Machine Learning Street Talk (MLST)

DeepMind Research Scientist / MIT scholar Dr. Timothy Nguyen discusses his recent paper on understanding transformers through n-gram statistics. Nguyen explains his approach to analyzing transformer behavior using a kind of "template matching" (N-grams), providing insights into how these models process and predict language.

MLST is sponsored by Brave:

The Brave Search API covers over 20 billion webpages, built from scratch without Big Tech biases or the recent extortionate price hikes on search API access. Perfect for AI model training and retrieval augmentated generation. Try it now - get 2,000 free queries monthly at http://brave.com/api.

Key points covered include:

A method for describing transformer predictions using n-gram statistics without relying on internal mechanisms.

The discovery of a technique to detect overfitting in large language models without using holdout sets.

Observations on curriculum learning, showing how transformers progress from simpler to more complex rules during training.

Discussion of distance measures used in the analysis, particularly the variational distance.

Exploration of model sizes, training dynamics, and their impact on the results.

We also touch on philosophical aspects of describing versus explaining AI behavior, and the challenges in understanding the abstractions formed by neural networks. Nguyen concludes by discussing potential future research directions, including attempts to convert descriptions of transformer behavior into explanations of internal mechanisms.

Timothy Nguyen's earned his B.S. and Ph.D. in mathematics from Caltech and MIT, respectively. He held positions as Research Assistant Professor at the Simons Center for Geometry and Physics (2011-2014) and Visiting Assistant Professor at Michigan State University (2014-2017). During this time, his research expanded into high-energy physics, focusing on mathematical problems in quantum field theory. His work notably provided a simplified and corrected formulation of perturbative path integrals.

Since 2017, Nguyen has been working in industry, applying his expertise to machine learning. He is currently at DeepMind, where he contributes to both fundamental research and practical applications of deep learning to solve real-world problems.

Refs:

The Cartesian Cafe

https://www.youtube.com/@TimothyNguyen

Understanding Transformers via N-Gram Statistics

https://www.researchgate.net/publication/382204056_Understanding_Transformers_via_N-Gram_Statistics

TOC

00:00:00 Timothy Nguyen's background

00:02:50 Paper overview: transformers and n-gram statistics

00:04:55 Template matching and hash table approach

00:08:55 Comparing templates to transformer predictions

00:12:01 Describing vs explaining transformer behavior

00:15:36 Detecting overfitting without holdout sets

00:22:47 Curriculum learning in training

00:26:32 Distance measures in analysis

00:28:58 Model sizes and training dynamics

00:30:39 Future research directions

00:32:06 Conclusion and future topics

Jay Alammar, renowned AI educator and researcher at Cohere, discusses the latest developments in large language models (LLMs) and their applications in industry. Jay shares his expertise on retrieval augmented generation (RAG), semantic search, and the future of AI architectures.

MLST is sponsored by Brave:

The Brave Search API covers over 20 billion webpages, built from scratch without Big Tech biases or the recent extortionate price hikes on search API access. Perfect for AI model training and retrieval augmentated generation. Try it now - get 2,000 free queries monthly at http://brave.com/api.

Cohere Command R model series: https://cohere.com/command

Jay Alamaar:

https://x.com/jayalammar

Buy Jay's new book here!

Hands-On Large Language Models: Language Understanding and Generation

https://amzn.to/4fzOUgh

TOC:

00:00:00 Introduction to Jay Alammar and AI Education

00:01:47 Cohere's Approach to RAG and AI Re-ranking

00:07:15 Implementing AI in Enterprise: Challenges and Solutions

00:09:26 Jay's Role at Cohere and the Importance of Learning in Public

00:15:16 The Evolution of AI in Industry: From Deep Learning to LLMs

00:26:12 Expert Advice for Newcomers in Machine Learning

00:32:39 The Power of Semantic Search and Embeddings in AI Systems

00:37:59 Jay Alammar's Journey as an AI Educator and Visualizer

00:43:36 Visual Learning in AI: Making Complex Concepts Accessible

00:47:38 Strategies for Keeping Up with Rapid AI Advancements

00:49:12 The Future of Transformer Models and AI Architectures

00:51:40 Evolution of the Transformer: From 2017 to Present

00:54:19 Preview of Jay's Upcoming Book on Large Language Models

Disclaimer: This is the fourth video from our Cohere partnership. We were not told what to say in the interview, and didn't edit anything out from the interview. Note also that this combines several previously unpublished interviews from Jay into one, the earlier one at Tim's house was shot in Aug 2023, and the more recent one in Toronto in May 2024.

Refs:

The Illustrated Transformer

https://jalammar.github.io/illustrated-transformer/

Attention Is All You Need

https://arxiv.org/abs/1706.03762

The Unreasonable Effectiveness of Recurrent Neural Networks

http://karpathy.github.io/2015/05/21/rnn-effectiveness/

Neural Networks in 11 Lines of Code

https://iamtrask.github.io/2015/07/12/basic-python-network/

Understanding LSTM Networks (Chris Olah's blog post)

http://colah.github.io/posts/2015-08-Understanding-LSTMs/

Luis Serrano's YouTube Channel

https://www.youtube.com/channel/UCgBncpylJ1kiVaPyP-PZauQ

Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

https://arxiv.org/abs/1908.10084

GPT (Generative Pre-trained Transformer) models

https://jalammar.github.io/illustrated-gpt2/

https://openai.com/research/gpt-4

BERT (Bidirectional Encoder Representations from Transformers)

https://jalammar.github.io/illustrated-bert/

https://arxiv.org/abs/1810.04805

RoPE (Rotary Positional Encoding)

https://arxiv.org/abs/2104.09864 (Linked paper discussing rotary embeddings)

Grouped Query Attention

https://arxiv.org/pdf/2305.13245

RLHF (Reinforcement Learning from Human Feedback)

https://openai.com/research/learning-from-human-preferences

https://arxiv.org/abs/1706.03741

DPO (Direct Preference Optimization)

https://arxiv.org/abs/2305.18290

Daniel Cahn, co-founder of Slingshot AI, on the potential of AI in therapy. Why is anxiety and depression affecting a large population? To what extent are these real categories? Why is the mental health getting worse? How often do you want an AI to agree with you? What are the ethics of persuasive AI? You will discover all in this conversation.

MLST is sponsored by Brave:

The Brave Search API covers over 20 billion webpages, built from scratch without Big Tech biases or the recent extortionate price hikes on search API access. Perfect for AI model training and retrieval augmentated generation. Try it now - get 2,000 free queries monthly at http://brave.com/api.

Daniel Cahn (who is also hiring ML engineers by the way!)

https://x.com/thecahnartist?lang=en

/ cahnd

https://thinkingmachinespodcast.com/

TOC:

00:00:00 Intro

00:01:56 Therapy effectiveness vs drugs and societal implications

00:04:02 Mental health categories: Iatrogenesis and social constructs

00:10:19 Psychiatric treatment models and cognitive perspectives

00:13:30 AI design and human-like interactions: Intentionality debates

00:20:04 AI in therapy: Ethics, anthropomorphism, and loneliness mitigation

00:28:13 Therapy efficacy: Neuroplasticity, suffering, and AI placebos

00:33:29 AI's impact on human agency and cognitive modeling

00:41:17 Social media's effects on brain structure and behavior

00:50:46 AI ethics: Altering values and free will considerations

01:00:00 Work value perception and personal identity formation

01:13:37 Free will, agency, and mutable personal identity in therapy

01:24:27 AI in healthcare: Challenges, ethics, and therapy improvements

01:53:25 AI development: Societal impacts and cultural implications

Full references on YT VD: https://www.youtube.com/watch?v=7hwX6OZyNC0 (and baked into mp3 metadata)

Prof. Subbarao Kambhampati argues that while LLMs are impressive and useful tools, especially for creative tasks, they have fundamental limitations in logical reasoning and cannot provide guarantees about the correctness of their outputs. He advocates for hybrid approaches that combine LLMs with external verification systems.

MLST is sponsored by Brave:

The Brave Search API covers over 20 billion webpages, built from scratch without Big Tech biases or the recent extortionate price hikes on search API access. Perfect for AI model training and retrieval augmentated generation. Try it now - get 2,000 free queries monthly at http://brave.com/api.

TOC (sorry the ones baked into the MP3 were wrong apropos due to LLM hallucination!)

[00:00:00] Intro

[00:02:06] Bio

[00:03:02] LLMs are n-gram models on steroids

[00:07:26] Is natural language a formal language?

[00:08:34] Natural language is formal?

[00:11:01] Do LLMs reason?

[00:19:13] Definition of reasoning

[00:31:40] Creativity in reasoning

[00:50:27] Chollet's ARC challenge

[01:01:31] Can we reason without verification?

[01:10:00] LLMs cant solve some tasks

[01:19:07] LLM Modulo framework

[01:29:26] Future trends of architecture

[01:34:48] Future research directions

Youtube version: https://www.youtube.com/watch?v=y1WnHpedi2A

Refs: (we didn't have space for URLs here, check YT video description instead)

How seriously should governments take the threat of existential risk from AI, given the lack of consensus among researchers? On the one hand, existential risks (x-risks) are necessarily somewhat speculative: by the time there is concrete evidence, it may be too late. On the other hand, governments must prioritize — after all, they don’t worry too much about x-risk from alien invasions.

MLST is sponsored by Brave:

The Brave Search API covers over 20 billion webpages, built from scratch without Big Tech biases or the recent extortionate price hikes on search API access. Perfect for AI model training and retrieval augmentated generation. Try it now - get 2,000 free queries monthly at brave.com/api.

Sayash Kapoor is a computer science Ph.D. candidate at Princeton University's Center for Information Technology Policy. His research focuses on the societal impact of AI. Kapoor has previously worked on AI in both industry and academia, with experience at Facebook, Columbia University, and EPFL Switzerland. He is a recipient of a best paper award at ACM FAccT and an impact recognition award at ACM CSCW. Notably, Kapoor was included in TIME's inaugural list of the 100 most influential people in AI.

Sayash Kapoor

https://x.com/sayashk

https://www.cs.princeton.edu/~sayashk/

Arvind Narayanan (other half of the AI Snake Oil duo)

https://x.com/random_walker

AI existential risk probabilities are too unreliable to inform policy

https://www.aisnakeoil.com/p/ai-existential-risk-probabilities

Pre-order AI Snake Oil Book

https://amzn.to/4fq2HGb

AI Snake Oil blog

https://www.aisnakeoil.com/

AI Agents That Matter

https://arxiv.org/abs/2407.01502

Shortcut learning in deep neural networks

https://www.semanticscholar.org/paper/Shortcut-learning-in-deep-neural-networks-Geirhos-Jacobsen/1b04936c2599e59b120f743fbb30df2eed3fd782

77% Of Employees Report AI Has Increased Workloads And Hampered Productivity, Study Finds

https://www.forbes.com/sites/bryanrobinson/2024/07/23/employees-report-ai-increased-workload/

TOC:

00:00:00 Intro

00:01:57 How seriously should we take Xrisk threat?

00:02:55 Risk too unrealiable to inform policy

00:10:20 Overinflated risks

00:12:05 Perils of utility maximisation

00:13:55 Scaling vs airplane speeds

00:17:31 Shift to smaller models?

00:19:08 Commercial LLM ecosystem

00:22:10 Synthetic data

00:24:09 Is AI complexifying our jobs?

00:25:50 Does ChatGPT make us dumber or smarter?

00:26:55 Are AI Agents overhyped?

00:28:12 Simple vs complex baselines

00:30:00 Cost tradeoff in agent design

00:32:30 Model eval vs downastream perf

00:36:49 Shortcuts in metrics

00:40:09 Standardisation of agent evals

00:41:21 Humans in the loop

00:43:54 Levels of agent generality

00:47:25 ARC challenge

Sara Hooker is VP of Research at Cohere and leader of Cohere for AI. We discuss her recent paper critiquing the use of compute thresholds, measured in FLOPs (floating point operations), as an AI governance strategy.

We explore why this approach, recently adopted in both US and EU AI policies, may be problematic and oversimplified. Sara explains the limitations of using raw computational power as a measure of AI capability or risk, and discusses the complex relationship between compute, data, and model architecture.

Equally important, we go into Sara's work on "The AI Language Gap." This research highlights the challenges and inequalities in developing AI systems that work across multiple languages. Sara discusses how current AI models, predominantly trained on English and a handful of high-resource languages, fail to serve the linguistic diversity of our global population. We explore the technical, ethical, and societal implications of this gap, and discuss potential solutions for creating more inclusive and representative AI systems.

We broadly discuss the relationship between language, culture, and AI capabilities, as well as the ethical considerations in AI development and deployment.

YT Version: https://youtu.be/dBZp47999Ko

TOC:

[00:00:00] Intro

[00:02:12] FLOPS paper

[00:26:42] Hardware lottery

[00:30:22] The Language gap

[00:33:25] Safety

[00:38:31] Emergent

[00:41:23] Creativity

[00:43:40] Long tail

[00:44:26] LLMs and society

[00:45:36] Model bias

[00:48:51] Language and capabilities

[00:52:27] Ethical frameworks and RLHF

Sara Hooker

https://www.sarahooker.me/

https://www.linkedin.com/in/sararosehooker/

https://scholar.google.com/citations?user=2xy6h3sAAAAJ&hl=en

https://x.com/sarahookr

Interviewer: Tim Scarfe

Refs

The AI Language gap

https://cohere.com/research/papers/the-AI-language-gap.pdf

On the Limitations of Compute Thresholds as a Governance Strategy.

https://arxiv.org/pdf/2407.05694v1

The Multilingual Alignment Prism: Aligning Global and Local Preferences to Reduce Harm

https://arxiv.org/pdf/2406.18682

Cohere Aya

https://cohere.com/research/aya

RLHF Can Speak Many Languages: Unlocking Multilingual Preference Optimization for LLMs

https://arxiv.org/pdf/2407.02552

Back to Basics: Revisiting REINFORCE Style Optimization for Learning from Human Feedback in LLMs

https://arxiv.org/pdf/2402.14740

Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence

https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/

EU AI Act

https://www.europarl.europa.eu/doceo/document/TA-9-2024-0138_EN.pdf

The bitter lesson

http://www.incompleteideas.net/IncIdeas/BitterLesson.html

Neel Nanda interview

https://www.youtube.com/watch?v=_Ygf0GnlwmY

Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet

https://transformer-circuits.pub/2024/scaling-monosemanticity/

Chollet's ARC challenge

https://github.com/fchollet/ARC-AGI

Ryan Greenblatt on ARC

https://www.youtube.com/watch?v=z9j3wB1RRGA

Disclaimer: This is the third video from our Cohere partnership. We were not told what to say in the interview, and didn't edit anything out from the interview.

Murray Shanahan is a professor of Cognitive Robotics at Imperial College London and a senior research scientist at DeepMind. He challenges our assumptions about AI consciousness and urges us to rethink how we talk about machine intelligence.

We explore the dangers of anthropomorphizing AI, the limitations of current language in describing AI capabilities, and the fascinating intersection of philosophy and artificial intelligence.

Show notes and full references: https://docs.google.com/document/d/1ICtBI574W-xGi8Z2ZtUNeKWiOiGZ_DRsp9EnyYAISws/edit?usp=sharing

Prof Murray Shanahan:

https://www.doc.ic.ac.uk/~mpsha/ (look at his selected publications)

https://scholar.google.co.uk/citations?user=00bnGpAAAAAJ&hl=en

https://en.wikipedia.org/wiki/Murray_Shanahan

https://x.com/mpshanahan

Interviewer: Dr. Tim Scarfe

Refs (links in the Google doc linked above):

Role play with large language models

Waluigi effect

"Conscious Exotica" - Paper by Murray Shanahan (2016)

"Simulators" - Article by Janis from LessWrong

"Embodiment and the Inner Life" - Book by Murray Shanahan (2010)

"The Technological Singularity" - Book by Murray Shanahan (2015)

"Simulacra as Conscious Exotica" - Paper by Murray Shanahan (newer paper of the original focussed on LLMs)

A recent paper by Anthropic on using autoencoders to find features in language models (referring to the "Scaling Monosemanticity" paper)

Work by Peter Godfrey-Smith on octopus consciousness

"Metaphors We Live By" - Book by George Lakoff (1980s)

Work by Aaron Sloman on the concept of "space of possible minds" (1984 article mentioned)

Wittgenstein's "Philosophical Investigations" (posthumously published)

Daniel Dennett's work on the "intentional stance"

Alan Turing's original paper on the Turing Test (1950)

Thomas Nagel's paper "What is it like to be a bat?" (1974)

John Searle's Chinese Room Argument (mentioned but not detailed)

Work by Richard Evans on tackling reasoning problems

Claude Shannon's quote on knowledge and control

"Are We Bodies or Souls?" - Book by Richard Swinburne

Reference to work by Ethan Perez and others at Anthropic on potential deceptive behavior in language models

Reference to a paper by Murray Shanahan and Antonia Creswell on the "selection inference framework"

Mention of work by Francois Chollet, particularly the ARC (Abstraction and Reasoning Corpus) challenge

Reference to Elizabeth Spelke's work on core knowledge in infants

Mention of Karl Friston's work on planning as inference (active inference)

The film "Ex Machina" - Murray Shanahan was the scientific advisor

"The Waluigi Effect"

Anthropic's constitutional AI approach

Loom system by Lara Reynolds and Kyle McDonald for visualizing conversation trees

DeepMind's AlphaGo (mentioned multiple times as an example)

Mention of the "Golden Gate Claude" experiment

Reference to an interview Tim Scarfe conducted with University of Toronto students about self-attention controllability theorem

Mention of an interview with Irina Rish

Reference to an interview Tim Scarfe conducted with Daniel Dennett

Reference to an interview with Maria Santa Caterina

Mention of an interview with Philip Goff

Nick Chater and Martin Christianson's book ("The Language Game: How Improvisation Created Language and Changed the World")

Peter Singer's work from 1975 on ascribing moral status to conscious beings

Demis Hassabis' discussion on the "ladder of creativity"

Reference to B.F. Skinner and behaviorism

In the coming decades, the technology that enables virtual and augmented reality will improve beyond recognition. Within a century, world-renowned philosopher David J. Chalmers predicts, we will have virtual worlds that are impossible to distinguish from non-virtual worlds. But is virtual reality just escapism?

In a highly original work of 'technophilosophy', Chalmers argues categorically, no: virtual reality is genuine reality. Virtual worlds are not second-class worlds. We can live a meaningful life in virtual reality - and increasingly, we will.

What is reality, anyway? How can we lead a good life? Is there a god? How do we know there's an external world - and how do we know we're not living in a computer simulation? In Reality+, Chalmers conducts a grand tour of philosophy, using cutting-edge technology to provide invigorating new answers to age-old questions.

David J. Chalmers is an Australian philosopher and cognitive scientist specializing in the areas of philosophy of mind and philosophy of language. He is Professor of Philosophy and Neural Science at New York University, as well as co-director of NYU's Center for Mind, Brain, and Consciousness. Chalmers is best known for his work on consciousness, including his formulation of the "hard problem of consciousness."

Reality+: Virtual Worlds and the Problems of Philosophy

https://amzn.to/3RYyGD2

https://consc.net/

https://x.com/davidchalmers42

00:00:00 Reality+ Intro

00:12:02 GPT conscious? 10/10

00:14:19 The consciousness processor thought experiment (11/10)

00:20:34 Intelligence and Consciousness entangled? 10/10

00:22:44 Karl Friston / Meta Problem 10/10

00:29:05 Knowledge argument / subjective experience (6/10)

00:32:34 Emergence 11/10 (best chapter)

00:42:45 Working with Douglas Hofstadter 10/10

00:46:14 Intelligence is analogy making? 10/10

00:50:47 Intelligence explosion 8/10

00:58:44 Hypercomputation 10/10

01:09:44 Who designed the designer? (7/10)

01:13:57 Experience machine (7/10)

Ryan Greenblatt from Redwood Research recently published "Getting 50% on ARC-AGI with GPT-4.0," where he used GPT4o to reach a state-of-the-art accuracy on Francois Chollet's ARC Challenge by generating many Python programs.

Sponsor:

Sign up to Kalshi here https://kalshi.onelink.me/1r91/mlst -- the first 500 traders who deposit $100 will get a free $20 credit! Important disclaimer - In case it's not obvious - this is basically gambling and a *high risk* activity - only trade what you can afford to lose.

We discuss:

- Ryan's unique approach to solving the ARC Challenge and achieving impressive results.

- The strengths and weaknesses of current AI models.

- How AI and humans differ in learning and reasoning.

- Combining various techniques to create smarter AI systems.

- The potential risks and future advancements in AI, including the idea of agentic AI.

https://x.com/RyanPGreenblatt

https://www.redwoodresearch.org/

Refs:

Getting 50% (SoTA) on ARC-AGI with GPT-4o [Ryan Greenblatt]

https://redwoodresearch.substack.com/p/getting-50-sota-on-arc-agi-with-gpt

On the Measure of Intelligence [Chollet]

https://arxiv.org/abs/1911.01547

Connectionism and Cognitive Architecture: A Critical Analysis [Jerry A. Fodor and Zenon W. Pylyshyn]

https://ruccs.rutgers.edu/images/personal-zenon-pylyshyn/proseminars/Proseminar13/ConnectionistArchitecture.pdf

Software 2.0 [Andrej Karpathy]

https://karpathy.medium.com/software-2-0-a64152b37c35

Why Greatness Cannot Be Planned: The Myth of the Objective [Kenneth Stanley]

https://amzn.to/3Wfy2E0

Biographical account of Terence Tao’s mathematical development. [M.A.(KEN) CLEMENTS]

https://gwern.net/doc/iq/high/smpy/1984-clements.pdf

Model Evaluation and Threat Research (METR)

https://metr.org/

Why Tool AIs Want to Be Agent AIs

https://gwern.net/tool-ai

Simulators - Janus

https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators

AI Control: Improving Safety Despite Intentional Subversion

https://www.lesswrong.com/posts/d9FJHawgkiMSPjagR/ai-control-improving-safety-despite-intentional-subversion

https://arxiv.org/abs/2312.06942

What a Compute-Centric Framework Says About Takeoff Speeds

https://www.openphilanthropy.org/research/what-a-compute-centric-framework-says-about-takeoff-speeds/

Global GDP over the long run

https://ourworldindata.org/grapher/global-gdp-over-the-long-run?yScale=log

Safety Cases: How to Justify the Safety of Advanced AI Systems

https://arxiv.org/abs/2403.10462

The Danger of a “Safety Case"

http://sunnyday.mit.edu/The-Danger-of-a-Safety-Case.pdf

The Future Of Work Looks Like A UPS Truck (~02:15:50)

https://www.npr.org/sections/money/2014/05/02/308640135/episode-536-the-future-of-work-looks-like-a-ups-truck

SWE-bench

https://www.swebench.com/

Using DeepSpeed and Megatron to Train Megatron-Turing NLG

530B, A Large-Scale Generative Language Model

https://arxiv.org/pdf/2201.11990

Algorithmic Progress in Language Models

https://epochai.org/blog/algorithmic-progress-in-language-models

Aidan Gomez, CEO of Cohere, reveals how they're tackling AI hallucinations and improving reasoning abilities. He also explains why Cohere doesn't use any output from GPT-4 for training their models.

Aidan shares his personal insights into the world of AI and LLMs and Cohere's unique approach to solving real-world business problems, and how their models are set apart from the competition. Aidan reveals how they are making major strides in AI technology, discussing everything from last mile customer engineering to the robustness of prompts and future architectures.

He also touches on the broader implications of AI for society, including potential risks and the role of regulation. He discusses Cohere's guiding principles and the health the of startup scene. With a particular focus on enterprise applications. Aidan provides a rare look into the internal workings of Cohere and their vision for driving productivity and innovation.

https://cohere.com/

https://x.com/aidangomez

Check out Cohere's amazing new Command R* models here

https://cohere.com/command

Disclaimer: This is the second video from our Cohere partnership. We were not told what to say in the interview, and didn't edit anything out from the interview.