Marius Kempf’s Post

View profile for Marius Kempf, graphic

Senior Engineering Manager & Program Lead - Mobile Apps & Service Channels

This presentation by Pip Klöckner is a great overview of current AI trends worth watching (slides/presentation in German!). It is very dense (153 slides in 40min) but I love this format – a very quick look into many peepholes and an invite to go deeper into a lot of AI related rabbit holes. Density over storytelling! 💰 Accenture makes more profit with AI than all AI startups combined. It has enabled more than 600k employees as companies want to “do something with AI” but need support. 📞 Klarna claims that AI allowed them to save 700 jobs in customer support. Zoom claims 90% of user requests are handled via AI and 400k h of human call center work was saved due to that. 🍞 LLMs will be commoditized. Making money with merely providing a good LLM will not work just as with other IT infrastructure in the past. Open source teams are capable to create good, competetive models, e.g. Mistral or Meta’s Llama 3 📈 We will see an inflation of AI generated content – to which AI generated content will react. Will Social networks be places where humans interact or machines? 🍼 AI models thirsty for more data. Yet, it is not too clear where that data will come from. The public internet is too limited and has been used to train models already. OpenAI transcribed YouTube videos into text to have more data for their model to learn from. Google pays reddit to exclusively use their data. 👨👩👧👧 The “Ai incest problem”: If AI will use more and more AI generated content in the web to learn models might collapse and deliver bad results. 💰 Whoever still has data that has not yet been used to teach models sits on a gold mine. For example, Google digitized 25m books over the last 20 years and can use these to train their AI. For comparison, GPT 3.5 has been fed only 5k books. 📦 Every country or culture will need its own, sovereign AI model due to difference in language, culture and rules. 🛠 Models will get smaller in a sense of less energy intensive to use. Furthermore, there will be compound models directing your request to the best and/or most cost/energy efficient AI to address it. 💀 OpenAI might kill (Google) search. And most AI tool related turnover will go to US companies – who will pay little tax in the EU. 🤓 AI tools will observe the way people work on their laptop, learn from it and might substitute their work in the future. ⚡ The energy and water hunger of AI is increasing fast. Crypto and AI will use 10% of global energy consumption in the next few years. The energy consumption of data centers of Google, Microstoft and Meta combined is already higher than that of BASF or Switzerland. Lack of electricity as key inhibitor for AI progress. 🏨 MedGemini reached 91% accuracy for answering medical question – more than experienced human doctors (87%). And the AI has been perceived as more empathetic. The potential for mental health is obvious! Slides:

To view or add a comment, sign in

Explore topics