Quick brief
AGI claims need a stated definition and evidence of broad, reliable capability; how much autonomy a system receives is a separate deployment question. Models, agents, speech systems, and specialized hardware can extend what people do, but each capability has limits that must be measured in its setting.
Full analysis
Originally published February 24, 2024. Updated October 18, 2026.
The thought behind artificial general intelligence is both simple and difficult to pin down: could a machine learn to handle a broad range of problems with the flexibility people bring to unfamiliar situations? Recent systems can write, code, interpret images, generate speech, plan some tasks, and use tools. Taken together, those abilities make the question feel closer without answering it.
The discussion below connects definitions of AGI to the model, agent, communication, decision, and hardware capabilities often presented as steps toward it. Examples are dated where product architecture, availability, or performance changes over time. The central point is simple: “AGI” needs a stated definition before anyone can say whether a system has reached it.
What is AGI and Why is it Significant?
There is no single operational definition of AGI that every researcher or organization uses. Some definitions emphasize performance across many cognitive tasks; others emphasize transfer to unfamiliar situations, the level of autonomy, or comparison with human performance. DeepMind’s published framework, for example, treats performance, generality, and autonomy as separate dimensions and proposes levels for discussion. It is a framework, not a universal certification test (DeepMind, “Levels of AGI”).
For this article, AGI means a system that can learn and perform across a broad range of tasks, including tasks it was not specifically built to solve, with enough reliability to be useful in varied settings. This contrasts with narrow AI, such as a facial-recognition or game-playing system designed for a defined task; AGI claims concern transfer across tasks, not just strong performance in one. This working definition does not require consciousness, human emotions, or a claim that a system is better than every person at every task. A system can be highly capable on many benchmarks and still fail when the task, data, instructions, or environment changes.
That distinction matters because a broad-purpose system could change how people do research, learn, design, and make decisions. It could help a scientist compare evidence across fields, help a teacher adapt practice material, or help a team find patterns in a large collection of records. These are plausible forms of assistance, not guaranteed outcomes. In medicine, climate science, or space exploration, a model’s output would still need domain evidence, measurement, and accountable human judgment.
It is also possible to imagine systems that act as collaborators: they could draft, compare alternatives, or carry out bounded tasks after receiving permission. The benefit would depend on the problem, the quality of available information, the authority a person grants, and the checks around the action. “Universal problem solver” is a useful thought experiment, but it is too strong as a description of any demonstrated system today.
Narrow superintelligence: exceptional depth without general intelligence
I find the question of narrow superintelligence useful because it separates two ambitions that are often bundled together: being extraordinarily good at something and being able to handle many different things. Here, I use the term to mean performance beyond the best human experts within a clearly defined domain. It does not imply broad judgment, consciousness, or permission to act independently.
The distinction has a research basis, although the terminology needs care. Google DeepMind’s Levels of AGI framework separates the breadth of tasks from the level of performance and places AlphaZero in its “Superhuman Narrow AI” category. In that framework, “Superhuman” means outperforming all humans on the task, a stronger claim than beating an average practitioner or a selected group of experts. This is the authors’ proposed classification, not a universal certification or a claim about every task an AI can attempt (Morris et al., Levels of AGI, 2024 revision).
AlphaZero makes the distinction concrete. The same learning approach achieved exceptional results in chess, shogi, and Go, but it was trained separately for each game. This was not one trained system picking up an unfamiliar game without further training. Its success within those formal environments does not establish an ability to negotiate a contract, understand a person’s circumstances, or decide which business objective deserves pursuit (Silver et al., Science, 2018; DeepMind, AlphaZero).
For an organization, the useful question may therefore be smaller than “When will AGI arrive?”: what could we do differently if one difficult part of our work became much better? Consider a hypothetical scheduling system that proposes better production plans than an expert team, given the same information, operating constraints, and agreed measures of plan quality. Before calling that superhuman performance, I would want a fair comparison on representative work, including disruptions, missing information, and the cost of correcting mistakes. That evidence would establish scheduling performance, not good judgment about whose priorities should prevail. Authority to promise a delivery date is a separate decision about delegation and accountability.
A small language model, an open-weight model, or a specialized combination of models and ordinary software might be sufficient for a particular task. None earns a superintelligence label simply by being specialized or inexpensive. The task, expert baseline, operating conditions, and evidence determine the claim. My practical interest is in what people can achieve with such systems: let the specialist extend our abilities, while people remain responsible for the objectives, tradeoffs, and commitments. That opportunity does not have to wait for AGI, and assembling several specialists does not by itself prove that AGI has arrived.
Current State of AI Technologies
Today’s AI covers a collection of systems with different training methods, inputs, strengths, and failure modes. Image recognition does not imply task planning, and a language model’s ability to call tools does not guarantee a correct result. Many products combine a model with ordinary software, databases, search, rules, and human review; that can be useful without being AGI.
Strengths
- Specialization and efficiency in narrow tasks. AI can perform strongly when a task is well specified and a suitable evaluation exists. Image and speech recognition, translation, text generation, and game-playing are familiar examples. DeepMind’s AlphaGo defeated Lee Sedol in a five-game Go match in 2016; that was a striking achievement in a defined environment, with explicit rules and a clear outcome (DeepMind, AlphaGo). Performance in a well-bounded task should be taken seriously, while keeping the boundary visible.
- Scale and data processing. Software can search, classify, summarize, or compare collections of material faster than a person could inspect each item manually. This can help in research and analysis, including medical or financial work, when data is suitable and results are checked. Throughput is not the same as sound judgment: a system can process more records and still reproduce a sampling problem, miss a relevant variable, or produce a confident but incorrect answer.
- Learning and adaptation. Some machine-learning systems are updated using new data or feedback, and some can adapt to examples in a conversation or task. These mechanisms differ. A deployed model does not ordinarily rewrite its own weights each time a user corrects it. A product may keep conversational context, save a preference, retrieve a record, run a separate learning job, or update a model after a controlled training process. Each has different privacy and reliability implications.
“Learning from experience” can refer to several technical processes, and a user should not assume that feedback changed the underlying model or that the change will persist. A product should say what it stores, for how long, and how users can inspect or remove it.
Limitations
- Generalization and transfer. Systems can transfer useful patterns to new inputs, but their reliability may change when a task or environment differs from the examples and evaluations used to build them. This is not the same as saying that current AI cannot generalize. The practical question is how performance holds up under distribution shift, novel combinations, incomplete context, and adversarial or simply unusual inputs.
- Reasoning, context, and reliability. Models can solve some problems that involve multi-step reasoning, but a correct answer on one prompt does not establish stable understanding or a dependable reasoning process. They may overlook causal structure, misread a reference, or invent a detail that fits the surrounding text. Evaluation should match the consequences of the task and include cases where evidence is incomplete or contradictory. Human language also carries ambiguity, culture, and emotion that cannot be reduced to fluent text output.
- Ethical and social considerations. Bias, privacy, misuse, security, access, and effects on work are system and deployment questions, not only properties of a model. A technically capable model can still be unsuitable for a use case if the data are not representative, the affected people have no recourse, or no one is accountable for the decision. Fairness and transparency require evidence and governance; they are not automatic results of adding an “ethical” module.
A fair assessment has to leave room for both sides. AI has moved beyond narrow pattern matching in some settings, and blanket claims that it cannot reason or transfer knowledge are too sweeping. At the same time, a polished demonstration is not evidence that performance will remain dependable across domains. The testing conditions and the limits of the claim need to travel with the result.
The Pathway to AGI: Integrating AI Models and Technologies
One approach to more capable systems is to combine methods, but it is not established that AGI must come from an assembly of multiple models. A single model may become more capable; a system may also combine specialized models, search, tools, and conventional software. The useful architectural choice depends on the task, evidence, latency, cost, and consequences of failure. Integration can improve a system, but it does not itself prove general intelligence.
Hybrid models. Symbolic methods can represent explicit rules or relations, while neural networks can learn patterns from data. Combining them may help when a task needs both pattern recognition and constrained operations. It also adds interfaces and failure points: the team must know which component produced a result, how conflicts are resolved, and what happens when the learned output does not fit the rule system.
Transfer and multitask learning. A model that handles several related tasks may reuse useful representations rather than start from scratch for each one. Transfer is valuable when the source and target tasks are meaningfully related, but success does not guarantee transfer to a distant domain. Evaluation should test the target use directly rather than infer its quality from a neighboring benchmark.
Learning from fewer examples. Few-shot and meta-learning research explores how systems can adapt from limited examples. A related line asks whether a model can assemble reasoning steps for a particular problem: Google DeepMind’s 2024 SELF-DISCOVER study tested this approach on benchmark tasks (DeepMind, “Large Language Models Self-Discover Reasoning Structures”). People also bring prior knowledge, physical experience, and goals to a new task, so the comparison is not simply “one example versus many.” A model’s apparent few-shot skill may depend on patterns learned during training or examples in the current prompt; evaluation should specify what information it received.
Alignment and social context. Safe behavior cannot be reduced to putting a list of values into a model. Teams need to define who can authorize an action, how the system handles uncertainty, how affected people can contest outcomes, and how behavior is monitored after release. Philosophy, psychology, law, and domain expertise can inform those choices, but governance must be tied to actual controls and evidence.
Building Blocks of AGI
The phrase “building blocks” is useful if it means capabilities that may contribute to more general systems. It should not suggest that the pieces form a known recipe or that a system becomes generally intelligent once each component is present. The sections below examine models, agents, communication, decisions, and hardware as separate capabilities.
1: The Foundation of AI Models
Mixture of Experts. A Mixture of Experts model contains expert subnetworks and a routing mechanism that selects which experts process an input. Sparse routing can increase the total number of parameters without activating every parameter for every token. It may help scale a model, but it is not a measure of general intelligence. The public GPT-4 technical report does not disclose enough architecture detail to establish that GPT-4 uses an MoE design (OpenAI, GPT-4 Technical Report). Mistral AI’s public MoE announcement describes Mixtral 8x7B, a specific model, rather than proving that every Mistral model or every leading model uses MoE (Mistral AI, Mixtral of Experts). Google’s February 2024 Gemini 1.5 announcement describes a MoE architecture for that model; this is a dated, model-specific statement (Google, Gemini 1.5).
Multimodal models. A multimodal system can accept or generate more than one type of input or output, such as text, images, audio, or video. In dated examples, GPT-4V was made available to analyze user-provided image inputs, and Google described Gemini 1.0 as natively multimodal across text, code, audio, image, and video (OpenAI, GPT-4V System Card; Google, Gemini 1.0 announcement). These specific capabilities illustrate multimodal systems; they do not establish human-like perception or AGI. Combining modalities can make interaction more natural and allow a system to relate information across them. Inputs may be missing, poorly captured, or interpreted incorrectly; a user should be able to see what evidence the system used when the stakes warrant it.
Context windows. The context window is the material a model can consider during a particular inference, subject to the product’s implementation. A large window can make it possible to provide a long document or conversation at once. This can be useful when a task depends on a full contract, a set of research papers, or an entire manuscript being available together. It is not automatically persistent memory, reliable retrieval, training, or a guarantee that every detail will be used correctly. Long input also needs good organization: an important fact can be present in the prompt and still be missed or misapplied.
Magic announced LTM-1 in 2023 with a 5-million-token context window, and Google described Gemini 1.5 Pro with a standard 128,000-token window and a 1-million-token private preview in February 2024 (Magic, LTM-1; Google, Gemini 1.5). These are dated provider specifications, not independently comparable measures of understanding. A token-to-page conversion depends on the material and formatting, so a page analogy cannot establish how much useful information a model will retrieve or apply. The practical question is what context can support, not a capacity ranking.
2: Autonomous AI Agents
An agent is often described as a system in which a model chooses some of its own next steps, including tool use, in pursuit of a goal. A workflow can also call a model at defined stages while ordinary code directs the overall sequence. The distinction is useful, although products mix both patterns. Anthropic’s description of agentic systems and OpenAI’s practical agent guide make these architectural choices explicit (Anthropic, “Building Effective Agents”; OpenAI, “A Practical Guide to Building AI Agents”).
Breaking a request into subtasks can help a system handle longer work: it might gather information, draft an answer, check a result, then ask before taking an external action. “Autonomous” should be tied to specific permissions: which files the system may read, which messages or records it may change, whether it can spend money or delete content, what requires approval, and how actions can be reversed. A model’s ability to propose or call a tool does not establish those limits; the surrounding application and access controls do.
In 2024, Google’s I/O materials described a long-term ambition to build a universal agent for everyday life. Rabbit presented r1 as a device that could move from words to actions, and MultiOn documented a browser-agent API. OpenAI later presented Operator as a computer-using research preview; Open Interpreter documents a desktop agent that acts within a selected workspace and permission model (Google I/O 2024 keynote; Google DeepMind, Project Astra announcement; Rabbit, r1 announcement; MultiOn Agent documentation; OpenAI, Operator System Card; Open Interpreter desktop documentation). These are examples of systems built to interact with software, not evidence of unrestricted autonomy or AGI. A benchmark or demonstration covers particular tasks and conditions, not every desktop, website, or consequential action.
Sora belongs in a separate category. Its 2024 research release described a video-generation model, not a general-purpose autonomous agent. Generating a video from instructions is a significant capability, but it is different from setting a goal, choosing tools, and carrying out actions in an external environment.
Interactions and Decision Making
Interaction and decision authority are related, but they are not the same. A system may converse naturally while having no permission to act. Another may execute a fixed operation without a natural-language interface. Good design tells people whether they are receiving information, a recommendation, a prepared action, or a completed action. It also makes the evidence and limits visible enough for a person to respond.
3: Enhancing Communication with AI
Speech gives people another way to provide information and receive a response. It can make systems easier to use, particularly when typing is difficult or the user is moving between tasks. Speech recognition converts audio into text or another representation; speech synthesis converts text into audio. Neither step establishes that a model understands the speaker’s intent or emotion.
From voice to text. OpenAI’s Whisper paper describes a model trained with large-scale weak supervision on 680,000 hours of audio paired with internet transcripts, including multilingual and speech-translation data. That is a historical training-data figure in the paper, not a current specification for every speech-recognition product (Radford et al., “Robust Speech Recognition via Large-Scale Weak Supervision”). In the paper’s additive-noise tests, Whisper degraded less than the compared LibriSpeech-trained systems under white and pub noise, especially at signal-to-noise ratios below 10 dB; that result is limited to those evaluations (Radford et al., “Robust Speech Recognition via Large-Scale Weak Supervision”). Recognition quality can still vary with noise, accents, specialist terms, microphone quality, and overlapping speakers. A transcript is a useful representation, not a perfect record of what someone said. In sensitive settings, people should be able to review the transcription before it is used to make a decision.
From text to voice. Modern speech synthesis can control pacing, intonation, and voice characteristics. In 2024, ElevenLabs’ multilingual v2 supported 29 languages, while latency depended on model and implementation. Its current documentation lists separate language coverage and latency characteristics for different models, including v4, v3, multilingual v2, and Flash variants (ElevenLabs model documentation; audio streaming and latency). The number is model-specific, and an inference-latency figure is not the same as the time a person waits for audio to begin in an application.
A natural voice can improve access and comfort, but it can also make generated content sound more authoritative than its evidence deserves. Voice cloning creates additional consent and impersonation risks. A deployment should identify generated speech where appropriate, limit who can create or use a voice, and provide a clear way to correct recognition errors. Conversational fluency is not emotional understanding.
4: AI’s Decision-Making Capabilities
Decisions with real-world consequences raise questions about long-term effects, ethical implications, and how information from different sources is combined. The examples below show capabilities within bounded tasks and conditions, not general judgment.
Game-playing systems show how a model can select actions in a defined environment and improve under a measurable objective. AlphaGo combined neural networks with search and defeated Lee Sedol 4–1 in 2016. AlphaZero later trained through self-play to play chess, shogi, and Go using the rules of each game (DeepMind, AlphaGo; DeepMind, AlphaZero). These results demonstrate powerful learning and planning within formal games. They do not establish that the same system can transfer its skill to unrelated tasks without new training or tools.
Autonomous vehicles are another example of perception, prediction, planning, and control interacting in a changing environment. A vehicle system’s capabilities are bounded by its operational design domain (ODD): the conditions and service areas for which it is designed (NHTSA, Automated Vehicles). “Self-driving” does not mean that every vehicle can safely handle every road or condition. The human role depends on the automation level: drivers remain responsible for monitoring Level 2 systems, while Level 4 systems handle driving only within limited service areas and do not need an engaged human driver. System designers also define operating conditions, transition behavior, and applicable local rules. A comparison to human decision-making should name the task and conditions rather than imply a general guarantee.
Enabling Technologies
AI systems depend on more than a model. Storage, data pipelines, networking, software libraries, evaluation, monitoring, and people who can intervene all affect what the system can do. Hardware matters because training and serving large models require computation, memory bandwidth, energy, and reliable connections among components. Improvements in one area can expose constraints in another.
5: Specialized AI Hardware
Specialized hardware can make particular workloads faster or more efficient; it does not provide a direct route from scale to AGI.
The Role of AI Chips and Hardware in Developing AGI
AI accelerators are designed to perform operations common in machine learning, often with different tradeoffs from general-purpose CPUs. The practical performance of a system depends on the model, numerical precision, batch size, memory capacity and bandwidth, interconnect, software stack, and workload. A peak throughput number does not predict how quickly a complete training or inference job will finish.
Compute capacity, memory, energy use, and access to hardware influence which experiments organizations can run and at what cost. Greater scale can support larger training runs, but it also raises questions about power, cooling, data-center capacity, supply chains, and who can afford to participate. None of those infrastructure measures is an AGI test.
Innovation in AI Hardware
Hardware development spans accelerators, memory, interconnects, packaging, cooling, and software. Neuromorphic research explores hardware inspired by aspects of neural computation, but it is a research direction rather than an established requirement for AGI. A chip’s similarity to a brain is not evidence that it has human-like cognition.
NVIDIA announced the H100 data-center accelerator in March 2022 as its first Hopper-based GPU (NVIDIA, Hopper architecture and H100 announcement). Accelerator performance comparisons depend on precision, workload, configuration, and metric; a peak number alone does not predict whole-job time. NVIDIA’s Grace Hopper systems combine a Grace CPU and Hopper GPU with a high-bandwidth NVLink-C2C connection; this is a specific CPU-GPU system design, not a generic description of every AI rack (NVIDIA, Grace performance tuning guide).
AMD’s Instinct accelerators and Microsoft’s Azure Maia are separate products. In 2023 AMD and Microsoft announced that MI300X accelerators would power an Azure virtual-machine series; Microsoft also announced Maia as its own custom AI-accelerator program (AMD, Microsoft Azure MI300X announcement; Microsoft Azure Maia announcement). The MI300X deployment and Maia program represent separate approaches to expanding Azure AI infrastructure.
Intel completed its acquisition of Barefoot Networks in 2019, adding programmable Ethernet switching technology to its portfolio (Intel, 2019 CSR report; Intel acquisition announcement). Intel also introduced Gaudi 3 in April 2024; Ponte Vecchio was its Data Center GPU Max Series product, launched in 2023; and Meteor Lake Core Ultra processors launched in December 2023 with an integrated neural processing unit (NPU) for power-efficient AI acceleration and local inference. These are examples from different product generations, not current product rankings (Intel, Gaudi 3; Intel, Data Center GPU Max 1100; Intel, Core Ultra launch). Google announced TPU v4 in 2021 and made Cloud TPU v4 capacity available through Google Cloud in 2022 (Google Cloud, TPU v4 announcement; Google Cloud TPU v4 availability). Google publishes research and software related to TPUs and offers cloud access, but this does not make the physical chip design open hardware.
Expanding semiconductor capacity calls for substantial capital and coordination across design, fabrication, packaging, power, cooling, and supply chains. Proposals and reported ambitions do not equal money raised, a fabrication contract, a completed plant, or delivered chip capacity. Those distinctions matter because access to compute affects which organizations can train and serve large models, while the cost and energy demands can limit who benefits.
Towards Achieving AGI
Research activity points to possible directions, not a dated countdown. There are active lines of work in multimodal models, memory and retrieval, planning, tool use, learning efficiency, evaluation, and hardware. Progress on one line may help with others; it may also reveal new failure modes. A stronger benchmark score can show improvement on that benchmark without resolving whether a system can adapt reliably across the range of tasks implied by AGI.
6: Combining All Elements
A system may combine a foundation model, tools, retrieval, planning, a speech interface, and conventional code. One design lets the model interpret a request, propose subtasks, search, and return a draft; another uses a fixed workflow and calls a model only for extraction or summarization. Either can work when the team knows which component chose each step and which one can change external state.
Broader inputs can improve communication; tools can extend what an agent can do; verifiers and human reviewers can catch some errors; and purpose-built accelerators can reduce run time. Each addresses a different limit. Together, these components still need a defined job, permission boundaries, testing, observability, and an owner who can respond when the system fails. A component list alone cannot show that they work reliably as a whole.
Integrating Technologies and Methodologies
An integration plan should begin with the job to be done and the consequences of an error. The team can then decide whether it needs one model, several specialists, a deterministic workflow, human review, or some combination. It should record the inputs available to each component, the authority to act, how the result is checked, and what evidence will count as success.
Modularity can make systems easier to replace and test. It can also make failures harder to trace when data passes through several services. A speech recognizer may mishear a name; retrieval may find a stale policy; a model may summarize it incorrectly; an agent may then submit a change. Logging the decision path, using explicit confirmation for consequential actions, and testing the full chain matter more than assuming each component is “intelligent.”
Challenges and Solutions
A system intended to be more general can create new value and new dependencies. Compatibility, data protection, bias, misuse, and changes to work require practical responses rather than a promise that governance will catch up later.
Integration Challenges
Complexity and compatibility. Different models, databases, and tools can use incompatible formats or make different assumptions. A component update can change outputs even when the overall application code stays the same. Versioning, interface tests, and end-to-end evaluations help identify what changed. Teams also need a fallback when a model or service becomes unavailable.
Data and privacy. A system can only use information it can access, but access should be limited to what the task needs. Personal data may travel from an interface to a model provider, logs, search indexes, or analytics systems. Teams need to understand retention, access, and deletion behavior, and they should avoid sending sensitive material when it is not necessary. Giving an agent broad computer or repository access increases the impact of prompt injection and mistaken actions.
Ethics and social effects. A model may reproduce bias, but the deployment can also create harm through poor workflow design, unreviewed automation, or a lack of appeal. A job-impact estimate should be treated as a question for evidence, not a fixed prediction. People affected by an automated decision need to know who is responsible and how to challenge it. High-stakes applications require domain-specific testing and oversight.
Potential Solutions
Interdisciplinary research and collaboration. Technical teams benefit from working with subject-matter experts and people affected by the system. A clinician can identify whether a summary omits a clinically important distinction; a teacher can judge whether a learning aid supports the intended lesson; a security specialist can probe what happens when retrieved content contains malicious instructions. Consultation is useful when it changes the tests, data, or controls, not as a label of approval.
Open standards and modular design. Documented interfaces, portable data formats, and replaceable components can reduce lock-in and make systems easier to test. Open standards do not automatically make a model’s data or hardware open, and modularity does not guarantee compatibility. Teams still need to verify behavior across versions and ensure that permissions follow the data and actions through every component.
Governance and evaluation. Governance should specify roles, permitted use, records, escalation, and recourse. The NIST AI Risk Management Framework offers a voluntary structure for organizing AI risk work; it does not certify a system or replace domain regulation (NIST AI RMF). Evaluation should reflect the intended users and conditions, test known failure cases, and monitor changes after release. For medical software, classification and obligations depend on the product’s function and context; FDA guidance distinguishes certain clinical decision-support functions and describes transparency considerations for machine-learning-enabled devices (FDA, Clinical Decision Support Software; FDA, Transparency for Machine Learning-Enabled Medical Devices).
Public engagement and education. People should be able to understand where AI is used, what it can change, and how to raise a concern. Training helps users recognize the limits of a system, but it cannot transfer accountability from the organization that selected and deployed it. Engagement should happen before design decisions are locked, especially when the affected group has little power to opt out.
More capable AI may expand what people can investigate and create, while increasing the importance of evidence, data stewardship, and accountability. The value will come from systems that solve specific problems well and make their boundaries clear. Whether any system meets a chosen AGI definition remains a separate question.
Is the object in the mirror closer than it appears?
Public discussion in 2024 included speculation that leading labs might be approaching or may already have reached AGI. A claim without verifiable evidence cannot establish that a system meets a defined capability threshold.
A serious claim should state its definition, the tasks tested, how much human intervention was involved, and what independent evidence is available. It should also show whether the system can generalize beyond the examples used in training and evaluation. A benchmark, product launch, or vivid demo can inform that assessment, but cannot settle it alone. Without an agreed test and stated threshold, a declaration that AGI has arrived would outrun the evidence.
Food for thought…
OpenAI’s February 2024 Sora research showcase. This curated set of prompt-and-video examples preserves the announcement’s historical context. The canvas, Minecraft and interpolation findings discussed below come from OpenAI’s linked technical report.
In February 2024, OpenAI presented Sora as a video-generation model and showed camera motion, scene consistency, object persistence, marks remaining on a canvas, a prompted Minecraft scene, and interpolation between two input videos to create transitions between different subjects and scene compositions. The research report described such examples as capabilities observed in video models trained at scale and discussed their possible relevance to world simulation (OpenAI, “Video generation models as world simulators”). These demonstrations showed visual regularities, not a measured physical model.
The report also described failures in basic physical interactions, including glass shattering, and inconsistencies in changes to a scene. A convincing camera move or persistent object does not prove a general causal understanding of physics, and a video sample cannot reveal the contents of a private training set.
In the Minecraft example, the model rendered a game world while controlling a player with a basic policy; that prompted video is not evidence of reliable gameplay or general physics. The original Sora product is no longer available as of April 26, 2026, a product-status change separate from the 2024 research findings (OpenAI, Sora).
What interests me here is whether a model can learn useful regularities about motion, including fluid motion, from visual examples. OpenAI reported that some simulation-like behaviors emerge without explicit 3D or object inductive biases, but its demonstrations do not isolate what was learned or test whether Sora learned fluid dynamics. The useful next step is to test which physical regularities it can carry into unfamiliar scenes, and where those expectations break down.
Discover more from Ajith Vallath Prabhakar
Subscribe to get the latest posts sent to your email.