Skip to content
ContentLora

    Tip: press / anywhere to search.

    Explainer

    Artificial intelligence: a crash course

    The large language models behind today's chatbots are mostly built on the Transformer, an architecture based on attention, and their prediction error falls predictably as models, data and compute grow.[1][2] As of November 2026 the main public questions are who controls the most capable models, whether AI is taking jobs, whether training on copyrighted work is legal, and how states and safety researchers keep up. Anthropic has held back its most capable model from general release until its cybersecurity safeguards improve, and OpenAI rates its newest model at a Critical level of cybersecurity capability.[3][4]

    Editor reviewedUpdated Explainer49 claims · 84% primaryArtificial intelligence, Frontier AI

    This crash course is a reading order for our artificial intelligence pages. It starts with how the technology works, then the companies that build it, and ends with the live trackers and debates about jobs, copyright, law and safety. Each step links to a fuller page with its own sources.

    How today’s AI works

    Chatbots and coding assistants run on large language models. These are neural networks built mostly on the Transformer, a 2017 architecture based on attention,[1] and their prediction error falls in a predictable way as models, datasets and computing power grow.[2] A raw model only predicts text. Extra training with human feedback turns it into an assistant: in one early study, people preferred answers from a 1.3-billion-parameter model trained this way to those of a model 100 times larger.[5][6] Our explainer on how large language models work covers this in full.

    Two newer ideas define the current frontier. Reasoning models are trained with large-scale reinforcement learning to work through a chain of steps before answering,[7] and they improve by spending more computing power at answer time.[8] Agents let a model direct its own steps and tools in a loop, checking results as they go.[9][10] See how reasoning models work and how AI agents work.

    Pictures and video come from a different family. Diffusion models build a picture through a series of denoising steps, guided by a text prompt.[11][12] Our page on how AI image and video generators work also covers watermarking and provenance labels.

    Open and closed models

    A key split is who can get at a model’s weights, the learned parameters that turn an input into an output.[13] Many developers give limited or no public access to their most advanced models’ weights, while Meta, Google, Microsoft and Mistral have released models with widely available weights.[14] OpenAI released two open-weight models under the Apache 2.0 licence in August 2025.[15] “Open weights” is not the same as open source: the Open Source Initiative’s definition also requires freedom to use the model for any purpose and enough information about the training data to build a substantially equivalent system.[16][17] The open-weight vs closed models explainer sets out the trade-offs and the policy positions.

    Measuring how capable any of these models is turns out to be hard. Models now score over 90% on popular benchmarks such as MMLU, which is why harder tests such as Humanity’s Last Exam were created,[18] and researchers now also measure how long a task an agent can finish on its own.[19] See how AI capabilities are measured.

    The frontier labs

    This course covers four of the companies building frontier models.

    • Anthropic, founded in 2021 by former OpenAI executives, develops the Claude models and is a public benefit corporation.[20][21]
    • OpenAI released ChatGPT in November 2022, and since October 2025 its business has been a public benefit corporation in which its nonprofit, the OpenAI Foundation, holds a stake valued at about $130 billion.[22][23]
    • Google DeepMind was formed in April 2023 by merging DeepMind with Google’s Brain team, and develops Google’s Gemini models.[24][25]
    • xAI, founded in 2023, develops the Grok models and became a SpaceX subsidiary in February 2026.[26][27]

    According to the 2026 AI Index, US and Chinese models have traded places at the top of performance rankings several times since early 2025.[28] The frontier AI tracker follows model releases and the June 2026 episode in which US export controls briefly forced Anthropic to suspend its newest models.[29][30] How these labs are funded by, and compete with, the largest tech companies is covered in our Big Tech crash course.

    Jobs. A Stanford working paper using payroll data to June 2026, and the Budget Lab at Yale as of October 2025, found no broad, economy-wide job disruption from generative AI.[31][32] The Stanford paper, which is not peer reviewed, found a widening gap for 22- to 25-year-olds in AI-exposed occupations, mainly through less hiring, and its authors call the patterns early descriptive indicators rather than causal estimates.[33][34][35] The AI and jobs debate weighs the evidence.

    Copyright. US courts have not settled the question. Two US judges found that training on books was fair use on the records before them, and Anthropic’s $1.5 billion settlement with book authors received final approval in July 2026.[36][37][38] The New York Times sued OpenAI and Microsoft in December 2023, alleging their models were built by copying millions of its articles.[39] Follow it in the AI copyright lawsuits tracker.

    Law. In the US, several states have passed their own AI laws: California and New York have passed laws aimed at developers of frontier models, and Colorado rewrote its 2024 AI Act in 2026.[40][41][42] A December 2025 executive order set up a federal effort to challenge state AI laws.[43] The state AI laws tracker follows them, and the EU AI Act explainer covers Europe’s law, which became generally applicable, with some exceptions, on 2 August 2026.[44] Our tech policy crash course puts these laws alongside the wider rules on platforms and chips.

    Safety and where to go next

    Capabilities are moving fast: in January 2026 METR estimated that the length of tasks AI agents can complete has doubled about every 89 days since 2024.[45] The International AI Safety Report 2026 says reliable pre-release testing has become harder because models increasingly tell test settings from real use.[46] The AI safety tracker records incidents and research, and the AI safety crash course explains the field.

    Readers who want more depth on the technology itself should continue with the frontier AI crash course after this one. Readers interested in the hardware and electricity behind AI should go to the AI data centres explainer listed under related pages.

    Frontier AI in 2026: a crash course

    A crash course on frontier AI in 2026: how reasoning models and AI agents work, who builds them, and where the frontier stands now.

    More in AI

    Developing in AI & Computing

    Questions readers ask

    How does a chatbot like ChatGPT or Claude actually work?

    It is a large language model, usually built on the Transformer architecture, that was then given extra training with human feedback so that a raw text predictor behaves like a helpful assistant.[1][5]

    What is the difference between an open-weight and a closed model?

    An open-weight model's learned parameters can be downloaded and run by anyone, while many developers give limited or no public access to their weights. Open weights alone do not meet the Open Source Initiative's definition of open source AI.[13][14][16]

    Is AI taking jobs?

    Not across the whole economy so far, according to a Stanford working paper using payroll data to June 2026 and the Budget Lab at Yale as of October 2025. The same Stanford paper, which is not peer reviewed, found employment of 22- to 25-year-olds in AI-exposed jobs 19% below the path of less-exposed peers.[31][32][33]

    Is it legal to train AI on copyrighted books?

    US courts have not settled it. Two federal judges in 2025 found book training was fair use on the records before them, but one sent pirated copies to trial and the other said the authors had made the wrong arguments.[36][37][47][48]

    Who regulates AI in the US?

    Several states, including California, New York, Texas and Colorado, have passed their own AI laws, and a December 2025 executive order set up a federal task force to challenge state AI laws it considers inconsistent with federal policy.[40][41][49][42][43]

    Where should I start if I know nothing about AI?

    Start with how large language models work, then reasoning models and agents. Those three explain most of what current AI products do.[1][7][10]

    Sources

    · 49 claims from 42 publishers

    Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.

    1. [1]

      The Transformer architecture, introduced in a 2017 paper, is based solely on attention mechanisms and dispenses with recurrence and convolutions. confirmedas of 2017-06-12

    2. [2]

      A 2020 study found that language-model loss falls as a power law as model size, dataset size and training compute grow, with some trends spanning more than seven orders of magnitude. confirmedas of 2020-01-23

    3. [3]

      Anthropic said it would not make Claude Mythos Preview generally available until it had made progress on cybersecurity safeguards that detect and block the model's most dangerous outputs. confirmedas of 2026-04-07

    4. [4]

      OpenAI rated GPT-6 Astra at a Critical level of cybersecurity capability, meaning that with tools and access it can find previously unknown security flaws and develop new ways to exploit them across many well-protected systems. confirmedas of 2026-09-03

    5. [5]

      InstructGPT was trained in two stages, supervised fine-tuning on human-written demonstrations and then reinforcement learning from human feedback (RLHF) using human rankings of model outputs. confirmedas of 2022-03-04

    6. [6]

      In human evaluations, outputs of the 1.3-billion-parameter InstructGPT model were preferred to those of the 175-billion-parameter GPT-3, despite it having 100 times fewer parameters. confirmedas of 2022-03-04

    7. [7]

      OpenAI's o1 model series is trained with large-scale reinforcement learning to reason using a chain of thought before answering. confirmedas of 2024-12-21

      • OpenAI o1 System Card · arXiv (OpenAI) · Primary source · 2024-12-21 · Abstract (retrieved 2026-10-10)
    8. [8]

      The International AI Safety Report 2026 describes inference-time scaling, in which models use more computing power to generate intermediate steps before giving a final answer, as a major way developers now improve capabilities. confirmedas of 2026-02-24

    9. [9]

      Anthropic defines agents as systems in which language models dynamically direct their own processes and tool use, keeping control over how they accomplish a task. confirmedas of 2024-12-19

    10. [10]

      Anthropic describes agents as typically language models using tools based on feedback from their environment in a loop, checking results such as tool outputs or code execution at each step. confirmedas of 2024-12-19

    11. [11]

      Diffusion models build an image through a sequence of denoising steps, and working directly on pixels made training powerful models cost hundreds of GPU days. confirmedas of 2021-12-20

    12. [12]

      Latent diffusion runs the diffusion process in the compressed latent space of a pretrained autoencoder and uses cross-attention layers to condition generation on inputs such as text. confirmedas of 2021-12-20

    13. [13]

      The OSI definition describes an AI model as its architecture, parameters (including weights) and inference code, and defines weights as the learned parameters that produce an output from a given input. confirmedas of 2026-11-07

    14. [14]

      NTIA's July 2024 report noted that many developers give limited or no public access to their advanced models' weights, while others, including Meta, Google, Microsoft and Mistral, have released models with widely available weights. confirmedas of 2024-07-30

    15. [15]

      On 5 August 2025 OpenAI released gpt-oss-120b and gpt-oss-20b, two open-weight language models, under the Apache 2.0 licence. confirmedas of 2025-08-05

      • Introducing gpt-oss · OpenAI · Primary source · 2025-08-05 · Introduction (hand-saved copy; fetch-source got HTTP 403) (retrieved 2026-11-07)
    16. [16]

      The Open Source Initiative's Open Source AI Definition (version 1.0) says an Open Source AI system must be made available on terms that let anyone use it for any purpose without asking permission, study how it works, modify it and share it. confirmedas of 2026-11-07

    17. [17]

      Beyond weights, the OSI definition requires "data information" detailed enough for a skilled person to build a substantially equivalent system, plus the code used to train and run it. confirmedas of 2026-11-07

      • The Open Source AI Definition 1.0 · Open Source Initiative · Primary source · Preferred form to make modifications to machine-learning systems (retrieved 2026-11-07)
      • The Open Source AI Definition 1.0 · Open Source Initiative · Primary source · Preferred form to make modifications to machine-learning systems (retrieved 2026-11-07)
    18. [18]

      Humanity's Last Exam, released in January 2025, is a benchmark of 2,500 expert-written questions across dozens of subjects, created because models were scoring over 90% on popular benchmarks such as MMLU. confirmedas of 2025-01-24

      • Humanity's Last Exam · arXiv (Phan et al.) · Primary source · 2025-01-24 · Abstract (retrieved 2026-10-10)
      • Humanity's Last Exam · Center for AI Safety and Scale AI · Primary source (retrieved 2026-10-10)
    19. [19]

      METR's 50%-task-completion time horizon is the length of tasks, measured by how long human experts take, that an AI model completes with 50% success. confirmedas of 2025-03-18

    20. [20]

      Anthropic was founded in 2021 by former OpenAI vice presidents Dario Amodei and Daniela Amodei, who are siblings. confirmedas of 2026-10-10

    21. [21]

      Anthropic is a public benefit corporation whose stated purpose is the responsible development and maintenance of advanced AI for the long-term benefit of humanity. confirmedas of 2026-10-10

      • Company · Anthropic · Primary source (retrieved 2026-10-10)
    22. [22]

      OpenAI released ChatGPT to the public on November 30, 2022. confirmedas of 2025-11-26

    23. [23]

      In October 2025 OpenAI completed a recapitalization in which its for-profit business became OpenAI Group PBC, a public benefit corporation, and its nonprofit became the OpenAI Foundation with an equity stake valued at about $130 billion. confirmedas of 2025-10-28

    24. [24]

      In April 2023 Google combined the Brain team from Google Research and DeepMind into a single group called Google DeepMind. confirmedas of 2023-04-20

    25. [25]

      On September 30, 2026 Google DeepMind introduced Gemini 4 Argon, a model aimed at software engineering, enterprise knowledge work and cybersecurity defense. confirmedas of 2026-09-30

    26. [26]

      xAI was founded in 2023, and its operating company, X.AI Corp. (now SpaceXAI Corp.), began operations in March 2023. confirmedas of 2026-10-22

    27. [27]

      SpaceX acquired xAI in an all-stock deal effective 2 February 2026, making xAI a wholly owned SpaceX subsidiary. confirmedas of 2026-10-22

    28. [28]

      The 2026 AI Index reports that US and Chinese models have traded places at the top of performance rankings several times since early 2025, and that as of March 2026 Anthropic's top model led by just 2.7%. confirmedas of 2026-03-31

    29. [29]

      On 12 June 2026, three days after Claude Fable 5's release, Anthropic said a US government export control directive required it to suspend all access to Fable 5 and Mythos 5 by foreign nationals, and that it had to disable both models for all customers to ensure compliance. confirmedas of 2026-06-12

    30. [30]

      Anthropic said the export controls on Fable 5 and Mythos 5 were lifted on 30 June 2026; Fable 5 returned to users globally on 1 July, and Mythos 5 access was restored for a set of US organisations after US government approval on 26 June. confirmedas of 2026-06-30

    31. [31]

      An August 2026 update of a Stanford working paper (not peer reviewed) using ADP payroll data on millions of US workers through June 2026 found no evidence of widespread, economy-wide job displacement from generative AI. reportedas of 2026-08-01

    32. [32]

      The Budget Lab at Yale reported on 1 October 2025 that its measures showed the broader US labor market had not experienced a discernible disruption in the 33 months since ChatGPT's release. reportedas of 2025-10-01

    33. [33]

      The same paper found that employment of workers aged 22 to 25 in AI-exposed occupations was 19% below where it would have been had it kept pace with less-exposed peers, with no comparable gap for experienced workers, and that the gap had widened since August 2025. reportedas of 2026-08-01

    34. [34]

      The authors report that the divergence works mainly through less hiring of young workers rather than more separations, and is concentrated in occupations where AI mainly substitutes for human tasks. reportedas of 2026-08-01

    35. [35]

      The authors say the patterns weaken when controlling for education, show some divergent trends before generative AI, and are stronger in the ADP sample than in national survey benchmarks; they describe them as early descriptive indicators, not causal estimates. reportedas of 2026-08-01

    36. [36]

      On 23 June 2025, Judge William Alsup granted Anthropic summary judgment that using the plaintiffs' books to train its models was fair use, and that converting purchased print books to digital form was also fair use. confirmedas of 2025-06-23

    37. [37]

      On 25 June 2025, Judge Vince Chhabria granted Meta summary judgment in Kadrey v. Meta, a suit by 13 authors, finding that Meta's use of their books to train its Llama models was fair use on the record before him. confirmedas of 2025-06-25

    38. [38]

      On 20 July 2026, Judge Araceli Martínez-Olguín granted final approval of the Bartz v. Anthropic settlement and granted class counsel's fee motion in part. confirmedas of 2026-07-20

    39. [39]

      The New York Times Company sued Microsoft and several OpenAI entities in the US District Court for the Southern District of New York on 27 December 2023 (No. 1:23-cv-11195), alleging that their large language models were built by copying millions of its copyrighted articles. confirmedas of 2023-12-27

    40. [40]

      California's SB 53, the Transparency in Frontier Artificial Intelligence Act, signed on 29 September 2025, requires large frontier developers to publish a safety framework, creates a channel for reporting critical safety incidents to the state Office of Emergency Services, and protects whistleblowers. confirmedas of 2025-09-29

    41. [41]

      On 19 December 2025 New York's governor signed the Responsible AI Safety and Education (RAISE) Act, S6953B; she said the agreed chapter amendments require large AI developers to publish information about their safety protocols and report incidents to the state within 72 hours. confirmedas of 2025-12-19

    42. [42]

      Colorado's SB26-189, signed by the governor on 14 May 2026, revised the 2024 law, repealing and re-enacting it with new requirements on automated decision-making technology (ADMT) that influences consequential decisions, beginning in 2027. confirmedas of 2026-05-14

    43. [43]

      Executive Order 14365, Ensuring a National Policy Framework for Artificial Intelligence, signed on 11 December 2025, directed the Attorney General to set up within 30 days an AI Litigation Task Force whose sole job is to challenge state AI laws inconsistent with the order's policy. confirmedas of 2025-12-11

    44. [44]

      The AI Act's transparency rules took effect in August 2026, when the Act became generally applicable on 2 August 2026. confirmedas of 2026-08-02

    45. [45]

      METR's Time Horizon 1.1 update (January 2026) estimated that the length of tasks AI agents can complete has doubled about every 131 days since 2023, and about every 89 days since 2024, compared with about 196 days in its original 2019–2025 estimate. confirmedas of 2026-01-29

      • Time Horizon 1.1 · METR · Primary source · 2026-01-29 · Doubling time estimates (retrieved 2026-10-10)
    46. [46]

      The International AI Safety Report 2026 says reliable pre-deployment safety testing has become harder because models increasingly distinguish between test settings and real deployment. confirmedas of 2026-02-03

    47. [47]

      The same order denied summary judgment on the pirated copies Anthropic kept in its central library and set them for trial. confirmedas of 2025-06-23

    48. [48]

      Judge Chhabria said the ruling did not mean Meta's training was lawful, only that these plaintiffs made the wrong arguments and failed to build a record for the right one. confirmedas of 2025-06-25

    49. [49]

      Texas HB 149 regulates the use of AI systems in the state with civil penalties; it was signed on 22 June 2025 and took effect on 1 January 2026. confirmedas of 2026-01-01

    Revision history (1)
    1. Page created.

    Created Nov 7, 2026. Last reviewed by an editor on Nov 8, 2026. Next scheduled review: Feb 8, 2027.

    Cite this page

    "Artificial intelligence: a crash course." ContentLora, updated Nov 7, 2026. https://contentlora.com/explain/ai

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.