Jan Maria Kowalski

A video discussing 5 levels of AI integration with Martin Jokub is out! https://lnkd.in/dXcqTkdg

9

New AI Project aims to mimic the Human neocortex The Gates-funded "Thousand Brains" project offers a fundamentally different approach to AI.The open-source initiative aims to partner with electronics companies, government agencies, and university researchers to explore potential applications for its new platform. But, for example, as they grow in size and power, they become more energy hungry—to train OpenAI’s GPT-3, a 2022 Nature study suggested the company spent US $4.6 million to run 9,200 GPUs for two weeks. Neural networks also often prove unstable. To overcome these challenges, the Thousand Brains Project aims to develop a new AI platform by reverse engineering the neocortex, which accounts for about 80 percent of the human brain’s mass.

22

2 komentarze

How effective are open-source language models for translation tasks: https://lnkd.in/dk2Y25wY

7

interesting read https://lnkd.in/em-pdteb

4

Unlocking the Future of Digital Selling with AI-Created Virtual Avatars In a thought-provoking article by Iva Filipović from EPAM Continuum, the rising trend of AI-generated virtual influencers is dissected, revealing their growing impact on the world of digital marketing. Filipović argues that virtual influencers, combining the appeal of traditional digital influencers and the novelty of computer-generated characters, represent a transformative shift in digital advertising. Virtual influencers like Lil Miquela and Aitana Lopez, who engage audiences through social media and specialized platforms, are setting new benchmarks for brand endorsements. These AI-powered personas not only add a futuristic edge to campaigns but also ensure a level of consistency and control that human influencers can't match. As virtual influencers make their mark in sectors ranging from fashion to beauty, their capacity to seamlessly integrate into digital ecosystems makes them invaluable. For instance, their success in China's dynamic market underscores their potential to dominate global digital trends. However, the integration of such technologies isn't without challenges. Brands must navigate the ethical implications and strive for transparency to maintain consumer trust. The strategic use of virtual influencers, when paired with responsible AI practices, can enhance customer engagement and foster deep connections, placing brands at the cutting edge of digital innovation. As we move forward, embracing virtual influencers could mean transforming basic marketing campaigns into rich, interactive experiences that captivate and resonate with audiences worldwide. For leaders in digital transformation, this isn't just an opportunity—it's the future knocking. #AI #DigitalMarketing #VirtualInfluencers #InnovationLeadership #marketing https://lnkd.in/gidR6yfp

3

Today's Shower Thought: Everybody just assumes that Generative AI follows an exponential curve when you add parameters, compute and training data. What if the curve is reverse of what we think? That we are only looking at 10-20% performance from 100-200% added parameters, compute and training data?

5

5 komentarzy

🚀 GPT-4 Breaks AI Barrier: Passes Turing Test! Researchers claim GPT-4, the AI behind ChatGPT, has achieved a historic milestone by becoming the first to pass the iconic Turing test. 🎓 📜 The Turing Test, proposed by Alan Turing in 1950, evaluates whether an AI's responses are indistinguishable from a human's during conversation. 🔬 In the study, 500 participants conversed with a human, ELIZA (a 1960s AI), GPT-3.5, and GPT-4 for 5 minutes each. They then judged whether each was human or AI. 📊 Results: GPT-4 was judged human 54% of the time, outperforming ELIZA (22%) but trailing actual humans (67%). GPT-3.5 scored 50%. 🌍 This suggests GPT-4's responses often can't be distinguished from humans - a significant AI milestone. But it also raises concerns about AI deception. 😮 Share your thoughts in the comments below! 👇 https://lnkd.in/eQvFMaA9

6

🤓 AI literacy primer ahead of the AI Act requirement: -- Feel free to correct if technically off! I aim to be correct, just to present from an unusual angle -- • “AI”: broad umbrella term for digital systems of statistics and automation • generative AI / GAI: a class of AI system designed to store and produce replacement grade instances of its training data. • GAI model: a lossy interleaved storage format for expressive media content as weights; can also be used to mathematically manipulate said media type. • prompting: approximate retrieval from an AI model • text-to-X: content retrieval from model by text search • inference: see above. • temperature: degree of pseudo randomness thrown in at inference time to hide source. • memorization: storage system working as intended at low temperatures. • open source: fully reproducible system • open weights: fully reproducible model • open data: fully transparent dataset • openAI: overhyped company engaging in none of the above • licensing: the legal way to obtain training data. • scraping: the convenient way to obtain training data • training data: convenient misnomer for other people’s propertiary, personal and private information. • “openly licensed”: fictional PR term for Creative Commons licensed works, all of which come with some rights reserved; not fit for unfettered commercial exploitation • public domain: legal concept of no property claims • publicly available -- possible to copy online, including from The Pirate Bay.

31

Wondering how to improve exponentially the performance of your LLMs???? Your solution could be CRITIC! 🤖🗣️🔧 CRITIC allows LLMs, which are essentially “black boxes” to validate and progressively amend their own outputs in a manner similar to human interaction with tools. More specifically, starting with an initial output, CRITIC interacts with appropriate tools to evaluate certain aspects of the text, and then revises the output based on the feedback obtained during this validation process. This offers an element of getting external feedback; in many cases this gives much better results than only using LLMs to critique itself. You should take a look to this cheatsheet by Val Andrei Fajardo, ML Engineer at LlamaIndex, that guides you through all the most relevant components of CRITIC. 📝 You can take a look to the paper below 👉 https://lnkd.in/dhskQ8rw #ai #genai

57

6 komentarzy

Bridging the Abstract and Concrete: The Role of Modern Language Models and Knowledge Graphs. Big idea: It is time to combine them at multiple levels. In an era where artificial intelligence (AI) and machine learning are transforming industries, the ability to connect abstract concepts with concrete data has become paramount. This connection is crucial for applications ranging from natural language processing to decision support systems. A recent diagram illustrates this interplay between the abstract and concrete, mediated by semantics, memory, and reinforcement learning. Coupled with insights from Anthropic's research on monosemanticity and the capabilities of modern fine-tuned language models (LLMs) and in-memory graph databases, we can explore how these connections enhance AI interpretability and functionality. A Conceptual Framework The diagram in question showcases a dynamic feedback loop that connects abstract ideas with concrete experiences through semantics. At its core, semantics acts as a bridge, enabling the translation of tangible, specific concepts (concrete) into generalized, intangible ideas (abstract), and vice versa. The process is underpinned by memory, which stores past experiences and knowledge, and reinforcement learning, which allows the system to adapt and improve through experience. Insights from Anthropic's Research Anthropic's work on monosemanticity provides a deeper understanding of how to make AI models more interpretable. By decomposing complex neural networks into more manageable units called features, Anthropic's approach allows for clearer and more consistent interpretation of model behavior. These features, which are patterns of neuron activations, serve as building blocks that connect abstract concepts with concrete data. For example, in a transformer language model, individual neurons might respond to a variety of unrelated contexts. However, when decomposed into features, these responses become more interpretable and context-specific. This decomposition facilitates a better understanding of how models process information, enabling targeted interventions and improvements in model behavior [[❞]](https://lnkd.in/gF3PgYHm) [[❞]](https://lnkd.in/g_J-Z2ZG). The Role of Modern Language Models and Knowledge Graphs Modern fine-tuned LLMs and knowledge graphs play a crucial role in operationalizing the abstract-concrete connection. LLMs, such as GPT-4, are capable of understanding and generating human-like text by leveraging vast amounts of data. When fine-tuned, these models can excel in specific domains, providing highly accurate and contextually relevant outputs. #ai #knowledge #graphs #semantics

2

6 komentarzy

The Future of AI Assistants Balancing Innovation and Ethics In the thought-provoking article by Iason Gabriel and Arianna Manzini, published on April 19, 2024, the authors delve into the complex terrain of advanced AI assistants—technologies that promise to redefine our interaction with machines. As AI becomes more integrated into daily life, handling tasks from scheduling to creative endeavors, the need for a proactive approach to the ethical dimensions of this technology becomes paramount. The paper highlights the immense potential of AI to augment human capabilities across various sectors including education, planning, and creativity. However, it also brings to the forefront the critical need for alignment with human values and interests, particularly as these systems gain autonomy. The potential for misuse, such as spreading misinformation or enabling cyber attacks, underscores the importance of rigorous safety protocols and ethical guidelines. Key takeaways from their research emphasize the dual focus on advancing AI technology while ensuring it supports beneficial outcomes for society. The authors call for more comprehensive evaluations and foresight activities to anticipate and mitigate the risks associated with these technologies. This paper serves as a call to action for policymakers, developers, and the public to engage in informed discussions and collaborative efforts to shape the development of AI assistants. As we stand on the brink of significant technological advancements, the choices we make today will determine the landscape of tomorrow. For leaders and professionals engaged in technology and ethics, this article provides a crucial perspective on the responsibilities and opportunities presented by the rise of AI assistants. It's a must-read for anyone committed to shaping a future where technology amplifies human potential while safeguarding societal values. #AI #EthicsInAI #DigitalTransformation #LeadershipInTech #AIAssistants #CAIO https://lnkd.in/gESdvJpk

3

🚀🏥 We are at the cusp of a transformative era in healthcare! At the #LondonBiotechnology Show 2024, Microsoft's #ThomasBalkizas unveiled how AI is poised to revolutionise the sector. From accelerating clinical trials to devising personalised medicine, AI is transforming from a helpful tool to a game-changer in healthcare. Balkizas emphasised the profound impact of Generative AI, enhancing everything from routine administrative tasks to complex clinical processes. This "tipping point" marks the dawn of a new era where AI becomes an integral part of healthcare innovation.🔗 #AI #HealthcareInnovation #Microsoft #Biotechnology #GenerativeAI #PersonalisedMedicine

10

Overview and basic terminology of LLM evaluations 🧑‍🏫 and benchmarking. Can recommend reading if you are new to these topics We will keep adding to this blog series as we move forward with the DevQualityEval 🏇

5

Why do modern AI systems produce false outputs and what should be done about it? Modern AI systems, as we have been warned, are prone to hallucination. Interesting article especially for those who have not yet understood how AI LLMs work https://lnkd.in/dWmXJj6s

2

3 komentarze

An interesting WSJ article today. The savings should be measured in potential efficiency gains, not just reducing parameters. TRAINING MODELS: Of course 10 billion parameters take less compute cost than a 1 trillion parameter LLM. But there is only 1% of trained knowledge by definition. If we aggregate for an apples-to-apples basis 🍎🍏, there is no savings on a “knowledge-adjusted” basis. INFERENCING: Inferencing off of a smaller model by definition allows for efficiency gains. It also allows for private data to be used internally and create more accuracy/precision, which will drive adoption. There are a wide range of issues of how to access the right model, at the right time, in an efficient manner to allow for the mass market general user who at one moment wants to say write a poem, and the next moment wants to write a thank you letter with a generative graphic. CONCLUSION: Comercial users will benefit from smaller model efficiency gains, but adoption still means massive compute capacity needs. Mass market users will require a complex back-end provisioning to marry “mass” knowledge with “small” models — a null set? What are your thoughts? Happy 4th 🇺🇸 all.

6

1 komentarz

Like with other technologies the US and Europe are not fully aligned when it comes to considering, implementing and adopting generative AI. Not to mention the diverse approaches within Europe. Yet, this is expected for two major regions where cultures, constraints and dependencies are different.

4

3 komentarze