
Lessons from Jared Kaplan
Jared Kaplan is the chief scientist and co-founder of Anthropic. Before working on AI he spent most of his career as a theoretical physicist; in his own words, "I actually have only been working on AI for about six years." — The Guardian: The biggest decision yet, Jared Kaplan on allowing AI to train itself (December 2025).
Part 1: The Physics Background
- On holography: Kaplan wrote "The basic idea of holography is that physics in d + 1 dimensions can be reproduced by a 'hologram at infinity', in d or fewer dimensions." — Jared Kaplan, Lectures on AdS/CFT from the Bottom Up (course notes; UCLA-hosted copy).
- On why he studied physics: Kaplan said "my mom was a science fiction writer and I wanted to figure out if we could build a faster than light drive and physics was the way to do that." — Y Combinator AI Startup School: Scaling and the Road to Human-Level AI, Jared Kaplan (June 2025)
- On taking ideas seriously: Kaplan wrote "It's always a good idea to take ideas very seriously, make them as concrete and specific as possible, and then see where you are led." — Jared Kaplan, Lectures on AdS/CFT from the Bottom Up (course notes; UCLA-hosted copy).
- On effective field theory: Kaplan wrote of "An incomplete theory of quantum gravity in AdS." — Jared Kaplan, Lectures on AdS/CFT from the Bottom Up (course notes; UCLA-hosted copy).
- On asking dumb questions: Kaplan said "As a physicist, that's what you're trained to do. You sort of look at the big picture and you ask really dumb things." — Y Combinator AI Startup School: Scaling and the Road to Human-Level AI, Jared Kaplan (June 2025).
- On the simplest question: Kaplan said "I was just sort of asking the dumbest possible question." — Y Combinator AI Startup School: Scaling and the Road to Human-Level AI, Jared Kaplan (June 2025).
- On precise trends: Kaplan said "We found that there's actually something very, very, very precise and surprising underlying AI training." — Y Combinator AI Startup School: Scaling and the Road to Human-Level AI, Jared Kaplan (June 2025).
Part 2: The Discovery of Scaling Laws
- On the core discovery: The abstract of the paper he led says "The loss scales as a power-law with model size, dataset size, and the amount of compute used for training, with some trends spanning more than seven orders of magnitude." — Kaplan et al., Scaling Laws for Neural Language Models (arXiv 2001.08361, 2020).
- On failing scaling laws: Kaplan said "my first inclination is to think if scaling laws are failing, it's because we've screwed up AI training in some way." — Y Combinator AI Startup School: Scaling and the Road to Human-Level AI, Jared Kaplan (June 2025)
- On scale and shape: The paper says "Performance depends strongly on scale, weakly on model shape." — Kaplan et al., Scaling Laws for Neural Language Models (arXiv 2001.08361, 2020).
- On crisp tasks: Kaplan said "it's easy right now and you see an explosion of progress for us to train AI models that can say write code that passes tests or that answer math questions correctly because it's very crisp what's correct and what's incorrect." — Y Combinator AI Startup School: Scaling and the Road to Human-Level AI, Jared Kaplan (June 2025)
- On architectural details: The abstract says "Other architectural details such as network width or depth have minimal effects within a wide range." — Kaplan et al., Scaling Laws for Neural Language Models (arXiv 2001.08361, 2020).
- On simple equations: The abstract says "Simple equations govern the dependence of overfitting on model/dataset size and the dependence of training speed on model size." — Kaplan et al., Scaling Laws for Neural Language Models (arXiv 2001.08361, 2020).
Part 3: Compute, Data, and Parameters
- On compute efficiency: The abstract says "Larger models are significantly more sample-efficient, such that optimally compute-efficient training involves training very large models on a relatively modest amount of data." — Kaplan et al., Scaling Laws for Neural Language Models (arXiv 2001.08361, 2020).
- On allocating compute: The abstract says "These relationships allow us to determine the optimal allocation of a fixed compute budget." — Kaplan et al., Scaling Laws for Neural Language Models (arXiv 2001.08361, 2020).
Part 5: Constitutional AI
- On the core mechanism: The Constitutional AI paper says "The only human oversight is provided through a list of rules or principles, and so we refer to the method as 'Constitutional AI'." — Bai et al., Constitutional AI: Harmlessness from AI Feedback (arXiv 2212.08073, 2022).
- On supervising AI with AI: The paper says "As AI systems become more capable, we would like to enlist their help to supervise other AIs." — Bai et al., Constitutional AI: Harmlessness from AI Feedback (arXiv 2212.08073, 2022).
- On self-critique: The paper says "In the supervised phase we sample from an initial model, then generate self-critiques and revisions, and then finetune the original model on revised responses." — Bai et al., Constitutional AI: Harmlessness from AI Feedback (arXiv 2212.08073, 2022).
- On RLAIF: The paper says "In the RL phase, we sample from the finetuned model, use a model to evaluate which of the two samples is better, and then train a preference model from this dataset of AI preferences." — Bai et al., Constitutional AI: Harmlessness from AI Feedback (arXiv 2212.08073, 2022).
- On fewer human labels: The paper says it trains a harmless AI assistant "without any human labels identifying harmful outputs." — Bai et al., Constitutional AI: Harmlessness from AI Feedback (arXiv 2212.08073, 2022).
Part 6: Predictability in AI Development
- On scaling laws in both phases: Kaplan said "it turns out that there are scaling laws for both of these phases of training." — Y Combinator AI Startup School: Scaling and the Road to Human-Level AI, Jared Kaplan (June 2025).
- On extrapolating trends: Kaplan said that when "you see something is true over many, many, many orders of magnitude" it is probably going to continue to hold for a long time further. — Y Combinator AI Startup School: Scaling and the Road to Human-Level AI, Jared Kaplan (June 2025).
- On what training is: Kaplan said "all there is to training these models is learning to predict the next word and then doing reinforcement learning to learn to do useful tasks." — Y Combinator AI Startup School: Scaling and the Road to Human-Level AI, Jared Kaplan (June 2025)
- On task length: Kaplan said that on the length of tasks AI models can do "it's doubling roughly every 7 months." — Y Combinator AI Startup School: Scaling and the Road to Human-Level AI, Jared Kaplan (June 2025)
- On turning the crank: Kaplan said "It's not that AI researchers are really smart or they suddenly got smart. It's that we found a very very simple way of making AI better systematically and and we're we're turning that crank." — Y Combinator AI Startup School: Scaling and the Road to Human-Level AI, Jared Kaplan (June 2025).
Part 7: Human-Level AI and Timelines
- On organizational knowledge: Kaplan said "One of the most important ingredients I think is relevant organizational knowledge." — Y Combinator AI Startup School: Scaling and the Road to Human-Level AI, Jared Kaplan (June 2025).
Part 8: AI Safety and The Ultimate Risk
- On letting AI train itself: The Guardian reported that "Humanity will have to decide by 2030 whether to take the "ultimate risk" of letting artificial intelligence systems train themselves to become more powerful." — The Guardian: The biggest decision yet, Jared Kaplan on allowing AI to train itself (December 2025).
- On letting go of control: Kaplan said that allowing AI to recursively self-improve "is in some ways the ultimate risk, because it's kind of like letting AI kind of go." — The Times of India: Anthropic chief scientist Jared Kaplan warns (December 2025), quoting his Guardian interview.
- On not knowing once no one is involved: Kaplan said "once no one’s involved in the process, you don’t really know." — Mint: Anthropic co-founder warns AI may design its own successor (December 2025), quoting his Guardian interview
- On a scary process: Kaplan said "It sounds like a kind of scary process. You don't know where you end up." — Mint: Anthropic co-founder warns AI may design its own successor (December 2025), quoting his Guardian interview.
- On the speed of improvement: Kaplan said "once no one's involved in the process, you don't really know." — Mint: Anthropic co-founder warns AI may design its own successor (December 2025), quoting his Guardian interview
- On losing control: Kaplan said "One is do you lose control over it? Do you even know what the AIs are doing?" — Mint: Anthropic co-founder warns AI may design its own successor (December 2025), quoting his Guardian interview.
- On misuse: Kaplan said "I think preventing power grabs, preventing misuse of the technology, is also very important." — Mint: Anthropic co-founder warns AI may design its own successor (December 2025), quoting his Guardian interview.
- On power grabs and misuse: Kaplan said "preventing power grabs, preventing misuse of the technology, is also very important." — Mint: Anthropic co-founder warns AI may design its own successor (December 2025), quoting his Guardian interview
- On the role of government: Kaplan said "We don't really want it to be a Sputnik-like situation where the government suddenly wakes up and is like: Oh, wow, AI is a big deal." — The Guardian: Scores of UK parliamentarians join call to regulate most powerful AI systems (December 2025).
- On a Sputnik-like situation: Kaplan said "We don’t really want it to be a Sputnik-like situation where the government suddenly wakes up and is like: Oh, wow, AI is a big deal." — The Guardian: Scores of UK parliamentarians join call to regulate most powerful AI systems (December 2025)