GPT-4.5 = Big Model Energy | YC Decoded

Key Concepts:

• GPT 4.5: OpenAI's latest large language model, emphasizing emotional intelligence, accuracy, and creative capabilities.
• Unsupervised Learning: Training AI models on unlabeled data, a key aspect of GPT 4.5's development.
• Hallucination Rate: The tendency of AI models to generate incorrect or nonsensical information.
• Reasoning Models: AI models designed for systematic problem-solving, like OpenAI's 01.
• Pre-training and Post-training: Stages in AI model development, involving initial training on large datasets and subsequent refinement.
• Vibes Testing: Subjective evaluation of AI model outputs by human testers.
• Token: A unit of text used for processing by language models.
• Inference Time: The time it takes for a trained model to generate a response or prediction.

GPT 4.5: An Overview

GPT 4.5 is presented as OpenAI's most advanced and human-like model to date, representing a significant step in scaling up unsupervised learning. It excels in natural conversation, creative tasks, and complex planning, while also exhibiting a reduced hallucination rate compared to previous models. Despite being an incremental improvement over GPT-4, its launch is considered an important milestone for future AI development.

Background and Development

Following the release of GPT-4 in early 2023, anticipation grew for its successor. Rumors circulated about internal projects named "Strawberry" and "Orion." Eventually, OpenAI revealed 01, a reasoning-focused model. Sam Altman later confirmed that "Orion" would be released as GPT 4.5. The model is potentially more than 10 times the size of GPT-4, representing advancements in both pre-training and post-training techniques.

Key Features and Capabilities

• Emotional Intelligence: GPT 4.5 is distinguished by its enhanced emotional intelligence, enabling deeper and more nuanced conversations. It demonstrates a better understanding of human intentions and desires.
  • Quote: "You can have much more deeper conversations about maybe more Curious facts that for all or even one or three many don't really you know know about but also we think that it has a much better understanding of what humans want that really gets what you mean when you ask for something and that's been really the the like magical experience for people at open for myself working with the model"
• Accuracy: GPT 4.5 achieves approximately 61.9% accuracy on benchmarks like Simple QA, significantly surpassing GPT 4.0's 38.4%.
• Reduced Hallucination: The model's hallucination rate is reduced to roughly 37%, a substantial improvement from GPT 4.0's 61.2%.
• Creativity: GPT 4.5 excels in creative tasks such as drafting emails, generating stories, telling jokes, and brainstorming ideas, producing more human-like prose than its predecessors.
• Persuasiveness: On benchmarks evaluating persuasive power (Make Me Pay and Make Me Say), GPT 4.5 outperforms models like GPT 4.0 and 01.
• Humor and Irony: Early testers have noted that GPT 4.5 demonstrates an understanding of humor and irony, a capability often lacking in other models.

Evaluation Methods

Researchers used "Vibes testing," relying on human testers to evaluate the model's outputs and provide feedback on its performance compared to GPT-4. This subjective evaluation focused on areas like writing quality, emotional intelligence, and overall model feel. The challenge lies in defining and quantifying subjective qualities like "good writing."

Limitations

• Cost: GPT 4.5 is significantly more expensive than other OpenAI models. Per input token, it is 30 times more expensive than GPT-4, and per output token, it is 15 times more expensive. This high cost may limit its suitability for large-scale deployments.
• Reasoning: Compared to specialized reasoning models like 01, GPT 4.5 falls short in structured reasoning domains, including complex STEM tasks, advanced math problems, and challenging coding tasks.

Future Directions

Sam Altman suggests that unsupervised pre-training models like GPT 4.5 and specialized reasoning-focused models like 01 will eventually converge into a unified architecture, potentially seen in GPT-5. The goal is to combine vast world knowledge, creative fluency, emotional nuance, and advanced reasoning into a single model.

• Quote: "we do think that reasoning is going to be a core capability of future models but these two paradigms they sort of they're not exclusive they they actually complement each other really well so you could imagine that a model that has the the knowledge and the intuition of GPT 4.5 but then combined with reasoning that would be really strong model"

Conclusion

GPT 4.5 represents a step forward in AI development, demonstrating improvements in accuracy, emotional intelligence, and creativity through scaling unsupervised learning. While it has limitations in cost and specialized reasoning compared to models like 01, it serves as a crucial bridge towards future AI systems that combine broad understanding with powerful reasoning capabilities.

Additional Information

The video concludes with an announcement of YC's first AI Startup School in San Francisco, featuring prominent AI experts and founders. The conference is free for computer science students and new grads in AI and AI research, with travel covered.