This Week In AI: (February 24th - March 2nd 2025)

This week in AI, OpenAI teases the release of GPT-4.5, outlines plans to integrate Sora into ChatGPT, and Meta prove they are once again behind in the AI race. In research, fractals have inspired a new generative model for image generation, the bridge between visual generation and understanding is slowly being removed, and researchers from Apple try to find out if loss prediction can act as an uncertainty estimator!

OpenAI Plans to Integrate Sora into ChatGPT

Rohan Sahai, OpenAI's product lead for Sora, has confirmed OpenAI plans to integrate the AI video generator directly into ChatGPT. This aims to enhance the user experience by bringing video creation to premium tiers, just as it has already brought image generation via DALL-E. Rohan noted during a Discord session that the integrated version might offer extra features that can not be found in the current Sora version. While no specific timeline has been given, we can speculate that it could potentially be in the first half of 2025 if a lead developer has begun discussing it online!

For those who are unaware, Sora is OpenAI’s cutting-edge AI model that generates high-quality, realistic videos from simple text prompts. Using diffusion-based techniques similar to those in AI image generation, Sora creates coherent, dynamic scenes up to one minute long. The model understands physics (some-what), object interactions, and cinematic styles, allowing it to produce complex animations, realistic landscapes, and human-like motion.

Meta Plans Standalone AI Chatbot App

Meta is reportdeky developing a standalone app for its AI assistant, Meta AI, to attempt to rival to current offerings from OpenAI's ChatGPT, Google's Gemini, and now...X's Grok. This move attempts to expand Meta's current AI ecosystem and is expected to launch within the second quarter of this year.

For those unfamiliar, Meta AI can currently answer questions, generate images, and edit photos. I am sure you can already tell, this seems a little behind all the other competitors. ChatGPT offers deep research and complex reasoning abilities, multiple models can access the internet directly, and much more. Once again, it seems like Meta is just playing catch-up with everyone else.

OpenAI tease ChatGPT 4.5

OpenAI has teased GPT-4.5, the latest iteration in its series of large language models (LLMs) in a new white paper. Described by CEO Sam Altman himself, GPT-4.5 is a "giant, expensive model", but aims to provide uses with more natural and human-like conversation with a bunch of new capabilites and upgrades!

Personally, the biggest upgrade for me is the reduced hallucination rate. Hallucinations are where an AI generates inaccurate or nonsensical information, and it can be a massive issue in LLMs as typically a user will not know when an AI is hallucinating so will blindly trust it. GPT-4.5 addresses this issue effectively, reducing the hallucination rate from 52% in its predecessor to 19%. This improvement enhances the reliability of the AI's responses, making it a more trustworthy tool for users.

Kerb your excitement for a second though, it has been announced that upon release, GPT-4.5 will only be available to ChatGPT pro users (200 dollars per month tier), with plans to then make it available to Plus tiers. And for those who care about cost, it is expensive! The model is priced at 75 dollars per million tokens, a significant increase from GPT-4's 2.50 dollars per million tokens. This can give us some insight into how expensiive this model is to infer upon, which doesn't seem promising to those who care about AI that is eco-friendly.

Hopefully, we will get to see more details as we get closer to the release of GPT-4.5, and then we will be able to do a proper deep dive into it!

Fractal Generative Models: A Recursive Approach to AI Creativity

A new paper, "Fractal Generative Models", introduces a novel approach to AI-driven content creation by incorporating fractal structures into generative models. The authors, inspired by self-similar patterns found in nature (like an artichoke), created a model that recursively applies smaller generative models within larger ones, forming a hierarchical structure.

The core idea is recursion- building larger generative structures by nesting smaller generative units within them. Just as fractals in mathematics emerge from simple rules applied repeatedly, this AI framework stacks multiple layers of generative models, each influencing and refining the next. This differs from conventional AI models, which operate at a single scale or rely on pre-defined hierarchical structures

One of the most impressive demonstrations of this fractal framework is in pixel-wise image generation, a notoriously challenging AI task. Conventional image-generation models often struggle with maintaining coherence and structure at different scales. By contrast, the fractal approach ensures that each level of the generative process inherits and refines features from the level above, leading to more detailed and realistic images.

Experiments on ImageNet show that the fractal model achieves competitive likelihood estimation and image quality, rivaling state-of-the-art autoregressive and diffusion-based models. The researchers also highlight its computational efficiency, as the recursive nature of the model allows it to generate images more effectively than traditional sequential approaches.

Unified Tokenizer for Visual Generation

Researchers from The University of Hong Kong introduce UniTok, a unified tokenizer designer to bridge the gap between visual generation and understanding- two tasks that have typically required seperate models to carry out.

Current AI models struggle with integrating visual understanding (as seen in models like CLIP) and high-quality image generation (which often uses vector quantization). Existing tokenizers are typically task-specific, meaning that a model trained for image generation often lacks strong semantic reasoning, and vice versa. Attempts to combine these approaches have led to training instability and performance trade-offs.

UniTok solves this problem by introducing multi-codebook quantization, which expands the representational space of discrete visual tokens without the instability caused by overly large codebooks. Instead of relying on a single codebook, UniTok splits visual data into multiple smaller sub-codebooks, allowing it to capture both fine details and abstract representations efficiently.

UniTok represents a major step forward in multimodal large language models (MLLMs) by enabling a single tokenizer to be used for both AI-generated images and text-based reasoning tasks. By effectively eliminating the trade-off between generation and understanding, UniTok could pave the way for more cohesive, general-purpose AI models.

When Does a Predictor Know Its Own Loss?

Researchers from Apple present a new paper, "When Does a Predictor Know Its Own Loss?", a study that examines a fundamental question in machine learning: can a model accurately predict the loss it will incur on an input? This problem of loss prediction is closely tied to uncertainty estimation and model calibration, both of which are crucial for applications in active learning, fairness auditing, and selective prediction.

At the heart of the research is the idea that a model's own self-entropy—a measure of uncertainty based on the probability distribution it assigns to different outcomes—acts as its built-in loss predictor. However, the researchers explore whether an external loss predictor can outperform this self-estimate, effectively giving a better understanding of where a model is likely to fail. Through a theoretical framework and empirical experiments, the study finds a strong connection between loss prediction and multicalibration, a fairness principle that ensures a model’s predictions are well-calibrated across different subgroups.

Interestingly, the researchers show that if an external loss predictor consistently outperforms the model’s own uncertainty estimate, then the model must be failing to satisfy multicalibration—meaning its predictions are systematically misaligned across certain subgroups. This insight transforms loss prediction from just an auxiliary tool into a direct way to audit model fairness and reliability.

The study provides experimental validation of its theoretical claims by training different loss predictors on UCI tabular datasets and testing them across a variety of machine learning models, including logistic regression, support vector machines, and deep neural networks. The key takeaway is that loss predictors consistently perform better when the base model has higher multicalibration errors.

By framing loss prediction as a fairness and reliability problem, this research provides a new perspective on AI accountability, pushing for more transparent and adaptable models that can better communicate their own limitations.