Vitaly Kleban

It is amazing to see the advancements in using AI in the medical field. Tools such as this are already impacting most industries. How are you integrating AI tools into your life? #ai #genai #aitools

1

1 Comment

NLP is the key to Making AI work for everyone LLMs might not reach superintelligence, but AI doesn’t need to. By pairing NLP with practical tools, including LLM training to remove "hallucinations", businesses can unlock AI’s potential without getting lost in the hype! This post shows front-end AI can be trusted to live harmoniously in business operations. #AI #AIAssistants #DigitalTransformation #Innovation #NaturalLanguageProcessing

3

2 Comments

NLP is the key to Making AI work for everyone. LLMs might not reach superintelligence, but AI doesn’t need to. By pairing NLP with practical tools, including LLM training to remove "hallucinations", businesses can unlock AI’s potential without getting lost in the hype! This post shows front-end AI can be trusted to live harmoniously in business operations. #AI #AIAssistants #DigitalTransformation #Innovation #NaturalLanguageProcessing

4

Researchers from Beihang and Microsoft introduce a novel method named MoRA, which employs a square matrix for high-rank updating while maintaining the same number of trainable parameters as LoRA. This approach leverages non-parameter operators to adjust the input and output dimensions of the square matrix, ensuring seamless integration with existing LLMs. The comprehensive evaluation of MoRA across various tasks, including instruction tuning, mathematical reasoning, and memory-intensive tasks, demonstrates its superiority in scenarios requiring significant memorization and comparable performance in other areas.   The study highlights the efficiency of PEFT methods like LoRA, which significantly reduce memory requirements and facilitate the storage and deployment of fine-tuned models. Despite LoRA's broad applicability, it struggles with tasks requiring the enhancement of knowledge and capabilities through fine-tuning. To validate these observations, a memorization task using pseudo-data revealed that LoRA performed significantly worse than full fine-tuning (FFT), even with a large rank.   Inspired by these findings, MoRA was developed to maximize the rank in the weight update matrix while maintaining the same number of trainable parameters. The method employs a square matrix and corresponding non-parameter operators to achieve high-rank updating, enhancing the model's capacity to store new information. In pretraining experiments, MoRA demonstrated superior performance compared to LoRA and ReLoRA, a variant that merges low-rank matrices into the model during training to increase the rank of the weight update matrix. The analysis of the spectrum of singular values for the learned weight update matrix further supports the effectiveness of MoRA's high-rank updating approach.   MoRA represents a significant advancement in PEFT methods for LLMs, addressing the limitations of low-rank updating and enhancing the model's ability to learn and memorize new knowledge. The method's ability to achieve high-rank updating while maintaining the same number of trainable parameters as LoRA allowing for LLM performance improvement across a large group of tasks    Arxiv: https://lnkd.in/enEWr5p2

1

Open-source AI models such as Llama2, Llama 3, and Mistral are making strides toward ensuring AI technology isn't monopolized by a few companies. However, the primary barrier remains the hardware requirements and associated costs needed to run these models. Although it's not yet fully democratized—thanks to a continuing hardware monopoly by companies like NVIDIA, reminiscent of the old adage, "sell shovels during the gold rush"—AI is becoming increasingly accessible to most developers through AI platforms like https://huggingface.co/, https://openai.com/, https://www.anthropic.com/, https://groq.com/, ...

51

2 Comments

LLM Unlocking Infinite Context: How HUMAN-LIKE EPISODIC MEMORY Brings Human-Like Memory to AI ... As AI continues to evolve, researchers are finding innovative ways to enhance language models' capabilities. A fascinating new paper from Huawei and UCL introduces EM-LLM, a novel approach that integrates human-like episodic memory into large language models. This advancement allows AI to process virtually unlimited context lengths while maintaining efficiency - a significant leap forward in natural language processing. 👉 The Power of Episodic Memory in AI Imagine if an AI could remember and process information like a human brain. That's exactly what EM-LLM aims to achieve. Here's how it works: 1. The model mimics how our brains segment experiences into discrete events. 2. It uses the concept of "surprise" to determine event boundaries, much like human cognition. 3. This approach enables efficient processing of extremely long contexts, far beyond current limitations. Think of it as giving AI the ability to create and recall "memories" of what it has processed, allowing for more coherent and contextually relevant outputs. 👉 Performance Gains The results speak for themselves: - 4.3% overall improvement on the LongBench dataset compared to state-of-the-art models - A remarkable 33% improvement on the PassageRetrieval task These numbers might seem modest, but in the world of AI, they represent significant progress. The PassageRetrieval improvement, in particular, demonstrates EM-LLM's enhanced ability to recall and utilize information from vast amounts of text - a crucial skill for many real-world applications. 👉 Bridging AI and Human Cognition What makes EM-LLM truly exciting is its potential to advance our understanding of human memory: - The model shows strong correlations between AI-perceived and human-perceived event boundaries. - This alignment opens new avenues for interdisciplinary research in AI and cognitive science. - We may gain new insights into how human memory works by studying these AI models. 👉 Future Implications and Applications The potential applications of this technology are vast: - More coherent long-form content generation and analysis - Improved information retrieval and question-answering over large documents - Enhanced personalized AI assistants that can maintain context over extended interactions Industries dealing with large volumes of text - from legal and medical fields to content creation and customer service - could see significant transformations as this technology develops. 👉 Looking Ahead EM-LLM represents a step forward in our quest to create AI systems that can process and understand information more like humans do. What are your thoughts on this research? How do you see technologies like EM-LLM impacting your industry or work? Let's discuss in the comments below.

86

5 Comments

NLP is the Key to Making AI Work for Everyone LLMs might not reach superintelligence, but AI doesn’t need to. By pairing NLP with practical tools, businesses can unlock AI’s potential without getting lost in the hype. This post shows front-end AI can live harmoniously in business operations. #AI #AIAssistants #DigitalTransformation #Innovation #NaturalLanguageProcessing 

5

Large language models have proven a critical point: training AI at ‘internet scale’ is not only possible, but also effective and lately getting affordable as well, completing the holy trinity. The ability to leverage vast datasets unlocks patterns and insights invisible at smaller scales—a fundamental shift in AI's potential and applicability. This breakthrough isn't an endpoint, but a launchpad. As we refine and build upon this capability, AI will revolutionize industries in ways we're only beginning to grasp. The long-term impact of AI, built on the foundation of internet-scale learning (we’ve never had anything like this before), will likely surpass even the most optimistic current predictions.

17

Compute Time - A new cartoon that illustrates how access to computer processing is becoming one of the most important currencies in research. NVIDIA recently launched Blackwell which with 20 Petaflops is 10x faster than its workhorse chip the H100 and places processing power traditionally only available to SuperComputers in the hands of consumers through applications like ChatGPT. Outside of these marketing headlines, there is a parallel battle going on to get access to these compute capabilities in the long-tail of academia and research. For example, universities that were instrumental in sustaining decades of R&D into AI when it was deeply unpopular are now edged out from access. This is undermining academia’s greatest strength, which lies in its ability to pursue long-term research projects and fundamental studies that push the boundaries of knowledge. Meanwhile, even within the tech companies leading the new wave of AI, there is competition for processing power. For example, OpenAI’s team that was responsible for analyzing the existential dangers of AI was promised 20% of the company's processing power, but the team have now all resigned or been re-assigned. As Jan Leike, the former DeepMind researcher who was this team’s colead, remarked, “we were struggling for compute and it was getting harder and harder to get this crucial research done.” #siliconvalley #nvidiablackwell #supercomputers

17

3 Comments

Many people ask me "How will Parallel Studios' AI Wayfinder (https://www.wayfinder.ai/) agent and Bittensor work together?" ... Here's how: Given the capabilities of both Wayfinder and Bittensor, there are several intriguing ways these two technologies could potentially collaborate or integrate. Here are five possibilities: 1. AI-Driven Subnet Management: Wayfinder's AI agents (Shells) could be used to autonomously manage and optimize Bittensor subnets. These agents could analyze subnet performance, adjust staking strategies, and even propose new subnet configurations based on real-time network data and the upcoming dTAO (Dynamic TAO) upgrade. 2. Cross-Chain Intelligence Marketplace: Leveraging Wayfinder's omni-chain functionality and Bittensor's decentralized intelligence market, a cross-chain AI services platform could be created. This would allow AI models on Bittensor to be accessed and utilized across multiple blockchains through Wayfinder's interface, expanding the reach and applicability of Bittensor's AI capabilities. 3. Enhanced Oracle Services: Bittensor's focus on creating decentralized AI oracles could be augmented by Wayfinder's AI agents. These agents could act as intelligent intermediaries, gathering, verifying, and delivering data from various sources to Bittensor's network, improving the accuracy and reliability of AI-driven oracle services. 4. Automated TAO Token Management: With the introduction of dTAO, Wayfinder's AI agents could be programmed to autonomously manage TAO token holdings. They could optimize staking, participate in governance decisions, and even engage in yield farming strategies across different subnets based on predefined parameters and market conditions. 5. AI-Powered Subnet Creation and Optimization: Wayfinder's AI could be used to analyze the Bittensor network and automatically propose and create new subnets that fill gaps in the ecosystem. These AI agents could continuously monitor subnet performance, suggest improvements, and even help implement changes through the RAO net upgrade, ensuring the Bittensor network remains efficient and innovative. These integrations could potentially create a powerful synergy between Wayfinder's AI-driven blockchain navigation capabilities and Bittensor's decentralized machine learning network, fostering innovation in both AI and blockchain technologies.

3

We've all experienced how a well-organized outline can make complex information easier to grasp. Interestingly, the same principle applies to large language models (LLMs) in artificial intelligence. A recent paper titled "Enhancing LLM’s Cognition via Structurization" goes into how organizing input data can enhance an AI model's understanding and performance. What's the Big Idea? The researchers drew inspiration from neurocognitive science - the study of how our brains process information. They proposed that by converting plain, unordered text into clear, hierarchical structures, AI models can process information more effectively. Key Findings: 1 - Improved Performance Across Tasks: Organizing information helped AI models perform better in various natural language processing tasks, including: Context-Based Question Answering: Providing more accurate answers by better understanding the context. Hallucination Evaluation: Reducing instances where the AI generates incorrect or nonsensical information. Passage-Level Retrieval: More efficiently finding relevant information within large texts. 2 - Efficiency with Smaller Models: An open-source model called LLaMA2-70B matched the performance of larger models like GPT-3.5-Turbo when using structured inputs. 3 - Distilling Capabilities into Smaller Models: The team developed StruXGPT-7B, a smaller AI model that retains 97% of the structuring abilities of its larger counterparts. This makes advanced AI more accessible and cost-effective. It highlights the importance of how we feed data into AI - not just how much data we provide. What are your thoughts on using structured information to enhance AI performance? #innovation #technology #future #management #startups

14

2 Comments

Did a quick GAT analysis of the AI startup ecosystem from 2013 and identified the top 10 most influential investors, including Sequoia Capital, Insight Partners, and NVIDIA, who shape the landscape through their investments and partnerships. Findings provide valuable insights for AI startup founders seeking to raise funds at different stages, with recommendations on which investors to focus on for Series A, Series B, and later rounds. Just a small experiment, have fun. https://lnkd.in/gMJkHqvq

68

5 Comments

🚀 See how Handshake cut their LLM GPU costs by 50% with Anyscale. Discover how they: 💰 Reduced LLM GPU costs by 50% or more. 📈 Seamlessly scaled large language models (LLMs) without compromising performance. ⏱ Enhanced operational efficiency, enabling faster development cycles. Check out the full story here: https://lnkd.in/gSiKJDaY

37

2 Comments

Agree with Rodney here that Gen AI is not a cure-all and might be overhyped atm. But at the same time - this is indeed the first time we have an opportunity to build atleast a semblance of ‘agency’ in our tools instead of nesting if-else’s. This reasoning capability, though still emerging and quite rough around the edges - can birth one of the biggest product category revolution of our times, if harnessed properly. It is a far more productive view to not look at the current large models as compressed (and natural language query-able) dbs but as an emerging dynamic and conversant reasoning engine. https://lnkd.in/gC2b4C8w

9

With increasing interest in #AIagents and systems, a myriad of definitions have emerged over time, especially after the introduction of large language models (#LLMs) and vision language models (#VLMs). As we advance this science, it is useful to have one encompassing definition that is extensible as well as expressive, building on commonly understood concepts from the past. Emergence's definition of an agent refers back to a few basic principles that embody an agent, using them to develop the notion of an “agent object” in the world of LLMs and LVMs. For an agent object to be programmed scalably to build robust systems, we believe it must embody the following primary concepts: ✔️ Autonomy: An agent is an autonomous, interactive, goal-driven entity with its own state, behavior, and decision-making capabilities. It has the capability of self-improvement when it sees that it is unable to meet the performance parameters for reaching its goal. ✔️ Reactivity and Proactivity: Agents can be reactive (understand their environment and respond to changes using actions) and be proactive (take initiative based on their goals). Agents can take actions that change the state of their environment, which can ensure their progress towards their goal. ✔️ Beliefs, Desires, and Intentions (BDI): A common model used in agent-oriented programming (AOP) is the BDI model, where agents are characterized by their beliefs (information about the world), desires (goals or objectives), and intentions (plans of action). ✔️ Social Ability & Communication: Agents have a communication mechanism and can interact with other agents or entities in their environment. This interaction can be highly complex, as it may involve negotiation, coordination, and cooperation. We believe that with the advent of LLMs, this communication can be completely based on natural language, which is both human- and machine-understandable. ✔️ Constitution: An agent needs to adhere to some regulations and policies depending on the imperatives of its task and goals. It needs to protect itself from being compromised or destroyed as well as be trusted not to harm other agents sharing the environment in which it is operating. ✔️ Memory: An agent must retain Long Term Memory (LTM) to ensure the successful completion of a task. When successfully employed by an agent, it can significantly reduce repetitive computing and reproduce human demonstrations. It also has short-term memory, which relies on any information based on the available context length. Learn more about the anatomy of #AI agents here 👉 https://lnkd.in/e4ctp8qC

33

 #Llama 3🦙Meta just released new best in-class open models https://lnkd.in/eWsF9Vwb Those new models are real improvement on the previous Llama 2 versions in terms of performance across various benchmarks: ### Llama 3 8B vs. Llama 2 70B - GPQA (0-shot): Llama 3 8B achieves a score of 34.2, while Llama 2 70B scores 21.0 - HumanEval (0-shot): Llama 3 8B scores 62.2, a substantial increase compared to Llama 2 70B's score of 25.6 - GSM-8K (8-shot, CoT): The Llama 3 8B model outperforms with a score of 79.6, compared to Llama 2 70B's 57.5 - MATH (4-shot, CoT): Llama 3 8B scores 30.0, while Llama 2 70B scores 11.6 ### Llama 3 70B vs. Llama 2 70B - GPQA (0-shot): Llama 3 70B scores 39.5, compared to Llama 2 70B's 21.0 - HumanEval (0-shot): Llama 3 70B significantly outperforms with a score of 81.7, while Llama 2 70B scores 25.6 - GSM-8K (8-shot, CoT): Llama 3 70B achieves a score of 93.0, which is much higher than Llama 2 70B's 57.5 - MATH (4-shot, CoT): Llama 3 70B scores 50.4, more than quadrupling Llama 2 70B's score of 11.6 ### Additional Improvements in Llama 3 - Context Window Size: Llama 3 has increased the context window size from 4k to 8k tokens, which allows for better understanding and generation of longer text passages - Training Data: Llama 3 models have been trained on a dataset of 15 trillion tokens, compared to Llama 2's 2 trillion tokens, which contributes to their improved performance - Tokenizer: A new tokenizer in Llama 3 expands the vocabulary size to 128,256 tokens, up from 32K in Llama 2, allowing for more efficient text encoding and potentially stronger multilingual capabilities - Grouped-Query Attention (GQA): The 8B version of Llama 3 uses GQA, which is an efficient representation that should help with longer contexts - Training Techniques: Llama 3 Instruct has been optimized for dialogue applications and was trained on over 10 million human-annotated data samples with a combination of advanced training techniques.

6

2 Comments

🎉 Congratulations the Team! 🎉 Our product, YouKnowMeBest Social, has been recognized in a cutting-edge study from TU Vienna on Social-Aware Fashion Recommendation Systems (SAFRS) lead by Julia Neidhardt’s team, Head of the CD Lab for Recommender Systems. This research highlights the power of #usercentric and #socialaware recommendations, and I couldn’t be prouder of our team for leading the way in innovation. 💪 Here’s to more groundbreaking achievements ahead! 🚀 YouKnowMeBest. Real-life shopping experience online!  youknowme.best #ecommerce #onlineshopping #aiforgood #digitalhumanism #recommendersystems #innovation #digitaltransformation #InnovaticGroup #YouKnowMeBest #SocialShopping #chatgpt #gpt4

5