How AI developed, how today’s systems work, what changed in 2025–2026, and what published studies report about AI and jobs, with workflow charts for reinforcement learning, fine-tuning and retrieval-augmented generation (RAG).
By Uzair Ahmed, Entrepreneur · Updated 22 September 2026
How to read this guide. It reports what the cited sources state and adds no forecasts or recommendations of its own. Figures come from each source’s own methods, samples and dates, so they are not directly comparable. Where sources disagree, both findings are shown. Numbers in brackets link to the source list.
Milestones as described in the original publications.
Alan Turing asks “Can machines think?” and replaces the question with an “imitation game,” now known as the Turing test [40].
The proposal for the 1956 Dartmouth summer project introduces the term “artificial intelligence.” It is based on the conjecture that any feature of intelligence can in principle be described precisely enough for a machine to simulate it [25].
Frank Rosenblatt describes the perceptron, an early model of learning in networks of simple units [33].
Rumelhart, Hinton and Williams popularise back-propagation. The method repeatedly adjusts a network’s connection weights to reduce the gap between its actual and desired outputs [34].
A large convolutional neural network, trained on GPUs, sharply lowers error rates in the ImageNet image-classification competition [21].
A deep reinforcement-learning agent (DQN) learns to play Atari games directly from pixels and game scores [27].
AlphaGo combines deep neural networks with tree search and defeats a professional Go player [37].
The Transformer architecture is introduced. It relies on attention mechanisms and drops recurrence and convolutions [41]. The same year, reinforcement learning from human preferences is demonstrated [9], and PPO is published [36].
BERT popularises the pre-train-then-fine-tune approach for language models [12].
GPT-3, a 175-billion-parameter language model, performs tasks from examples given in the prompt (“few-shot”) [5]. Scaling laws describe how performance changes with model size, data and compute [20]. Retrieval-augmented generation (RAG) is introduced [23].
InstructGPT applies reinforcement learning from human feedback (RLHF) to language models [29]. Other 2022 work: chain-of-thought prompting [42], ReAct’s reasoning-plus-acting loop [44], compute-optimal training [18], and Constitutional AI [3].
Language models are trained to call external tools (APIs) [35]. Direct Preference Optimization (DPO) offers an alternative to RL-based preference tuning [32].
The Model Context Protocol (MCP) is released as an open standard for connecting AI assistants to data sources and tools [1]. The EU AI Act is adopted [14].
DeepSeek-R1 reports that reinforcement learning can build up reasoning ability in large language models [10]. The Linux Foundation forms the Agentic AI Foundation with MCP, goose and AGENTS.md as founding projects [24].
The Stanford AI Index 2026 documents a sharp rise in investment and adoption, and uneven (“jagged”) capabilities [38]. The EU amends the AI Act through the “Digital Omnibus on AI” [15].
Most current AI systems are neural networks: layers of simple units joined by weighted connections. In training, the network’s output is compared with the desired output, and back-propagation works out how each weight should change to shrink that difference. The process repeats over many examples [34]. The perceptron is the historical starting point for this line of work [33]. Training much larger networks on GPUs was a key step in the 2012 image-recognition results [21].
Today’s large language models (LLMs) are built on the Transformer. It uses “attention,” which lets each position in a sequence weigh information from every other position, instead of processing text strictly in order [41].
GPT-3 is an autoregressive language model: it is trained on a large text corpus to predict the next token (a word or word-piece). After training, it can do new tasks from instructions or a few examples in the prompt, without further weight updates [5]. BERT uses a different pre-training objective and is then fine-tuned for specific tasks [12].
Kaplan et al. report that language-model loss follows power laws in model size, dataset size, and training compute [20]. Hoffmann et al. report that, for a fixed compute budget, model size and number of training tokens should grow in roughly equal proportion [18].
InstructGPT describes a three-stage process: (1) supervised fine-tuning on human-written demonstrations; (2) training a reward model on human rankings of outputs; (3) optimising the model against that reward model with the PPO reinforcement-learning algorithm. The authors report that labellers preferred outputs from a 1.3B-parameter InstructGPT model over those of the 175B GPT-3 [29]. Alternatives include Constitutional AI, which uses AI-generated feedback guided by a written set of principles [3], and DPO, which learns directly from preference pairs without a separate reward model or an RL loop [32].
Wei et al. report that prompting models to write out intermediate reasoning steps (“chain of thought”) improves performance on arithmetic, commonsense and symbolic reasoning tasks [42]. DeepSeek-AI reports that large-scale reinforcement learning, rewarding verifiably correct answers, can bring out long reasoning behaviour. Its R1-Zero variant was trained with RL without a supervised fine-tuning step first [10].
An “agent” is set up to take actions, not only produce text. The sources describe these building blocks:
Uneven capability. The 2026 AI Index reports that frontier models perform strongly on PhD-level science and mathematics questions, yet still struggle with reading analog clocks, video generation, and household robotics tasks (a 12% success rate) [38].
Unless another source is named, figures come from the AI Index 2026, published 13 April 2026.
| Area | Reported development |
|---|---|
| Investment | Global corporate AI investment reached $581.7 billion in 2025, up 130% from the previous year. The U.S. accounted for $285.9 billion [38]. |
| Competition | Chinese models nearly match U.S. models on performance. As of March 2026, the top model (Anthropic’s) led by 2.7% [38]. |
| Adoption | Generative AI reached 53% population adoption within three years, faster than the internet did [38]. |
| Energy & environment | Grok 4’s estimated training emissions were 72,816 tonnes of CO₂-equivalent. AI data-centre capacity reached 29.6 GW [38]. |
| Transparency | The average Foundation Model Transparency Index score fell from 58 to 40 [38]. |
| Education | Four out of five U.S. high-school and college students use AI for schoolwork. 6% report clear institutional policies [38]. |
| Healthcare | Clinical note-generation tools cut physician documentation time by up to 83%. Only 5% of clinical AI studies use real patient data [38]. |
| Public opinion | 59% of people surveyed feel optimistic about AI’s benefits, and 52% feel nervous about it. 33% of Americans expect AI to improve their job prospects [38]. |
| Reasoning via RL | DeepSeek-R1 (January 2025) reports reasoning capabilities built up mainly through reinforcement learning [10]. |
| Agent standards | In December 2025, the Linux Foundation formed the Agentic AI Foundation, co-founded by Anthropic, Block and OpenAI. MCP, goose and AGENTS.md are its founding projects [24]. |
| Regulation (EU) | Article 4 of the AI Act, on AI literacy for staff of providers and deployers, began to apply on 2 February 2025 [14]. The Act was then amended by the “Digital Omnibus on AI.” Consult the consolidated text for the current wording [15]. |
Published evidence in five strands: exposure estimates, employer surveys, controlled productivity studies, employment data, and usage data. They measure different things. “Exposure” means a job’s tasks overlap with what AI can do; it is not a measure of job loss.
| Source | Reported estimate |
|---|---|
| IMF staff note [8] | About 40% of global employment is exposed to AI, and about 60% in advanced economies. Women and college-educated workers are more exposed but also better placed to benefit. |
| ILO–NASK index [17] | One in four jobs worldwide is potentially exposed to generative AI, rising to 34% in high-income countries. Clerical occupations are the most exposed. The authors state that “full job automation, however, remains limited,” and that transformation is more likely than replacement. |
| Eloundou et al. [13] | About 80% of the U.S. workforce could have at least 10% of their tasks affected by LLMs, and about 19% could have at least 50% affected. The authors state that they make no predictions about adoption timelines. |
| Source | Reported finding |
|---|---|
| WEF Future of Jobs 2025 [43] | Survey of 1,000+ employers representing 14 million workers in 55 economies.
|
| Microsoft & LinkedIn Work Trend Index [26] | 66% of leaders say they would not hire someone without AI skills. 39% of AI users say they have received AI training from their company. |
| PwC AI Jobs Barometer [31] | Based on close to a billion job ads. Jobs requiring AI skills carry an average wage premium of 56%, up from 25% the year before. Postings requiring AI skills rose 7.5% while total postings fell 11.3%. |
These studies report both gains and losses, depending on the task and the worker.
| Study | Setting | Reported result |
|---|---|---|
| Brynjolfsson, Li & Raymond [7] | 5,179 customer-support agents | Issues resolved per hour rose 14% on average and 34% for novice and low-skilled workers, with minimal effect on the most experienced workers. |
| Noy & Zhang [28] | 444 professionals, writing tasks | Time taken fell by 0.8 standard deviations and quality rose by 0.4 SD. Lower-ability workers benefited more. |
| Dell’Acqua et al. [11] | 758 BCG consultants | On tasks inside the AI “frontier,” consultants completed 12.2% more tasks, 25.1% faster, at higher quality. On a task outside it, those using AI were 19% less likely to reach a correct solution. |
| Peng et al. [30] | Developers, one coding task | The group using GitHub Copilot finished 55.8% faster. |
| Becker et al. / METR [4] | 16 experienced open-source developers, 246 tasks in their own repositories | Allowing AI tools increased completion time by 19%. Developers had predicted a 24% speed-up, and afterwards believed they had been 20% faster. The authors caution against generalising beyond this setting. |
The Anthropic Economic Index (January 2026, data from 13–20 November 2025) reports [2]:
The sources above make these points directly. They are reported here, not endorsed.
Each chart follows the procedure described in the cited sources. Blue steps are where the model’s weights (or, for RAG, the retrieval index) are changed or used; orange steps are data or feedback inputs.
In reinforcement learning, an agent interacts with an environment. At each step it observes a state, takes an action, and receives a reward, and it learns a policy that maximises total reward over time [39]. RLHF applies this to language models: a reward model learned from human preferences replaces a hand-written reward [9], and the language model is then optimised against it with PPO [36] [29].
Fine-tuning takes a pre-trained model and trains it further on a smaller, task-specific dataset [12]. In full fine-tuning, all of the model’s weights are updated. Parameter-efficient methods such as LoRA freeze the pre-trained weights and train small low-rank matrices added to each layer. The LoRA authors report a roughly 10,000× reduction in trainable parameters for GPT-3 175B [19].
RAG combines a language model with a retriever. For each question, the system fetches relevant passages from an external document index and passes them to the generator alongside the question. Lewis et al. describe the index as “non-parametric” memory, meaning it can be updated without retraining the model [23]. A later survey divides the basic RAG pipeline into indexing, retrieval, and generation [16].
| Reinforcement learning / RLHF | Fine-tuning | RAG | |
|---|---|---|---|
| Learning signal | Reward or preference score [39] [29] | Fixed target outputs [12] | None; passages are retrieved at question time [23] |
| Model weights change? | Yes | Yes (all weights, or adapters only with LoRA [19]) | No; the index changes [23] |
| Typical use described | Aligning responses with human preferences; reasoning [10] | Adapting to a task, format or domain | Knowledge-intensive questions; updatable knowledge [16] |
Large language models are Transformer neural networks pre-trained on large text corpora to predict the next token (Vaswani et al., 2017; Brown et al., 2020). They are then post-trained, for example with supervised fine-tuning and reinforcement learning from human feedback, to follow instructions (Ouyang et al., 2022).
Fine-tuning trains a pre-trained model further on task-specific examples, so its weights change (Devlin et al., 2018; Hu et al., 2021). Retrieval-augmented generation leaves the model's weights unchanged and instead retrieves relevant passages from a document index at question time (Lewis et al., 2020).
In the InstructGPT method, humans rank model outputs, a reward model is trained on those rankings, and the language model is optimised against the reward model with the PPO algorithm (Ouyang et al., 2022; Schulman et al., 2017).
Estimates differ by method: an IMF staff note puts about 40% of global employment as exposed to AI (Cazzaniga et al., 2024), and the ILO–NASK index finds one in four jobs potentially exposed to generative AI (Gmyrek et al., 2025). Exposure measures task overlap, not job loss.
In the WEF Future of Jobs Report 2025, two-thirds of employers plan to hire talent with specific AI skills, and AI and big data top the list of fastest-growing skills. In the 2024 Work Trend Index, 66% of leaders said they would not hire someone without AI skills.
Applying for jobs? Check your resume against a job ad and see how applicant tracking systems read it, or browse all our free tools and data.