Opening Remarks:
Welcome to the UNSW and Google jointly organized event on artificial intelligence (AI).

Acknowledgement of Traditional Custodians:
Recognition of the Bedjigal people as traditional custodians of the land where the event is held. Paying respect to elders, past and present, and extending respect to Aboriginal and Torres Strait Islanders present.

Introduction of Professor Ian Jacobs:
Professor Ian Jacobs, President and Vice-Chancellor of UNSW Sydney, is invited to introduce the panellists.

Welcome and Acknowledgement:
Professor Jacobs echoes the respect and acknowledgement of the Bedugal people and indigenous attendees. He welcomes everyone to a landmark event featuring two eminent leaders and researchers in AI.

AI’s Promise and Concerns:
AI has captured the imagination with its promise of driverless cars and personal assistants. Concerns are raised about the appropriate use of AI, especially in warfare and job displacement. AI’s ability to predict human behavior better than humans themselves raises concerns for many.

Current State of Artificial Intelligence:
Rapid advancements in AI in recent decades. AI programs defeating humans in games like chess and Go. AI aiding medical diagnosis for years. Potential for AI and robots to replace complex cancer surgery. AI used in personal assistants, online shopping, and Google Maps.

UNSW’s Involvement in AI:
Strong history and international recognition for AI research. World Soccer Champions in robo-soccer. Center for Quantum Computation and Communication Technology developing computers with advanced capabilities.

Exciting Developments at Google:
Access to digital data, computing power, and better algorithms enables faster progress. Language understanding improved from single sentences to millions of pages. Image recognition at the level of humans for broad and specific tasks.

Deep Learning:
Deep learning contributes to applications in vision and natural language. Flexible models with the right level of complexity are key. Deep learning may evolve and incorporate new architectures.

Limitations of Deep Learning:
Not suitable for all tasks, such as long-term planning. Hybrid approaches combining deep learning and other techniques are necessary.

AlphaGo Example:
AlphaGo used computer vision techniques to assess the Go board and identify potential moves. Two key factors: determining the current position and finding interesting moves.

AI’s Four Storms:
AI has experienced four major storms, including the expert systems boom in the ’80s followed by an “AI winter” due to the complexity of the field. Recent years have seen a resurgence in AI research and funding, with a doubling rate every two years for the past six years.

Funding and Perception:
The current AI boom is supported by immediate returns and solutions-oriented approaches, driving continued investments. The risk of another “winter” lies in the perception that AI techniques excel at pattern recognition but lack full reasoning capabilities, leading to disillusionment.

Technical Breakthroughs:
Significant breakthroughs have occurred in AI, such as speech recognition and machine translation, but further advancements are needed. Incremental improvements in certain areas may be more challenging compared to the past.

Calming Public Fears:
The concerns about AI taking over should be addressed by acknowledging the complexity and long-term nature of the field, as illustrated by the story of John Henry.

AI and Repetitive Tasks:
AI is likely to take over repetitive tasks that are not very interesting and have a lot of data. Humans will still be needed for novel and new tasks that lack data. Humans and AI can work together effectively, with AI providing tools to enhance human productivity.

Google’s Use of AI:
AI is widely used at Google, in areas such as information organization, data center optimization, and energy efficiency. Google’s approach involves identifying problems, collecting data, and using machine learning and operations research to improve processes.

Learning from Google’s Mistakes:
Google is working on developing software tools for a smooth transition from hand-coded software to machine learning. Current tools for dealing with code are advanced, but there is a lack of tools for managing data. Data errors are more common than code errors, and major service disruptions are often caused by data-related issues.

Google’s Resiliency:
Google’s services are highly reliable, with minimal downtime. Unlike British Airways, Google’s systems rarely experience major outages.

Challenges in Implementing AI:
AI systems often fail due to human error, such as incorrect DNS entries, rather than code errors. Lack of strong typing for data compared to programs hinders the development of robust AI systems. The current toolkit for data management and verification is insufficient, leading to data-related challenges in AI implementations. There is a shortage of skilled personnel with expertise in AI and machine learning.

Addressing the Challenges:
Training initiatives have been implemented to address the skills gap, with 10,000 engineers receiving short courses on machine learning. Course materials are made available to universities and the public to encourage younger people to pursue computer science and machine learning. Aiming to make machine learning accessible to everyone, allowing them to use it effectively without deep expertise.

MOOCs and the Future of Education:
MOOCs (Massive Open Online Courses) have transformed the way people access education and acquire new skills. MOOCs offer opportunities for individuals to gain new skills and adapt to changing job requirements. The effectiveness of MOOCs in providing one-on-one tutoring and personalized learning experiences needs improvement.

AI Regulation and Societal Impacts:
Recognizing the need for regulation in the AI industry to address potential societal impacts. Learning from past technological innovations, such as the internal combustion engine, to avoid unintended consequences. Ongoing efforts to understand the effects of AI and develop appropriate management strategies.

Concerns about AI Algorithms:
Concerns are raised regarding algorithmic discrimination, particularly in the judicial system and search results. Lack of explanations and guarantees from AI algorithms poses challenges in addressing these concerns effectively.

Explanations Alone Are Not Enough:
Humans often generate explanations after making a decision, which may not be the true explanation. Machine learning systems may generate explanations that are not accurate representations of their decision-making process.

Need for Better Approaches to Explanations:
A whole field of study should be dedicated to developing better approaches to explanations in AI systems.

Other Ways of Monitoring Decision-Making:
Probing the system to see how changes in input affect the output can provide insights into the decision-making process. Guarantees and checks are needed to ensure fairness and prevent bias in AI systems.

Importance of Checks:
Explanations alone may not be enough to detect bias in AI systems. Looking at a wide variety of cases can help identify patterns of bias that may not be apparent from a single decision.

Conclusion:
Explanations, guarantees, and checks are all necessary for responsible AI decision-making.

The Echo Chamber Theory:
The idea that people are only exposed to information that confirms their existing beliefs, leading to political polarization and a lack of understanding of opposing viewpoints.

Segregation of News Sites:
A study found that the level of segregation of news sites that people visit is only slightly lower than random, suggesting that people are not as polarized in their media consumption as often assumed.

The Role of Big Media Sources:
The fact that there are only a few major media sources means that most people are exposed to a variety of viewpoints, even if they also consume content from fringe sources.

The Problem of Bots:
The Twittersphere is dominated by bots, which can distort the perception of public opinion and make it difficult for human voices to be heard.

The Need for Critical Thinking:
Education in critical thinking is essential to help people evaluate information and resist being swayed by echo chambers and bots.

The Impact of Technology on Journalism:
The decline of traditional revenue streams for journalism due to technology has led to layoffs and a reduction in the availability of fact-checked news.

The Role of Technology Companies:
Technology companies need to consider their impact on the media and explore ways to support quality journalism.

Benefits of AI:
Machine translation has made significant progress, enabling easier communication between people with different languages.

Challenges of AI:
News sources prioritize entertainment over international news to increase revenue. The spread of misinformation has become easier due to technology. Income inequality may be exacerbated as AI amplifies the differences between the rich and the poor. A.I. is not well suited for solving complex issues like the Middle East crisis.

AI in Education:
Private tutors powered by AI can provide personalized learning experiences. Developing explanations for AI’s conclusions is an ongoing area of research. It is important to invest resources in creating high-quality educational courses.

Jobs and AI:
A.I. may amplify income inequality by disproportionately benefiting the top earners. Efforts should be made to ensure that the benefits of AI are accessible to everyone.

Social Issues and AI:
Google is conducting research to address social issues using AI, but no specific details were provided.

AI’s Role in Addressing Societal Challenges:
Peter Norvig emphasizes the importance of AI in addressing societal issues like fake news and offensive content. Facebook uses AI to classify and remove inappropriate videos, including terrorist recruiting videos. Humans rate content to provide data for AI improvement. Public service announcements are displayed to counter harmful content.

Google’s Efforts to Democratize AI:
Google’s cloud computing offerings make AI accessible to small companies and individuals. AI services are provided on a pay-as-you-go basis, similar to electricity services. The availability of pre-trained models reduces the need for extensive data and computational resources.

Challenges in Broadening AI Access:
The current reliance on large amounts of data and specialized hardware may limit the accessibility of AI for some. Renting computational resources from Google may not be feasible for all applications.

Google’s Research Management Approach:
Google researchers are embedded in product teams, allowing for direct application of research findings. This approach differs from Microsoft’s more academic research lab model. It enables researchers to engage with the academic community and apply their findings to real-world problems.

Integrating Research and Engineering for Innovation:
Google’s employees are tasked with creating or improving products, not just publishing papers. Teams are formed with a mix of researchers, engineers, and managers, avoiding the segregation of research from production. This structure prevents researchers from getting sidetracked by short-term product demands, allowing them to pursue long-term research. New PhDs may initially struggle with this approach but often come to appreciate the rewards of creating products used by millions.

DeepMind’s Integration into Google:
DeepMind’s research-oriented approach led Google to grant them autonomy within the company. This arrangement has yielded positive results, with DeepMind achieving both theoretical advancements and practical applications.

Machine Translation Progress and Challenges:
Machine translation has made significant progress, approaching human-level performance in written form. However, human translators also make mistakes, so achieving human performance may not be the ultimate goal. Correcting human errors in translation is another area for improvement. Understanding the real world is crucial for accurate translation, as evidenced by examples like “I dropped the glass on the table, it broke.”

Machine Translation and Real-World Knowledge:
To achieve accurate machine translation, systems need to understand the world and its physical properties. Current training methods using text examples are insufficient for capturing real-world knowledge. Incorporating visual information and knowledge about objects and their interactions is necessary for comprehensive understanding.

Nuanced Speech Generation:
Speech recognition technology has made significant progress, but generating natural-sounding speech remains a challenge. Current systems can produce fluent speech, but struggle with conveying emphasis, pauses, and subtle nuances. Achieving nuanced speech requires further advancements in speech generation algorithms.

Machine Translation as the Last AI Problem:
Solving the machine translation problem requires a comprehensive understanding of language, the world, and the relationship between them. This makes machine translation one of the most challenging problems in AI, requiring significant advances in various fields. Until these challenges are addressed, perfect machine translation remains elusive.