{
  "name": "GPAI Wiki",
  "description": "A curated archive of research, roadmaps, implementations and open problems on the path to general-purpose artificial intelligence (GPAI).",
  "home_page_url": "https://gpai-wiki.github.io",
  "feed_url": "https://gpai-wiki.github.io/index.json",
  "generated": "2026-08-24T12:11:54+00:00",
  "sections": [
    { "label": "news", "url": "https://gpai-wiki.github.io/news/" },
    { "label": "papers", "url": "https://gpai-wiki.github.io/papers/" },
    { "label": "roadmaps", "url": "https://gpai-wiki.github.io/roadmaps/" },
    { "label": "systems", "url": "https://gpai-wiki.github.io/systems/" },
    { "label": "wiki", "url": "https://gpai-wiki.github.io/wiki/" },
    { "label": "open problems", "url": "https://gpai-wiki.github.io/open-problems/" },
    { "label": "pages", "url": "https://gpai-wiki.github.io/pages/" },
    { "label": "contribute", "url": "https://gpai-wiki.github.io/contribute/" }
  ],
  "items": [
    {
      "id": "https://gpai-wiki.github.io/open-problems/measuring-generality/",
      "uid": "OP-003",
      "collection": "problems",
      "title": "A measure of generality that benchmarks cannot be trained on",
      "url": "https://gpai-wiki.github.io/open-problems/measuring-generality/",
      "date_published": "2026-08-22T00:00:00+00:00",
      "summary": "Every fixed benchmark becomes a training target and stops measuring what it measured. Held-out task suites postpone the problem rather than solving it.\n",
      "tags": ["evaluation","generality"],
      "status": "contested",
      "kind": "",
      "source": "",
      "authors": ""
    },
    {
      "id": "https://gpai-wiki.github.io/wiki/scaling-laws/",
      "uid": "",
      "collection": "concepts",
      "title": "Scaling laws",
      "url": "https://gpai-wiki.github.io/wiki/scaling-laws/",
      "date_published": "2026-08-22T00:00:00+00:00",
      "summary": "Empirical power-law relationships between model size, dataset size, training compute and loss, which hold across many orders of magnitude and permit extrapolation.\n",
      "tags": ["scaling","generality"],
      "status": "",
      "kind": "",
      "source": "",
      "authors": ""
    },
    {
      "id": "https://gpai-wiki.github.io/systems/nanogpt/",
      "uid": "S-001",
      "collection": "systems",
      "title": "nanoGPT",
      "url": "https://gpai-wiki.github.io/systems/nanogpt/",
      "date_published": "2026-08-22T00:00:00+00:00",
      "summary": "A deliberately small, readable implementation of GPT training and fine-tuning — roughly 300 lines of model code and 300 of training loop. It is the standard reference point for understanding what a decoder-only Transformer actually does at the level of tensors.\n",
      "tags": ["architecture","training","reproduction"],
      "status": "maintained",
      "kind": "",
      "source": "",
      "authors": ""
    },
    {
      "id": "https://gpai-wiki.github.io/papers/language-models-are-few-shot-learners/",
      "uid": "P-0004",
      "collection": "papers",
      "title": "Language Models are Few-Shot Learners",
      "url": "https://gpai-wiki.github.io/papers/language-models-are-few-shot-learners/",
      "date_published": "2026-08-22T00:00:00+00:00",
      "summary": "Demonstrates that a 175B-parameter autoregressive language model performs a wide range of unseen tasks from instructions and a handful of examples in its context, with no gradient updates. This is the first large-scale evidence that generality can emerge from scale alone rather than from task-specific architecture.\n",
      "tags": ["scaling","in-context-learning","generality"],
      "status": "",
      "kind": "",
      "source": "",
      "authors": ["T. B. Brown","B. Mann","N. Ryder","M. Subbiah","J. Kaplan","et al."]
    },
    {
      "id": "https://gpai-wiki.github.io/open-problems/continual-learning-without-forgetting/",
      "uid": "OP-002",
      "collection": "problems",
      "title": "Continual learning at deployment scale",
      "url": "https://gpai-wiki.github.io/open-problems/continual-learning-without-forgetting/",
      "date_published": "2026-08-21T00:00:00+00:00",
      "summary": "Every known mitigation for catastrophic forgetting either accumulates stiffness, grows storage without limit, or grows parameters without limit. None survives a year-long stream.\n",
      "tags": ["memory","continual-learning"],
      "status": "open",
      "kind": "",
      "source": "",
      "authors": ""
    },
    {
      "id": "https://gpai-wiki.github.io/wiki/world-model/",
      "uid": "",
      "collection": "concepts",
      "title": "World model",
      "url": "https://gpai-wiki.github.io/wiki/world-model/",
      "date_published": "2026-08-21T00:00:00+00:00",
      "summary": "A learned predictive model of environment dynamics, used to evaluate candidate actions without executing them.\n",
      "tags": ["world-models","planning"],
      "status": "",
      "kind": "",
      "source": "",
      "authors": ""
    },
    {
      "id": "https://gpai-wiki.github.io/wiki/credit-assignment/",
      "uid": "",
      "collection": "concepts",
      "title": "Credit assignment",
      "url": "https://gpai-wiki.github.io/wiki/credit-assignment/",
      "date_published": "2026-08-21T00:00:00+00:00",
      "summary": "Determining which of an agent's earlier decisions were responsible for a later outcome.\n",
      "tags": ["credit-assignment","agents","learning"],
      "status": "",
      "kind": "",
      "source": "",
      "authors": ""
    },
    {
      "id": "https://gpai-wiki.github.io/roadmaps/the-next-decade-in-ai/",
      "uid": "R-002",
      "collection": "roadmaps",
      "title": "The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence",
      "url": "https://gpai-wiki.github.io/roadmaps/the-next-decade-in-ai/",
      "date_published": "2026-08-21T00:00:00+00:00",
      "summary": "Argues that scaling statistical learning alone will not produce robust intelligence, and proposes hybrid architectures with explicit symbolic machinery, large-scale knowledge, and cognitive models as prerequisites.\n",
      "tags": ["hybrid","symbolic","robustness"],
      "status": "contested",
      "kind": "",
      "source": "https://arxiv.org/abs/2002.06177",
      "authors": "Gary Marcus"
    },
    {
      "id": "https://gpai-wiki.github.io/papers/world-models/",
      "uid": "P-0003",
      "collection": "papers",
      "title": "World Models",
      "url": "https://gpai-wiki.github.io/papers/world-models/",
      "date_published": "2026-08-21T00:00:00+00:00",
      "summary": "Trains a compact generative model of an environment and then trains a controller almost entirely inside it. The demonstration that a policy learned in a learned simulation can transfer back to the real environment is the load-bearing result.\n",
      "tags": ["world-models","planning","agents"],
      "status": "",
      "kind": "",
      "source": "",
      "authors": ["D. Ha","J. Schmidhuber"]
    },
    {
      "id": "https://gpai-wiki.github.io/papers/overcoming-catastrophic-forgetting/",
      "uid": "P-0002",
      "collection": "papers",
      "title": "Overcoming Catastrophic Forgetting in Neural Networks",
      "url": "https://gpai-wiki.github.io/papers/overcoming-catastrophic-forgetting/",
      "date_published": "2026-08-21T00:00:00+00:00",
      "summary": "Proposes elastic weight consolidation, which slows learning on weights that a Fisher information estimate marks as important to previously learned tasks. It is the clearest early statement that catastrophic forgetting is a constraint problem, not a capacity problem — and the framing survives even where the method does not.\n",
      "tags": ["memory","continual-learning"],
      "status": "",
      "kind": "",
      "source": "",
      "authors": ["J. Kirkpatrick","R. Pascanu","N. Rabinowitz","J. Veness","G. Desjardins","A. A. Rusu","et al."]
    },
    {
      "id": "https://gpai-wiki.github.io/open-problems/stable-long-horizon-credit-assignment/",
      "uid": "OP-001",
      "collection": "problems",
      "title": "Stable credit assignment over month-long horizons",
      "url": "https://gpai-wiki.github.io/open-problems/stable-long-horizon-credit-assignment/",
      "date_published": "2026-08-20T00:00:00+00:00",
      "summary": "Discounting makes long-horizon consequences invisible; undiscounted objectives are unstable. No known method scales temporal credit assignment past a few thousand steps.\n",
      "tags": ["credit-assignment","agents"],
      "status": "open",
      "kind": "",
      "source": "",
      "authors": ""
    },
    {
      "id": "https://gpai-wiki.github.io/wiki/general-purpose-ai/",
      "uid": "",
      "collection": "concepts",
      "title": "General-purpose AI",
      "url": "https://gpai-wiki.github.io/wiki/general-purpose-ai/",
      "date_published": "2026-08-20T00:00:00+00:00",
      "summary": "A system whose competence is not bounded to the tasks it was trained on, and which can be directed at new tasks by description rather than by retraining.\n",
      "tags": ["generality","definitions"],
      "status": "",
      "kind": "",
      "source": "",
      "authors": ""
    },
    {
      "id": "https://gpai-wiki.github.io/wiki/continual-learning/",
      "uid": "",
      "collection": "concepts",
      "title": "Continual learning",
      "url": "https://gpai-wiki.github.io/wiki/continual-learning/",
      "date_published": "2026-08-20T00:00:00+00:00",
      "summary": "Learning from a non-stationary stream of experience without retraining from scratch and without losing what was already learned.\n",
      "tags": ["memory","continual-learning"],
      "status": "",
      "kind": "",
      "source": "",
      "authors": ""
    },
    {
      "id": "https://gpai-wiki.github.io/wiki/catastrophic-forgetting/",
      "uid": "",
      "collection": "concepts",
      "title": "Catastrophic forgetting",
      "url": "https://gpai-wiki.github.io/wiki/catastrophic-forgetting/",
      "date_published": "2026-08-20T00:00:00+00:00",
      "summary": "The abrupt loss of previously learned capability when a network is trained on new data, caused by the same weights being reused for both.\n",
      "tags": ["memory","continual-learning"],
      "status": "",
      "kind": "",
      "source": "",
      "authors": ""
    },
    {
      "id": "https://gpai-wiki.github.io/roadmaps/a-path-towards-autonomous-machine-intelligence/",
      "uid": "R-001",
      "collection": "roadmaps",
      "title": "A Path Towards Autonomous Machine Intelligence",
      "url": "https://gpai-wiki.github.io/roadmaps/a-path-towards-autonomous-machine-intelligence/",
      "date_published": "2026-08-20T00:00:00+00:00",
      "summary": "A modular architecture in which a configurator, a learned world model, a cost module and an actor are trained largely by self-supervised prediction in representation space rather than pixel space. The central proposal is JEPA: predict abstract representations of the future, not the future itself.\n",
      "tags": ["world-models","architecture","self-supervised"],
      "status": "active",
      "kind": "",
      "source": "https://openreview.net/forum?id=BZ5a1r-kVsf",
      "authors": "Yann LeCun"
    },
    {
      "id": "https://gpai-wiki.github.io/papers/attention-is-all-you-need/",
      "uid": "P-0001",
      "collection": "papers",
      "title": "Attention Is All You Need",
      "url": "https://gpai-wiki.github.io/papers/attention-is-all-you-need/",
      "date_published": "2026-08-20T00:00:00+00:00",
      "summary": "Introduces the Transformer, a sequence model built entirely from attention and feed-forward layers with no recurrence or convolution. Its significance for general-purpose systems is less the translation result than the property that made it the substrate for everything since: it parallelises over sequence length, so capability became a question of scale rather than of architecture search.\n",
      "tags": ["architecture","attention","scaling"],
      "status": "",
      "kind": "",
      "source": "",
      "authors": ["A. Vaswani","N. Shazeer","N. Parmar","J. Uszkoreit","L. Jones","A. N. Gomez","Ł. Kaiser","I. Polosukhin"]
    },
    {
      "id": "https://gpai-wiki.github.io/news/eu-ai-act-gpai-obligations/",
      "uid": "N-0003",
      "collection": "news",
      "title": "EU AI Act obligations for general-purpose AI models take effect",
      "url": "https://gpai-wiki.github.io/news/eu-ai-act-gpai-obligations/",
      "date_published": "2025-08-02T00:00:00+00:00",
      "summary": "The general-purpose AI provisions of Regulation (EU) 2024/1689 became applicable, imposing documentation, copyright-policy and training-data-summary duties on providers of GPAI models placed on the EU market, with additional obligations for models deemed to carry systemic risk.\n",
      "tags": ["policy","generality"],
      "status": "",
      "kind": "policy",
      "source": "https://eur-lex.europa.eu/eli/reg/2024/1689/oj",
      "authors": ""
    },
    {
      "id": "https://gpai-wiki.github.io/news/deepseek-r1-rl-reasoning/",
      "uid": "N-0002",
      "collection": "news",
      "title": "DeepSeek-R1 shows reasoning emerging from reinforcement learning alone",
      "url": "https://gpai-wiki.github.io/news/deepseek-r1-rl-reasoning/",
      "date_published": "2025-01-22T00:00:00+00:00",
      "summary": "Reports that long-form reasoning behaviour can be elicited by reinforcement learning on outcome rewards without a supervised fine-tuning stage, with weights released. The claim that matters is the negative one: the supervised reasoning traces widely assumed to be necessary appear not to be.\n",
      "tags": ["reasoning","training"],
      "status": "",
      "kind": "research",
      "source": "https://arxiv.org/abs/2501.12948",
      "authors": ""
    },
    {
      "id": "https://gpai-wiki.github.io/news/llama-3-herd-of-models/",
      "uid": "N-0001",
      "collection": "news",
      "title": "Llama 3.1 405B released with open weights",
      "url": "https://gpai-wiki.github.io/news/llama-3-herd-of-models/",
      "date_published": "2024-07-31T00:00:00+00:00",
      "summary": "Meta published the Llama 3 technical report alongside open weights for models up to 405B parameters. The significance for general-purpose work is distributional rather than architectural: capability near the contemporary frontier became something a research group could run and modify locally rather than only query through an API.\n",
      "tags": ["scaling","generality"],
      "status": "",
      "kind": "release",
      "source": "https://arxiv.org/abs/2407.21783",
      "authors": ""
    }
  ]
}