Applied AI Pod

Episode highlights:

• 01:00 - Conversational AI for the future of marketing and sales, focus on the real estate industry.
• 04:00 - How Structurely works and what it solves.
• 06:50 - Benefits to businesses utilizing AI within their companies.
• 10:55 - The future of real estate by use of machine learning.
• 16:10 - Creating a more promising future for AI as a tool for positive outcomes. E.g. Zillow.
• 23:00 - Conversational AI's next big challenges.

References:

• Nate's LinkedIn profile
• Nate's Twitter profile
• Structurely's Company Website

• 02:00 - Ada's performance, stories and metrics around. Size of the impact AI has in this space, as covered by Tradeshift.
• 05:35 - Working with AI/ML teams.
• 14:40 - Assessing how much data is needed for an AI project.
• 18:45 - Data risks.
• 24:25 - Is Agile good for AI teams?
• 27:30 - How much does UX matter in e-Invoicing and ML/Data projects?
• 36:35 - How can projects get derailed or fail? What should we watch out for.
• 40:05 - Funny fails.
• 41:50 - AI principles.

References:

• Lloyd's Linkedin Profile
• Tradeshift's Ada technology
• Tradeshift's surpass of $1 trillion in transactions processed on their platform.

• 02:35 - Why hasn’t voice AI taken off already?
• 22:50 - Can we fulfil an end to end new purchase naturally?
• 32:20 - How can we resolve the disambiguation problem in NLU?
• 37:20 - Context and memory perspectives.
• 43:20 - How do we make conversations natural?

References:

• Dustin's VUX World Podcast
• Dustin's Linkedin profile
• Hannes' LinkedIn profile
• Speechly's Twitter profile
• Speechly product search and checkout demo
• Speechly's Interspeech Research Paper 2021

• 01:15 - How does NLP work?
• 04:05 - How do Transformer-based NLP models work?
• 08:20 - How to look at unstructured data to take advantage of it more.
• 12:00 - How to leverage ML to bring more to unstructured data?
• 15:25 - Approach for low resources languages.
• 23:25 - Word embeddings for common reasoning needs.
• 26:55 - Techniques to follow to improve error and ambiguity in training data or for a model in general.
• 30:10 - Are GPTs leading effort in the field in a wrong direction?
• 34:15 -  Is DeepLearning the end of AI?
• 37:20 - What are some good NLP metrics to watch?
• 42:05 - How do we get past transactional queries to conversational queries?
• 52:00 - Is the Turing test still relevant for NLP or has it become obsolete?

References:

• AI-Powered Search referenced in respect of text not being unstructured.
• Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing
• Rethinking Search:Making Experts out of Dilettantes Common sense reasoning
• TWIML AI podcast 518 with Yejin Choi
• DARPA's Explainable AI Project
• EPITA is an engineering school in Paris.
• Marc's LinkedIn profile.

• 12:50 - Is the Turing test still relevant?
• 21:30 - Why it's important to use methodologies in AI projects and what are some best practices out there fit for AI projects.
• 28:00 - Falsehoods of methodologies in AI projects.
• 35:00 - Is Agile a good framework for AI/ML projects/products?
• 40:10 - How can projects get derailed or fail if you don't have a plan in place.
• 44:20 - The best compliment one can get after building an AI project or system.
• 47:25 - Is DL the end of AI?

References:

• CPMAI methodology
• Cognilytica's Voice Assistant Benchmark 1.0 and 2.0
• AI Today podcast show with Alexandra Petrus as guest
• AI Today podcast show

• 2:10 - Using AI to augment and reshape creativity in a modern world. Psychological creativity and story creativity - can an AI model help AI music artists, today, get off their creative blocks?
• 12:15 - Attempt to define ‘good’ music, using a cognitive music literature background.
• 17:00 - Are we better or worse off, for AI in audio/music? Is it sustainable for the effort input and cost, impact and efficiency output?
• 22:35 - ‘Deep Nostalgia” from myheritage initiative, and GPT-J - looking for strengths in the two approaches.
• 29:25 - The Sound of AI community - a HuggingFace version for audio?
• 31:15 - Train a DL - CNN sound classifier built with Pytorch and torchaudio on the Urban Sound 8k dataset.
• 35:00 - Is deep learning a dead end for artificial intelligence?
• 38:05 - Could someone that is a pure tech profile ever be in such an intersection in sync with the artistic world? Is it a pre-req to be domain savvy to build AI audio solutions?
• 42:10 - Helping music tech companies with a focus on audio (voice, speech, sound), the experience so far.
• 49:45 - Hard problems to solve when dealing with AI audio - Top three.
• 56:50 - First piece of music composed by a machine.

References:

• The Sound of AI YT Channel: https://www.youtube.com/c/ValerioVelardoTheSoundofAI/featured
• Sign up for The Sound of AI Slack Community
• PyTorch for Audio + Music Processing https://www.youtube.com/watch?v=gp2wZqDoJ1Y&list=PL-wATfeyAMNoirN4idjev6aRu8ISZYVWm
• Audio Signal Processng for ML https://www.youtube.com/watch?v=iCwMQJnKk2c&list=PL-wATfeyAMNqIee7cH3q1bh4QJFAaeNv0
• OpenSource Research project building a speech-operated neural synthesiser
• Deep Learning for Music https://github.com/ybayle/awesome-deep-learning-music
• Sweet Anticipation book: Music and the Psychology of Expectation by David Huron
• Valerio Velardo's LinkedIn
• The Frame Problem of AI

• 1:50 - Using AI for the environment
• 6:55 - AI spices for agriculture
• 12:15 - AI in outdoor uses
• 15:15 - Green AI in Seekar's work
• 22:15 - Training AI models for a green AI approach
• 27:10 - Seekar in the medical space, and covid19 opportunities
• 39:15 - NLP tradeoffs and takeaways
• 43:10 - Similarities in practicing jiu-jitsu and AI

References:

• Building AI models to be greener, and Seekar's Research Gate paper. This paper gives more insight into how Seekar was able to compress a large AI model down to a small enough size without compromising accuracy or performance.
• COVID-AI app from AppStore
• Exeda (Exploratory Emotional Detection Agent), mentioned in reference of using NLP for emotion recognition. Seekar's goal is to develop a psychological screening tool that can be downloaded as an app and used to check mental health daily through a 30-second voice recording in a similar manner as one brushes their teeth daily. 80% of personal communication happens through body language and Seekar’s products are utilizing this principle to better treat mental health. Research paper in progress.

• 01:25 - Do NLP models need someone that is not completely monolingual?
• 05:20 - Types of NLP  in marketing and/or e-commerce.
• 11:30 - Challenges in the e-commerce space: Behavioural data gathered by cookies has disappeared.
• 16:00 - Every 40 seconds, our attention breaks. Is that fact taken into account in NLP modeling for personalization?
• 18:20 - Models like GPT-3 open a whole new commercialization avenue in the marketing world, specifically for content creation. Impact of the wave.
• 21:50 - Is it fair to use an AI model for IP and content in such a way you influence millions of users on a website at once?
• 30:45 - Explainable models, debugging and how models could function.
• 37:00 - Provocative contexts for data scientists nowadays.
• 41:00 - Future of NLP.

Episode references:

• GPT3 the beginning of a new app ecosystem
• Amazon makes Alexa Conversations generally available to developers
• Copy.AI and Taglines.AI based on GPT3. Other spinoffs in the same space: Copy Shark; Snazzy AI; experiments using platforms like VWO.
• Explainable models by DARPA
• NLP in Marketing, part 1
• How virtual assistants (i.e. in your smartphone) understand you
• AI and NLP in marketing, webinar
• Katherine's Linkedin
• Katherine's Twitter
• Bucharest AI's meetup on Gender Imbalance, AI Mentorship & good delivery in AI

• 01:35 - Why did you decide to continue bootstrapping and decided to not opt for an investment.
• 06:50 - In the age of the million dollar supremacy how much money is a VC ready to invest.
• 08:56 Open source AI, good or bad idea? - VC and deep tech founder perspective.
• 14:15 - What’s the ideal shareholder split?
• 20:40 - Should one opt for Europe instead of Silicon Valley to raise capital faster?
• 23:10 - Effects of the pandemic on the deep tech investment space.
• 29:10 - Do VCs run their due diligence in their investment process + should VCs start considering checking reddit channels from now on?
• 32:45 - The gap between early stage deep tech startups and investments.
• 41:30 - Time, as an essential factor, in a deep tech startup  - time from idea to prototype.
• 49:45 - How is a founder coping with the long development cycle from a cost / business model perspective.
• 55:00 - Pre-seed to seed stage, where is the role of AI/ML: core, feature, end-to-end, black box.
• 59:10 - How much is reusing vs. proprietary AI work.
• 01:01:15 - What does a VC scout do?

Reference links:

Alexander Piskunov's LinkedIn

Amandine Flachs' LinkedIn

Amandine Flachs' Twitter

Venture Capital Scout Programs

• 02:43 - Motivation behind building a scaled MOOC AI course
• 06:40 - Effort behind an AI course to educate 1% of EU citizens
• 10:45 - Finland's heritage in education, and AI takeaways for course takers
• 20:55 - AI Challenge, or how are companies joining the AI education movement
• 25:45 - Digital spending priority: digital skills & education OR upgrading our health systems - Opinion
• 30:13 - Feels of a creator after building a popular AI course
• 36:55 - Ethics of AI course, and Elements of AI new chapters exploration

References:

• Elements of AI Romania
• Elements of AI global version 
• EU local Elements of AI Partners & movement
• Ethics of AI course
• Ready AI
• Finnish Center for AI
• Prof. Teemu Roos LinkedIn
• Prof. Teemu Roos Twitter
• Artificial Intelligence from Finland e-book

• 02:10 - Brief history of game development in relation to AI advancements
• 10:15 - Games driving advances in AI research: PR or reality?
• 15:50 - Latest AI technique popular in game development
• 20:55 - The role of Unity Game Simulation to reduce time & cost with games pre-launch testing
• 26:45 - What’s fancy in the games world
• 31:35 - Streaming a game vs. traditional edge processing, gamer’s lens
• 37:55 - What's next for games & AI

References:

• Unity ML-Agents Toolkit GitHub
• Jeff's Twitter handle @shihzy
• Jeff's LinkedIn

Host's notes:

• 2021 Update for AI advancements through game examples
• History of games at DeepMind
• Top AI Labs worldwide and AI's potential
• Facebook, Carnegie Mellon build first AI that beats pros in 6-player poker

• 2:00 - Hottest AI trends for 2021
• 5:35 - Open source for AI - paradigm shift
• 11:30 - AI model supremacy
• 21:10 - Authorship rights when AI contributes
• 31:00 - GPT encapsulating knowledge?
• 34:00 - Human consciousness replicable as computation
• 41:50 - Are we in a matrix?
• 42:30 - Cyberpunk 2077
• 50:50 - Can AI create emotion the way we cannot tell it is AI?

Conversation references:

• Book: "You look like a thing and I love you" - Janelle Shane
• Book: "Shadows of the Mind", Roger Penrose
• Chinese Room argument
• Manhattan project
• Art Breeder project
• The Origin of Circuits - re FPGA topic

Host's notes:

• Gartner Top Strategic Technology Trends for 2021 
• Jukebox - music-making tool by OpenAI. While the achievement is significant from a technological perspective, the results are unlikely to threaten the livelihoods of human musicians.
• DALL·E generates images in response to written inputs, and (whose name honours both Salvador Dalí and Pixar’s WALL·E) is a decoder-only transformer model. From Andrew Ng's 'The Batch' newsletter: OpenAI trained it on images with text captions taken from the internet. Given a sequence of tokens that represent a text and/or image, it predicts the next token. Then it predicts the next token given its previous prediction and all previous tokens. This allows DALL·E to generate images from a wide range of text prompts and to generate fanciful images that aren’t represented in its training data, such as “an armchair in the shape of an avocado.” WHY it matters? As Ilya Sutskever puts it ‘combining language and vision techniques could overcome computer vision’s need for large, well labeled datasets’.

1. Why do you do what you do?
2. Using big data and AI to connect big groups - how is that going and what are you current challenges?
3. How does a customer journey usually go. Take the example of the BeAI community, what would the journey look for us?
4. “Share your travel plans with your whole network or just a few selected friends and see if any of your plans match”, do you find it hard to resonate with people given the pandemic and limitation of travels? Have you pivoted on this USP?

Reference links:

• WIYO website

Notes:

1. The role of a PM in a research environment
2. Recurring skills needed for an AI PM to have successful products built and good communication with both researchers and business types
3. AI model governance and why is important in banking
4. Where can AI help in the banking industry
5. What is responsible AI
6. Borealis and RBC initiatives to help with social good causes and women in AI
7. Diversity and Inclusion - what it means and why it matters for product management
8. Top things for a PM in a research project journey

Reference links:

• Respect AI Initiative
• Borealis AI
• Wendy’s LinkedIn

1. 5000+ likes on Facebook, that is a good crowd for a startup, how did you build this?
2. Current tech stack and challenges.
3. Current increased online consumption and trends versus your solution - how do you see everything evolving?
4. What are the languages covered?

Reference links:

• EVAAI Website
• EVAAI Facebook Page

1. Why do you do what you do?
2. Are you a cybersec company?
3. Current AI used.
4. Problems addressed & industries targeted.
5. AI regulations - where to stand.

Reference links

• Factide Website
• Factide Facebook Page
• Factide LinkedIn Page
• Factide Twitter Page

1. Why do you do what you do?
2. Where and what do you use AI for?
3. How does it feel, for a computer science researcher, to build a research spin off startup in France? What do you struggle most with?
4. Who are your customers and users?
5. How does a customer journey feels like?

Reference links:

• Emoface Website 
• Sign up for beta

Notes:

1. Industries most interested during these times, and the ones taking a step back
2. Road to product market fit
3. Communicating with potential clients
4. Sales & growth team profile

Notes:

1. Deep Reinforcement Learning (DRL or DeepRL) applied to the automotive industry
2. Simulation platforms and the role of simulators in training agents
3. Obtaining data to prepare the autonomous vehicle
4. Methods to evaluate robustness of the solution
5. Deploying in real world
6. Startups to use DL or be at the forefront of DL
7. Techcrunch Disrupt Hackathon win & engineers at hackathons as a practice

We discuss:

• Can twins have identical typing patterns?
• Advantages and opportunities offered from early beginnings in Oradea, Romania
• Nailing a direction
• The generalist role
• Mixing behavioural biometrics with the AI technology