are getting bigger as per the scaling law, and the GPU poor consumers just can’t keep up the pace. This is potentially a viable solution.

🚨 Important Update from Microsoft 🚨 Starting July 10, 2024, Microsoft will discontinue the creation of GPTs. Additionally, from July 10, 2024, to July 14, 2024, all existing GPTs, including those created by customers, will be removed along with their associated GPT data. Stay informed and plan ahead! #Microsoft #GPT #Update #Technology #DataManagement

Today I earned my "Fundamentals of Computer Vision" badge! I’m so proud to be celebrating this achievement and hope this inspires you to start your own @MicrosoftLearn journey! #microsoft #microsoftlearn #Azure #MicrosoftLearn #Azurefundamentals #MSFTedu #MicrosoftTraining #LearnWithMicrosoft #MicrosoftCertification #MicrosoftSkills #MSLearningPath #MicrosoftEducation #TechWithMicrosoft #MicrosoftVirtualTraining #MSLearnCommunity #MicrosoftUpskill #LearnMicrosoft #MicrosoftAcademy #MSCertifications

🤖RDAgent (from Microsoft) RDAgent aims to automate the most critical and valuable aspects of the industrial R&D process This framework has two key components: 'R' for proposing new ideas and 'D' for implementing them https://lnkd.in/gzVAp-_e

Recently read about Swirl: https://lnkd.in/gu5TRgJk So without really training/loading your Enterprise data into an LLM, to be able to ask it Questions, you can use Swirl which connects your data from multiple sources to the LLM service of your choice. Thus reducing time/cost to load all that data and it also removes hallucinations by giving solid links. Once we get micro LLMs, or tech to better compress data, I am pretty sure everyone will have a google search type LLM on their phones (not even talking about laptops). Its such a good idea. Anyone reading this post used or using it ? #Swirl #LLM #Enterprise

A Major Move by Microsoft: RD-Agent Goes Open Source! Just reviewing Microsoft's latest open-source release, RD-Agent, and it feels like a game-changer for AI-driven R&D automation. Here's why I think it may me a major move: - Automates high-value R&D processes - Self-evolving AI systems that learn from feedback - Supports multiple scenarios including quantitative trading, medical research, and Kaggle competitions - Built-in copilot capabilities for research paper implementation So, the framework can both PROPOSE new ideas AND IMPLEMENT them - essentially creating a closed loop of continuous R&D improvement. Think of it as an AI researcher that never stops learning. Sure, we can geek out and it’s exciting, but here are some practicable real-world applications too: - Automated Quant Factory - Data Mining Agent - Research Copilot - Kaggle Competition Agent This could be a significant step on the paths to truly autonomous AI research and development. The implications for productivity and innovation are massive. And a tad scary. Currently at 1.1k stars on GitHub and growing fast! How you trues it yet? If so please share your thoughts? How do you see this impacting the future of R&D? #ArtificialIntelligence #Innovation #OpenSource #Microsoft #RD #DataScience #AIResearch

In 1995, a bug could remotely crash every Windows 95 system, but the impact was limited because computers weren’t as embedded in our lives as they are today. Fast forward to now, and we’re living in a world where a single faulty update can bring down entire systems—think about how much of our infrastructure relies on Microsoft’s cloud servers. This “excessive centralization” is a ticking time bomb. Now, let’s talk about AI. Imagine today’s AI as a new operating system, one that will become as intertwined with our economy as Microsoft is now. In 5-10 years, AI won’t just be a tool; it’ll be the backbone of industries like healthcare, transportation, military, and justice. But here’s the thing: with great power comes great vulnerability. Consider a scenario where a major AI system faces a Crowdstrike-like incident. The ripple effects could be catastrophic, exponentially worse than any IT outage we’ve seen. Why? - > Because unlike software bugs, AI systems are decision-makers. They learn, adapt, and can affect real-world outcomes in real-time. A compromised AI in healthcare could misdiagnose millions, a hacked AI in transportation could bring cities to a standstill, and an AI malfunction in the military could have dire consequences. This isn’t science fiction; it’s a near-future reality. The notion that “AGI will fix it” is not a safety net. AGI (Artificial General Intelligence) is far from being a cure-all, especially if we rush headlong into deploying AI without rigorous checks and balances. We need to think about AI safety not just as an afterthought but as a foundational principle. This means embedding redundancies, ensuring human oversight, and building systems that can fail gracefully rather than catastrophically. The lesson from 1995 isn’t just about bugs—it’s about understanding the far-reaching consequences of centralized power and the importance of cautious innovation. #Windows #AI #ITOutage