Trying fabric

extract-wisdom.md

fabric -y "https://www.youtube.com/watch?v=JTU8Ha4Jyfc" --stream --pattern extract_wisdom SUMMARY: François Chollet discusses intelligence, the limitations of large language models, and his work on measuring intelligence with the Abstraction and Reasoning Corpus (ARC) in an interview with Tim.

IDEAS:

• Intelligence is the ability to handle novelty and come up with models on the fly.
• Large language models fail at solving problems significantly different from their training data.
• The Abstraction Reasoning Corpus (ARC) is designed to be resistant to memorization.
• Introspection is effective for understanding how the mind handles system 2 thinking.
• Scale is not all you need in AI; performance increase is orthogonal to intelligence.
• Language models are interpolative databases that primarily memorize functions and programs.
• The kaleidoscope hypothesis: the world is made from repetition and composition of few atoms of meaning.
• Intelligence mines experience to identify repeated bits and extract unique atoms of meaning called abstractions.
• Synthesis combines building blocks to form a program matching the current situation.
• Abstraction generation looks at available information and distills it into reusable abstractions stored in memory.
• Working on automated theorem proving with deep learning was a catalyst for Chollet's ideas.
• Deep learning models are fundamentally limited; they are recognition engines, not system 2 thinkers.
• Spurious correlations are always available to explain something, no matter what you're looking at.
• Curves are a bad substrate to represent discrete computation; it's difficult to fit generalizable programs.
• Fitting a parametric curve via gradient descent is good for value-centric abstraction based on distance.
• Program-centric abstraction, based on exact graph matching, requires explicit step-by-step verification.
• Watching children shows learning is a series of feedback loops extracting skills from deliberate experiences.
• Children construct thoughts based on experiences, building the mind layer by layer from primitives.
• Different children construct similar models as they extract them via the same processes and experiences.
• Language models have near-zero intelligence; they fail to handle situations very different from training.
• The training process of language models is inefficient model building requiring dense sampling.
• The convex hull of a language model's training data may not capture all needed future novelty.
• Retraining language models on fresh data daily would still miss novel problems lacking online solutions.

INSIGHTS:

• Intelligence is separate from skill; systems can be skilled without being intelligent.
• Benchmarks measuring skill at memorization games do not measure intelligence.
• Interpolation and extrapolation within a dense training set do not equate to novel model building.
• Human intelligence relies on a large bank of abstractions that can be flexibly combined.
• Gradient descent is a weak, inefficient way to do program synthesis compared to human cognition.
• Children develop by setting goals based on what they already know, steadily building the mind's layers.
• Language models operate in the right direction as suggestion engines but lack human-level verification.
• The solution to building AGI must co-evolve with the benchmark defining the requirements for AGI.
• It may be as hard to specify a test for AGI as it is to build an AGI.
• Intelligence, agency, embodiment, and goal-setting are distinct and can be separated.
• AGI will be a powerful and valuable tool, but will not automatically make humans into gods.

QUOTES:

• "Intelligence is very specifically your ability to handle novelty, to deal with situations you've not seen before, and come up on the fly with models that make sense in the context of that situation."
• "If you're measuring something that's fundamentally driven by memory, then it makes sense that as you increase the amount of memory in the system, like the number of parameters, amount of train data, and compute is really just a proxy for that, you see a higher performance, because, you know, of course, if you can memorize more, you're going to do better at your memory game."
• "The kaleidoscope hypothesis is this idea that the world in general, and any domain in particular, follows the same structure, that it appears on the surface to be extremely rich and complex, and infinitely novel with every passing moment, but in reality, it is made from the repetition and composition of just a few atoms of meaning."
• "Everything you know, everything you think about, is built upon lower level primitives, which are built upon lower level primitives, and so on, and ultimately, it comes back to these extremely basic sensory-motor affordances that newborn children have."
• "Intelligence is a separate concept from skill, from behavior, that you can always be skilled at something without that necessarily being intelligent."

HABITS:

• Thinking deeply about the fundamental nature of intelligence and how to create it
• Developing clear mental models and architectures for intelligence and AI systems
• Critically examining the assumptions and limitations of current AI approaches like deep learning
• Observing child development closely to gain insights about learning and the mind
• Deconstructing skills and behaviors to identify the underlying components and process of intelligence
• Pursuing ideas and research directions even when they go against mainstream views in the field
• Engaging in public discourse and debate to share ideas and invite critical feedback
• Distilling complex topics into clear explanations using analogies and examples
• Updating beliefs and models based on new evidence and insights from experiments
• Collaborating with other researchers to push the boundaries of the field

FACTS:

• The current state-of-the-art performance on the Abstraction and Reasoning Corpus is 46%.
• Assembling all entries from the 2020 ARC competition yields a performance of at least 49%.
• Brute force program search with infinite compute could solve ARC using a DSL with 200 primitives.
• The best large language model, GPT-3.5, achieves only 21% on ARC using direct prompting.
• Babies in the womb are asleep 95% of the time, alternating between deep sleep and active sleep.
• Babies are sedated in the womb due to the low oxygen pressure environment.
• The placenta produces anesthetic products that keep babies in a dreamless sleep state.
• Children's consciousness peaks around age 9-10 and then starts gradually declining each year.
• The difference in consciousness between a 90-year-old, 10-year-old, and 3-year-old is very minor.
• Stories about the end times and our role in them are recurring memes throughout human history.

REFERENCES:

• Chollet's old blog posts and first edition of Deep Learning with Python book
• Automated theorem proving using deep learning research with Christian Szegedy
• Neural Turing Machines
• Elizabeth Spelke's work on core knowledge in child development
• Chollet's 2019 paper "On the Measure of Intelligence" introducing ARC
• The ARC challenge and current leaderboard on Kaggle
• Recent work on ARC from MosaicML (Jack Cole), Redwood Research (Ryan Greenblatt), MIT (Kevin Ellis)
• Dreamcoder paper from Josh Tenenbaum's group at MIT
• Chollet's Deep Learning with Python book chapter on model fine-tuning
• Shibani Santurkar's LLM modulo architecture
• Mark Solms' book The Hidden Spring on consciousness as prediction error
• Chollet's blog posts on intelligence as collective, situated and externalized
• The book The Language Game on language as a tool for thought
• Max Bennett's book on plasticity of semantic information sharing

ONE-SENTENCE TAKEAWAY: Intelligence is the efficient acquisition of skills to handle novel situations by flexibly combining abstractions.

RECOMMENDATIONS:

• Study child development to gain insights into the mechanisms of learning and intelligence.
• Build AI systems that combine large knowledge bases with flexible program synthesis and verification.
• Develop benchmarks for intelligence that focus on skill acquisition efficiency, not just task performance.
• Separate the notions of intelligence, skill, agency, embodiment, and goal-setting when analyzing AI systems.
• Be cautious of overinterpreting the capabilities of large language models; validate their outputs.
• Pursue research directions that may go against mainstream views if evidence and arguments are compelling.
• Engage in public discourse to share ideas, invite feedback, and build collective knowledge.
• Apply insights from cognitive science and psychology to inform AI architectures and training approaches.
• Measure consciousness in AI systems by assessing self-awareness, not just intelligent behaviors.
• Maintain a scientific mindset focused on understanding and engineering, rather than mythology and ideology.%