Skip to content
forward pass

· turning point

DeepSeek-R1

A Chinese laboratory releases an open reasoning model that matches OpenAI's o1, trained with pure reinforcement learning on restricted chips; a week later Nvidia loses $590 billion in a day.

category
model
significance
5 of 5
people
Liang Wenfeng
organisations
DeepSeek

what had to happen · 57 events back to 1943

Every event this one built on, transitively, in order. The organism has the same path lit. Direct influences are marked.

00 · One neuron · 1

  1. 1943A logical calculus of nervous activity

I · Foundations · 11

  1. 1948A mathematical theory of communication
  2. 1949Cells that fire together wire together
  3. 1950Programming a computer for playing chess
  4. 1950Computing machinery and intelligence
  5. 1958The perceptron learns
  6. 1959Samuel's checkers program coins 'machine learning'
  7. 1960ADALINE and the least-mean-squares rule
  8. 1965Moore's law
  9. 1966ELIZA
  10. 1969Perceptrons
  11. 1970Reverse-mode automatic differentiation

W1 · The first winter · 2

  1. 1974Werbos applies backpropagation to neural networks
  2. 1980The Neocognitron

II · Connection · 3

  1. 1982The Hopfield network
  2. 1985The Boltzmann machine
  3. 1986Backpropagation

W2 · The second winter · 6

  1. 1988Temporal-difference learning
  2. 1989Q-learning
  3. 1989LeNet reads handwritten postcodes
  4. 1990Finding structure in time
  5. 1991The vanishing gradient problem
  6. 1992TD-Gammon reaches world-class backgammon

III · Statistics and data · 9

  1. 1997Long short-term memory
  2. 1998MNIST and LeNet-5
  3. 1999The first GPU
  4. 2003A neural probabilistic language model
  5. 2006Deep belief networks and the word 'deep'
  6. 2007CUDA
  7. 2009Deep learning moves to GPUs
  8. 2009ImageNet
  9. 2010Rectified linear units

IV · Deep learning · 9

  1. 2012Google Brain's network discovers cats
  2. 2012Dropout
  3. 2012AlexNet wins ImageNet
  4. 2013Deep Q-networks play Atari
  5. 2014Attention
  6. 2014Sequence to sequence learning
  7. 2015Batch normalisation
  8. 2015Residual networks
  9. 2016AlphaGo beats Lee Sedol

V · Transformers · 8

  1. 2017Attention is all you need
  2. 2017Deep reinforcement learning from human preferences
  3. 2017AlphaGo Zero learns from nothing
  4. 2018GPT: generative pre-training
  5. 2019GPT-2 and the model too dangerous to release
  6. 2020Scaling laws for neural language models
  7. 2020GPT-3
  8. 2020Learning to summarise from human feedback

VI · Everyone · 8

  1. 2022InstructGPT
  2. 2022Chain-of-thought prompting
  3. 2022Chinchilla: the models were undertrained
  4. 2022ChatGPT
  5. 2023LLaMA leaks and open weights take off
  6. 2023GPT-4
  7. 2024o1 and reasoning modelsdirect
  8. 2024DeepSeek-V3 trained for $5.6 milliondirect

DeepSeek-R1 was released on 20 January 2025 with its weights, under an MIT licence, and a paper that explained how it had been made. Starting from the V3 base, the laboratory had applied reinforcement learning with rewards only for correct answers on maths and code, no human-written reasoning at all, and the model had learned by itself to think at length, to check its work and, in a passage the paper called an "aha moment", to stop and reconsider mid-solution. Its scores matched OpenAI's o1, whose method had been secret, and it cost a fraction as much to run.

The week that followed is the reason the event is at the top level of this timeline. DeepSeek's app reached the top of the American App Store; on Monday 27 January Nvidia's shares fell seventeen percent, erasing about $590 billion of value, the largest one-day loss any company had suffered, on the thought that if frontier models could be trained this cheaply the demand for chips had been overestimated. The thought did not last, and Nvidia was worth four trillion dollars by July, but the assumption that the frontier belonged to three American laboratories and their capital did not recover.

R1 also made the reasoning recipe public. Every laboratory's models thought out loud within months, and the distilled versions of R1 ran on laptops.

what it led to · 3 events downstream, through 2025

Built on it directly:

  1. 2025Nvidia is worth four trillion dollarsVII
  2. 2025Gold at the Mathematical OlympiadVII

And, through them, by era:

VII · Agents · 1
  1. 2025Gemini 3

sources · 2

Trace the lineage of this event on the timeline →Back to the ledger