115 recent and foundational AI, blockchain and technology research sources

A dated, editorially synthesised guide to recent and foundational papers, standards and institutional reports across artificial intelligence, blockchain and modern computing.

Coverage: Artificial Intelligence: 50 | Blockchain: 32 | Technology: 33

Selection method: Recent evidence with material decision value, historically influential work, field-defining system papers, widely used standards, or evidence-rich institutional reports with durable educational value.

Source policy: Every record links to a canonical author, publisher, standards body, university, or institutional source. Synopses and learnings are Rusaka editorial synthesis.

Limitations: This is a selective dated learning path, not a live feed, citation ranking or exhaustive systematic review. Reports, standards and papers are labelled separately; users should verify the source and current status before relying on it.

1. #1 Computing Machinery and Intelligence (1950)

   Artificial Intelligence | Paper | Alan M. Turing | Oxford Academic

   Turing replaces an abstract argument about whether machines can think with an operational imitation game, then examines the principal objections to machine intelligence.

   Key learnings

   • Useful AI evaluation starts with observable behaviour and a defined test.
   • Capability claims should be separated from philosophical claims about consciousness.

   Topics: AI history, evaluation, Turing test

   Read the canonical source

2. #2 The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain (1958)

   Artificial Intelligence | Paper | Frank Rosenblatt | American Psychological Association

   This early learning-machine paper formalises a trainable classifier inspired by neural organisation and explores how associations can emerge from examples.

   Key learnings

   • Learning from data can replace hand-written decision rules for bounded tasks.
   • Single-layer models have structural limits, making architecture choice consequential.

   Topics: neural networks, classification, machine learning

   Read the canonical source

3. #3 Learning Representations by Back-propagating Errors (1986)

   Artificial Intelligence | Paper | David E. Rumelhart, Geoffrey E. Hinton and Ronald J. Williams | Nature

   The paper demonstrates how gradients can propagate through hidden layers so multilayer neural networks learn internal representations from errors.

   Key learnings

   • Backpropagation made end-to-end representation learning practical.
   • Optimisation, architecture and data jointly determine what a network learns.

   Topics: backpropagation, deep learning, optimisation

   Read the canonical source

4. #4 Long Short-Term Memory (1997)

   Artificial Intelligence | Paper | Sepp Hochreiter and Jürgen Schmidhuber | MIT Press

   LSTM introduces gated memory cells that preserve and control information over long sequences, addressing unstable gradients in recurrent networks.

   Key learnings

   • Explicit memory gates improve learning across long temporal gaps.
   • Sequence models need mechanisms for selective retention and forgetting.

   Topics: LSTM, sequence modelling, recurrent networks

   Read the canonical source

5. #5 Gradient-Based Learning Applied to Document Recognition (1998)

   Artificial Intelligence | Paper | Yann LeCun, Léon Bottou, Yoshua Bengio and Patrick Haffner | IEEE

   This systems paper connects convolutional networks, gradient-based training and document-processing pipelines, showing strong handwritten-character recognition.

   Key learnings

   • Convolutions exploit spatial structure and parameter sharing.
   • End-to-end systems can outperform separately engineered recognition stages.

   Topics: convolutional networks, computer vision, LeNet

   Read the canonical source

6. #6 ImageNet: A Large-Scale Hierarchical Image Database (2009)

   Artificial Intelligence | Paper | Jia Deng et al. | IEEE Computer Society

   ImageNet establishes a large, hierarchically organised labelled image dataset and a benchmark that accelerated measurable progress in visual recognition.

   Key learnings

   • Shared datasets make model comparisons more reproducible.
   • Benchmark design and label quality shape the direction of a research field.

   Topics: ImageNet, datasets, computer vision

   Read the canonical source

7. #7 ImageNet Classification with Deep Convolutional Neural Networks (2012)

   Artificial Intelligence | Paper | Alex Krizhevsky, Ilya Sutskever and Geoffrey E. Hinton | NeurIPS

   AlexNet combines deep convolutions, GPU training, ReLU activations, dropout and data augmentation to deliver a major ImageNet performance jump.

   Key learnings

   • Compute, data and architectural choices can create step-change performance gains.
   • Regularisation and augmentation are essential when model capacity grows.

   Topics: AlexNet, GPU computing, computer vision

   Read the canonical source

8. #8 Efficient Estimation of Word Representations in Vector Space (2013)

   Artificial Intelligence | Paper | Tomas Mikolov et al. | arXiv

   Word2vec shows that simple predictive objectives can learn efficient dense word embeddings that encode useful semantic and syntactic relationships.

   Key learnings

   • Meaningful representations can emerge from distributional context.
   • Efficient training objectives can unlock value from very large corpora.

   Topics: word2vec, embeddings, natural language processing

   Read the canonical source

9. #9 Generative Adversarial Nets (2014)

   Artificial Intelligence | Paper | Ian Goodfellow et al. | NeurIPS

   GANs frame generation as a game between a generator and discriminator, enabling a model to learn complex data distributions without explicit likelihoods.

   Key learnings

   • Adversarial objectives can produce realistic synthetic data.
   • Training stability and evaluation remain central generative-model challenges.

   Topics: GAN, generative AI, adversarial learning

   Read the canonical source

10. #10 Sequence to Sequence Learning with Neural Networks (2014)

    Artificial Intelligence | Paper | Ilya Sutskever, Oriol Vinyals and Quoc V. Le | NeurIPS

    The encoder-decoder approach maps variable-length input sequences to variable-length outputs and demonstrates strong machine-translation performance.

    Key learnings

    • A shared learned representation can connect different sequence domains.
    • Input ordering and decoding strategy materially affect sequence quality.

    Topics: sequence-to-sequence, machine translation, encoder-decoder

    Read the canonical source

11. #11 Adam: A Method for Stochastic Optimization (2014)

    Artificial Intelligence | Paper | Diederik P. Kingma and Jimmy Ba | International Conference on Learning Representations

    Adam adapts learning rates using estimates of the first and second moments of gradients, providing an efficient optimiser for noisy, high-dimensional problems.

    Key learnings

    • Adaptive optimisation reduces manual tuning for many workloads.
    • Default convenience does not remove the need to monitor convergence and generalisation.

    Topics: Adam, optimisation, deep learning

    Read the canonical source

12. #12 Deep Residual Learning for Image Recognition (2015)

    Artificial Intelligence | Paper | Kaiming He et al. | IEEE CVPR

    Residual connections let very deep networks learn modifications to identity mappings, easing optimisation and improving visual recognition.

    Key learnings

    • Skip connections improve gradient flow through deep architectures.
    • Depth becomes useful only when the optimisation path remains tractable.

    Topics: ResNet, computer vision, deep learning

    Read the canonical source

13. #13 Mastering the Game of Go with Deep Neural Networks and Tree Search (2016)

    Artificial Intelligence | Paper | David Silver et al. | Nature

    AlphaGo combines policy networks, value networks and Monte Carlo tree search to master a domain previously considered resistant to brute-force search.

    Key learnings

    • Hybrid systems can unite learned intuition with explicit planning.
    • Self-play and simulation can generate valuable training experience.

    Topics: AlphaGo, reinforcement learning, planning

    Read the canonical source

14. #14 Attention Is All You Need (2017)

    Artificial Intelligence | Paper | Ashish Vaswani et al. | NeurIPS

    The Transformer replaces recurrent sequence processing with self-attention, enabling parallel training and stronger modelling of long-range dependencies.

    Key learnings

    • Attention provides a scalable mechanism for contextual representation.
    • Parallelisable architectures can compound gains from larger data and compute.

    Topics: transformer, attention, foundation models

    Read the canonical source

15. #15 BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2018)

    Artificial Intelligence | Paper | Jacob Devlin et al. | NAACL

    BERT pretrains deep bidirectional language representations and then fine-tunes them for many language-understanding tasks with minimal task-specific changes.

    Key learnings

    • General pretraining can transfer across numerous downstream tasks.
    • Bidirectional context is especially valuable for language understanding.

    Topics: BERT, pretraining, natural language processing

    Read the canonical source

16. #16 Improving Language Understanding by Generative Pre-Training (2018)

    Artificial Intelligence | Paper | Alec Radford et al. | OpenAI

    The first GPT paper pairs generative language-model pretraining with supervised fine-tuning, demonstrating broad transfer from unlabelled text.

    Key learnings

    • Unlabelled corpora can supply reusable language knowledge.
    • A consistent model architecture simplifies transfer across tasks.

    Topics: GPT, language models, transfer learning

    Read the canonical source

17. #17 Language Models are Unsupervised Multitask Learners (2019)

    Artificial Intelligence | Technical report | Alec Radford et al. | OpenAI

    GPT-2 shows that a larger next-token model can perform multiple language tasks in a zero-shot setting when prompted with natural-language context.

    Key learnings

    • Scale can reveal capabilities not explicitly supervised task by task.
    • Release strategy becomes part of responsible model governance.

    Topics: GPT-2, zero-shot learning, language models

    Read the canonical source

18. #18 Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer (2019)

    Artificial Intelligence | Paper | Colin Raffel et al. | Journal of Machine Learning Research

    T5 casts every language task as text-to-text and systematically studies objectives, datasets, architectures and scale in a unified framework.

    Key learnings

    • A common interface makes heterogeneous tasks easier to train and evaluate.
    • Data cleaning and objective design can matter as much as parameter count.

    Topics: T5, transfer learning, text-to-text

    Read the canonical source

19. #19 Language Models are Few-Shot Learners (2020)

    Artificial Intelligence | Paper | Tom B. Brown et al. | NeurIPS

    GPT-3 demonstrates that a very large autoregressive language model can perform many tasks from instructions and a few in-context examples without weight updates.

    Key learnings

    • In-context learning creates a new software interface based on examples and instructions.
    • Broad capability does not guarantee reliability, calibration or fairness.

    Topics: GPT-3, few-shot learning, large language models

    Read the canonical source

20. #20 Scaling Laws for Neural Language Models (2020)

    Artificial Intelligence | Paper | Jared Kaplan et al. | arXiv

    The study quantifies power-law relationships among model size, data, compute and loss, providing a predictive view of language-model scaling.

    Key learnings

    • Empirical scaling curves can guide compute allocation before expensive training.
    • Bottlenecks shift among model capacity, data and optimisation budget.

    Topics: scaling laws, compute, language models

    Read the canonical source

21. #21 Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020)

    Artificial Intelligence | Paper | Patrick Lewis et al. | NeurIPS

    RAG combines a neural retriever over external documents with a sequence generator, improving knowledge-intensive generation and source adaptability.

    Key learnings

    • External retrieval can make knowledge easier to update than model weights.
    • Retrieval quality and grounding need evaluation separately from fluent generation.

    Topics: RAG, information retrieval, grounded generation

    Read the canonical source

22. #22 Denoising Diffusion Probabilistic Models (2020)

    Artificial Intelligence | Paper | Jonathan Ho, Ajay Jain and Pieter Abbeel | NeurIPS

    The paper develops a probabilistic generative process that learns to reverse gradual noise corruption and produces high-quality image samples.

    Key learnings

    • Iterative denoising offers a stable alternative to adversarial generation.
    • Sampling quality, latency and compute form an important deployment trade-off.

    Topics: diffusion models, generative AI, image generation

    Read the canonical source

23. #23 Highly Accurate Protein Structure Prediction with AlphaFold (2021)

    Artificial Intelligence | Paper | John Jumper et al. | Nature

    AlphaFold integrates attention-based representations and geometric reasoning to predict protein structures with accuracy that transformed structural biology workflows.

    Key learnings

    • Domain structure can be built into a learning system without abandoning end-to-end training.
    • Scientific AI requires rigorous domain benchmarks and confidence estimates.

    Topics: AlphaFold, AI for science, protein structure

    Read the canonical source

24. #24 An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (2021)

    Artificial Intelligence | Paper | Alexey Dosovitskiy et al. | International Conference on Learning Representations

    Vision Transformer treats image patches as tokens and shows that Transformer architectures can rival or exceed convolutional networks with sufficient pretraining.

    Key learnings

    • Architectures can transfer across modalities when inputs are represented appropriately.
    • Large-scale pretraining can offset weaker built-in domain assumptions.

    Topics: Vision Transformer, computer vision, multimodal AI

    Read the canonical source

25. #25 Learning Transferable Visual Models From Natural Language Supervision (2021)

    Artificial Intelligence | Paper | Alec Radford et al. | OpenAI

    CLIP learns aligned image and text representations from internet-scale pairs, enabling flexible zero-shot visual classification from natural-language prompts.

    Key learnings

    • Natural language can act as an open-ended supervision interface.
    • Web-scale data increases capability while importing data-quality and bias risks.

    Topics: CLIP, multimodal AI, contrastive learning

    Read the canonical source

26. #26 Zero-Shot Text-to-Image Generation (2021)

    Artificial Intelligence | Paper | Aditya Ramesh et al. | International Conference on Machine Learning

    DALL·E models text and image tokens in a shared autoregressive framework, demonstrating compositional image generation from natural-language descriptions.

    Key learnings

    • Shared token modelling can connect language and visual generation.
    • Compositional demos do not replace systematic robustness evaluation.

    Topics: DALL-E, text-to-image, multimodal AI

    Read the canonical source

27. #27 LoRA: Low-Rank Adaptation of Large Language Models (2021)

    Artificial Intelligence | Paper | Edward J. Hu et al. | International Conference on Learning Representations

    LoRA freezes base-model weights and trains compact low-rank updates, greatly reducing the parameters and memory required for adaptation.

    Key learnings

    • Parameter-efficient tuning lowers the cost of specialised models.
    • Adapters create new versioning, provenance and deployment-management needs.

    Topics: LoRA, fine-tuning, parameter-efficient learning

    Read the canonical source

28. #28 On the Opportunities and Risks of Foundation Models (2021)

    Artificial Intelligence | Report | Rishi Bommasani et al. | Stanford Center for Research on Foundation Models

    This multidisciplinary report defines foundation models and maps their technical, social, legal and economic opportunities and risks across downstream uses.

    Key learnings

    • The same general model can propagate benefits and failures across many applications.
    • Evaluation and accountability must cover the full supply chain, not only the base model.

    Topics: foundation models, AI governance, risk

    Read the canonical source

29. #29 Training Language Models to Follow Instructions with Human Feedback (2022)

    Artificial Intelligence | Paper | Long Ouyang et al. | NeurIPS

    InstructGPT uses demonstrations, preference data and reinforcement learning from human feedback to make model behaviour better match user intent.

    Key learnings

    • Human preferences can materially improve perceived usefulness and safety.
    • Reward models are imperfect proxies and require ongoing adversarial evaluation.

    Topics: RLHF, alignment, InstructGPT

    Read the canonical source

30. #30 Training Compute-Optimal Large Language Models (2022)

    Artificial Intelligence | Paper | Jordan Hoffmann et al. | NeurIPS

    The Chinchilla study argues that many large models were undertrained and derives a more compute-efficient balance between parameter count and training tokens.

    Key learnings

    • Bigger models are not automatically the best use of a fixed compute budget.
    • Data quantity and quality are first-class scaling decisions.

    Topics: Chinchilla, scaling laws, compute efficiency

    Read the canonical source

31. #31 Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (2022)

    Artificial Intelligence | Paper | Jason Wei et al. | NeurIPS

    The paper shows that prompting large models with intermediate reasoning examples can improve performance on multi-step arithmetic, commonsense and symbolic tasks.

    Key learnings

    • Prompt format can change measured reasoning performance without weight updates.
    • Plausible reasoning traces can still support incorrect answers and need verification.

    Topics: chain of thought, prompting, reasoning

    Read the canonical source

32. #32 ReAct: Synergizing Reasoning and Acting in Language Models (2022)

    Artificial Intelligence | Paper | Shunyu Yao et al. | International Conference on Learning Representations

    ReAct interleaves language-model reasoning traces with actions in an external environment, improving interpretability and task completion on knowledge and decision tasks.

    Key learnings

    • Tool use can extend a model beyond static internal knowledge.
    • Agent systems need controls for action permissions, observations and recovery.

    Topics: AI agents, tool use, reasoning

    Read the canonical source

33. #33 Constitutional AI: Harmlessness from AI Feedback (2022)

    Artificial Intelligence | Paper | Yuntao Bai et al. | Anthropic

    Constitutional AI uses written principles and model-generated critiques to reduce reliance on direct human labels while training a more harmless assistant.

    Key learnings

    • Explicit principles make part of the alignment target inspectable.
    • A constitution embeds normative choices that still require human governance.

    Topics: Constitutional AI, alignment, AI safety

    Read the canonical source

34. #34 FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (2022)

    Artificial Intelligence | Paper | Tri Dao et al. | NeurIPS

    FlashAttention reorganises exact attention computation around memory hierarchy, reducing data movement and improving speed without approximate outputs.

    Key learnings

    • Systems-level memory efficiency can unlock model-level capability.
    • Algorithm analysis should account for hardware communication, not only arithmetic.

    Topics: FlashAttention, AI infrastructure, GPU optimisation

    Read the canonical source

35. #35 High-Resolution Image Synthesis with Latent Diffusion Models (2022)

    Artificial Intelligence | Paper | Robin Rombach et al. | IEEE CVPR

    Latent diffusion moves denoising into a compressed representation, cutting training and inference costs while retaining strong image quality and conditioning.

    Key learnings

    • Good latent representations make generative models more computationally accessible.
    • Compression introduces its own fidelity and bias constraints.

    Topics: latent diffusion, Stable Diffusion, image generation

    Read the canonical source

36. #36 Segment Anything (2023)

    Artificial Intelligence | Paper | Alexander Kirillov et al. | Meta AI Research

    The Segment Anything project combines a promptable segmentation model with a large segmentation dataset to support zero-shot transfer across visual tasks.

    Key learnings

    • Promptable interfaces can turn a narrow model into reusable infrastructure.
    • Large machine-assisted datasets need careful quality and licensing governance.

    Topics: Segment Anything, computer vision, foundation models

    Read the canonical source

37. #37 Direct Preference Optimization: Your Language Model is Secretly a Reward Model (2023)

    Artificial Intelligence | Paper | Rafael Rafailov et al. | NeurIPS

    DPO derives a simple classification objective that optimises language models directly from preference pairs without a separate reinforcement-learning loop.

    Key learnings

    • Preference optimisation can be simpler and more stable than full RLHF pipelines.
    • Results still depend on whose preferences are collected and how comparisons are sampled.

    Topics: DPO, alignment, preference learning

    Read the canonical source

38. #38 Llama 2: Open Foundation and Fine-Tuned Chat Models (2023)

    Artificial Intelligence | Paper | Hugo Touvron et al. | Meta AI Research

    The report documents pretrained and chat-tuned Llama 2 models, including training approach, safety evaluations, red teaming and responsible-use controls.

    Key learnings

    • Model documentation should cover data, tuning, evaluation and release controls together.
    • Open-weight access broadens experimentation while changing misuse and assurance risks.

    Topics: Llama 2, open models, large language models

    Read the canonical source

39. #39 GPT-4 Technical Report (2023)

    Artificial Intelligence | Technical report | OpenAI | OpenAI

    The report describes GPT-4 as a multimodal model, presents benchmark and safety results, and deliberately withholds several architecture and training details.

    Key learnings

    • Capability reports must be read alongside their disclosure limitations.
    • Pre-deployment safety work reduces but does not eliminate post-release risks.

    Topics: GPT-4, multimodal AI, model evaluation

    Read the canonical source

40. #40 Artificial Intelligence Index Report 2025 (2025)

    Artificial Intelligence | Report | Stanford Institute for Human-Centered Artificial Intelligence | Stanford HAI

    The eighth AI Index synthesises global evidence on technical performance, costs, investment, adoption, education, policy, responsible AI and public opinion.

    Key learnings

    • AI progress must be assessed with economic and social indicators as well as benchmarks.
    • Falling inference costs and rising adoption increase the importance of governance at deployment scale.

    Topics: AI Index, AI adoption, AI policy

    Read the canonical source

41. #41 Artificial Intelligence Risk Management Framework (AI RMF 1.0) (2023)

    Artificial Intelligence | Standard | National Institute of Standards and Technology | NIST

    NIST organises voluntary AI risk management around Govern, Map, Measure and Manage functions with characteristics for trustworthy systems.

    Key learnings

    • Risk management should be continuous across the AI lifecycle.
    • Context and impact determine which measurements and controls are proportionate.

    Topics: NIST AI RMF, AI governance, risk management

    Read the canonical source

42. #42 Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (2024)

    Artificial Intelligence | Standard | National Institute of Standards and Technology | NIST

    The Generative AI Profile adapts the AI RMF to risks such as confabulation, information integrity, dangerous content, privacy, intellectual property and value-chain dependencies.

    Key learnings

    • Generative AI risks need controls spanning data, model, interface and downstream use.
    • Testing should cover misuse and systemic effects, not only average task accuracy.

    Topics: generative AI risk, NIST, AI assurance

    Read the canonical source

43. #43 Gen-AI: Artificial Intelligence and the Future of Work (2024)

    Artificial Intelligence | Report | Mauro Cazzaniga et al. | International Monetary Fund

    The IMF examines occupational exposure, complementarity, inequality and policy readiness across advanced, emerging and developing economies.

    Key learnings

    • Exposure is not equivalent to displacement; complementarity shapes outcomes.
    • Digital infrastructure, skills and social protection affect how gains and risks are distributed.

    Topics: AI and jobs, economic policy, productivity

    Read the canonical source

44. #44 Generative AI and Jobs: A 2025 Update (2025)

    Artificial Intelligence | Report | Paweł Gmyrek et al. | International Labour Organization

    The ILO updates global occupational-exposure estimates with a refined task-level method that incorporates expert and model-based assessment.

    Key learnings

    • Most exposed occupations are more likely to be transformed than fully automated.
    • Exposure indicators are scenarios for capability, not forecasts of realised job losses.

    Topics: AI and jobs, labour markets, automation

    Read the canonical source

45. #45 AI Index Report 2026 (2026)

    Artificial Intelligence | Report | Stanford Institute for Human-Centered Artificial Intelligence | Stanford HAI

    The 2026 AI Index consolidates evidence on technical progress, research, responsible AI, economic activity, education, policy and public opinion into a documented annual snapshot.

    Key learnings

    • Annual indicators help separate measured changes in capability, adoption and investment from isolated product announcements.
    • Cross-domain evidence should be interpreted with each chapter’s definitions, methods, geographic coverage and reporting limits.

    Topics: AI Index, AI adoption, AI policy

    Read the canonical source

46. #46 Expanding the AI Evaluation Toolbox with Statistical Models (2026)

    Artificial Intelligence | Report | Andrew Keller, Kweku Kwegyir-Aggrey, Ryan Steed, Anita Rao, Julia Sharp and Amanda Bergman | National Institute of Standards and Technology

    NIST distinguishes fixed-benchmark accuracy from generalised accuracy and evaluates statistical models that quantify uncertainty, item difficulty and variance across frontier-model benchmarks.

    Key learnings

    • Evaluation reports should state their measurement target and assumptions instead of treating one benchmark score as universal performance.
    • Statistical modelling can reveal uncertainty and item effects that aggregate accuracy alone conceals from decision-makers.

    Topics: AI evaluation, benchmarks, measurement uncertainty

    Read the canonical source

47. #47 Challenges to the Monitoring of Deployed AI Systems (2026)

    Artificial Intelligence | Report | Anita Rao et al. | National Institute of Standards and Technology

    NIST synthesises literature and practitioner workshops into monitoring categories, implementation barriers and open questions for AI systems operating in real-world contexts.

    Key learnings

    • Pre-deployment evaluation must be complemented by repeated production measurement because context, inputs and system behaviour change.
    • Monitoring scope should cover functionality, operations, human factors, security, compliance and large-scale impacts with accountable ownership.

    Topics: AI monitoring, post-deployment assurance, NIST

    Read the canonical source

48. #48 Economic Policy for AGI (2026)

    Artificial Intelligence | Paper | Julian Jacobs and Alex Imas | Google DeepMind

    This paper evaluates twenty-four labour, welfare and taxation interventions across multiple technological-disruption scenarios using an automated multi-agent assessment methodology.

    Key learnings

    • Policy choices should be stress-tested across mild and severe disruption scenarios rather than anchored to one adoption forecast.
    • A phased least-regret strategy can preserve optionality while evidence about labour displacement and productivity continues to develop.

    Topics: AI economics, labour policy, scenario analysis

    Read the canonical source

49. #49 What Work Can Robots Do? (2026)

    Artificial Intelligence | Paper | Russell Legate-Yang and Maxim Massenkoff | Anthropic

    The study builds a task-level robot-exposure index, compares technical capability with cost competitiveness and backtests whether historical exposure predicted employment and wage changes.

    Key learnings

    • Technical exposure is not equivalent to economical automation because cost, environment, capability, preference and regulation constrain adoption.
    • Task-level measures offer more decision value than treating whole occupations as uniformly automatable or unaffected.

    Topics: robotics, future of work, automation economics

    Read the canonical source

50. #50 Automated Researchers Can Reliably Mitigate Alignment Failures (2026)

    Artificial Intelligence | Report | Anthropic Alignment Science | Anthropic

    Anthropic tests an automated research loop that proposes, trains and evaluates mitigations across ten alignment-failure categories, including withheld benchmarks and larger target models.

    Key learnings

    • Automated research claims require controls for benchmark leakage, capability regressions, monitorability and transfer to unseen evaluations.
    • Promising benchmark gains remain bounded evidence because proxy tasks and limited capability tests do not establish production-wide alignment.

    Topics: AI alignment, automated research, benchmark validation

    Read the canonical source

51. #51 How to Time-Stamp a Digital Document (1991)

    Blockchain | Paper | Stuart Haber and W. Scott Stornetta | Springer

    The paper proposes cryptographic chaining and distributed witness mechanisms that make document timestamps difficult to alter without detection.

    Key learnings

    • Tamper evidence can be achieved by linking records cryptographically.
    • Trust can be distributed across witnesses rather than concentrated in one archive.

    Topics: timestamping, cryptography, blockchain history

    Read the canonical source

52. #52 Hashcash: A Denial of Service Counter-Measure (2002)

    Blockchain | Paper | Adam Back | Hashcash

    Hashcash describes a computational proof-of-work stamp intended to price abusive email and service requests while remaining cheap to verify.

    Key learnings

    • Asymmetric cost can deter spam when verification is far cheaper than creation.
    • Proof of work originated as an anti-abuse mechanism, not a monetary policy.

    Topics: Hashcash, proof of work, anti-spam

    Read the canonical source

53. #53 b-money (1998)

    Blockchain | Proposal | Wei Dai | Wei Dai

    The b-money proposal sketches pseudonymous digital money maintained through collective accounting and computational commitments.

    Key learnings

    • Digital money designs predated Bitcoin’s specific consensus solution.
    • Identity, incentives and ledger agreement must be designed together.

    Topics: b-money, digital cash, blockchain history

    Read the canonical source

54. #54 Practical Byzantine Fault Tolerance (1999)

    Blockchain | Paper | Miguel Castro and Barbara Liskov | USENIX

    PBFT presents a replicated state-machine protocol that tolerates Byzantine participants with practical performance in asynchronous networks.

    Key learnings

    • Known-validator systems can reach finality without proof of work.
    • Fault thresholds and communication overhead constrain consensus design.

    Topics: PBFT, Byzantine fault tolerance, consensus

    Read the canonical source

55. #55 Bitcoin: A Peer-to-Peer Electronic Cash System (2008)

    Blockchain | White paper | Satoshi Nakamoto | Bitcoin.org

    Bitcoin combines a peer-to-peer ledger, proof of work, digital signatures and economic incentives to prevent double spending without a central operator.

    Key learnings

    • Consensus security emerges from cryptography, networking and incentives together.
    • Settlement assurance depends on assumptions about honest computational power.

    Topics: Bitcoin, proof of work, digital assets

    Read the canonical source

56. #56 Secure High-Rate Transaction Processing in Bitcoin (2013)

    Blockchain | Paper | Yonatan Sompolinsky and Aviv Zohar | IACR

    The GHOST protocol changes chain selection to account for work in stale branches, seeking stronger security when blocks arrive quickly.

    Key learnings

    • Block propagation and fork rates limit permissionless throughput.
    • Consensus rules can use more information than the single longest branch.

    Topics: GHOST, Bitcoin, consensus

    Read the canonical source

57. #57 Ethereum Whitepaper (2014)

    Blockchain | White paper | Vitalik Buterin | Ethereum Foundation

    Ethereum generalises blockchain state transitions through a programmable virtual machine, contracts and a shared execution environment.

    Key learnings

    • Programmable settlement turns a ledger into an application platform.
    • General-purpose execution expands both innovation and attack surface.

    Topics: Ethereum, smart contracts, Web3

    Read the canonical source

58. #58 Ethereum: A Secure Decentralised Generalised Transaction Ledger (2014)

    Blockchain | Technical paper | Gavin Wood | Ethereum Foundation

    The Yellow Paper formally specifies Ethereum’s state transition, virtual machine, transaction execution and fee mechanics.

    Key learnings

    • Protocol implementations need an unambiguous executable-state specification.
    • Fees are a resource-allocation and denial-of-service control.

    Topics: Ethereum Yellow Paper, EVM, protocol specification

    Read the canonical source

59. #59 Majority is not Enough: Bitcoin Mining is Vulnerable (2014)

    Blockchain | Paper | Ittay Eyal and Emin Gün Sirer | Springer

    The selfish-mining analysis shows how strategic block withholding can earn disproportionate rewards under some network and miner assumptions.

    Key learnings

    • Incentive compatibility must be tested against strategic participants.
    • Protocol security can fail before an attacker controls a simple majority.

    Topics: selfish mining, Bitcoin security, incentives

    Read the canonical source

60. #60 The Bitcoin Backbone Protocol: Analysis and Applications (2015)

    Blockchain | Paper | Juan Garay, Aggelos Kiayias and Nikos Leonardos | Springer

    The paper formalises the Bitcoin ledger and proves consistency and growth properties under explicit network and adversary assumptions.

    Key learnings

    • Security claims become meaningful when assumptions and failure bounds are explicit.
    • Confirmation depth is a probabilistic risk control rather than absolute finality.

    Topics: Bitcoin backbone, consensus security, formal analysis

    Read the canonical source

61. #61 Zerocash: Decentralized Anonymous Payments from Bitcoin (2014)

    Blockchain | Paper | Eli Ben-Sasson et al. | IEEE

    Zerocash applies succinct zero-knowledge proofs to hide transaction origin, destination and value while preserving ledger validity.

    Key learnings

    • Zero knowledge can separate transaction validity from transaction disclosure.
    • Privacy systems introduce setup, complexity and compliance trade-offs.

    Topics: Zerocash, zero knowledge, privacy

    Read the canonical source

62. #62 IPFS: Content Addressed, Versioned, P2P File System (2014)

    Blockchain | Paper | Juan Benet | arXiv

    IPFS proposes a peer-to-peer, content-addressed file system combining distributed hash tables, Merkle DAGs and versioned data.

    Key learnings

    • Content addressing verifies what was retrieved independently of where it came from.
    • Availability and persistence still require incentives or operational commitments.

    Topics: IPFS, decentralised storage, content addressing

    Read the canonical source

63. #63 The Bitcoin Lightning Network: Scalable Off-Chain Instant Payments (2016)

    Blockchain | White paper | Joseph Poon and Thaddeus Dryja | Lightning Network

    Lightning uses bidirectional payment channels and hashed timelock contracts to move frequent payments off the base chain while retaining settlement recourse.

    Key learnings

    • Layer-two systems trade global on-chain execution for local state updates.
    • Liquidity, routing and channel management become core user-experience constraints.

    Topics: Lightning Network, layer two, payments

    Read the canonical source

64. #64 Polkadot: Vision for a Heterogeneous Multi-Chain Framework (2016)

    Blockchain | White paper | Gavin Wood | Web3 Foundation

    Polkadot proposes shared security and cross-chain messaging for heterogeneous state-transition systems connected through a relay chain.

    Key learnings

    • Interoperability requires common security and message-validity assumptions.
    • Shared infrastructure can specialise execution while pooling validator security.

    Topics: Polkadot, interoperability, multichain

    Read the canonical source

65. #65 Cosmos: A Network of Distributed Ledgers (2016)

    Blockchain | White paper | Jae Kwon and Ethan Buchman | Cosmos

    Cosmos outlines independent application blockchains connected through hubs, zones and an inter-blockchain communication protocol.

    Key learnings

    • Sovereign chains gain flexibility but must manage their own security economics.
    • Cross-chain communication creates new verification and failure domains.

    Topics: Cosmos, IBC, interoperability

    Read the canonical source

66. #66 Algorand: Scaling Byzantine Agreements for Cryptocurrencies (2017)

    Blockchain | Paper | Jing Chen and Silvio Micali | ACM

    Algorand uses cryptographic sortition to select small, unpredictable committees for Byzantine agreement in a permissionless setting.

    Key learnings

    • Random committee selection can reduce communication while limiting targeted attacks.
    • Stake distribution and participation assumptions remain central to security.

    Topics: Algorand, proof of stake, consensus

    Read the canonical source

67. #67 Ouroboros: A Provably Secure Proof-of-Stake Blockchain Protocol (2017)

    Blockchain | Paper | Aggelos Kiayias et al. | IACR

    Ouroboros presents a proof-of-stake protocol with formal security analysis using epochs, slots and stake-weighted leader selection.

    Key learnings

    • Proof-of-stake protocols can be analysed under explicit cryptographic models.
    • Time synchronisation, randomness and stake assumptions are security-critical.

    Topics: Ouroboros, proof of stake, Cardano

    Read the canonical source

68. #68 Filecoin: A Decentralized Storage Network (2017)

    Blockchain | White paper | Protocol Labs | Filecoin

    Filecoin designs a market for storage and retrieval backed by proofs that providers hold data over time and by blockchain-based incentives.

    Key learnings

    • Useful-resource consensus can connect tokens to verifiable services.
    • Economic incentives do not by themselves guarantee retrieval quality or durability.

    Topics: Filecoin, decentralised storage, proof of spacetime

    Read the canonical source

69. #69 ChainLink: A Decentralized Oracle Network (2017)

    Blockchain | White paper | Steve Ellis, Ari Juels and Sergey Nazarov | Chainlink

    Chainlink proposes decentralised oracle networks that aggregate external data and bridge smart contracts to off-chain systems.

    Key learnings

    • Smart contracts are only as trustworthy as their external inputs.
    • Oracle design must address source quality, aggregation and operator incentives.

    Topics: Chainlink, oracles, smart contracts

    Read the canonical source

70. #70 Bulletproofs: Short Proofs for Confidential Transactions and More (2018)

    Blockchain | Paper | Benedikt Bünz et al. | IEEE

    Bulletproofs provide short, non-interactive zero-knowledge range proofs without a trusted setup, suited to confidential transaction amounts.

    Key learnings

    • Removing trusted setup changes operational trust assumptions.
    • Proof size and verification cost remain practical privacy-design constraints.

    Topics: Bulletproofs, zero knowledge, confidential transactions

    Read the canonical source

71. #71 Scalable, Transparent, and Post-Quantum Secure Computational Integrity (2018)

    Blockchain | Paper | Eli Ben-Sasson et al. | IACR

    The STARK paper constructs transparent proof systems using hash-based cryptography, aiming for scalable verification without trusted setup.

    Key learnings

    • Transparent proofs reduce setup risk and can offer post-quantum assumptions.
    • Larger proofs may be acceptable when scalability and auditability improve.

    Topics: zk-STARK, zero knowledge, post-quantum

    Read the canonical source

72. #72 HotStuff: BFT Consensus with Linearity and Responsiveness (2019)

    Blockchain | Paper | Maofan Yin et al. | ACM

    HotStuff provides a leader-based Byzantine consensus design with linear communication and optimistic responsiveness, influencing several modern blockchains.

    Key learnings

    • Linear view changes help BFT protocols scale to larger validator sets.
    • Liveness depends on leader rotation and realistic timing assumptions.

    Topics: HotStuff, BFT, consensus

    Read the canonical source

73. #73 Snowflake to Avalanche: A Novel Metastable Consensus Protocol Family for Cryptocurrencies (2018)

    Blockchain | Paper | Team Rocket | Ava Labs

    The Avalanche family uses repeated random subsampling and metastable preference dynamics to reach probabilistic agreement without a fixed leader.

    Key learnings

    • Probabilistic sampling can deliver fast convergence with low communication.
    • Parameter choices and network assumptions must be validated for the target threat model.

    Topics: Avalanche, consensus, probabilistic protocols

    Read the canonical source

74. #74 Flash Boys 2.0: Frontrunning in Decentralized Exchanges, Miner Extractable Value, and Consensus Instability (2020)

    Blockchain | Paper | Philip Daian et al. | IEEE

    The paper documents transaction-ordering bots and formalises miner extractable value, connecting application-layer arbitrage to consensus incentives.

    Key learnings

    • Public mempools create markets for ordering and information advantage.
    • Application design can feed destabilising incentives back into consensus.

    Topics: MEV, decentralised exchanges, market structure

    Read the canonical source

75. #75 SoK: Decentralized Finance (DeFi) (2021)

    Blockchain | Paper | Sam M. Werner et al. | ACM

    This systematisation categorises DeFi protocols, primitives, risks and open research problems across exchanges, lending, derivatives and asset management.

    Key learnings

    • Composability accelerates innovation and transmits failures between protocols.
    • Economic and governance risks are as important as smart-contract correctness.

    Topics: DeFi, systematisation of knowledge, smart-contract risk

    Read the canonical source

76. #76 3rd Global Cryptoasset Benchmarking Study (2020)

    Blockchain | Report | Cambridge Centre for Alternative Finance | University of Cambridge

    The study surveys cryptoasset firms globally to map industry structure, employment, regulation, mining, custody and service-provider practices.

    Key learnings

    • Industry evidence should distinguish self-reported survey data from protocol telemetry.
    • Compliance and institutionalisation vary sharply by region and business model.

    Topics: cryptoassets, industry benchmarking, digital assets

    Read the canonical source

77. #77 DeFi Risks and the Decentralisation Illusion (2021)

    Blockchain | Report | Bank for International Settlements | BIS

    The BIS analyses DeFi’s rapid growth, leverage, stablecoin dependencies and governance concentration, questioning claims of full decentralisation.

    Key learnings

    • Operational decentralisation can coexist with concentrated governance and economic power.
    • Stablecoins and oracles can become critical points of failure.

    Topics: DeFi risk, financial stability, BIS

    Read the canonical source

78. #78 Embedded Supervision: How to Build Regulation into Blockchain Finance (2019)

    Blockchain | Working paper | Raphael Auer | Bank for International Settlements

    The paper explores whether supervisors can use distributed-ledger data and economic consensus to monitor compliance with less reliance on repeated data collection.

    Key learnings

    • Machine-readable compliance can reduce reporting friction when data semantics are reliable.
    • Supervision cannot be automated safely without governance, identity and legal clarity.

    Topics: embedded supervision, regtech, blockchain finance

    Read the canonical source

79. #79 High-Level Recommendations for the Regulation, Supervision and Oversight of Global Stablecoin Arrangements (2023)

    Blockchain | Report | Financial Stability Board | FSB

    The FSB sets internationally coordinated recommendations covering governance, risk management, redemption, data, recovery and cross-border cooperation for stablecoins.

    Key learnings

    • Same-activity and same-risk regulation depends on clearly mapped functions and entities.
    • Credible redemption and reserve arrangements are fundamental to stablecoin trust.

    Topics: stablecoins, regulation, financial stability

    Read the canonical source

80. #80 Blueprint for the Future Monetary System: Improving the Old, Enabling the New (2023)

    Blockchain | Report | Bank for International Settlements | BIS

    The BIS proposes a unified-ledger concept connecting tokenised central-bank money, commercial-bank money and assets on programmable infrastructure.

    Key learnings

    • Tokenisation’s value depends on interoperability with trusted money and legal claims.
    • Programmability should preserve governance, privacy and monetary-system integrity.

    Topics: tokenisation, CBDC, future of money

    Read the canonical source

81. #81 CBDC Technologies: How Centralised is a CBDC? (2021)

    Blockchain | Working paper | Raphael Auer and Rainer Böhme | Bank for International Settlements

    The paper compares direct, intermediated, hybrid and distributed technical architectures for central bank digital currency.

    Key learnings

    • A digital currency’s policy model does not dictate a single ledger architecture.
    • Resilience, privacy and intermediaries must be co-designed with operational control.

    Topics: CBDC, digital currency architecture, central banking

    Read the canonical source

82. #82 Distributed Ledger Technology and Blockchain (2017)

    Blockchain | Report | World Bank Group | World Bank

    The World Bank note explains DLT concepts, permission models, use cases and constraints for development and financial-sector decision-makers.

    Key learnings

    • Blockchain is one database and coordination option, not a universal solution.
    • Institutional, legal and data-quality constraints often dominate technical feasibility.

    Topics: DLT, development finance, technology assessment

    Read the canonical source

83. #83 End-to-End Arguments in System Design (1984)

    Technology | Paper | Jerome H. Saltzer, David P. Reed and David D. Clark | ACM

    The paper argues that certain guarantees—such as correctness, security and duplicate suppression—can only be fully implemented at application endpoints.

    Key learnings

    • Lower layers can assist, but endpoints must verify the guarantees they need.
    • Placing functions at the right layer improves simplicity and generality.

    Topics: internet architecture, end-to-end principle, system design

    Read the canonical source

84. #84 Brewer’s Conjecture and the Feasibility of Consistent, Available, Partition-Tolerant Web Services (2002)

    Technology | Paper | Seth Gilbert and Nancy Lynch | ACM

    This formal treatment of the CAP conjecture proves limits on simultaneously providing strong consistency and availability during network partitions.

    Key learnings

    • Distributed-system guarantees must be stated under failure conditions.
    • CAP is not a licence to ignore latency, recovery or consistency levels outside partitions.

    Topics: CAP theorem, distributed systems, availability

    Read the canonical source

85. #85 The Part-Time Parliament (1998)

    Technology | Paper | Leslie Lamport | ACM

    Lamport presents Paxos through a parliamentary allegory to explain how unreliable participants can agree on ordered decisions.

    Key learnings

    • Consensus separates safety from liveness under failures.
    • A correct protocol still needs an understandable and operable implementation.

    Topics: Paxos, consensus, distributed systems

    Read the canonical source

86. #86 The Google File System (2003)

    Technology | Paper | Sanjay Ghemawat, Howard Gobioff and Shun-Tak Leung | ACM

    GFS designs a distributed file system for large files, streaming access and routine component failure on commodity clusters.

    Key learnings

    • Design for the workload and expected failures rather than abstract generality.
    • Replication, checksums and rapid recovery make failure a normal operating state.

    Topics: Google File System, distributed storage, cloud computing

    Read the canonical source

87. #87 MapReduce: Simplified Data Processing on Large Clusters (2004)

    Technology | Paper | Jeffrey Dean and Sanjay Ghemawat | USENIX

    MapReduce gives developers map and reduce functions while the runtime handles partitioning, scheduling, locality and fault recovery across a cluster.

    Key learnings

    • A constrained programming model can hide complex distributed execution.
    • Data locality is a core performance optimisation at scale.

    Topics: MapReduce, big data, distributed computing

    Read the canonical source

88. #88 Bigtable: A Distributed Storage System for Structured Data (2006)

    Technology | Paper | Fay Chang et al. | USENIX

    Bigtable introduces a sparse, distributed, persistent multidimensional map that scales structured data across commodity servers.

    Key learnings

    • A deliberately limited data model can enable predictable horizontal scale.
    • Storage layout and access patterns should be designed together.

    Topics: Bigtable, NoSQL, distributed databases

    Read the canonical source

89. #89 Dynamo: Amazon’s Highly Available Key-value Store (2007)

    Technology | Paper | Giuseppe DeCandia et al. | ACM

    Dynamo combines consistent hashing, version vectors, quorum-like reads and writes, and application-assisted conflict resolution for high availability.

    Key learnings

    • Some services should remain writable during failures and reconcile later.
    • Consistency is a business and application decision, not only a database setting.

    Topics: Dynamo, eventual consistency, distributed databases

    Read the canonical source

90. #90 The Chubby Lock Service for Loosely-Coupled Distributed Systems (2006)

    Technology | Paper | Mike Burrows | USENIX

    Chubby provides coarse-grained locking and reliable small-file storage so distributed applications can coordinate without implementing consensus themselves.

    Key learnings

    • A narrow coordination service can simplify many distributed applications.
    • Clients must handle sessions, leases and service unavailability explicitly.

    Topics: Chubby, coordination, distributed systems

    Read the canonical source

91. #91 Dapper, a Large-Scale Distributed Systems Tracing Infrastructure (2010)

    Technology | Paper | Benjamin H. Sigelman et al. | Google Research

    Dapper describes low-overhead request tracing across large service graphs using shared trace identifiers, sampling and central analysis.

    Key learnings

    • Observability context must propagate across service boundaries.
    • Sampling policy should preserve diagnostic value while controlling cost.

    Topics: distributed tracing, observability, microservices

    Read the canonical source

92. #92 Spanner: Google’s Globally-Distributed Database (2012)

    Technology | Paper | James C. Corbett et al. | USENIX

    Spanner combines global replication and distributed transactions with TrueTime, an API that exposes bounded clock uncertainty.

    Key learnings

    • Time can be an explicit system primitive rather than an invisible assumption.
    • Strong global consistency requires infrastructure and latency trade-offs.

    Topics: Spanner, distributed databases, TrueTime

    Read the canonical source

93. #93 In Search of an Understandable Consensus Algorithm (2014)

    Technology | Paper | Diego Ongaro and John Ousterhout | USENIX

    Raft decomposes consensus into leader election, log replication and safety, prioritising understandability while matching Paxos-style guarantees.

    Key learnings

    • Understandability is a reliability feature for implemented protocols.
    • Operational behaviour during elections and membership changes deserves explicit design.

    Topics: Raft, consensus, distributed systems

    Read the canonical source

94. #94 Conflict-Free Replicated Data Types (2011)

    Technology | Paper | Marc Shapiro et al. | Springer

    CRDTs define replicated data structures whose concurrent updates merge deterministically without central coordination across participating replicas.

    Key learnings

    • Convergence can be designed into data types rather than repaired after conflicts.
    • Coordination-free updates require accepting specific semantic constraints.

    Topics: CRDT, replication, offline-first

    Read the canonical source

95. #95 Large-Scale Cluster Management at Google with Borg (2015)

    Technology | Paper | Abhishek Verma et al. | ACM

    The Borg paper documents scheduling, admission control, isolation, availability and operational lessons from managing mixed workloads at enormous scale.

    Key learnings

    • Declarative desired state simplifies large-fleet operations.
    • Utilisation and reliability improve when workload diversity is scheduled together.

    Topics: Borg, containers, cluster management

    Read the canonical source

96. #96 Omega: Flexible, Scalable Schedulers for Large Compute Clusters (2013)

    Technology | Paper | Malte Schwarzkopf et al. | ACM

    Omega explores parallel, specialised schedulers that coordinate through optimistic concurrency over shared cluster state.

    Key learnings

    • One central scheduler can become a scalability and innovation bottleneck.
    • Shared-state concurrency needs conflict detection and carefully bounded retries.

    Topics: Omega, scheduling, cloud infrastructure

    Read the canonical source

97. #97 Architectural Styles and the Design of Network-based Software Architectures (2000)

    Technology | Dissertation | Roy Thomas Fielding | University of California, Irvine

    Fielding derives REST as a set of architectural constraints for scalable, evolvable networked systems and evaluates the Web through that lens.

    Key learnings

    • REST is a constraint system, not a synonym for JSON over HTTP.
    • Uniform interfaces and statelessness trade local efficiency for system-wide evolvability.

    Topics: REST, API design, web architecture

    Read the canonical source

98. #98 The Anatomy of a Large-Scale Hypertextual Web Search Engine (1998)

    Technology | Paper | Sergey Brin and Lawrence Page | Stanford University

    The paper describes Google’s early architecture, including crawling, indexing and PageRank-based ranking for a rapidly growing web.

    Key learnings

    • Link structure can serve as a signal of authority as well as navigation.
    • Search quality depends on an end-to-end data and serving system.

    Topics: PageRank, search engines, information retrieval

    Read the canonical source

99. #99 New Directions in Cryptography (1976)

    Technology | Paper | Whitfield Diffie and Martin E. Hellman | IEEE

    Diffie and Hellman introduce public-key cryptography and a key-agreement method that lets parties establish secrets over an insecure channel.

    Key learnings

    • Public-key methods transform key distribution at internet scale.
    • Authentication is still required to prevent active intermediaries.

    Topics: Diffie-Hellman, public-key cryptography, security

    Read the canonical source

100. #100 A Method for Obtaining Digital Signatures and Public-Key Cryptosystems (1978)

     Technology | Paper | Ronald L. Rivest, Adi Shamir and Leonard Adleman | ACM

     RSA provides a practical public-key construction for encryption and digital signatures based on the difficulty of integer factorisation.

     Key learnings

     • One-way mathematical problems can support identity and confidentiality protocols.
     • Key size, padding and implementation determine real-world security.

     Topics: RSA, digital signatures, cryptography

     Read the canonical source

101. #101 The Protection of Information in Computer Systems (1975)

     Technology | Paper | Jerome H. Saltzer and Michael D. Schroeder | IEEE

     The paper presents enduring security-design principles including least privilege, complete mediation, open design and psychological acceptability.

     Key learnings

     • Security principles remain useful across changing technologies.
     • Controls fail when they are hard to use correctly or easy to bypass.

     Topics: least privilege, security architecture, secure design

     Read the canonical source

102. #102 Protecting Respondents’ Identities in Microdata Release (2002)

     Technology | Paper | Latanya Sweeney | IEEE

     The k-anonymity paper formalises how released records can be made indistinguishable on quasi-identifiers to reduce re-identification risk.

     Key learnings

     • Removing direct identifiers is not sufficient for anonymity.
     • Privacy guarantees depend on attacker knowledge and data diversity.

     Topics: k-anonymity, data privacy, anonymisation

     Read the canonical source

103. #103 Calibrating Noise to Sensitivity in Private Data Analysis (2006)

     Technology | Paper | Cynthia Dwork et al. | Springer

     This foundational differential-privacy paper adds calibrated randomness so aggregate analyses reveal limited information about any individual record.

     Key learnings

     • Privacy can be expressed as a measurable bound rather than an informal promise.
     • A privacy budget creates an explicit utility-versus-disclosure trade-off.

     Topics: differential privacy, privacy engineering, data analysis

     Read the canonical source

104. #104 Spectre Attacks: Exploiting Speculative Execution (2019)

     Technology | Paper | Paul Kocher et al. | IEEE

     Spectre demonstrates that speculative execution can leak protected data through microarchitectural side channels across isolation boundaries.

     Key learnings

     • Performance optimisations can create security effects outside the architectural model.
     • Mitigation requires coordinated changes across hardware, operating systems and compilers.

     Topics: Spectre, side-channel attacks, hardware security

     Read the canonical source

105. #105 Meltdown: Reading Kernel Memory from User Space (2018)

     Technology | Paper | Moritz Lipp et al. | USENIX

     Meltdown exploits out-of-order execution to bypass memory isolation and infer kernel data through a cache side channel on affected processors.

     Key learnings

     • Hardware-enforced isolation assumptions need empirical adversarial testing.
     • Emergency mitigations can impose substantial performance and operational costs.

     Topics: Meltdown, processor security, side channels

     Read the canonical source

106. #106 The NIST Cybersecurity Framework (CSF) 2.0 (2024)

     Technology | Standard | National Institute of Standards and Technology | NIST

     CSF 2.0 organises cybersecurity outcomes under Govern, Identify, Protect, Detect, Respond and Recover for organisations of any size or sector.

     Key learnings

     • Cybersecurity governance belongs with enterprise risk governance.
     • Outcome-based profiles help prioritise controls without prescribing one technology stack.

     Topics: NIST CSF, cybersecurity, risk management

     Read the canonical source

107. #107 Zero Trust Architecture (2020)

     Technology | Standard | Scott Rose et al. | NIST

     NIST SP 800-207 defines zero trust as resource-focused access decisions based on continuously evaluated identity, device and contextual signals.

     Key learnings

     • Network location alone should never create implicit organisational trust.
     • Zero trust is an architecture and operating model, not a single product.

     Topics: zero trust, identity security, NIST

     Read the canonical source

108. #108 The NIST Definition of Cloud Computing (2011)

     Technology | Standard | Peter Mell and Timothy Grance | NIST

     NIST defines cloud computing through five essential characteristics, three service models and four distinct deployment models for shared computing resources.

     Key learnings

     • Elasticity, measured service and resource pooling distinguish cloud from ordinary hosting.
     • Clear service-model boundaries improve responsibility and risk allocation.

     Topics: cloud computing, NIST, service models

     Read the canonical source

109. #109 Accelerate State of DevOps Report 2024 (2024)

     Technology | Report | DORA | Google Cloud

     DORA synthesises survey research on software-delivery performance, organisational outcomes, platform engineering, AI adoption and developer experience.

     Key learnings

     • Delivery performance is shaped by technical practices and organisational conditions together.
     • AI tooling should be evaluated through system outcomes, not adoption counts alone.

     Topics: DORA, DevOps, software delivery

     Read the canonical source

110. #110 The RISC-V Instruction Set Manual, Volume I: User-Level ISA (2019)

     Technology | Paper | Andrew Waterman and Krste Asanović | RISC-V International

     The RISC-V specification defines an open, modular instruction-set architecture designed for implementation flexibility and long-term extensibility.

     Key learnings

     • Open standards can expand hardware innovation and reduce architectural lock-in.
     • A small base plus extensions supports diversity while preserving compatibility.

     Topics: RISC-V, computer architecture, open standards

     Read the canonical source

111. #111 In-Datacenter Performance Analysis of a Tensor Processing Unit (2017)

     Technology | Paper | Norman P. Jouppi et al. | ACM

     The TPU paper evaluates a domain-specific inference accelerator in production workloads and compares performance and energy efficiency with CPUs and GPUs.

     Key learnings

     • Domain-specific hardware can outperform general-purpose processors for stable workloads.
     • Total serving latency and energy matter more than peak arithmetic throughput.

     Topics: TPU, AI hardware, datacentres

     Read the canonical source

112. #112 The Transport Layer Security (TLS) Protocol Version 1.3 (2018)

     Technology | Standard | Eric Rescorla | Internet Engineering Task Force

     TLS 1.3 simplifies the protocol, removes obsolete cryptography, encrypts more of the handshake and reduces connection setup latency.

     Key learnings

     • Security improves when legacy options and negotiation complexity are removed.
     • Protocol deployment must balance compatibility with strong defaults.

     Topics: TLS 1.3, internet security, encryption

     Read the canonical source

113. #113 QUIC: A UDP-Based Multiplexed and Secure Transport (2021)

     Technology | Standard | Jana Iyengar and Martin Thomson | Internet Engineering Task Force

     QUIC integrates secure transport, multiplexed streams and connection migration over UDP to reduce latency and avoid head-of-line blocking between streams.

     Key learnings

     • Transport and cryptographic handshakes can be co-designed for lower latency.
     • Connection identifiers support mobility but require privacy-aware handling.

     Topics: QUIC, internet protocols, web performance

     Read the canonical source

114. #114 ENISA Threat Landscape 2024 (2024)

     Technology | Report | European Union Agency for Cybersecurity | ENISA

     ENISA analyses major cyber threats, incidents, motivations and sector patterns to support risk prioritisation across the European threat landscape.

     Key learnings

     • Threat intelligence becomes useful when mapped to assets, sectors and likely impacts.
     • Availability attacks, ransomware and social engineering require layered resilience.

     Topics: ENISA, threat intelligence, cybersecurity

     Read the canonical source

115. #115 Verizon 2025 Data Breach Investigations Report (2025)

     Technology | Report | Verizon Business | Verizon

     The DBIR analyses a large multi-contributor incident dataset to identify breach patterns, attack paths, affected assets and recurring human and third-party factors.

     Key learnings

     • Incident evidence should drive control priorities more than headline vulnerability lists.
     • Credential abuse, exploitation and human involvement demand complementary controls.

     Topics: DBIR, data breaches, cybersecurity

     Read the canonical source