A 2017 memo, explained and checked against 2026
In 2017, Dario Amodei, then at OpenAI and now CEO of Anthropic, wrote a memo arguing that intelligence mostly comes from a large, loosely organized mass of computation, shaped by a rich environment and the right training signal. Clever algorithms, he argued, matter much less than they appear to. Part one explains the argument and lets you try its parts. Part two sets it beside the Transformer and The Bitter Lesson. Part three checks it against what happened next.
Essay §1 · Introduction · Full text
People who expect AGI soon and people who expect it late tend to split on one question: does progress come mostly from raw compute or from new algorithms? Amodei puts himself near the compute end, but he thinks the question is badly posed. What counts as an algorithm? How must the compute be applied? Where do things that are neither, like the design of training environments, fit? And is lots of compute one sufficient path among several, or somehow necessary?
The hypothesis, in brief
Most of what it takes to get intelligent behavior is a big, loosely organized pool of computation. The environment it interacts with and the training it goes through give that pool its shape.
The levers that matter are high-level: what the training signal is, what the environment is like, and how much experience the system gets. Once those are set, learning a task is mostly a matter of having enough compute and memory.
Some algorithmic choices do matter, but they fall into a few plain categories. They exploit simple symmetries in the world, keep the numbers well behaved, or give the computation a smoother shape. Elaborate ideas rarely help. Supposedly deep divides, such as Bayesian versus non-Bayesian methods, model-based versus model-free reinforcement learning, or even ML versus non-ML, mostly relabel the same underlying computation. Designs that dictate which part of the system should do what tend to do worst of all.
Amodei says this one belief drives both his AI timelines and his view of safety. The memo was planned in six sections: the hypothesis, the evidence, then three on safety, including why it made him skeptical of MIRI’s “highly reliable agent designs” agenda and of framework-style safety plans in general. He calls Paul Christiano’s framework broadly compatible with his view and Eliezer Yudkowsky’s broadly incompatible. The surviving scan ends after section 3, so this page covers the first half. You can read those sections in full.
Essay §2 · More detailed formulation · Full text
Say you want a system to learn a task: recognizing images, playing video games, doing science, proving theorems, imitating a person, or even learning new things quickly. The memo says to worry about seven things. Everything else usually has a minor effect, or works at too low a level to generalize.
Enough computation for the task, flowing freely between what the system observes and what it does. It should be loosely structured and responsive to training. Its exact form is secondary.
NowFrontier runs reach about 10²⁶–10²⁷ operations on clusters of 100,000+ chips.
Room to store what it learns: lasting knowledge about the task (in a neural net, the weights) and short-term notes about the case at hand (the activations). For evolution, the genome and one brain’s synapses. Too little storage and the task won’t be learned well.
NowHundreds of billions to trillions of weights; million-token context windows serve as short-term storage.
Enough time acting in the environment. He says experience rather than data on purpose: a fixed dataset is a special case of interacting with a world. Compute can partly stand in for experience if the system can simulate some of it internally.
NowTens of trillions of tokens of pretraining, then reinforcement learning in software sandboxes and other environments.
The training mix must make the easiest way to do well in training also work on the real task. Either training and test cases come from the same distribution, or training varies enough that the cheapest thing to learn is what the cases share, such as the laws of physics. Then even very different tasks get solved. Ramping up difficulty may also help.
NowMuch of the public internet, plus code, math, and filtered or synthetic data.
Free-flowing numbers easily blow up or vanish. Many advances of the previous five years that were billed as algorithms, such as Adam, batch normalization, and residual connections, are simple ways of keeping quantities near 1 or expressing them as changes from 1. This plumbing has often beaten much fancier ideas.
NowEvery Transformer layer still pairs a residual path with a normalization step; Adam-style optimizers train nearly everything.
Build in the world’s symmetries to save parameters: convolutional nets share weights across space, recurrent nets across time, and memory architectures might exploit sparsity. A net could learn these itself, with more compute. Shape matters too: don’t squeeze information through a tiny layer or compress it before analyzing it. As tasks grow more general, built-in symmetries matter less.
NowAttention, sparse mixture-of-experts layers, and models that think step by step before answering. What shape means
What the system is graded on: a loss, a reward, a prediction error, an adversary, a fitness function, human judgment, a long interaction with a person, or a mix. A system learns whatever its training rewards in the long run, so the wrong process teaches the wrong task. His examples: dialogue systems trained to match human replies learn imitation rather than achieving what the person wants, and systems trained on one task were never trained to transfer to another. The target can be a multi-part process, as in learning from human feedback, where an agent learns to act so that a person would prefer short clips of its behavior.
NowNext-word prediction, then training on human preferences and on answers that can be checked automatically.
The memo’s rule of thumb: get knobs 1–6 right and the system will probably learn whatever knob 7 specifies, as long as it can be learned with tractable compute. Anything outside the seven rarely matters, though there are exceptions.
Load a case from the memo, or from the years after it, and then change the knobs. This is a cartoon of the rule above, not a simulation: knobs 1, 2, 3, 5, and 6 set how much the system can learn, knob 4 sets how far it generalizes, the weakest knob is the bottleneck, and knob 7 decides what it learns.
Wrong target, and too little blob besides. Fix knob 7 first.
Before 2012, vision systems chained hand-designed modules: edge detectors, contour and segmentation systems, pose estimators. The memo blames knobs 1, 2, 6, and 7.
Essay §2 · The snowflake model · Full text
To make the idea intuitive, the memo turns to snowflakes. Looking at all that fractal detail, you might decide that whoever made one did something impossibly intricate, and that building one piece by piece must be possible, since someone did it. Both conclusions are wrong. You make a snowflake by knowing the physics, supplying raw material and a big enough chamber, setting the temperature, pressure, and humidity, and waiting. Those settings are also your only control over its shape. Assembling one from chips of ice is hopeless.
He calls this the big blob of snow theory. The memo also credits Ilya Sutskever with a shorter version: “networks want to learn.”
Change the conditions and a different crystal grows. The growth rule itself is locked, the way physics is.
Model: Clifford Reiter’s hexagonal cellular automaton for snow-crystal growth (2005). Each cell holds water vapor. Vapor diffuses, collects on cells that touch the ice, and a cell freezes once its level reaches 1. The warm tint is vapor, the raw material; blue is ice.
Essay §3 · Evidence and examples · Full text
Amodei draws on three years of training AI systems and many years of studying the brain. Each case below sets a common belief against the BBOC reading. Pick a knob to see which cases turn on it.
Showing all 13 cases
Narrow tasks. On one Atari game, rather than video games in general, hand-engineering can produce what look like big algorithmic wins, at the cost of brittleness. This often-unnoticed overfitting creates an illusion of algorithmic progress and hides the simple relationship between compute and results. Large projects teach the opposite lesson. His first AI job, Baidu’s Deep Speech 2, used about 0.3 petaflop/s-days per trained model in 2015, which he believes was the most compute used in a single training run up to then. He expects BBOC to become more obvious as tasks become more general, because general tasks are algorithmically simpler to learn.
Tasks too narrow to need learning. Sorting, graph search, and alpha-beta pruning can be handled by a short program that generalizes perfectly within its task and not at all outside it. Pre-ML AI was mostly programming of this kind, which made algorithms look all-powerful. Generalizing even slightly further, say to arbitrary programming-interview questions, means searching over programs. That pulls in heuristics, learning, parameters, and lots of compute.
Part two
The memo sits between two texts that are now cited everywhere. The Transformer paper came out in June 2017 and supplied the architecture that scaled. Rich Sutton’s The Bitter Lesson came out in March 2019 and became the best-known statement of the memo’s core idea. The memo doesn’t mention the Transformer, and The Bitter Lesson came two years after it.
Neighbor one · June 2017
The Transformer was introduced in Attention Is All You Need in June 2017, the same year as the memo. It became the architecture behind GPT, BERT, and nearly every frontier model since. That makes it the natural test of the memo’s forecast that the remaining algorithmic changes would be simple ones that exploit symmetry or sparsity, or would come from the training process.
In the memo, shape is the layout of a computation rather than what it computes: how wide each stage is, how many steps run one after another, whether the system can loop over its own state, and which parts can reach which. The memo asks for three things. Don’t squeeze information through a narrow point, or force the system to compress what it sees before analyzing it. Don’t split the work into blocks that can’t share computation, which is how the memo explains the weak results of model-based RL. And give the network enough internal recurrence to iterate before it acts. Get the layout right and training decides what the computation does.
Each network predicts what comes next. The highlighted line traces how the first input reaches the output.
Recurrent net 2014
Everything about the past is squeezed into one fixed-size state and handed along, one step at a time. The first word reaches the prediction only through that chain, and training has to run the steps in order.
Transformer 2017
Each position reads every earlier position directly, so the first word is one hop from the prediction. All positions are computed at once, which lets far more compute flow through.
Thinking model 2024
The same network writes intermediate steps and reads them back, so it gets many passes before it answers. This is the memo’s idea of letting a policy iterate before it acts.
The first of those requests has a concrete history. In 2014, translation networks read a whole sentence into one fixed-size vector before writing the translation; Ilya Sutskever was an author of the best-known version. Bahdanau, Cho, and Bengio argued that the vector was a bottleneck and let the network look back at every input word instead. That was attention. The Transformer kept the attention and dropped the recurrence.
Same weights at every position
6Symmetry
Like convnets sharing weights across space and recurrent nets across time. The parameter count doesn’t grow with the length of the input.
Attention ignores word order; position encodings add it back
6Symmetry, broken on purpose
Shuffle the input and the outputs shuffle the same way. Relative encodings such as RoPE (2021) make attention depend only on the distance between tokens.
No locality built in
6Fewer symmetries for general tasks
Vision Transformers (2020) trailed convnets on ImageNet-sized data and beat them after pretraining on about 300 million images. Some of their attention heads learned local patterns on their own.
Each token looks up the others by content
26Memory, softly sparse
A learned lookup into everything seen so far, close to the memory-based models the memo had in mind. The lookup is still computed densely: every token checks every other one.
Every earlier token stays reachable, and all positions train at once
61Shape and compute flow
No fixed-size state to squeeze through and no step-by-step loop during training. This is the main reason it won.
Residual paths and normalization in every block
5Conditioning
The memo’s plumbing for keeping numbers near 1.
Mixture-of-experts layers and sparse attention, added to Transformers later
26Sparsity that saves compute
Each token uses a small slice of the weights. DeepSeek-V3 activates about 37 billion of its 671 billion parameters per token.
So the Transformer is partly a symmetry trick, only softly a sparsity trick, and mostly a fix to shape. That fits the memo’s forecast of simple changes. It was also simpler than the recurrent networks with attention that it replaced.
It complicates one of the memo’s claims. In section 3 the memo found recurrent architectures strikingly similar at equal compute, but the Transformer showed that a better shape can be worth a great deal at scale. The 2020 scaling-laws paper, which Amodei co-wrote, found that Transformers beat LSTMs mainly by making better use of long context. And the memo’s warning about alternatives to deep learning has held for the Transformer itself. Rivals proposed since 2017, from linear attention to state-space models like Mamba, have mostly matched it only at smaller scales or been folded in as hybrid layers alongside attention.
Neighbor two · March 2019
Two years after the memo, Rich Sutton, one of the founders of modern reinforcement learning and later a Turing Award winner, published a one-page essay that made the same core argument in public. Looking back over 70 years of AI, he found that general methods that use computation beat methods built on human knowledge by a large margin, because computation keeps getting cheaper. Building in what we know helps in the short run and is satisfying, then plateaus, and the breakthrough comes from scaling search and learning. His cases were chess, Go, speech recognition, and vision. The essay became the standard shorthand for the scaling view, while the memo stayed private until its scan surfaced in October 2026.
Side by side, they agree on the main point and on why researchers keep missing it. They differ on what may be built in, and only the memo is about training targets.
What wins
AgreeThe memoA big, loosely structured pool of compute, with enough parameters and experience.
The Bitter LessonGeneral methods that keep improving as computation grows.
Why researchers miss it
AgreeThe memoNarrow benchmarks reward hand-engineering, and pre-ML AI, which was mostly programming, made algorithms look all-powerful.
The Bitter LessonBuilding in knowledge pays off quickly and feels satisfying, and most research assumes computation stays fixed.
The engine
DifferThe memoLearning. Fixed search is a rigid subset of what learned systems will do, so learned search should win.
The Bitter LessonSearch and learning, the two methods that seem to scale without limit.
What to build in
DifferThe memoA few simple things: symmetries, normalization, and a sensible shape.
The Bitter LessonOnly methods that can find structure on their own. Symmetries are on his list of things not to build in, though he credits convolution in vision.
The training target
Memo onlyThe memoCentral. A system learns what its training rewards, so the wrong target teaches the wrong task.
The Bitter LessonNot discussed.
The scaling era from 2018 to 2024 vindicated both. Since then the differences have mattered more. The memo’s extra knob, the training target, moved to the center of the field after 2022 with human-feedback training and reinforcement learning on checkable answers. On symmetry they met in the middle. The memo expected built-in symmetries to matter less as tasks got more general, and the Transformer, which builds in less than a convnet, won on exactly those tasks.
The two authors also read language models differently. Amodei says the blob keeps working, through pretraining and reinforcement learning alike. In September 2025, Sutton argued on the Dwarkesh Podcast that large language models fall short of his own lesson, because they learn to predict what people would say from human-written text rather than learning from experience. Earlier that year he and David Silver had called for an era of experience. The memo had already treated static data as a special case of experience, which is roughly the path the field took: pretrain on human text, then learn from experience with reinforcement learning.
Part three
The memo was written before GPT-2, before scaling laws, and before any chatbot. Here is how its claims held up through early October 2026. The verdicts are judgment calls on the public record, and the memo is paraphrased throughout.
Knob 1 · Compute
The memo’s own footnote supplies a yardstick. Deep Speech 2, the speech system Amodei worked on at Baidu in 2015, used about 0.3 petaflop/s-days per trained model: a quadrillion operations per second, sustained for about seven hours. Epoch AI estimates that GPT-6 Astra, released by OpenAI in September 2026, took around 10²⁷ operations, or roughly 12 million petaflop/s-days.
Log scale: each gridline is 100× the one below. Hover, tap, or tab to a point.
Scorecard
Each row paraphrases a claim from the memo and sets it against what happened. The small numbers show which knobs a claim is about.
Timeline
Key events since the memo, tagged with the knob each one turned. Pick a knob to follow its thread.
Blind spots
The memo’s examples lean on game-playing agents, robots, and reinforcement learning. The engine of 2018 to 2024 was next-word prediction on internet text: the imitating-a-human task it listed but didn’t feature.
Chinchilla (2022) showed training data must grow along with model size, and by late 2024 high-quality text was a binding constraint. That pushed labs toward synthetic data and reinforcement learning, back toward the memo’s own word: experience.
The memo’s idea of letting a policy iterate before it acts turned into test-time compute. Models now spend more computation per answer, and accuracy rises with it.
Gigawatt data centers, chip export controls, and about $760 billion of planned capital spending by four companies in a single year. Compute is now discussed in units of power plants.
Flattery, reward hacking, and models gaming their tests are routine problems in training frontier systems. That is the failure mode knob 7 predicts, now showing up at scale.
Algorithmic efficiency improved fast: the compute needed for a given result halved about every 8 months. Each trick was simple, as predicted, but together they mattered more than the memo’s tone implies. Compute and data still explain most of the gains.
Still open
Nine years on, the argument is less about whether compute matters than about which knob to turn next. Two people the memo involves now take different sides.
Dario Amodei · February 2026
The hypothesis still holds. Pretraining still pays, and reinforcement learning now improves as steadily with compute as pretraining did. He puts industry compute at 10–15 gigawatts, roughly tripling each year, and expects AI as capable as a nation of top experts within a few years. He doubts that models must learn continually on the job; long context and broad training may be enough.
Dwarkesh Podcast interviewIlya Sutskever · November 2025
Scaling the familiar recipe is running out of road, because good data is finite. The central gap is generalization: people learn from a few examples and carry the lesson widely, while models ace benchmarks and still stumble on simple things. Big computers stay essential, but the next gains need new ideas.
Dwarkesh Podcast interviewBoth positions sit inside the memo’s frame. It never claimed that scaling 2017’s algorithms would reach AGI. It predicted that the necessary changes would be simple or would come from the training process, and that scale would be a huge part of the story. The move from pretraining to reinforcement learning on checkable tasks was exactly such a change: a new target (knob 7) and a new kind of experience (knobs 3 and 4).
One live example: knob 2 separates lasting storage (weights) from short-term storage (activations). The 2026 argument about continual learning is about whether models must move what they learn on the job from the second into the first.