AI-podden

Salman Avestimehr, USC professor and FedML co-founder, discusses AI, federated learning, and large language models. This week's topics include; AI definition, distributed and federated systems research, current challenges in generative AI, and trustworthiness in ML. Ather and Salman explore the future of accessible LLMs on personal devices, stress the importance of data, and discuss challenges like distinguishing machine-generated outcomes. The conversation touches on enforcing regulations in the era of open-source tools and recommends the book "Life 3.0"

Don't forget to subscribe and follow us on Linkedin.

In this episode of the AI Podcast, host Ather Gattami interviews Michal Sustr, co-founder of EquiLibre, a company that uses game theory and reinforcement learning for trading.

Michal discusses the company's approach to trading, which is based on the idea that markets can be modelled as games. He explains that EquiLibre uses game theory to develop algorithms which can predict the behaviour of other traders and make profitable trades.

Ather and Michal also discuss the future of AI, and Sustr predicts that large language models (LLMs) will continue to be a major area of research in the coming years. Michal believes that LLMs will eventually be able to reason and solve problems in a way that is indistinguishable from human intelligence.

Don't forget to subscribe and follow us on Linkedin.

Reinforcement learning is a type of machine learning that allows agents to learn by interacting with their environment. DeepMind researcher, David Abel is interested in using reinforcement learning to understand and build intelligent agents. One of the big questions in AI is what exactly "intelligence" is. Another big question is how to build intelligent agents that can reason and solve problems. Abel believes that it is important to achieve conceptual clarity in AI.

Abel's research focuses on AI's core scientific questions, emphasizing the need for conceptual clarity about intelligent behavior and agents. David express skepticism about creating AI with certain reward functions, suggesting that multi-criteria objectives could better align AI with human interests.

Don't forget to subscribe and follow us on Linkedin.

In this episode, Mimi and Ather delve into the impact of AI on elections and political communication as well as deepfakes and cybercrime.

They address OpenAI's suspension of a developer for creating a politician-impersonating chatbot and an AI-generated robocall falsely imitating President Biden, demonstrating AI's potential misuse in politics. They also discuss McAfee's Project Mockingbird, an AI technology introduced at CES 2024 to combat deepfake audio and cybercrime.

Additionally, the episode covers The New York Times' lawsuit against OpenAI, highlighting the intricate legal and ethical issues surrounding AI-generated content in the realms of copyright and media. Don't forget to subscribe and follow us on Linkedin.

Edward Hu, Researcher at OpenAI, known for his invention of LoRA, a widely used machine learning method to fine-tune Large Language Models (LLMs) which he co-invented at Microsoft, shares his views on LoRa. Together with Ather he also discusses the development of GFlowNets as a promising tool for the development of reasoning and planning AI systems.

Don't forget to subscribe and follow us on Linkedin.

Mimi Billing, European editor at Shifted and Ather highlight the latest news in AI in December and take a look back at 2023.

In this episode, you will hear the latest on the release of Google DeepMind's Gemini in its Ultra, Pro, and Nano versions. How European Union policymakers have reached a provisional agreement on AI rules under the EU AI ACT. Additionally, the episode also cover developments in robotics, provides a review of AI in 2023, and offers predictions for what we might see in the field of AI in 2024.

Don't forget to subscribe and follow us on Linkedin.

Andy Karvonen, Prof. of Urban Design and Planning at Lund University talks of the "Urban Brain", where we have some kind of smart city control room managing all aspects, from traffic to transportation and the management of facilities. Andy and Ather discuss the pros and cons of such a future, delving into the hard but important questions of privacy, surveillance, and so on. Andy and Ather also discuss the future of an AI-driven society, policymaking, and how to deal with the challenge of big tech dominating in the AI tech field.

Don't forget to subscribe and follow us on Linkedin

Mika Gustafsson, Prof., and David Martinez, PhD student, both at Linköping University conduct research in Bioinformatics and precision medicine. They are exploring how to use complex patient data, both clinical and environmental (such as the lifestyle) to develop personalized treatments and to help understand diseases more in depth in order to take preventive measures. David and Mika explain why bioinformatics and precision medicine is so complex, given the number of parameters affecting the human body and their complex interplay. The challenge of data acquisition is another challenge. For that challenge, they are working on federated learning to train AI systems without requiring private data to be shared. Ather, Mika, and David discuss the trade-offs of using federated learning, and the performance degradation that one might get for not sharing the data to a central unit directly. David and Mika go on and tell us about the greatest impacts of AI on medicine in the near future, from personalized medicine to helping less wealthy countries.

Don't forget to subscribe and follow us on Linkedin.

Let's start with the company behind chatGPT, OpenAI. No one has missed the last couple of weeks' happenings at OpenAI, where the CEO and cofounder Sam Altman was fired on Friday the 17th and then on Monday evening reinstated as CEO. There have been a lot of rumours of why he was fired in the first place but I think we need to focus on something different. Usually, when you kick out your CEO and cofounder, your investors get a heads-up at the very least. In the case of Open AI, the investors include Microsoft, Khosla Ventures, Andreessen Horowitz, Founders Fund and Sequoia — these are big firms. All of them were kept in the dark. The reason for this is that none of these investors sits on the OpenAI board of directors since the company has a different structure — it is run like a non-profit company. I believe this was set up as a part of safety measures since OpenAI is working on AGI (artificial general intelligence) and if the CEO diverged from the safest path, the board could fire him. So after that TDLR, is this a good way to govern an AI company? Amazon's new 2 trillion parameters LLM Olympus (double what GPT4 has) puts it in competition with OpenAI, Meta, Anthropic, Google, and others. Earlier this month, I read in Reuters that Amazon is investing millions in training an ambitious large language model (LLMs), hoping it could rival OpenAI, Google and Meta. The model, codenamed "Olympus", has 2 trillion parameters, sources said, which could make it one of the largest models being trained. OpenAI's GPT-4 model is reported to have one trillion parameters. So, it seems the more parameters the better, however, then I read about this Japanese LLM by NEC, which has reduced the size to "only" 13 billion parameters. This LLM is, which is said to achieve high performance while reducing the number of parameters through unique innovations. This not only reduces power consumption but also enables operation in cloud and on-premises environments due to its lightweight and high-speed. There is this understanding that the better the LLM is at language, the more persuasive it can be and also more innovative. Is this the reason why there is so much work being done on having LLMs taught on specific languages? Samsung AI race over Apple – how will the AI development be visible in our smartphones? https://www.theverge.com/2023/11/8/23953198/samsung-galaxy-ai-live-translate-call Some say that AI-powered features seem like they're becoming the next battleground for smartphone makers. And Samsung has come out this month with a feature that use artificial intelligence to translate phone calls in real-time, it is calling it "AI Live Translate Call," and will be built into the company's native phone app. Samsung says "audio and text translations will appear in real-time as you speak". But Samsung is not alone, Google, for example, has a suite of AI-powered tools to help you edit and improve photos with its Pixel 8 lineup. Apple is reportedly spending a lot of money every day to train AI, and I have to imagine all that investment will show up in some AI-powered features for iPhones. So, what will this mean for our smartphones?

Don't forget to subscribe and follow us on Linkedin.

Agnes Lindell, Head of Data Driven Business at Elvenite shares her experience in building data driven tools in the food industry, all from some of their most successful use cases to how they managed to do it. She mentioned the biggest challenges and how they managed to tackle them. A few examples include inventory optimization with over 10% of savings, to using AI to efficiently grow various types of seeds in different environments. Agnes Lindell walks us through simple steps that she follows in order to become data driven and check the AI readiness, that is all from the level of digitalization to the availability of data, checking data quality, starting small to build a proof of concept, etc. She does also share her favorite AI tools (Hint: it's not a Chatbot!).

Don't forget to subscribe and follow us on Linkedin.