Uploaded November 2023 | Updated September 2026, 2 weeks ago
Learn about the limitations of state-of-the-art generative music systems, and procedural issues we're facing in generative music research. You'll also learn about possible solutions.
Website of the Generative Music AI Workshop in Barcelona:
https://www.upf.edu/web/mtg/generative-music-ai-workshop
Get the lecture slides:
github.com/musikalkemist/generativemusicaicourse/blob/main/08.%20Limitations%20and%20future%20vision/Slides/8.%20Limitation%20and%20future%20vision.pdf
Sign up to The Sound of AI Slack Community to join the discussion:
valeriovelardo.com/the-sound-of-ai-community
======================================
Interested in music AI consulting?
thesoundofai.com/consulting.html
Interested in music AI recruitment?
thesoundofai.com/recruitment.html
Become a Python ninja with my Advanced Python Programming course:
the-sound-of-ai-academy.teachable.com/p/advanced-python-programming
Connect with Valerio on LinkedIn:
linkedin.com/in/valeriovelardo
Follow Valerio on Twitter:
twitter.com/musikalkemist
======================================
Content
0:00 Intro
0:41 Text-to-music limitations
3:32 The curse of DL
4:53 Deep learning limitations
9:48 Solving the curse of DL
11:26 Music representation limitations
13:20 Music representation: Way forward
14:12 User interfaces
15:39 Problems with research procedure
17:56 Research proedure: Way forward
20:09 Engineer / domain knowledge divide
21:33 Key takeaways
24:05 What's up next?
24:37 Workshop info
Learn about the limitations of state-of-the-art generative music systems, and procedural issues we're facing in generative music research. You'll also learn about possible solutions.
Website of the Generative Music AI Workshop in Barcelona:
https://www.upf.edu/web/mtg/generative-music-ai-workshop
Get the lecture slides:
github.com/musikalkemist/generativemusicaicourse/blob/main/08.%20Limitations%20and%20future%20vision/Slides/8.%20Limitation%20and%20future%20vision.pdf
Sign up to The Sound of AI Slack Community to join the discussion:
valeriovelardo.com/the-sound-of-ai-community
======================================
Interested in music AI consulting?
thesoundofai.com/consulting.html
Interested in music AI recruitment?
thesoundofai.com/recruitment.html
Become a Python ninja with my Advanced Python Programming course:
the-sound-of-ai-academy.teachable.com/p/advanced-python-programming
Connect with Valerio on LinkedIn:
linkedin.com/in/valeriovelardo
Follow Valerio on Twitter:
twitter.com/musikalkemist
======================================
Content
0:00 Intro
0:41 Text-to-music limitations
3:32 The curse of DL
4:53 Deep learning limitations
9:48 Solving the curse of DL
11:26 Music representation limitations
13:20 Music representation: Way forward
14:12 User interfaces
15:39 Problems with research procedure
17:56 Research proedure: Way forward
20:09 Engineer / domain knowledge divide
21:33 Key takeaways
24:05 What's up next?
24:37 Workshop info
![GPT-4: The Good, the Bad and the Ugly [Paper Breakdown]](https://i.ytimg.com/vi/wYqVE5_8Nqc/mqdefault.jpg "GPT-4: The Good, the Bad and the Ugly [Paper Breakdown]
GPT-4 is the most competent general-purpose AI trained to date. It can reason, it shows an advanced understanding of the world, and it has common sense. We know OpenAI used a transformer architecture to build GPT-4, but what else do we know about the model? What are the shortcomings and threats of this technology? Why is OpenAI research procedure problematic?
Join The Sound of AI Slack Community:
https://valeriovelardo.com/the-sound-of-ai-community/
GPT-4 Technical Report:
https://arxiv.org/abs/2303.08774
Music AI talent recruitment:
https://thesoundofai.com/
Interested in hiring me as a consultant/freelancer?
https://thesoundofai.com/consulting.html
The Sound of AI Academy:
https://the-sound-of-ai-academy.teachable.com/
Advanced Python Programming:
https://the-sound-of-ai-academy.teachable.com/p/advanced-python-programming
Connect with Valerio on Linkedin:
https://www.linkedin.com/in/valeriovelardo
Follow Valerio on Facebook:
https://www.facebook.com/TheSoundOfAI
Follow Valerio on Twitter:
https://twitter.com/musikalkemist
Content
0:00 Intro
1:30 What we know about GPT4
3:35 Capabilities
9:21 Earth, chicken nuggets, and intelligence
12:28 Intelligence = language + multimodality?
13:29 Limitations and safety concerns
16:26 Procedural no-gos
19:24 The Sound of AI Slack Community")
![This AI Can Solve 604 Tasks [Paper Analysis of Gato by DeepMind]](https://i.ytimg.com/vi/zO49vZ31xb0/mqdefault.jpg "This AI Can Solve 604 Tasks [Paper Analysis of Gato by DeepMind]
DeepMind published a revolutionary paper 🔥 They introduced Gato, a generalist AI agent that can carry out more than 600 tasks with a single transformer neural architecture. The tasks are varied, from playing Atari games to providing captions to images.
This paper demonstrates that:
📌 Generalist agents can perform reasonably well on many tasks / embodiments / modalities
📌 Generalist agents have the potential to learn new tasks with few data points
📌 By scaling up the parameter size, we can build a general-purpose agent
This work shocked me. I’ve always tackled AI from the perspective of Narrow Intelligence: build a specialised model that does well on a single - quite constrained - task.
👉 Gato paves the way for Artificial General Intelligence (AGI). In so doing, it opens new ethical dilemmas that should at least spark discussions in the AI community.
Since I’ve finished reading this paper, I can’t stop asking a question: is it ethical to push this research line given the grave dangers which may come with quasi-AGI agents?
Would you like to learn more? Check my last video, where I provide a breakdown of the paper, and analyse its ethical implications.
A Generalist Agent by DeepMind:
https://www.deepmind.com/publications/a-generalist-agent
Interested in hiring me as a consultant/freelancer?
https://valeriovelardo.com/
Join The Sound Of AI Slack community:
https://valeriovelardo.com/the-sound-of-ai-community/
Connect with Valerio on Linkedin:
https://www.linkedin.com/in/valeriovelardo
Follow Valerio on Facebook:
https://www.facebook.com/TheSoundOfAI
Follow Valerio on Twitter:
https://twitter.com/musikalkemist
Content:
0:00 Intro
1:11 General vs Narrow intellicence
3:06 Research hypotheses
4:25 Idea to approach AGI
8:23 Benefits of single network for many tasks
10:04 Datasets used
11:59 Data preparation
18:24 Model architecture
20:31 Training
22:48 Loss function
27:25 Recognising a task
30:50 Inference
33:37 How does the model perform?
39:00 Scale analysis
40:18 Can the model tackle unseen tasks?
44:14 Key discoveries
47:40 Ethical implications")
