Papers
One foundational paper per page. Each covers the problem it solved, what it claims, the key idea to take away, and what the field has since moved past.
42 papers in 7 topics, from 1967 to 2026
Agreeing on one truth
What it costs for two machines to believe the same thing, and what you get back for paying it.
Time, Clocks, and the Ordering of Events in a Distributed System
Massachusetts Computer Associates, Inc.
"Before" made precise without a clock: a partial order built out of the messages a system already sends.
8 boards · 3 labs · 4 videos
2001Paxos Made Simple
Microsoft Research
Consensus derived from its own invariants in nine pages - and the log everyone actually builds sketched in the last two.
5 boards · 3 labs · 3 videos
2002Brewer's Conjecture and the Feasibility of Consistent, Available, Partition-Tolerant Web Services
MIT Laboratory for Computer Science
Three words given definitions precise enough to prove they are incompatible - and a proof short enough to hold in your head.
6 boards · 4 labs · 3 videos
2006The Chubby lock service for loosely-coupled distributed systems
Google Inc.
Consensus sold as an operable service rather than a library, because teams discover they need availability after the program is written.
5 boards · 2 labs · 2 videos
2007Dynamo: Amazon's Highly Available Key-value Store
Amazon.com
A store that always accepts a write, and hands the resulting disagreement back to the application to sort out.
4 boards · 5 labs · 3 videos
2010ZooKeeper: Wait-free coordination for Internet-scale systems
Yahoo! Research
Do not implement the lock. Ship a wait-free store with ordering and a notification, and let clients assemble the primitives.
6 boards · 2 labs · 3 videos
2011Conflict-free Replicated Data Types
INRIA, Université Pierre et Marie Curie and Universidade do Minho
Concurrent updates converge with no coordination at all, provided the merge is a join and the operations commute - and provided no invariant crosses a replica.
7 boards · 3 labs · 3 videos
2012Spanner: Google's Globally-Distributed Database
Google, Inc.
Transactions ordered by real time across continents, paid for with atomic clocks and a deliberate pause on every commit.
6 boards · 5 labs · 3 videos
2014In Search of an Understandable Consensus Algorithm
Stanford University
Consensus, re-derived with understandability as the design goal rather than a nice property of the write-up.
4 boards · 3 labs · 2 videos
2019Zanzibar: Google's Consistent, Global Authorization System
Google, LLC
One relation-tuple graph answering every permission check at Google, evaluated at a snapshot the caller is allowed to choose.
6 boards · 3 labs · 1 video
Storing it, and reading it back
Disks are fast in one direction and slow in the other. Every design here is a way of only writing the fast way.
The Log-Structured Merge-Tree (LSM-Tree)
UMass Boston, Digital Equipment Corporation and Oracle
Buy cheap writes with expensive reads on purpose, and settle the trade with a cost model rather than a benchmark.
7 boards · 3 labs · 1 video
2003The Google File System
A file system designed from four written-down assumptions about the workload, one of which is that its own machines are always broken.
5 boards · 3 labs · 2 videos
2006Bigtable: A Distributed Storage System for Structured Data
Google, Inc.
A sparse, distributed, persistent multidimensional sorted map - and the sort order of the row key is the entire schema design.
5 boards · 4 labs · 2 videos
2011Kafka: a Distributed Messaging System for Log Processing
LinkedIn Corp.
The log as the primitive: a broker that remembers nothing about its readers, and consumers that pull at their own pace.
9 boards · 5 labs · 5 videos
Spreading the work
Ten thousand machines, one job - and the arithmetic that says how much of it was ever going to help.
Validity of the Single Processor Approach to Achieving Large Scale Computing Capabilities
International Business Machines Corporation
The part of a job that cannot be divided sets a ceiling on the whole job, and a decade of trying had not made that part smaller.
5 boards · 3 labs · 2 videos
1997Consistent Hashing and Random Trees
MIT Laboratory for Computer Science
A hash whose answers barely change when the set of machines does - invented for web caches, not for databases.
5 boards · 3 labs · 4 videos
2004MapReduce: Simplified Data Processing on Large Clusters
Google, Inc.
Take away almost everything a program can say, and what is left can be parallelised, scheduled and re-run by a library nobody has to write twice.
5 boards · 3 labs · 2 videos
2010Large-scale Incremental Processing Using Distributed Transactions and Notifications
Google, Inc.
Transactions and triggers bolted onto a store built without either, so an index can be updated one page at a time instead of rebuilt whole.
7 boards · 2 labs · 3 videos
2012Resilient Distributed Datasets: A Fault-Tolerant Abstraction for In-Memory Cluster Computing
University of California, Berkeley
Fault-tolerant shared memory on a cluster is cheap if nobody may write to it one cell at a time, because then you can log how data was made instead of the data.
3 boards · 1 lab
2013The Tail at Scale
Google, Inc.
At a thousand-way fan-out your typical request is your server's one-in-a-thousand bad day, so stop making components fast and start making slow ones not matter.
6 boards · 2 labs · 4 videos
2013Scaling Memcache at Facebook
Facebook, Inc.
A single-machine hash table turned into a multi-region cache tier, one named race and one named overload shape at a time.
6 boards · 3 labs · 4 videos
2015Large-scale cluster management at Google with Borg
Google, Inc.
A decade of running everything Google runs on shared machines, written up with the regrets left in.
6 boards · 2 labs · 5 videos
2015The Dataflow Model: A Practical Approach to Balancing Correctness, Latency, and Cost in Massive-Scale, Unbounded, Out-of-Order Data Processing
Stop treating a stream as a dataset that will eventually be complete: ask what, where, when and how of every pipeline, and a batch job becomes the case where the answers are all of it, at the end, once.
3 boards · 1 lab
Where the rule belongs
Papers about the shape of a system rather than any part of it. Both of them outgrew the field they were written for.
A Relational Model of Data for Large Shared Data Banks
IBM Research Laboratory, San Jose
Programs should name data by what it is and never by where it is stored, so that the arrangement underneath can change without breaking a single one of them.
3 boards · 1 lab
1984End-to-End Arguments in System Design
MIT Laboratory for Computer Science
A function that depends on the application's own definition of correct cannot be completely implemented by any layer that does not have it.
6 boards · 3 labs · 3 videos
2017Attention Is All You Need
Google Brain, Google Research and the University of Toronto
A translation model with the recurrence taken out, which turned out to matter far more than the translation.
3 boards · 3 labs · 3 videos
What these models are
Six essays from the people building the models and the people studying them, on what feeds a model, what it can be, and who is holding the pen. They disagree with each other, and that is the reading.
The Bitter Lesson
University of Alberta and DeepMind
Seventy years of AI research say the same thing every time: a general method that rides falling compute cost beats a method built from what humans know, and the people who built in the knowledge never see it coming.
5 boards · 3 labs · 3 videos
2023Centaurs and Cyborgs on the Jagged Frontier
The Wharton School, with Harvard Business School and Boston Consulting Group
AI's ability is a wall with an invisible outline: 758 consultants did far better on the tasks inside it and worse than no AI at all on the one task outside, because they could not see which side they were on.
7 boards · 1 lab · 3 videos
2025Welcome to the Era of Experience
Google DeepMind
The high-quality human text is nearly used up, so the next generation of agents will have to learn the way animals do - from a long stream of their own actions, scored by the world rather than by a human rater.
6 boards · 2 labs · 3 videos
2025Large AI Models Are Cultural and Social Technologies
Johns Hopkins University, University of California Berkeley, Carnegie Mellon University and the University of Chicago
A large model is not a mind in the making. It belongs with writing, print, markets and bureaucracies - technologies that let people use what other people know - and it should be studied and governed as one.
7 boards · 2 labs · 3 videos
2026Why AI Needs a Sense of Smell
Noema Magazine
Vision, language and sound each had a digital format waiting for them. Smell has none - no wheel, no frequency, no map - and an intelligence built only from the senses that were easy to digitise is missing the oldest one.
6 boards · 1 lab · 3 videos
2026A Framework for Frontier AI and the Dawning of a New Age
Google DeepMind
AGI is a few years out, the race is outrunning the understanding, and the fix proposed is a standards body modelled on the one that polices stockbrokers: voluntary at first, funded by the labs it tests.
6 boards · 1 lab · 3 videos
How a model gets made
The technical papers under the models people use: what each one changed about training, tuning or serving, what it measured, and which of its conclusions the field quietly revised.
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Facebook AI Research, University College London and New York University
Give a generator a searchable library of plain text, train the pair together, and its answers get more specific and more factual - and change when you swap the library, with no weight retrained.
5 boards · 2 labs · 3 videos
2021LoRA: Low-Rank Adaptation of Large Language Models
Microsoft
Freeze the model and train a thin correction beside each weight, because the change fine-tuning makes turns out to need only a few directions.
1 board · 1 lab · 1 video
2022Training language models to follow instructions with human feedback
OpenAI
Fine-tune GPT-3 on written answers, then on which of several answers people preferred, and a 1.3 billion parameter model is preferred to the 175 billion parameter one it came from.
3 boards · 2 labs · 2 videos
2022Training Compute-Optimal Large Language Models
DeepMind
For a fixed training budget, grow the model and the data in equal proportion, and a 70 billion parameter model on 1.4 trillion tokens beats the 280 billion one trained with the same compute.
3 boards · 2 labs · 1 video
2023Direct Preference Optimization: Your Language Model is Secretly a Reward Model
Stanford University and CZ Biohub
The best model for the RLHF objective has a closed form, so the reward model and the reinforcement learning loop collapse into one classification loss on the preference pairs.
1 board · 1 video
2023ReAct: Synergizing Reasoning and Acting in Language Models
Princeton University and Google Research, Brain team
Let a language model write a thought as one of its actions, interleaved with real actions and their results, and it plans better than acting alone and invents facts less than reasoning alone.
5 boards · 2 labs · 3 videos
2023Efficient Memory Management for Large Language Model Serving with PagedAttention
UC Berkeley, Stanford University and UC San Diego
Store each conversation's attention cache in small fixed-size blocks behind a table, the way an operating system pages memory, and two to four times as many tokens come off the same GPU.
5 boards · 2 labs · 1 video
Inside one database
How one database picks a plan, survives a crash, and decides what two transactions may see at once. Three papers every engine since was built on, or argued with.
Access Path Selection in a Relational Database Management System
IBM Research Division, San Jose
Price every way of answering a query in one made-up unit, from five numbers about the data, and keep only the cheapest plan for each set of tables and each useful order.
4 boards · 1 lab
1992ARIES: A Transaction Recovery Method Supporting Fine-Granularity Locking and Partial Rollbacks Using Write-Ahead Logging
IBM Almaden Research Center and IBM Santa Teresa Laboratory
Put one log sequence number on every page, replay history for every transaction before undoing the losers, and log each undo so a rollback can never be undone.
4 boards · 2 labs
1995A Critique of ANSI SQL Isolation Levels
Microsoft, Sybase and the University of Massachusetts Boston
The SQL standard's isolation levels are ambiguous and incomplete; read broadly they are locking in disguise, and the snapshot level many products ship is not serializable.
4 boards · 2 labs