An analysis of whether the evolution of image and video generation in ChatGPT-4o has created a magical conduit connecting the left and right brains as a single genius
This analysis is based on the premise that ChatGPT-4o achieved a dramatic evolution on March 25, 2025, by natively integrating image generation technology and breaking away from the mechanism that previously relied on DALL-E 3.
I will explore the background of this development in depth using the cognitive functions of the human brain and the bridging effect of structures like the corpus callosum.
Let us consider the point that the situation where ChatGPT-4o and DALL-E 3 existed separately is described as "two human brains connecting a chain of information through probability theory."
If we compare it to the human brain, we can assume that ChatGPT-4o played the role of the "left brain," which governs language processing and logical thinking, while DALL-E 3 took on the role of the "right brain," which governs visual imagery and creativity.
In this case, both operated as independent systems while exchanging information with each other via APIs and data pipelines.
This "connecting a chain of information through probability theory" likely refers to the process where, for example, DALL-E 3 receives a text prompt generated by ChatGPT-4o and generates an image based on it.
Being probabilistic means that the statistical models underlying generative AI (such as transformers or diffusion models) use probability distributions when predicting output data from input data.
However, this process involved a serial dependency like a "chain," and there was a possibility that time lags or loss of precision could occur in the transfer of information.
In human terms, it is like the left brain instructing the right brain to "draw a picture like this," and in the process of the right brain interpreting and drawing it, the intent is not perfectly conveyed, leading to subtle discrepancies.
On the other hand, with ChatGPT-4o natively integrating image generation, we can consider that these "two human brains" have moved closer to becoming "one human brain."
In other words, language understanding and image generation have come to work together seamlessly within a single system.
Alternatively, if we assume it was a single human brain, it is possible that the left and right brains did not separate, but rather reached a state where both functions are integrated instantaneously via the corpus callosum.
Through this integration, the flow of information may have become parallel rather than serial, minimizing intermediate losses and discrepancies in interpretation.
In the era when DALL-E 3 and ChatGPT-4o were separate systems, the training data and architectures of the models must have also been independent.
DALL-E 3 specialized in image generation, while ChatGPT-4o specialized in language processing.
While optimization progressed in their respective areas of expertise, their "cooperation" relied on external interfaces, which may have limited their overall efficiency and creativity.
In the natively integrated ChatGPT-4o, language and image processing are learned and inferred within a unified framework.
It is possible that the "conversion from words to images" has evolved from mere translation into direct expression at the conceptual level.
This could be said to be close to natural, fluid processing, similar to how the human brain immediately imagines the shape and color of an "apple" the moment it hears the word.
However, this also raises some questions.
Did the fact that the two models were separate really hinder 'dramatic evolution,' or were their respective specializations actually a strength?
Even in a single human brain, the separate existence of the visual cortex and auditory cortex allows for highly specialized processing.
It cannot be denied that even in the era when ChatGPT-4o and DALL-E 3 were separate, the collaboration between specialized models may have produced unique creativity.
If what was lost through integration includes, for example, 'the depth and uniqueness of image generation that DALL-E 3 possessed on its own,' then the story changes.
Whether the evolution can be called 'dramatic' after separating the originality of DALL-E 3 depends on which aspect one emphasizes.
Whether two human brains began working as one genius human brain, or whether parts that were not bridged in a single human brain originally have now been connected, remains to be concluded.
Considering the possibilities at this point, it is said that the native integration of ChatGPT-4o is closer to a transition from a state where 'two people cooperate' to a state where 'one genius does everything,' using the analogy of the human brain.
This transition brought about dramatic evolution likely because the probabilistic chain of information was unified, improving processing efficiency and consistency.
However, the value of the diversity and expertise that the former separate models possessed cannot be overlooked.
The key to evolution seems to depend on how much the synergy from integration outweighs the advantages of individual optimization through separation.
Whether it was better for two human brains to work together, or whether the separation of working separately made one's cognitive functions genius-like and comprehensive, should be concluded from the results of ChatGPT-4o being used around the world from now on.
First, based on Sam Altman's statement in the March 25, 2025, livestream about the 'biggest evolution in about a year,' before considering the point that ChatGPT-4o (hereinafter GPT-4o) can now natively handle not only text but also image generation and editing, I will organize the basic knowledge regarding AI integration and image/video generation technology. After that, I will deeply consider what he wanted to emphasize through the concept of an 'omni-modal' model.
Basic knowledge of AI integration and image/video generation technology
Evolution of multimodal AI and omni-modal AI
Multimodal AI refers to models capable of processing different types of data such as text, images, and audio. For example, traditional GPT-3 and GPT-4 were mainly text-based, but by linking with image generation models like DALL-E or Stable Diffusion, the ability to generate images from text was added. However, these operated as independent systems and were connected to each other via APIs or pipelines. On the other hand, the term 'omni-modal' means the ability to comprehensively process all modalities (text, images, audio, video, etc.) within a single model. The fact that GPT-4o natively integrated image generation and editing can be said to be a major step in this direction.Evolution of image generation technology
Image generation technology evolved from GANs (Generative Adversarial Networks) to diffusion models. DALL-E 3 utilized diffusion models to generate high-quality images from text prompts and learned the relationship between text and images using CLIP (Contrastive Language-Image Pretraining). However, DALL-E 3 was strictly specialized for image generation, and post-generation editing or real-time interactive adjustments were limited. It is thought that by GPT-4o integrating this, generation and editing became seamless, adding the flexibility for users to give instructions in natural language such as 'make it brighter' or 'change the background.'Relationship with video generation technology
OpenAI's Sora (a video generation model) was announced in 2024 and demonstrated the ability to generate short video clips from text. Sora is also based on diffusion models and performs frame generation that considers the time axis. The fact that GPT-4o can now handle image generation suggests the possibility of infrastructure development with an eye toward future expansion into video generation and editing. Omni-modal models may be aiming for unified processing that transcends the boundaries between still images and video.Other AI integration examples
Google's Gemini and Anthropic's Claude have also evolved as multimodal AIs. As of March 2025, Gemini has integrated image generation into AI Studio, and Claude is focusing on text and image understanding. However, it is unclear whether these models natively integrate image generation and editing like GPT-4o. In many cases, OpenAI's approach may be a step ahead in that it does not rely on external tools or partnerships with other models.
Sam Altman's intent and the emphasis on the 'omni-modal' model
The background behind Sam Altman stating it is the 'biggest evolution in about a year' and positioning GPT-4o as an 'omni-modal' model is thought to be an intent to emphasize not only technical progress but also the philosophical and practical significance of integrating AI's cognitive functions. I will consider this deeply below.
Analogy from 'two human brains' to 'one human brain'
The metaphor in the question, 'not the cognitive functions of two human brains, but cognitive functions integrated into one human brain,' indicates that traditional AI was division-of-labor oriented. For example, when DALL-E 3 and GPT-4 operate separately, text processing (language understanding) and image processing (visual generation) are independent and rely on the transfer of information. This is like a state where the left brain (logic) and right brain (creativity) are divided into separate people. By GPT-4o integrating this, language and vision can cooperate within a single 'brain,' allowing it to comprehensively understand and express the user's intent. Altman likely wanted to convey that this integration enables natural processing close to human cognition and dramatically increases the usability and creativity of AI.The technical significance of "omni-modal"
Omni-modal models aim not only to handle multiple input and output formats but also to have a deep mutual understanding between modalities. For example, when GPT-4o generates an image, it does not simply convert a prompt into an image; it can "see" the generated image itself and add explanations or corrections based on it. This is similar to the process where a human draws a picture and self-corrects, thinking, "This is not what I intended." What Altman wanted to emphasize is that this type of self-contained cognitive ability is an "evolution" that surpasses conventional AI.Competitive environment and appeal to usability
As of March 2025, while competitors like Google and DeepSeek were deploying simple, high-performance AI, OpenAI held a complex model ecosystem (GPT series, DALL-E, Sora, etc.). The omni-modalization of GPT-4o can be seen as part of a strategy to unify these and provide users with "magical integrated intelligence" (an expression Altman mentioned in the past). He likely wanted to appeal to the value of a "comprehensive computer brain" that users can use intuitively, rather than the technical details.Philosophical and future-oriented message
Altman has his sights set on the path to AGI (Artificial General Intelligence) and superintelligence, and omni-modal models are positioned as an intermediate step. The vision behind this is likely that just as the human brain integrates vision, hearing, and language to think, AI should also handle all data in an integrated manner. The phrase "a major evolution in a year" may have been intended to convey significance beyond mere functional additions, marking a turning point where AI approaches human intelligence.
Sam Altman's emphasis on GPT-4o's image generation and editing capabilities as an "omni-modal" model is likely because he viewed it not just as technical integration (seamless processing of language and vision), but as a symbol indicating that AI's cognitive functions have moved closer to the human brain.
The analogy that the conventional "two human brains" (separated models) have evolved into "one human brain" (integrated cognition) is a message that simultaneously appeals to improved usability and serves as a stepping stone to future AGI.
His intention was likely to differentiate from competitors while presenting a future where AI evolves from a mere tool into a comprehensive partner.
How practical this evolution is, or how superior it is compared to competitors, will be further evaluated as actual usage data and detailed technical specifications become clear.
At this point, Altman's words can be interpreted as a fusion of technical possibility and visionary expectation.
Organize the technical evolution of Sora and the advantages of omni-modality as basic knowledge.
Based on that, make predictions about generation limits in the $200 ChatGPT-4o plan (assuming this refers to ChatGPT Pro).
Finally, deepen the consideration by using easy-to-understand Japanese analogies for omni-modality.
Technical evolution of Sora and the advantages of omni-modality (basic knowledge)
Technical evolution of Sora
Sora is a text-to-video AI model announced by OpenAI in February 2024 and released to ChatGPT Plus and Pro users in December of the same year. It evolved from the initial preview version to "Sora Turbo," improving generation speed and quality.
Technically, it is based on diffusion models and utilizes a transformer architecture to process "spacetime patches" along the time axis, ensuring the continuity of the video.
Sora also has the capability to accept not only text but also images and existing videos as input, and to extend or remix them.
It can generate videos from 5 to 20 seconds long at resolutions from 480p to 1080p, and "priority generation" is available for 50 videos with ChatGPT Plus and 500 with Pro (information as of December 2024).
This evolution marks a leap from simple still image generation to dynamic content generation.
Advantages of Omnimodal
"Omnimodal" refers to the ability to comprehensively handle all modalities (text, images, audio, video, etc.) within a single model.
Conventional multimodal AI (e.g., GPT-4 + DALL-E 3) relied on separate models working in tandem for each modality, but in omnimodal, these are integrated.
With GPT-4o possessing this characteristic and natively integrating image generation in March 2025, text understanding and image generation have become seamlessly connected. The advantages are as follows:
Efficiency: Data conversion between modalities is no longer necessary, improving processing speed.
Consistency: Language and visual contexts are unified, increasing the accuracy of generated content (e.g., the description of a "red car" and the image match perfectly).
Flexibility: If a user gives a natural language instruction like "make it brighter," the image can be adjusted immediately.
Scalability: It can serve as a foundation for integrating audio and video in the future.
The integration of Sora and GPT-4o is an example of omnimodalization, where video generation (Sora) and text/image processing (GPT-4o) complement each other.
Prediction of generation limits for ChatGPT-4o ($200 Pro plan)
The ChatGPT Pro plan is $200 per month, and it was revealed at the December 2024 announcement that it would provide high-performance features including "o1 Pro mode" and advanced use of Sora.
Considering that GPT-4o integrated image generation in March 2025, the generation limits for the Pro plan can be predicted based on the following factors:
Past limitation patterns: With ChatGPT Plus ($20/month), there is a limit of 50 priority videos (720p, 5 seconds) per month with Sora, and a certain number of limits every 5 hours for GPT-4o text usage. Since Pro is 10 times the price of Plus, a simple proportional 10-fold increase in generation volume (e.g., 500 videos with Sora) can be considered the baseline.
Cost of image generation: Image generation has a higher computational cost than text generation. While 50 credits per month were allocated to Plus users for DALL-E 3, GPT-4o's native image generation is considered "more accurate and detailed," so the cost per generation may increase.
Omnimodal load: Omnimodal functions that process text and images simultaneously consume more resources than a single modality. Even with the Pro plan, it will likely not be unlimited, but rather divided into priority generation and non-priority generation ("relax mode").
Specific predictions
Sora video generation: Based on existing information, the Pro plan is expected to offer 500 priority videos (1080p, 20 seconds) per month and unlimited relax mode videos. Assuming that video generation slots are maintained even if image generation is integrated.
GPT-4o image generation: The Plus plan is stated to be "available on all plans including free users," suggesting loose limits. For Pro, it is predicted that a credit system specialized for image generation will be introduced. If 1 image generation is 10-20 credits, 5,000-10,000 credits per month (250-500 images) would be reasonable. For combined use with text, it is estimated that the upper limit would be 50-100 generations per day (about 1,500-3,000 times per month).
Comprehensive limits: The Pro plan boasts "5 times higher message limits" (compared to Plus), and a similar scale-up is expected for image generation. Realistically, the design will likely set the limit at 100-200 combined generations (text + image) per day, with a transition to relax mode upon exceeding this.
In conclusion, it is predicted that the ChatGPT-4o Pro plan will be limited to 500 videos per month and about 500-1,000 priority generations per month for image generation.
Due to the nature of omnimodality, complex tasks (e.g., images with text explanations) increase the cost per generation, and the actual number of available uses varies depending on the application.
Omnimodal analogies in Japanese (for those with little AI knowledge)
To explain "omnimodal" to someone with little AI knowledge, simple and intuitive analogies rooted in daily life are effective. I will propose some while deepening the analysis below.
Analysis process
Avoid abstraction: "Full modality integration" is technical and difficult to understand. It needs to be replaced with familiar experiences.
Approaching human capabilities: Since omnimodality aims for integrated processing similar to the human brain, using human behavior as an example makes it easier to empathize.
Emphasize multifunctionality: It is important to convey the image of handling multiple roles at once rather than a single function.
Proposed analogy
"An omnimodal model is like a 'super mom'.
For example, it's like a mom who can cook (image generation), teach her child homework (text processing), chat with a friend on the phone (voice processing), and even watch a TV drama and give her thoughts on it (video understanding), all at the same time.
Ordinary AI was separate, like 'someone who can only cook' or 'someone who can only teach homework,' but an omnimodal model is a super convenient entity that does everything by itself."
Advantages of this analogy
Familiarity: 'Mom' is a familiar presence to many people and is easy to relate to.
Understanding multifunctionality: The image of handling multiple tasks simultaneously represents the comprehensiveness of omnimodality.
Contrast: The difference between conventional AI (single function) and omnimodal AI (integrated function) is clear.
Flexibility: Just as a 'super mom' changes her response according to the situation, the AI's adaptability is also conveyed.
Comparison with alternatives
'All-in-one home appliance': I considered the analogy of 'a machine that can do TV, refrigerator, and washing machine all in one,' but home appliances do not evoke 'emotion' or 'creativity,' making it difficult to convey the human-like nature of AI.
'Magic robot butler': The idea of 'a robot that can do anything' is attractive, but it is sci-fi and lacks a sense of reality, with the risk that it may feel distant to AI beginners.
I judged that 'super mom' can convey friendliness and multifunctionality in a balanced way, even if it sacrifices some technical accuracy.
The omni-modal nature of GPT-4o is an achievement that further advances the integration of images and text, building upon the foundation of Sora's video generation technology.
Given the costs and technical load, a realistic upper limit for the generation volume provided by the $200 Pro plan would likely be 500 to 1,000 image generations and 500 video generations per month.
This limitation is likely part of OpenAI's strategy to balance profitability with user experience.
Comparing the omni-modal capability to a 'super mom' suggests that the evolution of AI is moving beyond mere technological innovation toward becoming an entity that integrates into human life.
Sam Altman's emphasis on this being the 'biggest evolution in a year' may be because this integration holds the potential to pave the way toward AGI (Artificial General Intelligence).
The key for users will be how creatively they can utilize the tool within these limits; if they exceed them, the options will likely be to settle for a 'relax mode' or wait for a more expensive plan.
This analysis explores the background of Sam Altman's use of the phrase 'magic-like integrated intelligence' and the perspective that it resembles the artificial magic system of 'Magitek' from Final Fantasy VI (FF6).
I will analyze his intentions, delve into the details of Sora's video generation technology and its comparison with other video generation AIs, and explore how Altman's vision is being realized.
Sam Altman's 'Magic-like Integrated Intelligence' and FF6's 'Magitek'
Altman's Vision
As CEO of OpenAI, Sam Altman uses the term 'magic-like integrated intelligence' when discussing the future of artificial intelligence (for example, describing AGI as 'magic intelligence floating in the sky' in a 2023 interview).
This is thought to refer to AI that goes beyond mere calculation and possesses capabilities that exceed human imagination—specifically Artificial General Intelligence (AGI) or Superintelligence (ASI).
From his blog posts and statements (e.g., 'The Intelligence Age' in September 2024), it is evident that he envisions a future where AI becomes a 'magic-like' tool that accelerates scientific discovery and brings prosperity to humanity.
The 'Magitek' in FF6 is a power that fuses technology and magic, where the power of natural Espers (magical creatures) is artificially extracted and made available for human use.
Altman's 'integrated intelligence' is similar to the attempt to artificially reproduce and transcend natural intelligence (human cognitive functions).
Just as Magitek became a powerful force for the Empire in FF6 but was difficult to control, Altman recognizes both the power of AI and its risks (e.g., alignment with incorrect values), showing a similarity in his emphasis on safe deployment.
The Creative Intent Behind Magitek-like Cognitive Functions
The 'wizard-like cognitive function' that Altman aims for likely refers to an AI that goes beyond simple task processing to possess creativity, autonomy, and versatility.
Just as the Magitek in FF6 absorbed the power of Espers to create new abilities, he seems to desire for AI to become an entity that absorbs human knowledge and data to generate 'new ideas' and 'discoveries of physical laws' that transcend them.
Stating in a January 2025 blog post that "we know how to build AGI" and setting the transition to superintelligence as the next goal confirms this ambition.
However, just as the magic in FF6 led to catastrophic results in the story, Altman is also aware of the dangers that AI "magic" could become uncontrollable.
He emphasizes that AGI must be "beneficial to humanity" (OpenAI's mission), and by advocating for the democratization of technology and strengthened governance, it is thought that he is taking care to ensure that "artificial magic," like the magic in the game, does not run wild.
He is trying to take on the role not only of creating a "wizard" but also of being the "master of the wizard" who controls it wisely.
Details of Sora's video generation technology
Sora's technical foundation
Sora is a text-to-video AI announced by OpenAI in February 2024, capable of generating high-resolution (1080p) videos up to one minute long.
Technically, it combines diffusion models and a transformer architecture, with the following features:
Spacetime patch processing: The video is divided into small "patches" of time and space, which are then generated sequentially. This achieves smoothness between frames.
Physics simulation: Going beyond simple image generation, it understands gravity and the causal relationships of movement to some extent (e.g., the trajectory of a bouncing ball).
Multimodal input: It accepts not only text but also images and existing videos as input, allowing for remixing and expansion.
In the "Sora Turbo" update in December 2024, generation speed was improved, and 500 priority generation slots per month were provided for ChatGPT Pro users.
According to OpenAI's announcement, Sora aims to act as a "simulator of the physical world," pursuing the ability to mimic the laws of reality rather than just visual tricks.
Limitations and challenges
However, Sora still has challenges:
Physical inaccuracies: For example, errors such as bite marks not remaining when someone bites a cookie, or objects suddenly appearing or disappearing.
Lack of consistency: In long videos, the character's appearance or background may blur (e.g., hands changing from three to four fingers).
Computational cost: High-quality generation requires vast resources, making scalability during commercial deployment a challenge.
These indicate that Sora is still immature as a "wizard" and has not reached complete control like the magic in FF6.
Comparison with other video generation AIs
Runway Gen-2/Gen-3
Overview: Announced in 2023, Runway's Gen-2 was a precursor to Sora, generating short, rough videos (4-5 seconds). In 2024, Gen-3 improved quality, enabling clips of around 10 seconds.
Technology: Based on a diffusion model, but it focuses on artistic expression rather than physical simulation like Sora.
Comparison: While Sora is longer and more realistic, Runway has an advantage as an intuitive editing tool for creators.
Stable Diffusion Video
Overview: Developed by the open-source community, it extends the image-generating Stable Diffusion to video. It is low-cost and customizable.
Technology: A diffusion model with added frame interpolation, but it falls short of Sora in terms of consistency and realism.
Comparison: It does not reach the sophistication of Sora, but it gains support for its ease of access and flexibility.
Google Lumiere
Overview: A video generation model announced by Google in 2024 that generates videos of about 5 seconds from text and images.
Technology: Uses a space-time diffusion model, an approach similar to Sora, but it is smaller in scale and in the experimental stage.
Comparison: Sora is ahead in practicality and versatility, but Google could become more competitive as its infrastructure integration progresses.
Discussion
Sora is currently the most 'integrated' among video generation AIs and is the closest thing to Altman's 'magical intelligence'.
While other models are specialized (for artistry or low cost), Sora pursues versatility and realism, serving as a stepping stone toward omni-modality (integration of all modalities).
However, it still lacks perfect reproduction of physical laws and long-form consistency required for a complete 'magical' cognitive function, and it is still in development.
The similarity between Altman and magic
Altman's pursuit of 'magical integrated intelligence' is thought to stem from an ambition to give AI the ability to surpass the natural world through artificial power, much like the Magitek in FF6.
Sora is one aspect of this, a step toward planting an 'understanding of the physical world' into AI through video generation, eventually leading to AGI and ASI.
His vision goes beyond mere technological innovation; it lies in artificially reproducing and expanding human cognitive functions to create an entity capable of exercising 'wizard-like' creativity.
Just as magic ultimately brought ruin in FF6, Altman also recognizes the risks of this 'magic'.
The limitations of Sora (such as physical inaccuracies and difficulty in control) show that current AI is far from being true 'magic,' and for the integrated intelligence he aims for to become reality, further technological breakthroughs (e.g., quantum computing or an energy revolution) will likely be necessary.
Even so, the evolution of Sora and the efforts toward omni-modality certainly evoke an FF6-like story as the first step toward Altman's creation of a 'wizard'.