The Artificial Intelligence Show

Salesforce AI Cloud, White House Action on AI, AI Writes Books in Minutes, ChatGPT in Cars, and More

Tue Jun 27 2023

AIMarketingSalesGenerative AISynthetic MediaGovernment RegulationResponsible AI DevelopmentAdvancements in AI ModelsData Annotation

Description

The episode covers various topics related to AI, including building foundational models, marketing applications, sponsorship and events, Salesforce AI Cloud, generative ad capabilities, AI-generated avatars and synthetic media, government regulation on AI, responsible AI development, advancements in synthetic media, advancements in AI models, AI in marketing and sales, DeepMind and OpenAI's perspective on AI safety, and data annotation for AI training.

Insights

AI models and their limitations

AI models are becoming our interface to the world, but they may still have limitations such as hallucinations or making things up.

Government regulation on AI

Government regulation related to AI is being explored in the United States Senate and the European Union Parliament. OpenAI, Microsoft, and Google have lobbied for looser regulations, while OpenAI has expressed concerns about the potential impact on their business in Europe.

Advancements in synthetic media

Advancements in synthetic media include features like eye contact in post-production, intelligent script matching, and generative AI audio. The technology of replicating voices will become extremely commonplace within the next one to two years.

AI models and responsible development

Building large language models is complex and requires significant resources. Responsible AI development involves addressing challenges like hallucinations and ensuring ethical use of AI.

AI in marketing and sales

AI is being integrated into marketing and sales software platforms, creating new possibilities for personalized customer experiences. Companies like Mercedes Benz and HubSpot are exploring AI voice assistants and GPT-like capabilities for data querying.

DeepMind, OpenAI, and AI safety

Mustafa Suleyman, co-founder of DeepMind and OpenAI, emphasizes the importance of regulations and AI safety. OpenAI has raised significant funding for training models and envisions a future with personalized AIs accessible across platforms.

Data annotation for AI training

Data annotation is a crucial step in training AI systems, but it can be tedious and isolating for workers. The future of data annotation may involve automation by artificial intelligence.

Summary

Transcript

Building Foundational Models in AI

00:00 - 07:01

Building foundational models in AI is complex and requires significant resources.

The Marketing AI Show

00:00 - 07:01

The Marketing AI Show is a podcast that explores the practical applications of artificial intelligence in business.
The hosts, Paul Ritzer and Mike Kaput, have been traveling but are now back to their regular weekly format.

Sponsorship and Events

00:00 - 07:01

The episode is sponsored by Brando, a platform that offers various tools for brand marketing performance.
The fourth annual Marketing AI Conference (MAKON) is happening soon in Cleveland.

Salesforce AI Cloud

00:00 - 07:01

Salesforce has launched AI Cloud, a platform that allows enterprises to use enterprise-grade AI products safely.
AI Cloud includes Salesforce's own AI-powered products as well as language models from other providers like Amazon.
Salesforce has also doubled its venture capital fund for generative AI startups to $500 million.

Generative Ad Capabilities

06:39 - 14:08

Microsoft, Google, Salesforce, and others will infuse generative ad capabilities into their platforms in 2023, accelerating adoption.

AI-generated Avatars and Synthetic Media

06:39 - 14:08

Cynthia, a company using AI-generated avatars to lower video production costs, raised $90 million and is valued at $1 billion.
Synthetic media creation, such as deep fakes and virtual videos, is expected to see rapid deployment.
Sintasia aims to disrupt video production by creating AI-generated avatars based on real-life actors speaking in front of a green screen.

Government Regulation on AI

13:45 - 21:59

Government regulation related to AI is being explored in the United States Senate through educational sessions for senators.
The initiative seeks to help lawmakers understand AI's implications and create optimal legislation while mitigating risks.
The government is just starting to address the basics of AI regulation.
The generative AI age is only seven months old, and people are still grasping its significance.
The European Union Parliament has taken steps towards implementing comprehensive restrictions on AI.
The proposed EU AI Act aims to safeguard consumers from harmful applications of AI.
The legislation is not yet law but may be approved by the end of 2020.
Provisions in the legislation include a risk-based approach to AI and requirements for labeling and data transparency.
OpenAI has expressed concerns about the potential impact on their business in Europe.
OpenAI CEO Sam Altman privately lobbied for lighter restrictions in the EU's AI Act.
Time reports that OpenAI's lobbying efforts seem to have been successful in influencing the final draft of the act.
Microsoft and Google have also lobbied EU officials for looser regulations on large AI providers.
US President Joe Biden and his advisors are prioritizing AI regulation through executive orders and new policies.
The White House is focusing on areas like cybersecurity, consumer protection, and economic transformation in relation to AI.
Initiatives include collaborating with AI companies, releasing guidance for federal agencies, developing international norms, and building upon an earlier AI bill of rights.

Responsible AI Development

29:02 - 35:47

Major initiatives are being undertaken by organizations to address important AI topics and make a positive impact.
Aiden Gomez, CEO of Cohere, shared his perspective on various AI issues in an interview with the Financial Times.
Google Brain was once the dominant player in AI research, but it struggled to turn its advancements into real products and experiences.
Building large language models is complex and requires significant resources, creating a potential moat for companies like Cohere.
AI models will become our interface to the world, although they may still have limitations such as hallucinations or making things up.
Some tasks currently performed by humans may be completely replaced by AI, while others will never be replaced.
Aiden Gomez disagrees with the idea of pausing AI development and sees it as impractical.
Descript, an AI video editing tool, announced new features including Descript for web and an eye contact feature.

Advancements in Synthetic Media

35:25 - 42:55

The feature called eye contact uses AI to make sure your eyes are always looking at the camera in post-production.
Replace Selection or Record in Replace allows you to record over an existing script and intelligently match the new transcript to the original.
Recording scene by scene enables manual step-through of scenes for recording, even if they are different from the original script.
Regenerate is a generative AI audio feature that seamlessly improves edits and removes unexpected background noise.
Script and Runway take a smarter approach to product development with AI as their first approach.
Meta AI researchers have developed VoiceBox, a generative AI model for speech that can generate high-quality audio clips across six languages.
VoiceBox outperforms other models in terms of error rate and audio quality, but Meta is not releasing it publicly due to concerns about misuse.
The technology of replicating voices will become extremely commonplace within the next one to two years, with both positive and negative implications for society.

Advancements in AI Models

42:26 - 49:30

Meta's chief AI scientist, Jan Lecun, introduced a new AI model called I-J-E-P-F that learns an internal model of the world by creating abstract representations of images.
The model demonstrates strong efficiency and performance, opening up possibilities for applications in video and image artificial intelligence.
Jan Lecun believes that large language models are not a path to general intelligence; instead, AI needs to experience the world around it to achieve human-like intelligence.
AI can now write entire novels in minutes based on an initial prompt using GPT author, an open source project by OtherSideAI.
A novel produced by GPT author can cost as little as $4 and is written in just a few minutes.
The concept of AI writing novels raises questions about copyright concerns and ethical implications.
OpenAI is considering launching a marketplace where customers can sell customized AI models to other businesses, making it easier for companies to access and deploy more customized AI systems.

AI in Marketing and Sales

49:08 - 56:08

Mercedes Benz is beta testing chat GPT as a voice assistant in its cars
The current voice assistant system in Tesla cars is considered horrible and rudimentary
Combining AI tools with existing marketing and sales software platforms creates the concept of 'composability'
Chat spot for HubSpot applies GPT-like capabilities to querying data in the HubSpot CRM
Inflection AI has announced a new large language model called inflection 1 that outperforms GPT 3.5, Meta's llama model, and Google's Palm model

DeepMind, OpenAI, and AI Safety

55:48 - 1:02:16

Mustafa Suleyman, co-founder of DeepMind and OpenAI, teamed up with Reed Hoffman to create a major player in the world of AI.
Suleyman believes that AI safety is a concern on a multi-decade time scale.
He emphasizes the importance of regulations, including watermarks for synthetic media and third-party red teamers.
OpenAI has raised $225 million and plans to raise another $675 million for training models.
Suleyman envisions a future with many specialized AIs that are personalized and accessible across platforms and devices.
The revenue model for OpenAI's personal AI assistant is still undecided.
Suleyman acknowledges the need for governing what the AI can say and be an arbiter of truth.
He believes hallucinations can be eliminated by June 2025.
While job loss may occur, Suleyman sees AI as ultimately positive for humanity.

Data Annotation and AI Training

1:01:58 - 1:07:16

Notable names on this year's list include Open AI, cohere, Descript, Inflection AI, Jasper, and Midjourney.
Coca-Cola has appointed a global head of Generative AI in their marketing transformation offices.
A new report highlights the hidden work of data annotation behind AI systems.
Data annotation is a necessary step in training AI systems and involves labeling data sets.
The work of data annotation can be tedious, repetitive, and isolating for workers.
There is secrecy around the industry of data annotation, leading to a lack of understanding about the information shaping AI systems.
The article provides background information on how AI models are trained through human labeling.
The job of annotating data is part of an emerging economy that many people are unaware of.
The future of data annotation may involve automation by artificial intelligence.