TEDAI Vienna – Part Two

Day two of the TEDAI sessions, held at the beautiful Volkstheater in Vienna continued to reveal a collective of intellectual speakers and participants. The ornate Volkstheater opened in 1889 and is spectacular inside, multileveled with gold-gilded side balconies. During the break, we watched a Boston Robotics dog nimbly manage stairs and chatted with speakers and other attendees. The following article is a continuation of TEDAI Vienna – Part One.

Rama Akkiraju

Rama Akkiraju, VP of Enterprise AI & Automation at NVIDIA, held our attention and walked us through the complexities of Chatbots and why they sometimes don’t get the answers quite right. Being a Chatbot builder, I appreciated this request for empathy and understanding. Creating a Chatbot for enterprises requires understanding company policy, access control and working with a variety of data sources. Ensuring Intellectual Property is secured, along with ensuring the flow of conversation is helpful and not annoying to users is paramount.

NVIDIA is deploying Chatbots across its enterprise to assist its internal help desk operations through to end customers.

Rama introduced us to the acronym FACTS to be applied to Chatbots

Freshness – RAG vector database supporting information presented to users
Architecture – flexibility for the interchange of models
Cost Economics – SLM verses using an LLM
Testing Cycle – use a LLM to test the LLM
Security – ACL sensitivity on data such as categorization of documents to be used

Vít Růžička

Vít Růžička, an AI Researcher at Oxford University, showed us how machine learning was being applied for Climate Change applications, namely in methane detection using satellite data. Methane gas leaks can last for up to 40 years and we have the technology to identify these leaks which are invisible to the human eye.

Vít has developed a Neural Network model and deployed it to the International Space Station (ISS) to detect methane gas leaks across the globe. The device takes hyperspectral images, which are very large (2Gb), so transferring them back to Earth is challenging and the transfer has to occur when the ISS is in alignment with ground receivers. To overcome this, his model automatically detects the images of interest before transferring. The process of detecting the leaks has gone from weeks to minutes.

As natural gas is primarily methane, early detection is beneficial not only to the environment but also to oil and gas companies.

Hilke Schellmann

Hilke Schellmann has spent the last six years exploring the viability of AI automation of Human Resources systems for the hiring process. Using creative measures to test the systems, she found that every one of them displayed bias and very strange behaviours. She used a German Wikipedia page to apply for a job and received a 73% match. Using the content from a Chinese page, an improved 80%, and note that both transcripts were gibberish and irrelevant to the job requirements. This did bring some chuckles across the theatre.

Using Artificial Intelligence systems to scan, filter and select candidates is excluding the people companies should be hiring. Diversity in hiring is increasingly recognized as a critical asset for organizations, offering numerous advantages that extend beyond ethical considerations. Unfortunately, Human Resources departments are often hiring for positions they have no skills in and do not understand, and think that applying automated systems will save them time. In reality, AI systems cost businesses in the long run by excluding the right candidates. A diverse workforce also ensures your products are designed and curated for the wider population.

Raia Hadsell

‍Raia Hadsell, VP of Research at Google DeepMind shared that there are currently two million scientists across 190 countries using Alphafold. AlphaFold is an advanced artificial intelligence system developed by Google DeepMind that predicts the three-dimensional structures of proteins from their amino acid sequences. This has been a long-standing challenge for biologists. Understanding protein folds has revolutionised medical research.

Applications include rapid disease screening – such as Alzheimer’s, Parkinson’s, and cystic fibrosis, accurate protein structure analysis for drug design and immune response simulations.

DeepMind has also made significant advances in developing models to address weather and climate prediction challenges. Utilising machine learning techniques, trained on a dataset spanning 40 years of weather information and over a million multimodal data points, they created GenCast.

GenCast can generate a global 15-day forecast in approximately 8 minutes when run on a Cloud TPUv5 device.

Raia described the possibility of modelling Earth to combine forest, ocean, agricultural, and acoustic data into a single model. They could detect bat calls when experimenting with bird songs even though the model was not trained on sounds from bats. I have found that models trained on specific supplier forms will also correctly extract some of the data points from other company’s proprietary document layouts.

For DeepMind’s Earth model, the real challenge ahead is the collection and organisation of data, particularly when this data is unstructured.

Ramin Hasani

Ramin Hasani, Co-Founder and CEO at Liquid AI, shared their company’s Liquid Foundation Model (LFM), which is a general all-purpose low energy consuming model. Direct from their chatbot: “Liquid AI focuses on developing universally capable general intelligence, creating efficient AI systems that can handle large amounts of multimodal data and make reliable decisions at every scale.”

The model is multimodal and applicable across a variety of applications. LFM can run on edge devices – including handheld devices, Raspberry Pi, autonomous vehicles and CPUs, thus addressing privacy concerns and poor cloud connectivity.

Liquid AI aims to introduce new and improved capabilities across various industries, such as financial services such as anomaly detection, human vital sign monitoring, as well as consumer electronics. They currently have a 40B and 3B parameter model, available to try at playground.liquid.ai.

Riccardo Loconte

Can you spot a liar?

Apparently, we are not too good at this, and we all lie daily – Riccardo Loconte, a Psychologist and a PhD Student in Cognitive, Computational and Social Neuroscience at IMT School of Advanced Studies Lucca (Italy), told us that on average we lie twice a day.

How good are people at detecting lies? Well, our accuracy is a mere 50% in detecting if a stranger is lying to us – so we have a fifty-fifty chance. Humans who are “Expert in lie detection” are only marginally better with 55%. Riccardo’s talk outlined how using AI to recognize deception could improve this detection.

The benefits of deploying this at scale would allow the detection of fake news, lies from Politicians (this did bring a laugh from the audience), interviews and increased border control safety to name just a few applications.

But what are the consequences if we place faith in Artificial Intelligence to do this for us? If we rely on AI to detect lies, we will lose trust in ourselves and others?

Juergen Schmidhuber

The TEDAI Vienna session day wrapped up with the final speaker – Juergen Schmidhuber, Scientific Director at Swiss AI Lab IDSIA. Noted by the New York Times in their famous line “When A.I. Matures, It May Call Jürgen Schmidhuber ‘Dad’.” And from Bloomberg “Jürgen Schmidhuber says he’ll make machines smarter than us. His peers wish he’d just shut up.”

Juergen has been instrumental in the advancement of Artificial Intelligence, especially Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM), Residual Neural Network (ResNet) and meta-learning. The LSTM is the dominant technique for natural language processing tasks – including speech and translation implemented in Siri and Google translation to name just a few services.

He had the idea at the age of 15 to build an Artificial Intelligence machine smarter than himself and then retire. His talk covered exponential acceleration in the most important events of history, starting with the Big Bang 13.8 billion years ago, taking a quarter of that to 3.5 million years ago – life emerged on this planet. Then take a quarter of that, we get 900 million years – of mobile animal life. Take a quarter again brings us to 220 million years ago – mammals emerged. Take a quarter of that, 55 million years ago for the first primates. Again, take a quarter of that time and we are now at 13 million years ago when the first hominids emerged, our ancestors. Again, take a quarter of that time we are now 3.5 million years ago and into the Stone Age. Juergen continued over further developments, which have all been a quarter of the previous time.

So the future? Cheap AIs will surpass human brain power soon. Exploration of space is a real thing, spreading and self-replicating AI machines through the Solar system and beyond. Humans have just been a small and necessary stepping stone in evolution.

Some of his previous talks describe the next major development occurring between 2040 and 2050. At TEDAI Vienna, he mentions 2042 as the next big date. I would like to think he settled on 2042 to reference the Hitchhiker’s Guide to the Galaxy.

Will humans be replaced by robots? Is that the future path of evolution?

I caught up with Juergen briefly at the after-party but was so exhausted from very full days that, was not in a brain-fit state to converse, so rather cheekily just asked for a selfie.

Juergen Schmidhuber indulgingly posing for a selfie with Melinda Macey

Speakers Included Here

Vít Růžička, AI Researcher – Oxford University

Rama Akkiraju, VP of Enterprise AI & Automation – NVIDIA

Hilke Schellmann, Journalist and AI Expert – New York University

‍Raia Hadsell, VP of Research – Google DeepMind

Ramin Hasani, Co-Founder and CEO – Liquid AI

Riccardo Loconte, Psychologist and a PhD Student in Cognitive, Computational and Social Neuroscience at IMT School of Advanced Studies Lucca (Italy)

Juergen Schmidhuber, Scientific Director – Swiss AI Lab IDSIA

https://tedai-vienna.ted.com