From the 2017 breakthrough in language processing to the systems of 2026 that write, search, code and use tools: how research, computing power, capital and human labor built the AI industry, and who controls its operation and development.
Prefer reading in Arabic? Read the Arabic version.
This essay covers events up to 17 October 2026.
A small window onto a wide world
We type a request into a chat window and get back a message, an image or a program. We can also give the system a longer task: it opens pages, gathers sources, edits files and uses other tools to get the job done. The interface is almost the same, but what happens behind it is completely different: generating an answer is not the same as making a change in a bank account or editing a project's files. Between the two lie decades of research, thousands of workers, and computing infrastructure that demands huge investment.
How did we reach this point? A list of model names and release dates is not enough to answer. We need to understand what changed in the way machines learn and compute, how research moved from the lab to the market, and why using advanced capabilities became much easier than owning the means to make them.
AI research began long before this era. Alan Turing asked early questions about machine behavior and intelligence, and the Dartmouth conference of 1956 gave the name "artificial intelligence" to a young research project. Efforts split between programming explicit rules that tell the machine what to do, and building systems that learn patterns from examples. Neither approach replaced the other; their relative importance shifted as algorithms, data and hardware advanced.
In neural networks, the system learns by adjusting internal numbers called weights. These weights set how strongly signals affect one another. The network's output is compared with the desired result, the errors are calculated, and backpropagation is used to work out how the weights should change to reduce them. A 1986 paper was a landmark in spreading this method. But a way to learn is not enough on its own: it needs many examples, and the computing power to repeat training a huge number of times.
Images, text, search and translation provided growing amounts of data. Graphics processing units (GPUs) made it possible to run many mathematical operations in parallel, the kind of operations neural networks need. AlexNet in 2012 showed the effect of this combination clearly: it made a large leap in image classification accuracy, thanks to the network's design, the availability of training data, and processing power. The advance came not from one algorithm, but from bringing together an architecture, training data and hardware at sufficient scale.
Language was harder than recognizing a word. The meaning of a word depends on where it appears and the words around it. Systems began to learn numerical representations of words that place words used in similar contexts close together, and then recurrent neural networks and LSTMs were used to keep information from earlier words while processing a sentence. In translation, sequence-to-sequence models used one network to understand the sentence and another to produce its translation. Bahdanau and colleagues added a mechanism that lets the model look back at different parts of the source text when generating each part of the translation. Here the idea of attention appeared, the idea that would become central a few years later.
On a different track, AlphaGo combined learning with search through possible moves to defeat top players of Go. Its success shows that AI research was not waiting for the language assistant; it was already achieving major results in varied ways and on different tasks.
Part One: How scaling became a research strategy
2017: Processing language in parallel
In 2017, researchers presented the paper Attention Is All You Need, which introduced the Transformer architecture. Recurrent networks processed a sentence during training in successive steps, so processing the next word depended on the results of the previous step. The Transformer replaced this sequence, when computing text representations, with self-attention: for each position in the text, the system computes how relevant the other positions are, and then uses these relations to build a richer representation of the context.
Take a sentence such as "Sarah put the book on the table, then took it." To understand what "it" refers to, the system needs to link the pronoun to "the book", not just read the words one by one. Attention helps build links like this. Information about word order is also added to the computation, so that two sentences are not treated as the same just because they contain the same words.
The decisive effect was also an engineering one. Many positions of the text could now be processed at the same time during training, making use of parallel processors. This did not make computation free: attention itself is costly, and generating text in the usual way remains sequential; the model produces a new piece after the previous ones. But better use of the hardware opened the way to training larger networks on more data.
Hardware was developing at the same time. NVIDIA introduced the Volta architecture and Tensor Cores, which specialize in matrix operations, and Google developed TPUs and systems that combine large numbers of them. Then the speed of the links between chips, and the way work is split across them, became part of the design of training systems. Meanwhile AlphaZero kept developing the approach of learning and search for games; not all AI research was on the road of language models.
2018–2019: Different ways to teach machines language
Researchers pursued several competing approaches. ELMo made a word's representation change with its context. GPT-1 was trained to predict the next word and then adjusted to perform specific language tasks. BERT learned from text in both directions, using the words before and after to predict hidden words. XLNet tried a different formulation for using context during training.
These choices served different goals. Finding an answer inside a paragraph, classifying a message, translating a sentence and writing a new text are not one task. A model can excel at one and lag at another. That is why the competition between BERT and GPT was not an early announcement of a final winner; it was a test of different training paths. The GPT-1 paper included limited experiments without extra training examples, but its main results came from adjusting the model to specific tasks.
In 2019, GPT-2 raised a new question alongside quality: who gets the model itself? OpenAI released smaller versions first and delayed the largest version, with about 1.5 billion parameters, then published it in November 2019. The company cited concerns about the risks of use; at the same time, the decision showed that whoever trains a model can decide when it becomes available and who may use it.
2019–2020: Data and computing become an industry
T5 brought a practical idea: turning many kinds of tasks into the form "text in, text out". This framework can be used for translation, summarization and question answering by changing the wording of the request and the data. The project relied on the C4 dataset, built from web pages that were collected, filtered and cleaned. Training data was not neutral raw material: choosing pages and languages, and excluding repeated or unwanted content, changes what the model learns.
In July 2019, OpenAI announced a $1 billion investment from Microsoft and a partnership to use Azure. In the same years, techniques such as Megatron-LM and ZeRO appeared to spread work and memory across large numbers of processors. A large model needs more than computing power; it needs a way to store its weights, its intermediate results and its updates across many machines.
Then came scaling laws. Researchers studied how a model's error changes as the number of parameters, the amount of data and the training compute increase. They found empirical relationships that help estimate the return from a larger model before spending the full training budget. The laws were not a guaranteed recipe for general intelligence, but they made the decision to spend on larger models easier to calculate.
GPT-3 gave a striking example of the result. Instead of adjusting the model's weights for each new task, it became possible to give it instructions and examples inside the request itself, such as showing three translated sentences and then asking for the fourth. In June 2020 OpenAI launched an API that lets developers use its models remotely. A company no longer had to buy hardware to train a large model, but it came to depend on a service that sets the prices, terms and limits of use. In September 2020 Microsoft announced an exclusive commercial license to GPT-3 technology, before the partnership's arrangements changed in later years.
Part Two: How did models become an industry?
What parameter counts leave out
The news often presents a new model with a huge number of "parameters", but the parameter count alone does not decide what can be run. In training, the weights are stored, the updates are computed, and the intermediate results that backpropagation needs are kept. This data can exceed the memory of one machine many times over. Even when it is spread across many machines, moving it between processors becomes a bottleneck: a processor waiting for data cannot keep computing.
That is why fast networking technologies, including Mellanox products, processors such as the A100, and distributed training software moved to the core of the model industry. ZeRO offered a way to reduce the duplicate copies of the model's state during training, meaning the weights, gradients and optimizer states, that the machines keep. Every bit of memory saved and every transfer avoided allows a larger model to be trained, or fewer machines to be used. The economic question became tied to the scientific one: how much does it cost to train the model, and how much to run it for each user afterward?
Alongside the large commercial labs, other groups built their own models and datasets. The Pile helped provide a broad text collection for research, followed by models such as GPT-Neo, GPT-J and GPT-NeoX. The large labs continued their scaling experiments, including Gopher, Megatron-Turing and Switch Transformer. Switch Transformer used a mixture-of-experts design: the model contains many specialized components, but activates only selected ones for each input. In this way the total size of the model can grow without a matching increase in the computation used at each step.
In 2020 an important research formulation appeared: retrieval-augmented generation (RAG). The idea is simple: instead of asking the model to answer only from what it learned during training, the system first searches for documents related to the question, then places excerpts from them in the model's input. If an employee asks about a company policy that changed yesterday, the system can retrieve the new version instead of relying on old knowledge stored in the weights. This does not guarantee a correct answer automatically; it provides material that can be checked. The original paper came years before the wave of assistants connected to files and search engines.
2021: AI enters work tools
With GitHub Copilot, code generation was no longer limited to a technical demo or a chat page. The suggestion now appeared inside the code editor itself, while the programmer was writing a function, a comment or a test. The division of labor changed: the tool could generate code quickly, while the programmer still had to verify that it worked, met the requirements and did not introduce security flaws.
The Codex paper showed the importance of multiple attempts. A single suggestion may fail to solve a programming problem, while one of dozens of attempts succeeds when tested. But testing a hundred suggestions and picking the correct one is different work from getting a correct answer on the first try. That is why the rate reported on a test does not describe the experience of all users.
At the same time, ways to adapt models improved. LoRA offered a way to adjust a relatively small number of added parameters instead of retraining all the weights of a large model. FLAN focused on training with varied instructions to improve the model's ability to follow new requests, and LaMDA advanced the path of models built for dialogue. In images, CLIP linked text descriptions with images, while DALL·E showed that images could be generated from a written description. The building blocks of the products of the following years were already taking shape.
The human work behind the assistant that understands instructions
There is a difference between a model that learned to complete text and a model that answers the user's request clearly and keeps to its format. In the InstructGPT research, human workers wrote examples of desired answers, then compared different answers and ranked them. These judgments were used to build an evaluation system that learns the raters' preferences, and then to improve the assistant's answers through reinforcement learning: the model receives a reward signal for higher-rated answers and updates its behavior to make similar answers more likely. In the research tests, the raters preferred the answers of a smaller model that went through this process over those of a larger model that did not go through it in the same way.
This explains part of the shift the public would see with ChatGPT: the model was no longer merely generating fluent text; it had become better at following requests, acknowledging mistakes, refusing inappropriate instructions and sustaining a natural conversation. Behind this was human work in writing examples and judging quality, not independent machine computation alone. And what the raters considered a suitable answer was shaped by the instructions and standards applied to them.
In December 2022, Anthropic offered another path in its research on Constitutional AI. The model reviews its answers against a set of written principles and rewrites them, and then evaluations generated by a model are used to improve its training. The aim is to reduce the need for direct human evaluation of every answer, while the choice of the principles themselves remains a human decision.
It is necessary here to separate the different jobs. The workers who wrote examples and compared answers in the InstructGPT project are not necessarily the same people covered by the TIME investigation into labeling harmful texts in Kenya. The investigation presented testimonies about exposure to harsh content while safety data was being prepared, together with the company's position on these operations. What links the two paths is that models depend on people behind the interface, but the nature of the work, its conditions and its responsibilities differ.
2022: The biggest model is not always the best
The prevailing trend was to increase the number of a model's parameters, but Chinchilla showed the value of splitting the training budget differently. Instead of training a very large model on a relatively limited amount of text, the same budget can be spent on a smaller model that reads more data during training, and gets a better result on the measures the researchers tested. The result was not to stop scaling models; it was to recalculate the balance between model size, amount of data and compute.
In 2022 hardware capacity grew with the H100 and TPU v4 systems. FlashAttention offered an important improvement that does not rest on inventing a new "attention", but on organizing operations so that less data moves between fast and slow memory. In many computations, moving data is slower and more expensive than the math itself. Reducing this movement speeds up execution and lowers memory use, while the model's basic computing requirements remain.
Efforts to make models available for research and experiment also widened, including OPT and BLOOM, which researchers from many institutions helped develop. On another track, Chain-of-Thought studies tested whether models can benefit from written intermediate steps before the final answer, and ReAct offered an approach in which the system alternates between reasoning and practical steps, such as using a tool or consulting a source. These ideas came before commercial AI agents, and their early experiments were not at the level of the products that came later.
2022: Images leave the lab
The transformation was not limited to text. Services such as Midjourney made it possible to produce images from short descriptions, while the release of the Stable Diffusion weights made it possible to run and modify an image-generation model outside the original service. This came after a line of research that included DDPM, CLIP, DALL·E, DALL·E 2 and Latent Diffusion techniques.
To understand the idea, imagine an image that gradually turns into noise. Diffusion models learn the reverse process: they start from noise and remove it in stages until an image appears that matches the requested description. Instead of running all the steps on the full image, Stable Diffusion works on a compressed representation of it, which reduces the computation needed. The release of the version 1.4 weights let developers build interfaces and extra tools and run the model on suitable hardware, under the terms of its license.
This changed the relationship between users and the technology. Some users could now keep a copy of the model and change the way it runs, instead of sending every request to a remote service. At the same time, disputes arose over the use of artists' work in training data, creators' rights, and the attribution of the resulting works. Artists were not just an audience receiving the technology; they were part of the material the systems learned from, part of the market that used them, and part of the debate over their rules.
In speech recognition, OpenAI released Whisper, which helped turn speech into text across many languages. Images, speech and text were converging in products, even though the technologies that handle each of them differ.
2022: Chips become an instrument of state policy
In October 2022, the United States imposed new controls on exporting certain advanced processors and semiconductor manufacturing equipment, with different effective dates on 7, 12 and 21 October. The ability to buy the tools for training large models was no longer a question of budget and specifications alone; it now included restrictions tied to the country, the end user and the intended use.
Computing chips and the means of making them became part of the industrial and security competition between states. The restrictions affected the options open to companies and researchers in the markets they covered. But a restriction on a specific chip does not, on its own, prove that the design of a particular model was a direct result of it. This distinction will matter when we come to the story of DeepSeek.
Part Three: What did the public do with the models?
November 2022: Conversation becomes the front door
On 30 November 2022, OpenAI launched ChatGPT as a conversation that could be used for free during its research phase. The language model was not an invention that began that day, but the way to reach it changed radically. Users no longer needed to write a program that connects to an API, or understand the settings of a research model. It was enough to type a question in ordinary language and continue the conversation, as one would with a person helping on a task.
Easy access drove the public breakthrough. A programmer could ask for an explanation of an error message, a teacher could try preparing a lesson, and a writer could revise a paragraph, all through the same interface. At the same time, it became clear early on that fluency in an answer does not mean it is correct, and that the quality of performance varies with the language and the type of task.
On 2 December 2022, a programmer wrote about an experience while solving a challenge in the Advent of Code contest. They asked ChatGPT for a table to help calculate the results of rock-paper-scissors games. The reply looked neat, but the table treated some different moves as ties. The programmer noticed the error because they knew the rules of the game, objected, and the system admitted its mistake and gave a corrected table. The tool was useful for a small job, but finding its error took a person who knew what to look for.
On 27 December 2022, a user on the OpenAI developer forum posted that they had asked for a story in Arabic and found many spelling errors, and offered to help improve the service. And in February 2023, an Arab blogger told how at first they could not reach the service from their country. When they were able to try it and asked for an article about adolescence, their son, a school student, saw the result and said it would do his homework. The blogger worried about the effect on education; the son saw a shortcut for schoolwork. Together, these accounts reveal what headline adoption figures concealed: useful answers that still make mistakes, languages that do not get the same quality, and users who differ in what they think should be done with the tool.
Then the product expanded commercially. The ChatGPT Plus subscription launched, and APIs for chat models appeared, including gpt-3.5-turbo, which came after the original launch model. An estimate spread that the service reached one hundred million monthly active users in January 2023; it is an estimate by UBS analysts based on Similarweb data.
Early 2023: AI enters the search engine
The big companies began to reshape their products. The chat-based Bing experience appeared, linked to the name Sydney during testing, then Google's Bard and Anthropic's Claude, alongside PaLM 2. The question was no longer only how the model answers, but how it changes the experience of search and existing services, and how it is built into products that millions of people already use.
In March 2023 OpenAI announced GPT-4. The technical report drew wide attention to its test results, but it did not disclose key details about the number of parameters, the full model design, the training data and the compute used. Here the gap widened between the capability available for use and the knowledge available to researchers outside the company to rebuild it or test its causes. These missing details cannot be extracted from performance results alone.
Conversations with Bing/Sydney drew extra attention because of replies that seemed personal or emotional. But they remain the outputs of a chat system, not evidence that the model has desires or consciousness. More important in practice was the move of models from producing text to proposing the use of other services. OpenAI introduced plugins and then function calling, which lets the model identify the tool needed and draft structured inputs for it. The program around the model is still what decides whether to actually carry out the request, and what permissions it has.
2023: Running the model on your own machine changes the rules
Another path widened in 2023 with Meta's LLaMA, the llama.cpp project, and experiments such as Alpaca. This work showed that using a language model was no longer always conditional on connecting to a commercial provider. Models could be run on personal machines, tools could be built around them, and specific modifications could be tried, even though restrictions tied to licenses and the origin of the weights remained.
Quantization was a key part of the story. Instead of storing each weight in a large number format, many weights can be represented with fewer bits, while balancing the effect on quality. This lowers the memory needed to run some models. llama.cpp in March 2023 showed that quantized models could run on the central processor (CPU). The GGUF format common today, and some later quantization methods, came in later stages.
Adaptation tools also expanded. LoRA had appeared before this wave, and then QLoRA offered a way to lower memory needs when adapting large models. For serving models to large numbers of users, vLLM and the PagedAttention technique improved the management of the memory used to keep earlier computation results while text is generated. Then came Llama 2 and Mistral, and in images, tools such as ControlNet spread for controlling the composition of a generated image. Owning the weights gives a developer more freedom to experiment and run; but it does not mean that training a model from scratch became within everyone's reach.
For anyone choosing a machine to run models locally today: You Can’t Buy “The Best Local AI PC” in 2026.
Late 2023: Data, ownership and power over platforms
The more valuable texts and images became as training material, the sharper the competition over the right to access them. Reddit changed the terms of access to its API and introduced fees for certain uses, and Twitter, which later became X, announced changes to access to its data. But paying a fee to use an API does not automatically mean obtaining a license to train models on everything it shows, nor does it necessarily stop every other means of collecting data.
In December 2023, The New York Times filed a lawsuit against OpenAI and Microsoft over the use of its material. The paper set out its claims in the complaint, and OpenAI published its response. This came alongside cases and claims by artists, authors and image companies, and new licensing deals. Training data moved from being a technical resource gathered by researchers to the subject of a wide commercial and legal dispute.
The year also saw shifts in the structure of the market. OpenAI held its DevDay developer conference and announced there the Assistants API and custom GPTs, then went through the board crisis, which raised again questions of influence inside the company and its ties to its funders. In public policy, updates to US chip export controls and the US executive order on AI appeared. Google introduced the first Gemini family while continuing to develop TPU chips, and Mistral AI released the Mixtral model. Competition was no longer a race between labs on a single test; it had spread to distribution, ownership, computing and the governance of companies.
In images, DALL·E 3 came to ChatGPT in October 2023, so an image could be generated and its description revised within the same conversation, instead of writing separate commands for each attempt.
2024: Text, image and voice in one service
In 2024, products moved toward accepting images, voice and longer files, keeping some information between sessions, and integrating with the tools people work with. The year saw the GPT Store and Sora, Gemini 1.5, Claude 3 and Llama 3, then GPT-4o and voice interaction, Apple Intelligence and Artifacts. With every announcement there was a difference worth tracking between showing a technology on stage and the date it actually became available to users.
The length of the context window drew a lot of attention. It means how much text or data the model can take in a single request, usually measured in units called tokens: a token can be a short word, part of a word, or a punctuation mark. A larger window allows longer documents to be entered, but it does not guarantee that the system will find every detail in them or connect them accurately. In the Gemini 1.5 report, the researchers distinguished between retrieving one piece of information hidden in a long context and harder tasks that require retrieving many pieces. Window size does not equal quality of understanding.
Competition outside closed services continued in 2024. Meta released Llama 3.1 in sizes up to 405 billion parameters, and the Qwen team released Qwen2.5 in various sizes, with models specialized for coding and math. These releases widened the options for running and adapting weights, without making the training of the largest models possible without a huge computing infrastructure.
Behind the interface, the engineering problems were getting more complex: how can intermediate results be stored as a conversation grows longer? How is running the model spread across several chips? And how can speech be processed and produced with a short enough delay for natural conversation? Development of Blackwell systems expanded, while DeepSeek advanced on another track through DeepSeekMath and the GRPO method for improving reinforcement training, then DeepSeek-V2 and the multi-head latent attention (MLA) mechanism to reduce some memory costs. These innovations would become central to understanding what happened with DeepSeek-R1.
In the same year, the EU Artificial Intelligence Act entered into force on 1 August 2024, with its provisions starting to apply in later stages. It was a shift from debating principles of regulation to setting legal obligations that differ by the type of system and its risks.
Part Four: From producing the answer to carrying out the task
2024: More computation for a better answer
Increasing capability was no longer limited to training a larger model and then running it in the usual way. In 2024 a trend appeared that gives the model extra time and computation during inference, that is, while it processes the user's request and produces the answer. OpenAI introduced the o1 family within this trend, with tests on solving problems that need several steps, especially in math and programming.
To understand the reported results, we need to tell the test procedures apart. Did the model answer in a single attempt? Were many answers produced and then voted on? Was a system used to pick the best answers after they were generated? OpenAI's results show these setups separately. Reaching a high rate when dozens of attempts are checked does not mean reaching the same rate when a user asks one question and waits for the first answer.
Economically, long reasoning adds a new item to the cost. A single request may use more computation, but it can save the user hours of trial and work if it actually solves the task. Comparing message prices or the results of one test is not enough to know the real cost; the more important measure is the cost of completing the task successfully, including review, correction and operation.
Late 2024: The model starts to use tools
Systems took another step from writing answers to carrying out parts of a task. Anthropic showed a computer use capability that lets the model work with a computer's interface, and launched MCP, while OpenAI released new versions of o1 and showed a preview of o3, and Google announced Gemini 2.0 and the Astra and Mariner projects. These examples differ in maturity and availability: some are capabilities inside a product, and some are research demos or specifications for connecting tools.
MCP is a protocol that helps applications make data and tools available to models in a relatively standard way. Instead of writing a special connection for every model and every source, an application can use a shared format to define and call tools. But the protocol does not decide on its own whether a command to delete a file or send a message should be carried out. That is the job of the permission and verification system around the model.
Here the kind of error changes. A wrong answer in a chat window can be ignored or corrected. But sending a message to the wrong person, changing a production file, or exposing data through a connected tool can leave a mark outside the conversation. That is why managing permissions, logging actions, and the ability to undo or stop execution became basic conditions for designing agents, not cosmetic additions to a language model.
DeepSeek: What does it mean for training to get cheaper?
In late 2024 the DeepSeek team published the V3 report, and then in January 2025 R1 drew global attention to its reasoning abilities and the efficiency of model development. In media coverage, the figure for the training cost got mixed up with the cost of building the whole project, although they are two different things.
The V3 report stated the use of about 2.788 million H800 GPU hours in the reported training, and estimated their cost at about $5.576 million on the basis of $2 per GPU hour. This is an estimate of computing for a defined scope of training, and it does not include all the earlier research, the experiments not counted, staff salaries, or the costs of buying hardware and running the organization. So the figure does not mean that any company can build an equivalent model from scratch for the same amount.
The R1 paper also presented two different pictures of training. R1-Zero used reinforcement learning without a traditional first stage of guided reasoning examples, while the final R1 relied on starter data and extra stages to improve its outputs. And not all the ideas began with R1: GRPO had appeared earlier in DeepSeekMath, and MLA in DeepSeek-V2. The achievement was in developing an accumulated engineering and research chain, not in a sudden discovery that canceled what came before. V3 was trained on H800 processors, and here the distinction from the chips story returns: the restrictions are context for this innovation, and the report gives no evidence that they were its direct cause.
DeepSeek's rise coincided with sharp swings in AI-related stocks. In the same month, on 21 January, OpenAI and its partners announced the Stargate project with an investment plan of up to $500 billion over four years. This was a stated investment target. The two events reveal two parallel trends: making models more efficient on one side, and building larger computing centers to serve growing demand on the other. A lower cost for some operations does not automatically mean lower total spending if uses and required scale grow.
For another perspective on the economics of keeping AI infrastructure up to date: The Bill for Not Falling Behind.
2025: Agents take on multi-step tasks
In 2025, agent capabilities began reaching users through commercial products. OpenAI introduced Operator as an experiment in using the browser, then Deep Research for research across multiple steps and sources. Anthropic launched Claude Code with Claude 3.7, while OpenAI introduced the Responses API and Agents SDK for building applications that use tools and manage successive steps.
Suppose a user asks for a bug on a website to be fixed. Writing a new line of code is only part of the task. The system has to read the project, locate the bug, edit the right files, run the tests, and then review the effect of the change. This ability to manage a sequence of actions distinguishes an agent from a model that merely suggests code. Even so, permission and successful execution are two separate matters.
Reasoning and personalization models kept developing, including Gemini 2.5, o3 and o4-mini. In April, Google introduced A2A to make communication between agents easier, while MCP serves to connect applications with models, tools and data. The difference matters: an agent knowing how to send a task to another agent does not mean that either of them has permission to carry it out, or that the result is correct.
2025: Model names hide whole systems
Releases followed one another: Llama 4, Qwen3, Claude 4, GPT-5, Sonnet 4.5 and GPT-5.1. But the name a user sees no longer always describes a single model doing all the work alone. Behind the interface, a mechanism may route the request to a faster model or to a model that spends longer reasoning, alongside search and execution tools and usage limits.
OpenAI described GPT-5 as a system that combines fast responses, deeper thinking, and routing between them. In this way the management of models became part of a product's quality and cost, alongside the quality of the trained weights. Two users may get different results because their tasks, tools and allowed limits differ, even if they see the same brand.
In November 2025, Google presented interactive interfaces generated on demand. The answer was no longer always a paragraph or a fixed image; the system might make an interactive diagram, a simple simulation, or an interface tailored to the question. In the evaluation published with the research, sites designed by human experts were still preferred over the generated interfaces. It showed how AI-generated answers were beginning to take new, interactive forms.
Meanwhile, alternative architectures such as RWKV and Mamba continued, and Grok and others kept widening the competition. A different research problem also appeared: what happens if new models are trained again and again on the outputs of earlier models? The result depends on the quality of the synthetic data, its share and the way it is selected; so research on model collapse does not mean that all generated data is useless. And alongside legitimate uses, the technology was increasingly used for fraud, disinformation and non-consensual images, which made safety part of the history of adoption itself.
2026: Who decides access and the limits of use?
In April 2026 DeepSeek announced the V4 preview with a context window of up to one million tokens and a focus on running efficiency. In May Google presented Gemini Omni and Gemini 3.5 as part of its work on multimodality and interaction with tools. On 7 October OpenAI announced wider availability of GPT-6 and Intelligent UI, after introducing models of the family in September. The interactive interface itself was becoming part of the answer, instead of a fixed screen for showing text.
Access was no longer governed by prices and commercial policies alone. In June 2026 Anthropic announced that it had suspended access to Fable 5 and Mythos 5 after a US government directive, then announced on 30 June the lifting of the restrictions, availability of Fable 5 to users around the world through its own services from 1 July, and limited access to Mythos 5 for selected US organizations. This event shows that a regulatory decision can change access to an existing service in a short time. It describes a sequence the company announced.
The relationship between model developers and infrastructure providers also changed. On 27 April 2026 OpenAI announced an amendment to its agreement with Microsoft that makes Microsoft's license to OpenAI technology non-exclusive, with Microsoft remaining a primary cloud partner under the agreement. So the future of models is not decided by research alone; it is also decided by licensing contracts, the ability to buy chips, the supply of energy, and the terms of running a service in different markets.
During September, several tiers of the GPT-6 family appeared, from Astra to Sol and Luna, then GPT-6.1 Sol. The trade-off between speed, capability and running cost became a visible part of product design. Despite better model efficiency, building the largest systems still requires data centers, networks, cooling, energy and specialized teams. This material world does not disappear when the interface becomes simpler.
What is now within our reach?
The modern journey began with a better way to compute the relations between words, then widened to models that learn from text, images and sound, retrieve information, write programs, use tools and build interactive interfaces. Access to these capabilities became easier and cheaper for many uses, and models with published weights opened new paths for independent work. But training the most advanced models and running them at huge scale still needs capital, hardware, data and expertise that are not evenly distributed.
The industry's central imbalance is now clear: the circle of people who can use AI widened, while the ability to build its largest systems and set the terms of access to them stayed concentrated in a small number of companies and institutions. Publishing weights opened an important door outside closed services, but it did not remove the gap between running an existing model and developing a frontier model from scratch.
That is why the question "which model is best?" is not enough to understand this period. We also have to ask: what tasks does it actually get done? What does it cost to get them done after review and correction? Who provided the training data? Who did the human work behind the results? Who holds the authority to carry out actions? And can the user examine the system, replace it, or run an alternative?
These questions link the history of algorithms to the history of the economy, institutions and individual rights. From 2017 to October 2026, it was not only the machines that advanced; the way digital knowledge is built, the way computing power is sold, and the way power over what users can do is distributed all changed too. This is the AI era as it took shape: a major scientific achievement, a resource-heavy industry, and tools whose adoption is spreading faster than the ability to build them.
Top comments (0)