A world model in artificial intelligence is a machine learning system that builds an internal representation of an environment. The model predicts how that environment changes over time in response to actions. Researchers design world models to help agents plan, reason, and act without constant real-world trial and error. World models differ from systems that merely classify or generate outputs. They simulate dynamics such as physics, object interactions, and causality. Early ideas date to the 1990s. Modern versions power robots, autonomous driving, and interactive video generation. == History == Jürgen Schmidhuber introduced the term world model in machine learning in 1990. He proposed recurrent neural networks that predict future states from observations and use those predictions to train agents. David Ha and Schmidhuber revived the concept in a 2018 paper. Their agents learned to drive virtual cars and play video games inside self-generated simulations. Yann LeCun advanced the idea in a 2022 position paper titled "A Path Towards Autonomous Machine Intelligence". He argued that intelligence requires predictive models of the world rather than pure pattern matching. LeCun proposed the joint embedding predictive architecture (JEPA) as a practical foundation. LeCun and collaborators developed several JEPA variants. V-JEPA 2 reached state-of-the-art performance on video understanding and physical reasoning at the time. It supports zero-shot robot control in unfamiliar environments. Introduced in March 2026, LeWorldModel trains stably end-to-end from raw pixels and uses two loss terms and avoids hand-crafted heuristics. LeCun founded Advanced Machine Intelligence Labs in 2026 to further develop world models. Google DeepMind introduced Genie in 2024. The model learned interactive environments from unlabeled internet videos. Genie 2 followed in late 2024 and added three-dimensional generation. The Genie series set benchmarks for general-purpose simulation. Genie 3 was introduced in August 2025. It produces photorealistic, real-time interactive worlds from text prompts which are displayed at 24 frames per second and explored in real time with text or image prompts. The model supports persistent three-dimensional worlds and real-time interaction. Waymo adopted Genie 3 in February 2026 and used it to create a specialized world model for autonomous driving simulation, called the Waymo World Model. It produces synchronized camera and lidar outputs and creates edge cases that real robotaxis rarely encounter. The edge cases were reported to be unusual by PCMag. General Intuition announced a $133.7 million seed round. World Labs raised $1 billion. AMI raised $1.03 billion. In April 2026, Alibaba announced Happy Oyster, its world model designed for real-time and “flowy” world model. It includes a directing mode for world building based on text and image prompts and a wandering mode for exploring the resulting world. It can generate 3-minute in-world video clips. Also in April, World Labs, co-founded by Li Fei Fei, unveiled Spark 2.0, an open-source 3D Gaussian splatting rendering engine that targets smartphone-class devices. In June 2026, Nvidia released Cosmos 3, a family of open-weight models. It combines previously independent physical reasoning, world simulation, and action generation. Cosmos 3 integrates can process and generate text, image, video, audio, and action sequences. The model employs a Mixture-of-Transformers" (MoT) approach. An autoregressive (AR) transformer handles reasoning and next-token prediction, while a diffusion transformer (DT) does multimodal generation. Encoders (ViT for vision, VAE for visual/audio, and domain-specific for actions) and generate a shared representation space using 3D multi-dimensional rotary position embedding (mRoPE) for spatial and temporal information. The family includes Cosmos3-Nano (16B parameters) for workstations; Cosmos3-Super (64B parameters) for research. == Architecture == World models process raw sensory data such as video frames or lidar scans. They compress this input into compact latent representations. The system then predicts future representations rather than pixel-by-pixel reconstructions. Many modern world models use joint embedding predictive architecture (JEPA). An encoder turns observations into embeddings. A predictor estimates one or a suite of embeddings from the current one and an action. In some cases a critic chooses one embedding as the best result. A regularizer keeps embeddings well-behaved. The model trains by minimizing prediction error in embedding space. This approach avoids the high cost of generating every detail. Some architectures add explicit components. A fast reactive path handles immediate responses. A slower deliberative path performs longer-horizon planning. Video prediction accuracy or robot success rates are key metrics, but do not always predict real-world performance. Generative world models such as Genie 3 combine these with a simulator. They accept text prompts or layouts and output consistent video, lidar, or three-dimensional scenes. World models often train with self-supervised learning. They use large unlabeled datasets of video or robot interactions. Self-supervised learning can speed learning. Reinforcement learning can fine-tune a model for specific tasks. == Applications == World models support robot learning. Agents train inside simulations and transfer skills to the physical world. This reduces the need for dangerous or expensive real-world trials. Autonomous vehicles use world models to test rare events. Waymo's system simulates tornadoes or unusual pedestrian behavior. Companies train planners without putting vehicles on public roads. Interactive entertainment benefits from world models. Genie 3 lets users generate playable environments from simple descriptions. Game studios prototype levels faster. Scientific simulation gains from these models. Researchers model physical systems or biological processes at scale. Planners in logistics or urban design test strategies inside accurate digital twins. == Comparison with large language models == Both world models and large language models (LLMs) use inferencing on their inputs to make predictions. LLMs operate on textual inputs. They predict the next token in text sequences. They excel at language-oriented tasks such as translation or summarization. However, they lack understanding of physics. World models operate on sensor inputs such as pixels. They predict state changes in that data in latent space. This design supports planning and causal reasoning. LLMs generate fluent text but often fail at consistent physical predictions. Their architecture employs transformers with refinements such as mixture of experts. World models divide an inferencing task into work performed by encoders, predictors, simulators, and other pieces. They typically handle multimodal inputs such as video, lidar, radar, and audio, guided by textual prompting. LLMs power chatbots and code assistants. World models drive embodied agents that act in dynamic environments, such as autonomous driving. The two may be combined in hybrid systems. For example, a LLM handles instructions, while a world model manages low-level control. World model proponents such as LeCun claim that because LLMs are trained only on text, they have no ability to predict anything beyond text, such as real-world events. == Benchmarks == World model benchmarks test physical understanding, long-term consistency, planning, and generalization from sensor data. Meta introduced three benchmarks for V-JEPA 2. IntPhys 2 measures a model's ability to detect physics violations. It presents pairs of videos that diverge when one breaks physical rules. Humans score near 100% accuracy. V-JEPA 2 achieves little better than random chance on many conditions. Minimal Video Pairs (MVPBench) tests physical understanding through multiple-choice questions based on short video clips. It probes object interactions and causality. Something-Something tests action recognition. Epic-Kitchens-100 tests human action anticipation. DeepMind benchmark: Interactive evaluation measures consistency over minutes of interaction, memory of off-screen objects, and response to user actions or text prompts. Waymo benchmark: Output generation quality: Metrics include realism, controllability (via text prompts), and usefulness for training planners in simulated worlds. However, pixel reconstruction error rate with episodic rewards often fails. Other: Epic-Kitchens-100 (often measured with Recall@5) Ego4D 50 Salads, Breakfast, etc. Potential benchmarks: Zero-shot transfer to robots Long-horizon planning Implausible prediction rate
The Best Free AI Presentation Maker for Beginners
Shopping for the best AI presentation maker? An AI presentation maker is software that uses machine learning to help you get more done — it keeps getting smarter as the underlying models improve. Pricing, accuracy, and the size of the model behind the tool are the three factors that most affect daily usefulness. Whether you are a beginner or a pro, the right AI presentation maker slots into your workflow and pays for itself fast. Below we compare features, pricing, and real output so you can choose with confidence.
Alexander Y. Tetelbaum
Alexander Y. Tetelbaum (born August 16, 1948) is a Ukrainian American computer scientist, inventor, and academic who has contributed to electronic design automation (EDA) and artificial intelligence (AI) since the late 1960s; and holds 46 U.S. patents in EDA and related fields. Tetelbaum is the founding president of International Solomon University, the first Jewish university in Ukraine, established during a period of renewed efforts to address antisemitism in Ukraine. == Early life and education == He graduated from a Kyiv mathematical high school with a silver medal in 1966. Tetelbaum enrolled at the Kyiv Polytechnic Institute (KPI), now National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" in 1966, graduating in 1972 with an MS in Electronics with honors. He earned his PhD in Electrical and Computer Engineering from KPI in 1975, with a dissertation on electronic design automation, and his Doctor of Engineering Science in 1986. == Academic career == Tetelbaum began his academic career at KPI in 1973 as a junior scientist, becoming a professor in the Computer and Electrical Engineering Department in 1980. Later, he founded and served as president of International Solomon University in Kyiv from 1991 to 1996, the first Jewish university in Ukraine. The university became a major academic center for computer science and Jewish studies in the post-Soviet era. He was a visiting and adjunct professor at Michigan State University from 1993 to 1996. == Professional career == Tetelbaum worked as an engineer at the Kiev Institute of Cybernetics from 1972 to 1973, and later, he led the Design Automation Lab at Kyiv Polytechnic Institute from 1975 to 1987. In the United States, he served as EDA manager at Silicon Graphics Corporation from 1996 to 1998 and principal engineer at LSI Corporation from 1998 to 2012. He founded and served as CEO of Abelite Design Automation, Inc., from 2012 to 2022. == Contributions in computer science == Tetelbaum has contributed to electronic design automation (EDA) and artificial intelligence (AI) since the 1960s. His early work included methods for EDA, particularly physical design automation and mathematical optimization; and he developed force-directed placement and topological routing methods. Tetelbaum generalized Rent's rule for hierarchical systems and large blocks, proposing a graph-based framework that extends applicability to arbitrary partition sizes with improved accuracy. Additional IEEE and related conference contributions from the mid-1990s include: "Path Search for Complicated Function", 1995 IEEE International Symposium on Circuits and Systems "A Performance-driven Placement Approach of Standard Cells" (International Conference on Intelligent Systems, 1995) "Framework of a New Methodology for Behavioral to Physical Design Linkage" (38th Midwest Symposium on Circuits and Systems, 1996) Statistical timing design and variations Test Methodologies These and other works and patents contributed to timing-driven placement, crosstalk reduction, clock tree synthesis, and interconnect optimization in VLSI design. == Patents == Tetelbaum holds 46 U.S. patents in EDA and related fields. Notable examples include: For the full list of patents, see Justia Patents or Google Patents. == Publications == === Early publications in the Soviet Union === Before the appearance of American books on electronic design automation (EDA), Tetelbaum published several scientific books and monographs on the subject in Russian/Ukrainian. Electronic Design Automation, Kiev: Znanie Publisher, 1975. Planar Design of Electronic Circuits, Kiev: Znanie Publisher, 1977. Formal Design of Computer Systems, Moscow: Sovetskoe Radio, 1979. CAD of Electronic Equipment: Topological Approach, Kiev: Vyssha Shkola, 1980; 2nd ed. 1981. Automated Design of Electronic Circuits (1981) CAD of VLSI Circuits, Kiev: Vyssha Shkola, 1983. Topological Algorithms of Multilayer Printed Circuit Boards Routing, Moscow: Radio i Svyaz, 1983. CAD of VLSI Circuits on Master Slice Chips, Moscow: Radio i Svyaz, 1988. Increasing the Effectiveness of CAD Systems, Kiev: UMKVO, 1991. === Scientific Monographs (English) === Minimum Number of Timing Signoff Corners (2022) Interviewing AI (2026) The AI Debate (2026) New Nostradamus Predictions: 2026: The Next Decade & Beyond (2035–2050+) (2026) For a consolidated record of Tetelbaum's publications, see Alexander Y. Tetelbaum, Wikidata Q4720205. === Other publications === Tetelbaum also published educational books on problem-solving methods: Yes-No Puzzles-Games Puzzle Games for Kids Solving Non-Standard Problems Solving Non-Standard Very Hard Problems Additionally, Tetelbaum published three thrillers: Omerta Operations Executive Director Eruption Yacht Finally, he published his memoir and an entertaining book: Unfinished Equations Artificially Intelligent Humor
AI washing
AI washing is a deceptive marketing tactic that consists of promoting a product or a service by overstating the role of artificial intelligence (AI) and the integration of it. Companies often involve in the practice to mislead customers to boost their offerings, and to secure funding from investors. The practice raises concerns regarding transparency, and legal issues. == Definition == AI washing is a deceptive marketing practice. It involves promoting a product or a service by overstating the role of artificial intelligence (AI) and its integration in the design and manufacture of the same. The practice raises concerns regarding transparency, compliance with security regulations, and consumer trust in the AI industry potentially hampering legitimate advancements in AI. The term was first defined by the AI Now Institute, a research institute based at New York University in 2019. The term is derived from greenwashing, another deceptive marketing technique that misrepresents a product's environmental impact in a similar manner. AI washing might involve a company claiming to have used AI in the development or enhancement of its products or services without its actual involvement, or using buzzwords such as "smart" or "AI-powered" without the product actually offering it or making use of it. A company may overstate the usage of AI or misuse the term, which is also construed as AI washing. In 2026, The Washington Post defined AI washing as "a trend for bosses to blame layoffs on the productive capabilities of AI and its ability to replace workers, even when job cuts may have little to do with the technology". == Usage and effects == AI washing can lead to deception of customers and misleading of investors. It is also an illegal and unethical practice that lacks transparency regarding disclosing the details of a product or a service. Companies get involved in such a practice often in response to competition who might have used AI in their offerings. It might also be used as a ploy to secure funding and investment, assuming that it will attract them towards it. AI washing has been compared to dot-com bubble, when businesses appended "dot-com" to the end of the business name to boost their valuation. In September 2023, Coca-Cola released a new product called Coca-Cola Y3000, and the company stated that the Y3000 flavor had been "co-created with human and artificial intelligence". The company was accused of AI washing due to no proof of AI involvement in the creation of the product, and critics believed that AI was used as a way to grab consumer attention more than it was used in the actual product creation. In 2026, mass tech layoffs were attributed to AI washing from AI innovation instead of balance sheet restructuring. == Mitigation == Companies are expected to be transparent and clearer in communicating the usage of AI in their products or services. Consumers can mitigate the same by requesting for hard evidence from the companies regarding the usage of AI tools. Customers should evaluate the product or service as a whole rather than being swayed by the usage of AI. Informed decision making and purchasing can keep them from falling for such marketing gimmicks. The United States Securities and Exchange Commission (SEC) imposes penalties for companies indulging in such practices. In March 2024, the SEC imposed the first civil penalties on two companies for misleading statements about their use of AI, and in July 2024, it charged a corporate executive from a supposed AI hiring startup with fraud for the usage of buzzwords related to AI.
AI browser
An AI browser is a web browser with integrated artificial intelligence capabilities, such as automatically summarizing web page content or answering questions about it. A more specialized type is an agentic browser, based on the concept of agentic AI, which can take actions – such as navigating webpages or filling out forms – on behalf of the user. Several agentic browsers emerged in 2025, including ChatGPT Atlas (macOS only), Comet, and Dia. As of 2025, this is a recent development in the browser market, including new entrants from OpenAI, Opera and Perplexity. The designation of 'AI browser' also includes established browsers that later added non-agentic AI features, such as Microsoft Edge with the Copilot chatbot, Google Chrome with the Gemini chatbot (for Windows desktop users in the US with their language set to English), and Firefox with multiple chatbot providers (such as ChatGPT, Claude, Copilot, Gemini, and Le Chat). AI browsers have been noted to be susceptible to prompt injection attacks. == Browser extensions and integrations == Rather than creating entirely new browsers, some AI browsing solutions integrate with existing browsers through extensions or companion applications. These tools add agentic capabilities to established browsers without requiring users to switch platforms. Examples include Composite, which functions as a cross-browser agent that works with Chrome, Edge, and other browsers to automate web-based tasks for workers. == Cloud-based implementations == Cloud-based implementations of AI browsers allow users to run automated browsing agents without local installation. These systems operate on remote servers using frameworks such as Puppeteer or Playwright. Examples include Browserbase, Browser-use and AI Browser. The AI typically parses the Document Object Model (DOM) to locate and interact with page elements, and may also analyze browser screenshots to interpret layout and structure. == Criticisms and dangers == AI browsers have been noted to be susceptible to being vulnerable to prompt injection attacks, in which the content of websites can be used to hijack the control of the browser. Multiple organisations have argued against using AI browsers due to this vulnerability. The United Kingdom national cyber security centre and Gartner consider them to be too risky for adoption by most organisations. A study by the CISPA Helmholtz Center and Saarland University concluded that this vulnerability makes them easy targets for malware, fraud, automated defamation, disinformation and biased outputs.
Inductive probability
Inductive probability attempts to give the probability of future events based on past events. It is the basis for inductive reasoning, and gives the mathematical basis for learning and the perception of patterns. It is a source of knowledge about the world. There are three sources of knowledge: inference, communication, and deduction. Communication relays information found using other methods. Deduction establishes new facts based on existing facts. Inference establishes new facts from data. Its basis is Bayes' theorem. Information describing the world is written in a language. For example, a simple mathematical language of propositions may be chosen. Sentences may be written down in this language as strings of characters. But in the computer it is possible to encode these sentences as strings of bits (1s and 0s). Then the language may be encoded so that the most commonly used sentences are the shortest. This internal language implicitly represents probabilities of statements. Occam's razor says the "simplest theory, consistent with the data is most likely to be correct". The "simplest theory" is interpreted as the representation of the theory written in this internal language. The theory with the shortest encoding in this internal language is most likely to be correct. == History == Probability and statistics was focused on probability distributions and tests of significance. Probability was formal, well defined, but limited in scope. In particular its application was limited to situations that could be defined as an experiment or trial, with a well defined population. Bayes's theorem is named after Rev. Thomas Bayes 1701–1761. Bayesian inference broadened the application of probability to many situations where a population was not well defined. But Bayes' theorem always depended on prior probabilities, to generate new probabilities. It was unclear where these prior probabilities should come from. Ray Solomonoff developed algorithmic probability which gave an explanation for what randomness is and how patterns in the data may be represented by computer programs, that give shorter representations of the data circa 1964. Chris Wallace and D. M. Boulton developed minimum message length circa 1968. Later Jorma Rissanen developed the minimum description length circa 1978. These methods allow information theory to be related to probability, in a way that can be compared to the application of Bayes' theorem, but which give a source and explanation for the role of prior probabilities. Marcus Hutter combined decision theory with the work of Ray Solomonoff and Andrey Kolmogorov to give a theory for the Pareto optimal behavior for an Intelligent agent, circa 1998. === Minimum description/message length === The program with the shortest length that matches the data is the most likely to predict future data. This is the thesis behind the minimum message length and minimum description length methods. At first sight Bayes' theorem appears different from the minimimum message/description length principle. At closer inspection it turns out to be the same. Bayes' theorem is about conditional probabilities, and states the probability that event B happens if firstly event A happens: P ( A ∧ B ) = P ( B ) ⋅ P ( A | B ) = P ( A ) ⋅ P ( B | A ) {\displaystyle P(A\land B)=P(B)\cdot P(A|B)=P(A)\cdot P(B|A)} becomes in terms of message length L, L ( A ∧ B ) = L ( B ) + L ( A | B ) = L ( A ) + L ( B | A ) . {\displaystyle L(A\land B)=L(B)+L(A|B)=L(A)+L(B|A).} This means that if all the information is given describing an event then the length of the information may be used to give the raw probability of the event. So if the information describing the occurrence of A is given, along with the information describing B given A, then all the information describing A and B has been given. ==== Overfitting ==== Overfitting occurs when the model matches the random noise and not the pattern in the data. For example, take the situation where a curve is fitted to a set of points. If a polynomial with many terms is fitted then it can more closely represent the data. Then the fit will be better, and the information needed to describe the deviations from the fitted curve will be smaller. Smaller information length means higher probability. However, the information needed to describe the curve must also be considered. The total information for a curve with many terms may be greater than for a curve with fewer terms, that has not as good a fit, but needs less information to describe the polynomial. === Inference based on program complexity === Solomonoff's theory of inductive inference is also inductive inference. A bit string x is observed. Then consider all programs that generate strings starting with x. Cast in the form of inductive inference, the programs are theories that imply the observation of the bit string x. The method used here to give probabilities for inductive inference is based on Solomonoff's theory of inductive inference. ==== Detecting patterns in the data ==== If all the bits are 1, then people infer that there is a bias in the coin and that it is more likely also that the next bit is 1 also. This is described as learning from, or detecting a pattern in the data. Such a pattern may be represented by a computer program. A short computer program may be written that produces a series of bits which are all 1. If the length of the program K is L ( K ) {\displaystyle L(K)} bits then its prior probability is, P ( K ) = 2 − L ( K ) {\displaystyle P(K)=2^{-L(K)}} The length of the shortest program that represents the string of bits is called the Kolmogorov complexity. Kolmogorov complexity is not computable. This is related to the halting problem. When searching for the shortest program some programs may go into an infinite loop. ==== Considering all theories ==== The Greek philosopher Epicurus is quoted as saying "If more than one theory is consistent with the observations, keep all theories". As in a crime novel all theories must be considered in determining the likely murderer, so with inductive probability all programs must be considered in determining the likely future bits arising from the stream of bits. Programs that are already longer than n have no predictive power. The raw (or prior) probability that the pattern of bits is random (has no pattern) is 2 − n {\displaystyle 2^{-n}} . Each program that produces the sequence of bits, but is shorter than the n is a theory/pattern about the bits with a probability of 2 − k {\displaystyle 2^{-k}} where k is the length of the program. The probability of receiving a sequence of bits y after receiving a series of bits x is then the conditional probability of receiving y given x, which is the probability of x with y appended, divided by the probability of x. ==== Universal priors ==== The programming language affects the predictions of the next bit in the string. The language acts as a prior probability. This is particularly a problem where the programming language codes for numbers and other data types. Intuitively we think that 0 and 1 are simple numbers, and that prime numbers are somehow more complex than numbers that may be composite. Using the Kolmogorov complexity gives an unbiased estimate (a universal prior) of the prior probability of a number. As a thought experiment an intelligent agent may be fitted with a data input device giving a series of numbers, after applying some transformation function to the raw numbers. Another agent might have the same input device with a different transformation function. The agents do not see or know about these transformation functions. Then there appears no rational basis for preferring one function over another. A universal prior insures that although two agents may have different initial probability distributions for the data input, the difference will be bounded by a constant. So universal priors do not eliminate an initial bias, but they reduce and limit it. Whenever we describe an event in a language, either using a natural language or other, the language has encoded in it our prior expectations. So some reliance on prior probabilities are inevitable. A problem arises where an intelligent agent's prior expectations interact with the environment to form a self reinforcing feed back loop. This is the problem of bias or prejudice. Universal priors reduce but do not eliminate this problem. === Universal artificial intelligence === The theory of universal artificial intelligence applies decision theory to inductive probabilities. The theory shows how the best actions to optimize a reward function may be chosen. The result is a theoretical model of intelligence. It is a fundamental theory of intelligence, which optimizes the agents behavior in, Exploring the environment; performing actions to get responses that broaden the agents knowledge. Competing or co-operating with another agent; games. Balancing short and long term rewards. In general no agent will always provi
Contrastive Language-Image Pre-training
Contrastive Language-Image Pre-training (CLIP) is a technique for training a pair of neural network models, one for image understanding and one for text understanding, using a contrastive objective. This method has enabled broad applications across multiple domains, including cross-modal retrieval, text-to-image generation, and aesthetic ranking. == Algorithm == The CLIP method trains a pair of models contrastively. One model takes in a piece of text as input and outputs a single vector representing its semantic content. The other model takes in an image and similarly outputs a single vector representing its visual content. The models are trained so that the vectors corresponding to semantically similar text-image pairs are close together in the shared vector space, while those corresponding to dissimilar pairs are far apart. To train a pair of CLIP models, one would start by preparing a large dataset of image-caption pairs. During training, the models are presented with batches of N {\displaystyle N} image-caption pairs. Let the outputs from the text and image models be respectively v 1 , . . . , v N , w 1 , . . . , w N {\displaystyle v_{1},...,v_{N},w_{1},...,w_{N}} . Two vectors are considered "similar" if their dot product is large. The loss incurred on this batch is the multi-class N-pair loss, which is a symmetric cross-entropy loss over similarity scores: − 1 N ∑ i ln e v i ⋅ w i / T ∑ j e v i ⋅ w j / T − 1 N ∑ j ln e v j ⋅ w j / T ∑ i e v i ⋅ w j / T {\displaystyle -{\frac {1}{N}}\sum _{i}\ln {\frac {e^{v_{i}\cdot w_{i}/T}}{\sum _{j}e^{v_{i}\cdot w_{j}/T}}}-{\frac {1}{N}}\sum _{j}\ln {\frac {e^{v_{j}\cdot w_{j}/T}}{\sum _{i}e^{v_{i}\cdot w_{j}/T}}}} In essence, this loss function encourages the dot product between matching image and text vectors ( v i ⋅ w i {\displaystyle v_{i}\cdot w_{i}} ) to be high, while discouraging high dot products between non-matching pairs. The parameter T > 0 {\displaystyle T>0} is the temperature, which is parameterized in the original CLIP model as T = e − τ {\displaystyle T=e^{-\tau }} where τ ∈ R {\displaystyle \tau \in \mathbb {R} } is a learned parameter. Other loss functions are possible. For example, Sigmoid CLIP (SigLIP) proposes the following loss function: L = 1 N ∑ i , j ∈ 1 : N f ( ( 2 δ i , j − 1 ) ( e τ w i ⋅ v j + b ) ) {\displaystyle L={\frac {1}{N}}\sum _{i,j\in 1:N}f((2\delta _{i,j}-1)(e^{\tau }w_{i}\cdot v_{j}+b))} where f ( x ) = ln ( 1 + e − x ) {\displaystyle f(x)=\ln(1+e^{-x})} is the negative log sigmoid loss, and the Dirac delta symbol δ i , j {\displaystyle \delta _{i,j}} is 1 if i = j {\displaystyle i=j} else 0. == CLIP models == While the original model was developed by OpenAI, subsequent models have been trained by other organizations as well. === Image model === The image encoding models used in CLIP are typically vision transformers (ViT). The naming convention for these models often reflects the specific ViT architecture used. For instance, "ViT-L/14" means a "vision transformer large" (compared to other models in the same series) with a patch size of 14, meaning that the image is divided into 14-by-14 pixel patches before being processed by the transformer. The size indicator ranges from B, L, H, G (base, large, huge, giant), in that order. Other than ViT, the image model is typically a convolutional neural network, such as ResNet (in the original series by OpenAI), or ConvNeXt (in the OpenCLIP model series by LAION). Since the output vectors of the image model and the text model must have exactly the same length, both the image model and the text model have fixed-length vector outputs, which in the original report is called "embedding dimension". For example, in the original OpenAI model, the ResNet models have embedding dimensions ranging from 512 to 1024, and for the ViTs, from 512 to 768. Its implementation of ViT was the same as the original one, with one modification: after position embeddings are added to the initial patch embeddings, there is a LayerNorm. Its implementation of ResNet was the same as the original one, with 3 modifications: In the start of the CNN (the "stem"), they used three stacked 3x3 convolutions instead of a single 7x7 convolution, as suggested by. There is an average pooling of stride 2 at the start of each downsampling convolutional layer (they called it rect-2 blur pooling according to the terminology of ). This has the effect of blurring images before downsampling, for antialiasing. The final convolutional layer is followed by a multiheaded attention pooling. ALIGN a model with similar capabilities, trained by researchers from Google used EfficientNet, a kind of convolutional neural network. === Text model === The text encoding models used in CLIP are typically Transformers. In the original OpenAI report, they reported using a Transformer (63M-parameter, 12-layer, 512-wide, 8 attention heads) with lower-cased byte pair encoding (BPE) with 49152 vocabulary size. Context length was capped at 76 for efficiency. Like GPT, it was decoder-only, with only causally-masked self-attention. Its architecture is the same as GPT-2. Like BERT, the text sequence is bracketed by two special tokens [SOS] and [EOS] ("start of sequence" and "end of sequence"). Take the activations of the highest layer of the transformer on the [EOS], apply LayerNorm, then a final linear map. This is the text encoding of the input sequence. The final linear map has output dimension equal to the embedding dimension of whatever image encoder it is paired with. These models all had context length 77 and vocabulary size 49408. ALIGN used BERT of various sizes. == Dataset == === WebImageText === The CLIP models released by OpenAI were trained on a dataset called "WebImageText" (WIT) containing 400 million pairs of images and their corresponding captions scraped from the internet. The total number of words in this dataset is similar in scale to the WebText dataset used for training GPT-2, which contains about 40 gigabytes of text data. The dataset contains 500,000 text-queries, with up to 20,000 (image, text) pairs per query. The text-queries were generated by starting with all words occurring at least 100 times in English Wikipedia, then extended by bigrams with high mutual information, names of all Wikipedia articles above a certain search volume, and WordNet synsets. The dataset is private and has not been released to the public, and there is no further information on it. ==== Data preprocessing ==== For the CLIP image models, the input images are preprocessed by first dividing each of the R, G, B values of an image by the maximum possible value, so that these values fall between 0 and 1, then subtracting by [0.48145466, 0.4578275, 0.40821073], and dividing by [0.26862954, 0.26130258, 0.27577711]. The rationale was that these are the mean and standard deviations of the images in the WebImageText dataset, so this preprocessing step roughly whitens the image tensor. These numbers slightly differ from the standard preprocessing for ImageNet, which uses [0.485, 0.456, 0.406] and [0.229, 0.224, 0.225]. If the input image does not have the same resolution as the native resolution (224×224 for all except ViT-L/14@336px, which has 336×336 resolution), then the input image is first scaled by bicubic interpolation, so that its shorter side is the same as the native resolution, then the central square of the image is cropped out. === Others === ALIGN used over one billion image-text pairs, obtained by extracting images and their alt-tags from online crawling. The method was described as similar to how the Conceptual Captions dataset was constructed, but instead of complex filtering, they only applied a frequency-based filtering. Later models trained by other organizations had published datasets. For example, LAION trained OpenCLIP with published datasets LAION-400M, LAION-2B, and DataComp-1B. == Training == In the original OpenAI CLIP report, they reported training 5 ResNet and 3 ViT (ViT-B/32, ViT-B/16, ViT-L/14). Each was trained for 32 epochs. The largest ResNet model took 18 days to train on 592 V100 GPUs. The largest ViT model took 12 days on 256 V100 GPUs. All ViT models were trained on 224×224 image resolution. The ViT-L/14 was then boosted to 336×336 resolution by FixRes, resulting in a model. They found this was the best-performing model. In the OpenCLIP series, the ViT-L/14 model was trained on 384 A100 GPUs on the LAION-2B dataset, for 160 epochs for a total of 32B samples seen. == Applications == === Cross-modal retrieval === CLIP's cross-modal retrieval enables the alignment of visual and textual data in a shared latent space, allowing users to retrieve images based on text descriptions and vice versa, without the need for explicit image annotations. In text-to-image retrieval, users input descriptive text, and CLIP retrieves images with matching embeddings. In image-to-text retrieval, images are used to find related text content. CLIP’s ability to connect vis