Vandana Premkumar

Here is my column on the #TiEcon2024 conference in #indicanews:

61

4 Comments

Stay tuned !! To learn more about impact/use-cases of GenAI in Manufacturing

52

1 Comment

🎙️ AI's New Superpower: Multimodal Understanding 🧠 As someone who's been in the tech trenches for a while, I'm genuinely excited about the leap we're seeing in AI voice tech. Building on Ethan Mollick’s post and why it matters: 1. Context is King: ChatGPT's Advanced Voice isn't just processing words – it's grasping the full context, including emotion and tone. This is huge for natural interactions. 2. Beyond Text-to-Speech: We're moving past robotic voices. This tech can respond with varied voices and sounds, opening doors for more engaging AI assistants. 3. Holistic AI: Multimodal models process information more like humans do – through multiple "senses." This could revolutionize how AI understands and interacts with our world. 4. Expanded Capabilities: With more inputs and outputs, these models gain a richer understanding, potentially leading to more creative and nuanced AI applications. 5. Industry Shake-up: While OpenAI is public about this, you can bet Google and Meta are working on similar tech. The race is on! The potential here is massive. Imagine AI that truly "gets" us, leading to more intuitive interfaces and powerful assistive technologies. #AI #Tech #Innovation

1

💥💥💥 Mechanistic Unlearning: Robust Knowledge Unlearning and Editing via Mechanistic Localization Phillip Guo, Aaquib Syed, Abhay Sheshadri, Aidan Ewart, Gintare Karolina Dziugaite Abstract Methods for knowledge editing and unlearning in large language models seek to edit or remove undesirable knowledge or capabilities without compromising general language modeling performance. This work investigates how mechanistic interpretability -- which, in part, aims to identify model components (circuits) associated to specific interpretable mechanisms that make up a model capability -- can improve the precision and effectiveness of editing and unlearning. We find a stark difference in unlearning and edit robustness when training components localized by different methods. We highlight an important distinction between methods that localize components based primarily on preserving outputs, and those finding high level mechanisms with predictable intermediate states. In particular, localizing edits/unlearning to components associated with the lookup-table mechanism for factual recall 1) leads to more robust edits/unlearning across different input/output formats, and 2) resists attempts to relearn the unwanted information, while also reducing unintended side effects compared to baselines, on both a sports facts dataset and the CounterFact dataset across multiple models. We also find that certain localized edits disrupt the latent knowledge in the model more than any other baselines, making unlearning more robust to various attacks. 👉 https://lnkd.in/dpfrUDNr #machinelearning

1

Here is a challenge for NVIDIA to solve. Complex solutions like Fourcastnet and ModulusAI need substantial investment from a startup. But during planning and modelling phases other than webinar & vague communities there is no one to address any queries -- Downcasting of meterological data or climate model data using Fourcastnet is an important topic to discuss with solution providers needing support. Not everyone is working on a single purpose cheap application which has been the trend of past economy. With recognition of complex systems theories and nature's behaviour, folks are working on "microclimate solutions" or regional boundary systems of complex earth systems behaviour. So if NVIDIA wants to succeed in democratizing its work done on various platforms, much more is needed. You need folks with whom one can talk to to resolve queries to find out how much money is to be raised also. Due to this challenge the motivation to attend GTC or NVIDIA conferences is also minimised as one does not know if answers to queries can be found. NVIDIA has built the greatest force of positive change in technology but more needs to be done to harness the community which can be built around it. Nidhiya V Raj Anima Anandkumar Jensen Huang Bertrand Le Saux #fourcastnet #climatemodelling #AI #nvidia

4

1 Comment

Some amazing good use cases of #AI in some famous companies, e.g. JPMorgan Chase & Co. using 400+ good use cases of #AI in PROD. there ! Barron's article - https://lnkd.in/g4nJ3REA (need subscription) "Other industries will take longer to realize the impact of AI, but the productivity boost is coming. Financial-services firms are automating more back-office compliance functions, as well as complex underwriting processes, despite the regulated industry’s aversion to risk. JPMorgan Chase has a $15 billion annual technology investment budget, 2,000 AI and machine-learning experts on staff, and a newly created position of chief data & analytics officer reporting to CEO Jamie Dimon. “We have been actively using predictive AI and ML for years—and now have over 400 use cases in production in areas such as marketing, fraud, and risk—and they are increasingly driving real business value across our businesses and functions,” wrote Dimon in his annual letter earlier this month."

9

💥💥💥 $100K or 100 Days: Trade-offs when Pre-Training with Academic Resources Apoorv Khandelwal, Tian Yun, Nihal V. Nayak, Jack Merullo, Stephen H. Bach, Chen Sun, Ellie Pavlick Abstract Pre-training is notoriously compute-intensive and academic researchers are notoriously under-resourced. It is, therefore, commonly assumed that academics can't pre-train models. In this paper, we seek to clarify this assumption. We first survey academic researchers to learn about their available compute and then empirically measure the time to replicate models on such resources. We introduce a benchmark to measure the time to pre-train models on given GPUs and also identify ideal settings for maximizing training speed. We run our benchmark on a range of models and academic GPUs, spending 2,000 GPU-hours on our experiments. Our results reveal a brighter picture for academic pre-training: for example, although Pythia-1B was originally trained on 64 GPUs for 3 days, we find it is also possible to replicate this model (with the same hyper-parameters) in 3x fewer GPU-days: i.e. on 4 GPUs in 18 days. We conclude with a cost-benefit analysis to help clarify the trade-offs between price and pre-training time. We believe our benchmark will help academic researchers conduct experiments that require training larger models on more data. 👉 https://lnkd.in/d5Vh3ycg #machinelearning

1

Here's a 60-second clip from my recent API Days talk where I talk about how all the amazing capabilities of ChatGPT are strictly emergent from the data. This is still my single favorite visualization about LLMs because it seems to reveal something about the structure and hierarchy of intelligence. For example, as you scale parameters you can go from simple language understanding to translation, and somehow then end up with joke explanations. It also makes me think about how vast the theoretical tree of intelligence might be, which could be beyond human comprehension. #genai #llm #intelligencerevolution #agi

9

2 Comments

The results of our AI in the Enterprise survey make for very interesting reading. I can’t say I found them surprising, as the challenges Enterprises face in adopting and achieving value with AI is something I’ve been helping them with for years. As the AI hype-cycle we’re living through drops away from its peak, we’re already hearing lots of stories that AI is failing. On the other side of the coin, we’re increasingly hearing the success stories: how more Enterprises are embedding AI and GenAI into their ways of working, building new services and changing existing processes with AI to bring value. The survey results show that achieving this value is hard, but that there is still a significant desire to do so and it’s recognised that succeeding with AI can be transformational. The reality is that achieving value with AI is far from just putting some new technology live: success with AI is at least as much a hearts and minds piece for an Enterprise as it is a technology and process change. Enterprises want to be Responsible with AI – in industries currently unregulated for AI this requires new ways of working. And the survey results affirm the shortage of key personnel across industry globally. If you are yet to commence your AI journey and want to, my advice is to start, and to start small. Use Proof of Concepts (PoCs) across your business to discover and learn how you can embed and utilise AI responsibly. While you’re doing this, take the learnings from these PoCs and conversations across your business to strategically plan, and tie these plans to your vision for what you want your business to achieve. Make sure you baseline and measure how you are achieving value and embedding change. And data – did some mention data? AI is one of the fundamental reasons why so much investment has been made into Enterprise data over the past decade… is yours ready for AI? It's a journey – make sure you celebrate success along the way! UST #AI #GenAI #ResponsibleAI #AIStrategy

25

AI Ramblings AI Musings Have you ever wondered about the reliability of large language models? Cleanlab, a dynamic AI startup originating from MIT's quantum computing lab, introduces an innovative solution - the Trustworthy Language Model. LLM's are renowned for their creative abilities but have often raised concerns about accuracy. With the introduction of the "Trustworthy Language Model", users can gauge the reliability of generated outputs on a scale from 0 to 1. The tool should empower users and businesses to make informed decisions, distinguishing between trustworthy responses and those that may not be accurate. In essence, it's like having a BS-o-meter for your chatbot! "Addendum: it's worth noting that the example in the story is from GPT3.5 and that GPT4 is much better, although still nowhere near perfect!" #AI #TrustworthyAI #Innovation #artificialintelligence #aimusings #airamblings #cleanlab

1

2 Comments

If you are still thinking that AI is just a hype, you should listen the conversation between Mukesh Ambani and NVIDIA CEO Jensen Huang. Why ? Mukesh Ambani bullishly talks about Generative AI models, GPUs, need of open source models etc. He takes up the challenge to surprise world with AI capabilities in coming years. He is not someone who has been leading IT/Software companies day in day out like Narayan Murthy/ Shiv Nadar/Bill Gates etc. What it means? AI has already arrived and every big firm, irrespective of their traditional business, wants to build business and create impact using this. AI is not just another cool tech anymore. It is the industrial revolution changing every facet of our life. Embrace it or get ready for classic "Kodak Case study" in your career and business. Watch here: https://lnkd.in/ejvnCMHZ

10

Education is an industry in serious need of disruption if it is going to stay relevant in the future. Balaji Srinivasan, a crypto entrepreneur, started a new program called the Network School. It is based on his book The Network State, where people can come together to basically create their own countries, but connected through a network instead of only physical land. Highly recommend you read the book. The premise is that there is a lot of talent worldwide that can't afford traditional university education. He uses the example of the current US educational system, where getting a four degree at a top university is now $100K or more a year. If you're from a third world country, the price of a four year education could be more than what people make in a lifetime. There is also the fact that the current education system was originally made in era where farming was predominant, so most of the summer months are down months. You can't spend 3 months on break and expect to stay up to date anymore. The Network School costs $1,000 a month to live with roommates or $2,000 if you want to live solo. And that's it. The program emphasizes continuous education, offering mini-classrooms that you can be in for the day to solve a specific problem. If you solve it, you basically create a cryptocredential, which is verifiable proof. These cryptocredentials are like the 21st century version of a resume. The Network School isn't just focused on education. The second pillar of the school is focused on exercise and nutrition. Every student gets a daily workout with a trainer, and there is a special nutrition program that is used based on Bryan Johnson's Blueprint program. The third pillar called Earn, which is a way for students to earn money for solving problems. A $1,000 bounty is given out each day to whoever solves the problem. The last pillar is fun, and it is unstructured. The school's physical location is based an hour from Singapore, making it accessible to many activities and even as a hub to visit other countries. This school solves the program of educational cost as well as encouraging individuals that have talent throughout the world. It reminds me a modern version of the Ancient Greek Lyceum, where education wasn't just teaching how to think but also incorporated exercise. I'll be interested to see how this school does and if successful will inspire more of them around the world. https://lnkd.in/gjeQFf3G

7

Ola Krutrim , India’s first AI unicorn, recently launched its LLM, Krutrim-7B-chat. This new model is trained in 10 Indian languages and is available on Databricks Marketplace. It can generate text in Hindi, English, Tamil, Telugu, Marathi, Malayalam, Gujarati, Bengali, and Kannada. This is the first of several India-specific models trained by Ola Krutrim. It has been trained on a massive dataset of Indic languages— “Krutrim delivers text infused with Indian sensibilities and cultural awareness.” Check out the model here. The announcement came to light at the Data Intelligence Day in Mumbai on May 10, 2024, by Gautam Bhargava, VP & Head of AI Engineering at Krutrim. “We’re excited to be at the forefront of building foundational LLMs trained on Indian languages to better serve our customers in India. We have been working very closely with the Databricks team to pre-train and finetune our foundational LLM,” said Bhargava. He added, “We are confident that Krutrim, our base model, will be able to provide high-quality responses in Indian languages and dialects.” According to Bhargava, Krutrim was trained using Databricks Mosaic AI Training’s GPU orchestration and system optimisations, enabling state-of-the-art generative AI solutions for customers globally. The model is now available on the Databricks Marketplace.

49

1 Comment

AI Reading for Tuesday June 4! https://lnkd.in/eGcaVjUV In: Partnering with Apple. Out: Building God. - Intelligencer OpenAI Insiders Warn of a ‘Reckless’ Race for Dominance. - NY Times Musk demanded Nvidia redirect GPUs from Tesla to xAI. Conflict of interest much? Shareholders take note when voting proxies, class-action lawyers take note. - CNBC Intel talks at Computex. - Yahoo Finance A deeper dive into Lunar Lake. Integrating RAM on the chiplet stack should give I/O edge vs. AMD which still uses traditional DDR memory sticks. - The Verge Lunar Lake looks competitive with what AMD is shipping now. but not a leapfrog and not available until later this year. Gaudi 3 GPUs look very price-competitive with Nvidia but have little traction in marketplace. Data center CPUs I don't think are competitive with 384-thread AMD Epycs on density. Raspberry Pi adds AI with $70 AI chip kit that can run (I think) 2GB models for e.g. image classification, on about 8W of power at 13 TOPS. - PCMAG Moving from RAG to systems of multiple RAG agents. - VentureBeat Adobe removes art created 'in the style of Ansel Adams' after estate complains. - Artnet Artists abandon Instagram over AI training on their images. - Creative Bloq AI weapons don't feel stability-enhancing, kill chains and OODA loops speed up, use'em or lose'em, “A single pixel is enough to confuse a bomber with a dog, a civilian with a combatant,” and random radicals get access to robot drone weapons. - Defense News Nobody wants to get left behind on AI weapons, and Silicon Valley wants to cash in. - The New Republic Amazon is now using gen-AI and computer vision to scan shipments for problems and reduce returns. - Fast Company Millions of horrified ChatGPT users forced to use their brains as ChatGPT takes part of the morning off. - The Register Today in dumb deepfakes, Taylor Swift TikToks and a Trump barbecue - NBC News

1

Introducing Llama 3.1: Our most capable models to date: "Llama 3.1 405B is the first openly available model that rivals the top AI models when it comes to state-of-the-art capabilities in general knowledge, steerability, math, tool use, and multilingual translation." => https://lnkd.in/gwUtn93C 👍 -------------------------———- #ai #llm #meta #ppensource

1

Navigating Your Tech Career in the Age of AI 🤖 Generative AI is reshaping the tech landscape, prompting tech professionals to adapt and evolve. Despite concerns, AI hasn't led to widespread job losses but requires a strategic shift. Here are essential strategies for success: [1] Understand AI's Limits: AI can’t replicate human creativity and contextual understanding. Focus on tasks AI can’t perform, like innovating new features and solutions. [2] Think Strategically: Understand the “why” behind features by analyzing user needs and market trends. Collaborate with product owners to learn their decision-making processes. [3] Solve Industry Challenges: Proactively identify and brainstorm solutions to longstanding industry problems. [4] Push New Ideas: Advocate for emerging technologies and their potential impacts, showcasing their value through presentations. Shifting to this new focus involves gaining team and managerial support. AI’s rise is inevitable, but by integrating it thoughtfully, you can leverage it to advance your career. 💡 #TechCareer #AI #Innovation #StrategicThinking

9

"The pseudo-acquisition emerges as an off-ramp for AI companies, as some companies struggle to find a viable business model as staying at the frontier proves costly." A very interesting read... TaPod - We Talk Talent Craig Watson #ai #talentacquisition #recruitment

4