Nick Black’s Post

AI Daily Digest - Today's Strategic Insights (03.10.2024) Here are some of the AI headlines that caught our attention on October 3, 2024. ▶️ MIT spinoff Liquid debuts non-transformer AI models Liquid AI, a startup co-founded by former MIT researchers, introduced the Liquid Foundation Models (LFMs), a new series of AI models that depart from the traditional transformer architecture. The models outperform similar-sized transformers like Meta’s Llama 3.2 and Microsoft’s Phi-3.5, especially in memory efficiency and performance. Designed for tasks spanning multiple data types like text, audio, and video, Liquid’s models have shown exceptional results in fields like financial services, biotech, and consumer electronics. However, they aren’t open-source, available only via specific platforms. ▶️ Google developing reasoning AI to rival OpenAI Google is accelerating its efforts in AI reasoning models to compete with OpenAI’s o1 system. These models use a technique called chain-of-thought prompting, originally pioneered by Google, to solve complex, multi-step problems. Although Google has adopted a more cautious approach to AI product releases, its models like AlphaProof and AlphaGeometry 2 are already showing advanced reasoning abilities. As the race for AI dominance continues, Google’s next moves could significantly influence the landscape. ▶️ France appoints first AI minister In a major push to lead in AI, France has appointed its first AI minister, Gabriel Attal, to spearhead the country’s digital transformation. Attal introduced "Albert," an AI system designed to streamline bureaucratic tasks like handling tax queries and environmental project submissions. With this initiative, France aims to enhance the efficiency of public services while positioning itself as a global AI leader. ▶️ Nvidia launches open-source AI model NVLM 1.0 Nvidia has unveiled NVLM 1.0, a new open-source AI model that is challenging proprietary systems from OpenAI and Google. With 72 billion parameters, the NVLM-D-72B model excels in both vision and language tasks, even outperforming models like GPT-4 in some areas. This move signals a new era of open AI, potentially reshaping the industry by granting developers unprecedented access to advanced AI systems. #AIFuture #AIAdvancement #AIandBusiness #AIForEveryone #AI

Discover the potential of Google's latest AI model, Gemma 2, designed to empower developers and researchers with exceptional performance and versatility. Gemma 2 is available in two sizes – 9B and 27B parameters. Despite being significantly smaller than models like Llama3 \(70B\) and NVIDIA's Nemotron \(340B\), the 27B version has demonstrated impressive capabilities, even surpassing larger models in specific tasks. One standout feature of Gemma 2 is its compatibility with multiple backends such as TensorFlow, JAX, and PyTorch. It also integrates smoothly with Keras 3, enabling efficient development of applications like chatbots. This flexibility allows developers to leverage various tools for diverse text-based tasks. Google has prioritized safety and bias mitigation in developing Gemma 2. The training process involved filtering pre-training data and rigorous testing against comprehensive metrics to identify and mitigate biases. This commitment ensures a more reliable model for users. Gemma 2's open-source nature makes it an invaluable resource for the AI community. Developers can access it on platforms like Google Colab and Kaggle or use it as an NVIDIA NIM inference microservice optimized with TensorRT-LLM. Benchmark results highlight Gemma 2's strengths: scoring top marks in MMLU \(75.2\), BBH \(74.9\), HellaSwag \(86.4\), GSM8K \(74.0\), MATH \(42.3\), and HumanEval pass@1 \(51.8\). These scores indicate its robustness across general knowledge, reasoning, math, and code generation tasks compared to other models like Llama3 and Grok-1. In summary, whether building efficient chatbots or conducting advanced research, Gemma 2 offers a high-performing solution that balances practicality with robust performance across various applications. #AI #MachineLearning #OpenSource

I've been saying it for a year the plateau is bypassed by focusing on algorithms cut your losses (CCE) is one such algo, so it loqt (ah yes, a point i didn't get to, is hardware, all Nvidia is is a massive matrix multiplication piece of hw, that can be improved upon as well) The current AI landscape is dominated by tech giants like OpenAI, Google, and xAI, which are engaged in an "arms race" to develop the most powerful large language models (LLMs) in pursuit of Artificial General Intelligence (AGI). However, this approach has led to a lopsided ecosystem where only a few players can afford to develop and train these expensive models. In 2025, a new approach is expected to emerge, focusing on developing lightweight, ultra-fast, and ultra-cheap models that are "good enough" for commercial applications. This approach will prioritize affordability and usability over pursuing AGI. Key points: 1. _High inference costs_: Current LLMs are expensive to use, making it challenging for developers to create affordable AI-powered apps. 2. _New law for AI inference_: The cost of inference is expected to drop by a factor of 10 per year, driven by advances in AI algorithms, inference technologies, and chip prices. 3. _Lightweight models_: A new generation of AI startups will focus on developing lightweight models that are cheaper to train and use, making AI more accessible to developers and businesses. 4. _Vertical integration_: Companies like (link unavailable) are achieving significant cost savings by vertically integrating model development, inference engine creation, and app development. This shift in approach is expected to lead to a proliferation of AI-powered apps in the next two years, driving innovation and adoption in the AI ecosystem.

H(AI)ghs & L(AI)es #4 A short & biased round up of (mostly) AI developments in the past week. 1. Robot gives me an apple & explains why: Watch the video at (https://lnkd.in/dcnuJ2td). That robot wasn’t coded to do any of the things in the video specifically. It was trained on a lot of video data, and is now able to respond to verbal requests to do physical things in real-time thanks to that training and a partnership with OpenAI. I'd written about Figure earlier, the company behind the bot, when they raised a $675m round at a buzzy $2.6B valuation. 2. Meet your new software engineer: Cognition, a 2 month old AI start-up, 'released' Devin (who doesn't need coffee or meal breaks, works 24/7) last week. You tell Devin something you want to code — a website, a game, debugging your code - and it plans out the required steps, executes them in sequence, and pings you when it needs something. As someone who doesn’t know how to code but would love to, I welcome Devin. Cognition AI is an interesting start-up - it is full of sport-coders - the 10 employees have won 10 gold medals at IOI (Int. Olympiad in Informatics), and CEO, Scott Wu says this background gives them an edge - now you can understand how they went from idea to launch in 2 months. 3. From the Highs to the Lies now: Science journals are in for a rough time - with fake & spurious (LLM hallucinations) content in the submissions - the two clearest examples so far this year are a now retracted paper with a big-balled rat and a new lithium battery paper, with an opening sentence clearly written by an LLM. Plenty of such examples surfacing now - this is a potential crisis for journal publishers, since reputation is all they have got. GenAI-created "stuff" could quickly overwhelm their review processes, leading to reputation losses. Everyone in science loses. From AI to Fake Science (FS). 4. Taylor goes on The Eras Tour; Jensen goes to AI Woodtsock: Over 10000 customers, partners and fans of Nvidia descended on the SAP Center, the home of the NHL’s San Jose Sharks, for Huang’s keynote speech at the annual Nvidia conference earlier this week. Pro wrestling’s WWE Monday Night RAW event took place there in February. Justin Timberlake is scheduled to play at the arena in May. Huang unveiled the Blackwell B200 generation of GPUs - a new chip with 208-billion-transistors. What is also interesting is that there are reports estimating that the new chips cost twice as much as the H100 to manufacture, but Nvidia is increasing prices much less than expected meaning somewhat lower margins going forward. Price competition from AMD? Microsoft, Google, OpenAI building their own chips? Moats, what moats? 5. Chocolate or Chips? I couldn't resist this - cheaper chips & bitter chocolate on the horizon. So far this year Cocoa Futures are up 93.6%, besting AI bellwether Nvidia. Better buy that chocolate now, hold off on the B100's. I'm telling you! Have a sweet weekend, folks.

At Mesozoic Technologies, we believe that the future of AI lies in the responsible use of synthetic data. The unique properties of synthetic datasets—no PII risks, no licensing issues, and the ability to create expansive datasets tailored to specific needs—position us to lead in this space. In a recent analysis by NVIDIA, the effectiveness of synthetic data in reducing bias in AI training datasets was emphasized, showing an improvement of over 30% in fairness metrics across various models. This is a pivotal moment in AI, as we strive for systems that are not only efficient but also equitable. #AI #SyntheticData #BiasReduction Link https://lnkd.in/e57hNfNu

RAG and LLM are two emerging trends that are set to transform the way businesses use large language models (LLMs) in context of Generative AI. Retrieval augmented generation (RAG) is a powerful tool that combines LLMs with retrieval-based techniques to improve data retrieval. By retrieving contextually relevant data from enterprise systems, RAG can significantly improve the accuracy, timeliness and relevance of LLM's responses. One of the most significant benefits of RAG is that it enables companies to efficiently access relevant internal data and use it for various use cases. This means that LLMs can be used for business-critical tasks without any accuracy or relevance problems. However, companies face several challenges in implementing these technologies, such as lack of skilled people, limited budgets, legacy compute systems, and silos of legacy data storage that don’t keep powerful AI-optimized GPUs fully productive. To address these challenges, Pure provides a solution with AIRI (AI Ready Infrastructure) that helps companies and institutions be fast if they invest in compute layers with GPUs from Nvidia. With Pure, businesses can build a larger ecosystem around AI for themselves, necessary to support these emerging technologies, from small to large. The combination of RAG+LLM and Pure AIRI creates a powerful tool that can help companies improve their performance and accuracy, making them more competitive in today’s fast-paced business environment. And do not be afraid if the GPU eat up your Power Budget. We help to balance this out for very large Datasets that reside in one 42RU Rack with the Power Consumption equal to 7 Hairdryers. The Future will be easy by switching to Pure.

Today, AI at Meta released Llama 3.1 405B 🦙, their most powerful open #LLM that can be used for synthetic data generation, improving model accuracy. 👀 ➡️ https://nvda.ws/3zTl5qJ Technical deep dive on how to use synthetic data to improve LLMs and evaluate RAG.

This paper has been good read for order preserving RAG models. Would like to add in my RAG pipeline and see the results for myself 😁 #genai #rag #llm

I think RAG has become more useful using langgraph and its possibilities of cyclic processes like Self Reflection, but still a difficulty in how to extract, chunking, vectorize information in a better way from documents (mainly pdfs with flow diagrams and larger contexts of a page). How to perform knowledge database pre-processing in a better way, in the context of complex PDFs?

𝐃𝐞𝐞𝐩𝐒𝐞𝐞𝐤’𝐬 𝐧𝐞𝐰 𝐀𝐈 𝐦𝐨𝐝𝐞𝐥 𝐚𝐩𝐩𝐞𝐚𝐫𝐬 𝐭𝐨 𝐛𝐞 𝐨𝐧𝐞 𝐨𝐟 𝐭𝐡𝐞 𝐛𝐞𝐬𝐭 ‘𝐨𝐩𝐞𝐧’ 𝐜𝐡𝐚𝐥𝐥𝐞𝐧𝐠𝐞𝐫𝐬 𝐲𝐞𝐭 DeepSeek, a Chinese AI firm, has unveiled DeepSeek V3, a cutting-edge AI model that sets a new benchmark for open AI systems. Released under a permissive license, DeepSeek V3 empowers developers to use, modify, and integrate the model into various applications, including commercial ones. The model’s capabilities are impressive, handling a diverse range of text-based tasks like coding, translation, and content creation with high precision. According to DeepSeek’s internal benchmarks, DeepSeek V3 surpasses notable competitors like Meta’s Llama 3.1 405B and OpenAI’s GPT-4o in tasks such as coding competitions and integration challenges on platforms like Codeforces and Aider Polyglot. DeepSeek V3 boasts 671 billion parameters, making it one of the largest open AI models available. Its training process utilized 14.8 trillion tokens and was completed in just two months on Nvidia H800 GPUs, showcasing exceptional efficiency with a modest $5.5 million training budget. While the model excels in technical tasks, it does have limitations regarding politically sensitive topics due to regulatory requirements in China. Nevertheless, its development is a testament to the rapid progress being made in AI technology. With backing from High-Flyer Capital Management, DeepSeek continues to push the boundaries of what’s possible in the AI domain. High-Flyer’s vision of creating “superintelligent” AI systems aligns with DeepSeek’s ambitious trajectory. Learn more about DeepSeek V3 and its groundbreaking capabilities here: https://lnkd.in/gTRRZgRF Follow me for more updates on the latest in software, AI, and technology! #ai #opensource #deepseek #machinelearning #softwareengineering #technologynews #artificialintelligence #datascience