Breandan Considine–AI Timelines, Coding AI, Neuro Symbolic AI

Video quality	The size	Download

Information Breandan Considine–AI Timelines, Coding AI, Neuro Symbolic AI

Title	:	Breandan Considine–AI Timelines, Coding AI, Neuro Symbolic AI
Lasting	:	1.44.14
Date of publication	:
Views	:	1,1 rb

OUTLINEbrbr00:00 Introductionbr01:16 Do We Need Symbolic Reasoning to Get To AGI?br05:41 Merging Symbolic Reasoning & Deep Learning for Powerful AI Systemsbr10:57 Blending Symbolic Reasoning & Machine Learning Elegantlybr15:15 Enhancing Abstractions & Safety in Machine Learningbr21:28 AlphaTensor's Applicability May Be Overstatedbr24:31 AI Safety, Alignment & Encoding Human Values in Codebr29:56 Code Research: Moral, Information & Software Aspectsbr34:17 Automating Programming & Self-Improving AIbr36:25 Debunking AI "Monsters" & World Domination Complexitiesbr43:22 Neural Networks: Limits, Scaling Laws & Computation Challengesbr59:54 Real-world Software Development vs Competitive Programmingbr1:02:59 Measuring Programmer Productivity & Evaluating AI-generated Codebr1:06:09 Unintended Consequences, Reward Misspecification & AI-Human Symbiosisbr1:16:59 AI's Superior Intelligence: Impact, Self-Improvement & Turing Test Predictionsbr1:23:52 AI Scaling, Optimization Trade-offs & Economic Viabilitybr1:29:02 Metrics, Misspecifications & AI's Rich Task Diversitybr1:30:48 Federated Learning & AI Agent Speed Comparisonsbr1:32:56 AI Timelines, Regulation & Self-Regulating Systems
Comment from : The Inside View

going out of sync by 31:30, not sure what's going on, back in sync around 53:52 and no, it doesn't work
Comment from : Caleb Morgan

i love that he is wearing a helmet the whole time
Comment from : Comedians guide to true crime

Why is your guest wearing a helmet?
Comment from : Thomas Hunter

It was great to listen to this interview with Breandan and he brings some new interesting perspectives on this topic I especially added some reading materials for my own knowledge on Differentiable programming, SAT solver for boolean satisfiability and symbolic reasoning I did write down some thoughts on Breandan's points:brbrBreandan speaks about code as a general tool for assigning values I understand where he is coming from, especially from a theoretical computer science perspective However, both computer code and legal code as two examples have engineering applicable issues brbr1) Computer code: I'll use the concept of formal verification as a generalization of the issue with the meta argument Formal verification is a great idea It's even gotten a bit of use in the industry, though on toy projects or in limited security settings However, vast majority of code does not use Formal verification And even systems built with Formal verification run in such a way that they are dependent on other non formal systems Deep learning is very new to formal methods and I can't think of real shipped projects that use such things The broader meta problem is that even in such situations, we have a really really hard time with the formal specification step, much less scaling issues etcbrbr2)Law was tangentially referenced as something for assigning values However, legal systems and law practice require judges, lawyers, juries, etc We are not close to formally specifying legal code anywhere I wish we could apply the same rigor to engineered systems like the legal code or computer programs as we do in set theory via Zermelo–Fraenkel I claim we do not in almost all human engineered systemsbrbrBreandan discusses that it takes a large amount of computation to make single cell organism I think that is an interesting point, but is the optimization algorithm reality used as effective as gradient descent in our current systems? Given what we can produce in 1 year with GPT models, with potential hardware overhangs via purely throwing compute at transformers, I think "intelligent design" of systems will result to be a better optimizerbrbr44:00 brAmdahl's law / bigger graphs just make for parallelization constraintsbrSure, that is true assuming you have the best algorithm But do we in every case? Let's say we're calculating fibbonaci(100,000,000,000) You use a simple recursion with memoization algorithm I have 1000 humans and spend a bit of time to use some linear algebra and get a second order difference equation I have a better algorithm, I win Especially if it is a zero sum gamebrbrParallelizing tasks is a function of intelligence, but more generally intelligence servers predict/approximate the future state of the world Better algorithms apply generally and specifically in ways that are beyond paralyzationbrbrAGI can also have external direct memory access to data, greatly outperforming humans/brbrre: chess, sure there are limits to parallelization / Amdahl's law again However, imagine the goal is to win a chess tournament Having 150 more elo gets us from 50-50 to 70-30 in a single game, or 96 chance of winning a b07 50 elo gives a 96 chance of winning at bo31brbrLet's imagine another example where diminishing returns for parallelism/speed yield great results If we both run HFT hedge funds and you have 250ms latency to the exchange and I have 5ms latency to the exchange, you will soon run out of money Amdahl's law is a limit, but being able to spin 1,000 copies of yourself to generate different ideas and another 10,000 copies to check the best idea still yields incremental improvement, and potentially breaks Amdahl's law if new paradigms / algorithms are created A 10x engineer that is able to make substantial algorithm improvements or invent new ways of doing things can make a 100x contributionbrbr1:21:00brGenerally I think if we sample the space of unimaginable things in any random fashion, the probability of humanity existing in them is infinitesimalbr1:22:00brTuring test: While a very discriminating tester can rely on sidechannel attacks (ie, timing) or exploit the context window or other limits of current models, the GPT models are getting very close to passing the turing test already It seems with 32k context window with long backstory prompts the rate will go up By 2025, I am not sure Breandan will win his bet Additionally, what level of type II errors do we allow? The turing tests that fail GPT now probably fail a large fraction of humans toobrbr1:24:00brWhile it does seem that a lot of individual curves are s curves, the reality of tech improvements is many many s curves stacked on top of each-other that approximate an exponential-esque curve Ie, that's how moore's law works in practicebrbrMichael makes some good points on speed improvement causing hard takeoff For example, Breandan makes a point about fair cake cutting algorithms and self regulating systems I cut the cake, you choose it Here's two ways to break it randomly chosen from the space of unimaginable things: 1) I secretly have a cake factoring making 10e15 cakes for me 2) I command my nanobots to influence your brain to choose the tiny cake slice, and report back to everyone that we split the cake fairly That's hard takeoffbrbrGreat conversations!
Comment from : Alexey Malafeev

Filmed at MILA for the Montreal AI Symposium, Sep 2022
Comment from : Neuromancer420

DETAILED OUTLINEbrbr00:00 Introductionbr01:16 Applying Machine Learning to SAT Solversbr05:41 Using Symbolic Reasoning To Get Powerful AI Systems Is An Open Problembr06:40 Deep Learning Is Limited Because Of Computational Complexitybr09:22 Representational Capacity Outstrips Machine Learning Model Abilitybr09:52 Combining Symbolic Reasoning and Machine Learning for Safety and Constraintsbr10:57 Symbolic Reasoning And Machine Learning Can Be Blended Elegantlybr12:16 Automatic Differentiation: A Bridge Between Symbolic and ML Domainsbr13:13 Symbolic Abstraction and Neural Networks: A New Layer of Safetybr15:15 Using Domain-Specific Languages for Program Abstractions in Machine Learningbr17:56 Probabilistic Programming Languages for Inference and Verification in Neural Networksbr20:08 Scalability of Type-Safe Programming Languages in Neural Networksbr21:28 AlphaTensor's Applicability May Be Overstatedbr24:31 AI Safety and Alignment: Balancing Human Values and AI Efficiencybr26:48 What AI Alignment Failure Might Actually Look Likebr28:05 The Importance of Encoding Human Values in Codebr29:05 Various Programming Models Serve As Ways To Transfer Values To Computersbr29:56 Code Research Encompasses Moral, Information, And Software Aspectsbr31:13 Building Legal Code, DAOs, Autonomous Organizationsbr34:17 Automating Programming, Self-Improving AIbr36:25 "Monsters In Code" Might Not Be Possiblebr38:32 Taking Over The World Might Be Too Complex For A Single Agentbr39:24 Organizations Are More Efficient Than Superhuman Systemsbr40:16 Intelligence May Not Ensure Survivalbr43:22 Large Neural Networks May Not Necessarily Lead To Self-Destructive Behaviorbr45:16 Larger Graphs Won't Solve All The Harder Problems Of Computer Sciencebr47:47 Parallel Computation Might Help A Lot For Human-Brain Tasksbr50:46 Diminishing Returns For Parallelism: Amdahl's Lawbr52:36 Algorithm Design Faces Bottlenecks Due To Sequential Tasksbr53:44 Computational Irreducibility and Scaling Laws Hypothesisbr55:54 Limits of Neural Networks and Universal Function Approximatorsbr57:28 Surprising Success of AlphaCode and Codexbr59:54 Differences Between Competitive Programming and Real-world Software Developmentbr1:00:42 Developers At The Cutting Edge Are Already Becoming More Productivebr1:01:28 Programmers Are Not Islandsbr1:02:59 Challenges in Measuring Programmer Productivity and in Evaluating AI-generated Codebr1:04:46 AI Progress Accelerating Workflow in Various Domains, Cooking Robotsbr1:06:09 Unintended Consequences and Reward Misspecificationbr1:07:13 Using Symbolic Reasoning To Help With Specification Designbr1:09:10 Not Treating Computers As Servantsbr1:10:49 Using AI Tools Will Involve Some Level Of Interactionbr1:12:29 Symbiosis And Interactions Between AI and Human Systemsbr1:14:07 AI-Human Collaboration and Convergent Goalsbr1:16:59 AI's Vastly Superior Intelligence and The Limitations of Human Intelligencebr1:18:00 People's Integration of AI Technology With Moral and Ethical Understanding Drives Its Usagebr1:16:59 AI Technology Can Become Unbounded and More Powerful Than Humansbr1:20:13 Self-Improvement in AI is Possible but Its Impact is Uncertainbr1:21:50 Passing the Turing Test by 2025 Seems Unlikelybr1:23:12 Human-Level AGI Could Be Reached by 2030br1:23:52 All Exponential Curves Taper, and AI Scaling Will Reach an Inflection Pointbr1:24:59 No Free Lunch Principle Implies Trade-offs In Optimizationbr1:25:29 Economically Viable Tasks Are A Limited View Of AI Potentialbr1:27:55 Metrics Can Be Gamed And May Not Capture Rich Aspects Of Lifebr1:29:02 Focusing Too Much On Numerical Metrics Can Lead To Misspecificationsbr1:30:08 AI Will Be Used For A Rich Diversity Of Tasks, Not Just Economically Valuable Onesbr1:30:48 Federated Learning Scenario With Multiple AI Agentsbr1:31:48 How Fast Will The First AI Be Compared To Othersbr1:32:56 Extending AI Timelines By Slowing Down Sciencebr1:33:40 Regulating Compute Usage In AI Developmentbr1:34:19 Heavy-Handed Regulation Might Lead To Unintended Consequencesbr1:35:24 On The Physical Constraints Of An AI Singularitybr1:36:22 Can Nuclear Regulations Inform AI Regulationsbr1:39:20 AI Development Requires Clear Indicators of Resource Usagebr1:40:06 Would A Laissez-Faire AI Economy Lead To A Takeoverbr1:41:29 Self-Regulating Systems Might Be More Robustbr1:42:42 People Will Develop Reasonable AI Solutions Once They're Comfortable With It
Comment from : The Inside View

Transcript: theinsideviewai/breandan
Comment from : The Inside View

Guess which part of the video is generated by AI thread
Comment from : The Inside View