Introduction

Fifi Lee, co-director of Stafford Institute of human-centered AI, welcomes the attendees to the event and introduces Dennis Hasabis, founder and CEO of Alphabet Deep Mind.

Welcome to Stafford High's Headquarters Space

• Fifi Lee welcomes President Mark Desi Leving for attending the event.

• She introduces Dennis Hasabis as the guest of honor.

About Dennis Hasabis

• Dennis started his career in video gaming before launching his own company.

• He pursued a PhD in cognitive Neuroscience at MIT to better understand the architecture of the brain.

• He has been a creative thinker and leader in thinking about the relationship between machine intelligence, human brain, and more.

• In 2010 he co-founded Deep Mind which was acquired by Google in 2014.

Deep Mind's Achievements

• Deepmind has made breakthroughs in data center energy consumption, generative degenerative eye conditions, and protein folding through Alpha fold work.

• It has made significant advances in deep learning and reinforcement learning.

• It has beaten world champions at Go game.

• The company is looking to apply techniques from Alpha fold to nuclear fusion to halt climate change.

Artificial General Intelligence

• Demis teams are exploring ideas around artificial general intelligence which he calls an epoch-defining technology that will change human lives.

Talk by Demis Hasabis

Demis Hasabis talks about using AI to accelerate scientific discovery itself.

Using AI for Scientific Discovery

• Demis is passionate about using AI to accelerate scientific discovery itself.

• He discusses how they have been at the forefront of where AI field was going since 2010 when they founded DeepMind.

• They expected things would go like this but it has been amazing for them to live through it.

• Demis talks about the difficulty of raising seed money in 2010 and how things have changed since then.

Current Work on Generative Models

• Demis discusses the current work on generative models and large language models.

• He talks about how they are exploring ideas around artificial general intelligence.

• The talk ends with a Q&A session.

Introduction to DeepMind

In this section, Demis Hassabis, the co-founder of DeepMind, talks about the mission and goals of the company. He also discusses how they started with deep reinforcement learning systems and their first big result in 2013.

The Mission of DeepMind

• The mission statement of DeepMind is to solve intelligence by fundamentally understanding its nature and recreating it in an artificial construct.

• They aim to use this technology to advance science and benefit humanity.

• They set up DeepMind in 2010 because they felt that there was a lot of different techniques coming together that they could see deep learning had just been invented, and they wanted to bring those things together along with some understanding they had about the human brain.

First Big Result: Atari DQN

• In 2013, they managed to scale up what then became called deep reinforcement learning systems to something that was actually significant.

• They built Atari DQN, their first successful system that learned how to play games like Space Invaders by only being given the raw pixels on the screen.

• This was probably the first example of an end-to-end learning system on something as challenging as an Atari game.

AlphaGo Program

• AlphaGo is a program developed by DeepMind that cracked the game of Go through self-learning systems.

• Go is a super complex game played in Asia with more possible positions than there are atoms in the universe.

• Unlike chess players who can distill Grandmasters' knowledge into a set of rules and program expert system chess computers, Go players cannot explicitly explain the rules they follow.

The Promise of Learning Systems

In this section, the speaker discusses how learning systems like AlphaGo can come up with solutions to problems that humans may not know how to solve. He also talks about the different levels of creativity in these systems.

Levels of Creativity

• There are three levels of creativity in learning systems.

• The first level is interpolation, which involves averaging things together.

• The second level is extrapolation, where the system discovers new strategies never seen before.

• The third level is invention or out-of-the-box thinking, where the system invents something entirely new.

True Creativity and Self-play Systems

In this section, the speaker talks about true creativity and self-play systems like AlphaGo and AlphaZero.

Self-play System

• A self-play system starts by playing randomly and generates a dataset from it.

• This dataset is then used to train a new version of the program that predicts what sorts of moves are likely to be good.

True Creativity

• True creativity involves inventing something entirely new.

• Current AI systems cannot take on abstract conceptual instructions yet.

Face-Off Match

In this section, the speaker explains how to train a system to play games using a face-off match.

Training Process

• The training process involves playing a hundred game match between V1 and V2.

• If V2 beats V1 by above a 55% win rate threshold, it is assumed that it is significantly better.

• The Master System is then replaced with the new V2 system.

• This process is iterated by playing another 100,000 games with V2, which makes it slightly stronger.

• The generated data from these games is used to train V3, which is then faced off against V2.

Creating Models for Games

In this section, the speaker explains how creating models can help narrow down search options in games.

Using Models for Tree Search

• Creating models helps model the dynamics and strategies of a game like Go.

• It allows tree search on top of that in an attractive way.

• Instead of exploring all possibilities like in Gray here, you use the model to narrow down your search to just the most reasonable options.

• This allows you to find near-optimal or very strong moves within a certain amount of constant thinking time.

Breakthroughs in AI Gaming

In this section, the speaker talks about breakthroughs made in AI gaming over the last decade.

Landmark Results

• Over the last decade, there have been many big breakthroughs made in different games such as Atari and AlphaGo.

• AlphaZero was created to generalize these breakthroughs to every two-player game.

• AlphaStar was created to beat Grand Masters players at Starcraft 2.

Using Games for AI Training

In this section, the speaker talks about using games as a means to an end in AI training.

The End Goal

• Although games have been used in every way possible, they have always been a means to an end.

• The end goal was never just to win at Go or Starcraft but to use games as a convenient proxy to test out algorithmic ideas.

• This allowed them to apply these ideas to important real-world problems.

AlphaFold Program

In this section, the speaker talks about the AlphaFold program and its significance in predicting protein folding.

Protein Folding Problem

• The protein folding problem is an important problem in biology that involves predicting the 3D structure of a protein from its amino acid sequence.

• AlphaFold was created to predict protein folding directly from its one-dimensional sequence.

• It has made significant breakthroughs in predicting protein structures accurately.

Protein Folding Problem

In this section, the speaker discusses the protein folding problem and its significance in biology. The speaker also talks about how this problem can be solved computationally using AI.

Significance of Protein Folding Problem

• It takes several years to experimentally determine one protein structure.

• Determining the 3D structure of a protein is a significant challenge in biology that has been unsolved for over 50 years.

• The protein folding problem is a suitable problem for AI to solve.

Difficulty of Protein Folding Problem

• There are 10^300 possible shapes that an average-sized protein can take, making it difficult to determine its structure experimentally.

• Nature spontaneously folds proteins in milliseconds, which raises questions about how this process occurs.

CAST Competition

• The CASP competition is like the Olympics of protein folding and tests the best computational systems every two years.

• Experimentalists give their predictions to the competition before publishing their results, making it a double-blind experiment for testing computational systems.

Alpha Folder Project

• Alpha Folder was started by DeepMind in 2016 as their next big project after AlphaGo.

• The winning scores of the top team in each competition have been increasing since 2006.

Protein Folding Problem

In this section, the speaker discusses the protein folding problem and how it is useless for experimentalists.

Accuracy of Prediction

• Proteins are known to be truly unknown from their sequence.

• Experimentalists need predictions that are accurate to within the width of an atom, which is less than one angstrom error.

• A 90 GDT score is required for predictions to be useful to experimentalists.

Alpha Fold One

• Alpha Fold One won the competition by a huge margin in 2018.

• Cutting-edge machine learning techniques were used as the core of a protein folding system.

• Alpha Fold Two needed a completely different approach and architecture to achieve a 90 GDT score.

Innovations in Alpha Fold Two

• Physics constraints and chemical constraints were built into the network properties of Alpha Fold Two.

• Prior domain knowledge was combined with a learning system without interfering with it.

• The organizers declared that the problem had been solved after seeing the results from Alpha Fold Two.

Impact of Alpha Folding

• The methods were published open source for maximum impact and benefit for humanity.

Alpha Fold Protein Folding

In this section, Dr. Demis Hassabis talks about how Alpha Fold was able to generate all proteins known to science and the importance of creating a new database for it.

Generating Proteins with Alpha Fold

• During Christmas holidays, the team generated the whole human proteome and 20 model organisms.

• High accuracy predictions were made for every protein known to science.

• This is important because less than one percent of proteins in some organisms like plants are known experimentally.

• The coverage for humans was doubled with high accuracy predictions.

Creating a New Database

• The team decided to create a new database instead of doing their own.

• They teamed up with emblem ebi European bioinformatics Institute to create a new database and host it.

• All alpha fold predictions were put onto that database and plugged into other databases such as genetics databases.

Safety and Ethics

In this section, Dr. Demis Hassabis talks about how they consulted experts in various fields to ensure that releasing Alpha Fold would not be dangerous.

Consulting Experts

• Over 30 experts were consulted from biologists, pharma, biosecurity, human rights, etc.

• They unanimously agreed that the benefits far outweighed any potential risks involved.

Uses of Alpha Fold

In this section, Dr. Demis Hassabis talks about some examples of how Alpha Fold has been used by scientists around the world.

Examples of Use

• John Mcgean and team from Portsmouth are using it to design plastic eating enzymes.

• It was used to help determine the nuclear pore complex, which governs a kind of gateway to let nutrients in and out of the cell nucleus.

• Feng Zang at the broad Institute is using it for creating a molecular syringe as a new drug delivery mechanism.

Science at Digital Speed

In this section, the speaker talks about the concept of science at digital speed and how Alpha Fold is an example of it. He also discusses how digital speed can accelerate the pace of progress in science.

Science at Digital Speed

• The speaker defines science at digital speed as being able to do science at the same speed as digital technology.

• Alpha Fold is an example of science at digital speed because it can fold proteins in milliseconds instead of taking years of experimental work.

• There are two ways in which Alpha Fold demonstrates digital speed:

• It can fold proteins much faster than traditional methods

• The solution can be disseminated quickly and easily to biologists, drug discovery teams, and other researchers around the world.

• The speaker believes that we are at the dawn of a new era of "digital biology" where AI will be used to describe biological systems.

Isomorphic Labs

• The speaker mentions that they have spun out a new Alphabet company called Isomorphic Labs to further develop these technologies in the biochemistry space. They hope to reimagine drug discovery using AI and computational techniques.

Environment and Objective Function

In this section, the speaker discusses how to use a model to guide a search process according to some objective function. This approach can be applied to many problems, including drug discovery.

Using Models and Objective Functions

• The environment and context of a problem can be learned from data or simulated data.

• A model is used to guide a search process according to an objective function.

• This approach can be applied to many problems, such as biological or chemical problems.

• An example of this approach being used is in drug discovery.

Drug Discovery

In this section, the speaker explains how the general algorithmic solution discussed earlier can be applied to drug discovery.

Searching Through Chemistry Space

• Chemical compounds can be represented as nodes in a tree structure.

• A generative process is used to explore through chemistry space with an underlying model of biochemistry.

• The objective function for drug discovery includes properties such as no side effects or solubility.

• If successful, this approach would revolutionize the drug discovery process.

Recent Advances in AI Science

In this section, the speaker highlights recent advances in AI science across various domains.

Examples of Recent Advances

• Recent advances include quantum chemistry, pure mathematics, fusion reactors, and meteorology.

• Building a general algorithmic solution allows for its application across various domains.

• Collaboration with domain experts is crucial to identifying the right problem and answering it effectively.

• Applications of AI science include energy efficiency in data centers.

Text-to-Speech Systems and Large Models

In this section, the speaker discusses the impact of text-to-speech systems and large models on various applications.

Impact of Text-to-Speech Systems

• Modern text-to-speech systems are based on the speaker's work.

• These systems have a significant impact on applications such as YouTube videos and recommendation systems.

• They can reduce YouTube video size by 4% and improve recommendation systems across the board.

Large Models

• The speaker's team has done a lot of work in the area of large models.

• Chinchilla is one of their famous results that looks at scaling laws of large language models.

• Alpha code is their system that can program at median human programmer level.

• Flamingo is their system to describe what's happening in visual images.

• Gatto is one of their most general agents out there so far. It can do pretty much all things like controlling robot arms with the same system.

Innovations Needed for AGI

• The speaker believes that scale is incredibly important for large pre-trained models but they are not sufficient in themselves for AGI.

• Innovations are needed in areas like grounding, factuality, planning, missing memory, reasoning, etc.

Sparrow: A Dialogue Agent

In this section, the speaker talks about Sparrow - an AI dialogue agent designed to be helpful more often than not.

About Sparrow

• Sparrow is an AI dialogue agent designed to answer questions with responses that are useful and evidence-based.

• It uses search and retrieval to partly fact-check and ground its responses.

• It aims to be more accurate and grounded than other systems out there, with less than 1% errors even under adversarial testing.

Trade-Offs

• There is a trade-off between accuracy and safety versus fun and engagement.

• The speaker's team is working on finding the best of both worlds for Sparrow.

Ethics and Responsibility in AI

In this section, the speaker discusses the importance of deploying AI responsibly and safely for the benefit of everyone. He emphasizes that AI is a powerful technology that should be treated with respect and caution.

The Importance of Ethics in AI

• The speaker highlights that advancing science is important but it should be done responsibly.

• DeepMind has had an Ethics Charter from the very start, which is now part of Google's AI principles.

• The speaker mentions that they have been discussing these things for many years and continue to provide leadership on topics such as AI strategy, coordinating safety work, risks, ethics, and engaging with the wider community.

• It's important to ensure that AI is used for the benefit of everyone.

Moving Forward Responsibly

• The speaker believes we are approaching a critical moment in human history with AGI (Artificial General Intelligence).

• He advises against moving fast and breaking things when it comes to something as powerful as AGI.

• Instead, he advocates for using the scientific method to gain a better understanding of systems before deploying them at scale.

• We need to be thoughtful about empirical testing in the wild and use rigorous analysis techniques to understand potential unintended consequences.

• As we approach AGI, we need to treat it with respect and precautionary principles.

Conclusion

• The speaker concludes by emphasizing that technology like AGI demands respect and caution. We need to deploy it responsibly for the benefit of everyone.

The Intersection of AI and Science

In this section, Demis Hassabis talks about his passion for science and how he believes that AI can help us answer some of the deepest questions we have about the human condition.

Exciting Scientific Questions

• Demis Hassabis is excited about using AI to revolutionize human health and disease, as well as sustainability.

• As a scientist, he is fascinated by the brain and how it works. He believes that attempting to build AI in a way that mimics the human mind will ultimately give us answers to some of the deepest questions we have about consciousness, creativity, dreaming, and more.

• He is also interested in physics and understanding the nature of reality.

Using AI to Understand Neuroscience

• Demis Hassabis believes that AI can be used as a tool to help us understand neuroscience better. By building a general learning system that mimics the human mind, we can compare it to our own brains and gain insights into how they work.

• He is most excited about finding out these deep questions ultimately related to the nature of reality with our AIs.

Personal Interests

• Demis Hassabis has always been interested in physics. His favorite subject at school was physics, and he enjoys reading books by scientists like Stephen Weinberg and Richard Feynman.

Physics, Neuroscience, and AI

In this section, the speaker discusses the potential for AI to generate a new set of Newtonian laws for physics and neuroscience. They also discuss the role of humans in scientific discovery going forward.

AI's Potential for Scientific Discovery

• The speaker believes that AI could potentially generate a new set of Newtonian laws for physics and neuroscience.

• They believe that AI systems will be able to do a lot of the drudgery work in science such as pattern matching and searching literature.

• However, they note that coming up with the right questions to ask is still the hardest thing in science and that this is where human scientists will still play an important role.

AlphaFold and Turing Machines

• The speaker sees themselves as a champion of Turing machines and classic computers, believing that they can do more than we previously thought.

• They discuss how AlphaFold has effectively modeled quantum systems in a tractable way using classical computing methods.

• The speaker is thinking about what this means for P = NP and other complex questions in computer science.

Human Scientists' Role Going Forward

• While current AI systems cannot decide what experiments to conduct or generate hypotheses, the speaker believes that they will become more advanced over time. In the meantime, they see these systems as incredibly useful tools for human scientists.

Safety and Norms in AI

In this section, the speaker discusses the importance of safety and norms in AI during a risky moment. They suggest that there needs to be more rigorous red teaming of these systems, better analysis tools, advanced interpretability, and more focus on safety and analysis.

Importance of Safety Features

• The speaker suggests that rules adherence is an important safety feature for AI systems.

• There can be unintended consequences of implementing certain rules, so it's important to carefully consider what rules should be put in place.

• Data creation and filtering inappropriate inputs are also crucial for ensuring safe AI systems.

Partnerships for Investigating AI Systems

• The speaker suggests that partnerships with external collaborators could help investigate emerging AI systems before they're released to the public.

• It's difficult to test these systems manually, so better automated testing tools are needed.

Society's Role in the Future of AI

In this section, the speaker discusses some of the societal issues surrounding AI. They suggest that society needs to debate and discuss values, rules, and deployment strategies for AI. Additionally, they mention how political biases can impact accuracy in answering questions.

Societal Issues Surrounding AI

• Society needs to debate values and rules surrounding the use of AI.

• Political biases can impact accuracy when answering questions about controversial topics.

• These debates go beyond just technical considerations but will have a significant impact on how we deploy and use AI going forward.

Starting with a Student Question

In this section, a PhD student asks about the future of Robotics and what DeepMind is doing to push towards that future.

Future of Robotics

• The speaker believes that Robotics will be an important application of general AI systems.

• However, he is not sure if it's the fastest path to get to AGI by insisting on embodiment.

• Robotics is seen as an important industrial application area with huge implications and impact.

• It's also used in fundamental research as a grounding problem and as a challenge task for low data regimes and transfer learning.

Technical Monoculture

• The speaker agrees that there is a technical monoculture where increasingly large resources are devoted to an increasingly small number of future directions.

• DeepMind has historically been multi-disciplinary, but they have had to slightly change their emphasis to put more emphasis on large models and pre-trained models due to compute resources.

• The speaker thinks that exploring large models is necessary, but more than half the organization is still exploring other innovations that might be needed.

Current Techniques in AI Research

The speaker discusses the current techniques used in AI research.

Approaching the Right Question

• The ability to ask the right question is crucial in research.

• DeepMind's approach involves honing this ability and training oneself to be a broad scientist.

• Picking a problem that is not too easy or too hard, but rather in the "sweet spot" of the S curve, is important.

• Surrounding oneself with smart and multi-disciplinary people can help brainstorm ideas for tackling a problem.

AI for Scientific Research

The speaker discusses how AI can be used for scientific research.

Foundational Model for Scientific Research

• DeepMind's work points towards the need for a foundational model for scientific research.

• Their group is exploring building an AI-based model to assist with literature search and more.

Encouraging Research Direction

In this section, the speaker encourages researchers to continue working on their projects.

Encouraging Research Direction

• The speaker encourages researchers to continue working on their projects.

Knowing When to Double Down or Give Up

In this section, a PhD student asks how to know when it's time to double down or give up on a project.

Knowing When to Double Down or Give Up

• A PhD student asks how to know when it's time to double down or give up on a project.

• The speaker suggests getting a different set of people in the room to look at the problem and assess new ideas.

• The quality and difficulty of implementing new ideas is important in assessing whether to double down or give up.

• The speaker recommends looking for orthogonal signs that you might be on the right track and multidisciplinary approaches can help with this.

Overcoming Challenges in Research

In this section, the speaker discusses overcoming challenges in research.

Overcoming Challenges in Research

• Researchers often face difficult problems but should not panic and should look for orthogonal signs that they are on the right track.

• Multidisciplinary approaches can help researchers find new solutions.

• The speaker gained confidence in continuing research by seeing amateur gamers solve protein folding puzzles using intuition and pattern matching skills similar to those used in AlphaGo.

Bringing in Marginalized Communities into AI Research

In this section, the speaker discusses how technologists and scientists can bring in the perspectives and ideas of people from marginalized communities into AI research.

Importance of Multi-Disciplinary Perspectives

• The speaker emphasizes the importance of bringing in many perspectives into AI research, including those of ethicists, social scientists, and philosophers.

• DeepMind works with external experts to ensure that they are paying attention to diverse perspectives when designing and deploying systems.

• The community as a whole could do better at incorporating diverse perspectives.

Democratizing Resources for AI Research

• Stanford and ATI have been lobbying for a bill called National AI Research Resource to democratize resources for AI research.

• One key aspect of this effort is reaching out to traditionally underrepresented communities.

Core Founding Philosophies at DeepMind

In this section, the speaker discusses some of the core founding philosophies that enabled DeepMind to continuously innovate at the frontier for as long as it has.

Key Techniques

• DeepMind bet on deep learning early on instead of expert systems like other companies were doing at the time.

• They also understood that compute would become a huge factor in their work and backed GPUs early on.

• Using simulations and games as a test platform rather than robotics allowed them to make faster progress.

Geographical Diversity

• Being based in Europe allowed DeepMind to tap into untapped talent there.

The Importance of Not Moving to Silicon Valley

In this section, the speaker talks about how moving to Silicon Valley was recommended to them but they resisted and why they are pleased with that decision.

Resisting the Move to Silicon Valley

• The speaker was advised by early funders and others to move to Silicon Valley.

• However, the speaker resisted this advice for various reasons.

• The speaker is pleased with their decision not to move as it has given them a unique perspective.