maybe this means that intelligence is task-specific and different species have different performance constraints. We can compare intelligence across species but our confidence in the quest for a single benchmark for intelligence for multiple species is misplaced.
same conclusions for ConceptARC paper. There are different performance constraints on machines and humans (machines are faster, etc.). Machines may reason in ways different to humans. They may also use different concepts. Ultimately our hope that machines may reason in the same way that we do is also misplaced (cite Rich Sutton essay Bitter Lessons ...). We made the same mistake with chess: we thought that we will imbue machines with the same strategies that we use while playing chess. The winning techniques were approaches that used brute-force search with engineering-ey approaches.
this is especially releavnt since current approaches to ARC use a lot of effort and energy and computational resources. Thousands of dollars per task. We need to come up with a theory of intelligence and energetics that spans biological intelligence and artificial engineered inteligence (such as LLMs). it will also allow us to create a theory of how energetics and evolutionary constraints shape intelligence.
Some equations for this theory given below: optimal foraging theory of infomation (trainng cost, data acquisition, Inference cost, number of parameters, computational budget per task, number of tokens per task) .
treat intelligence less as a single score and more as a family of efficiency frontiers under task and resource constraints
this ties into scaling lecture in class
same conclusions for ConceptARC paper. There are different performance constraints on machines and humans (machines are faster, etc.). Machines may reason in ways different to humans. They may also use different concepts. Ultimately our hope that machines may reason in the same way that we do is also misplaced (cite Rich Sutton essay Bitter Lessons ...). We made the same mistake with chess: we thought that we will imbue machines with the same strategies that we use while playing chess. The winning techniques were approaches that used brute-force search with engineering-ey approaches.
this is especially releavnt since current approaches to ARC use a lot of effort and energy and computational resources. Thousands of dollars per task. We need to come up with a theory of intelligence and energetics that spans biological intelligence and artificial engineered inteligence (such as LLMs). it will also allow us to create a theory of how energetics and evolutionary constraints shape intelligence.
Some equations for this theory given below: optimal foraging theory of infomation (trainng cost, data acquisition, Inference cost, number of parameters, computational budget per task, number of tokens per task) .
treat intelligence less as a single score and more as a family of efficiency frontiers under task and resource constraints
🤔❓more derivations here Energetic Niche Theory of Intelligence
For biology, those costs are metabolic, attentional, developmental, and social. For AI, they are FLOPs, latency, tokens, memory bandwidth, training data, and dollar cost.
performance–budget frontier for substrate s.
Then “more intelligent” means not “higher on one benchmark,” but “better frontier shape” across the relevant task distribution.
practical and theory on all of the costs of LLMs
data, training, inference, curation, alignment, RLHF, guardrails, human costs
scaling papers with notebooks CAISH
scaling curves papers
[!NOTE] Learning Goals
By the end of this topic, you should be able to:
- Explain what scaling laws are and why they matter.
- Understand the relationship between model size, data, and compute.
- Describe the transition from the Kaplan scaling laws to the Chinchilla scaling laws.
- Discuss practical bottlenecks to continued scaling.
- Critically evaluate whether scaling alone is sufficient for achieving more general intelligence.
The modern era of AI has been shaped by a simple observation: larger models trained on more data with more compute tend to perform better in predictable ways. Understanding these relationships has become essential for understanding the development of Large Language Models (LLMs).
Scaling laws help us answer questions such as:
Why read it?
This short essay provides the intellectual backdrop for modern AI. Sutton argues that throughout AI history, methods that leverage increasing computation repeatedly outperform approaches built around handcrafted human knowledge.
đź”— https://www.incompleteideas.net/IncIdeas/BitterLesson.html
[!TIP] Discussion Question
Is the success of modern LLMs evidence that Sutton was right? What examples from AI history support or challenge his argument?
Key idea: Language model performance follows remarkably smooth power-law relationships with:
This paper established scaling laws as a central framework for understanding AI progress.
đź”— https://arxiv.org/abs/2001.08361
[!IMPORTANT] One of the most surprising findings is that performance improvements remain highly predictable across many orders of magnitude.
The GPT-3 paper demonstrated that simply scaling a language model can lead to qualitatively new capabilities such as:
đź”— https://arxiv.org/abs/2005.14165
[!NOTE] This paper sparked renewed interest in the possibility that new capabilities may emerge from scale alone.
This paper challenged prevailing assumptions about model scaling.
The central argument:
đź”— https://arxiv.org/abs/2203.15556
[!WARNING] Bigger models are not necessarily better models. Training strategy matters.
An important follow-up study that revisits and tests the Chinchilla conclusions using a much larger collection of models.
đź”— https://epoch.ai/publications/chinchilla-scaling-a-replication-attempt
[!TIP] Ask yourself:
- Are scaling laws universal?
- How robust are these results?
- What assumptions underlie these analyses?
[!NOTE] These resources are ideal for tutorials, lab sessions, and self-study.
A highly accessible introduction to scaling and capability development.
Features:
đź”— https://ai-safety-course.github.io/chapters/chapter-1/
An interactive simulator that allows you to explore:
đź”— https://epoch.ai/latest/introducing-the-distributed-training-interactive-simulator
[!TIP] Classroom Activity:
Have students estimate the infrastructure required to train a GPT-4-scale model and compare assumptions.
Investigates a major question for the future of scaling:
What happens when we run out of high-quality human-generated text?
[!IMPORTANT] Future AI progress may be constrained by data availability rather than compute.
Students can:
đź”— https://github.com/mmcmanus1/empirical-scaling-harness
[!TIP] Recommended mini-project:
Replicate a scaling-law curve using a subset of experiments and compare your fitted exponent with the published result.
A research-oriented notebook exploring scaling behaviour in reasoning tasks.
Students can:
đź”— https://github.com/WANGXinyiLinda/reasoning-scaling-law
An accessible explanation of the practical consequences of compute-optimal scaling.
Includes:
đź”— https://www.alignmentforum.org/posts/6Fpvch8RR29qLEWNH/chinchilla-s-wild-implications
A concise pedagogical introduction with exercises and review questions.
đź”— https://www.aisafetybook.com/textbook/scaling-laws
A practical guide to large-scale training systems.
Topics include:
đź”— https://jax-ml.github.io/scaling-book/
[!QUESTION]
- Why do power laws appear so frequently in modern AI?
- Is scaling discovering intelligence or merely exploiting statistics more efficiently?
- What might eventually limit further scaling?
- Does scaling alone lead to reasoning and abstraction?
- How do scaling laws relate to François Chollet’s arguments in ARC and the measurement of intelligence?
- Are there domains where scaling may fail entirely?
[!SUMMARY]
The central story of modern AI can be viewed as:
The Bitter Lesson → Scaling Laws → GPT-3 → Chinchilla → Data and Compute Limits
Understanding this progression provides a foundation for understanding both the successes and the future challenges of large language models.